hyperspectral applications of precision agriculture for

i

Hyperspectral Applications of Precision Agriculture for Field Crops

in Drylands

Thesis submitted in partial fulfillment of the requirements for the degree of

“DOCTOR OF PHILOSOPHY”

By

Ittai Herrmann

Submitted to the Senate of Ben-Gurion University of the Negev

May, 2012

Beer-Sheva

ii

Hyperspectral Applications of Precision Agriculture for Field Crops in

Drylands

Thesis submitted in partial fulfillment of the requirements for the degree of

“DOCTOR OF PHILOSOPHY”

By

Ittai Herrmann

Submitted to the Senate of Ben-Gurion University

of the Negev

Approved by the advisors:

Prof. Arnon Karnieli ___________

Dr. David J. Bonfil ____________

Approved by the Dean of the Kreitman School of Advanced Graduate Studies:

Prof. Michal Shapira: __________

May, 2012

Beer-Sheva

iii

This work was carried out under the supervision of:

Prof. Arnon Karnieli

The Remote Sensing Laboratory

Jacob Blaustein Institutes for Desert Research

Ben-Gurion University of the Negev

Dr. David J. Bonfil

Field Crops and Natural Resources

The Institute of Plant Sciences

Agricultural Research Organization, Gilat Research Center

iv

Acknowledgments

First, I would like to thank my two advisors, Prof. Arnon Karnieli and Dr. David J. Bonfil, for

their valuable and educating guidance indoors and outdoors, in every aspect of this research as

well as its surroundings and much more. These interactions were at close range and also

remotely sensed when needed.

I would like to thank Dr. Natalia Panov for her assistance dealing numbers and Mr. Alexander

Goldberg (aka Sasha), the tireless, for his help and endless ideas in the field as well as in the lab.

The former and current citizens of the Remote Sensing Laboratory do deserve my gratitude for

their help and friendship but the list of students in the past six and a half years is too long to be

mentioned here. I would like also to thank from the bottom of my heart to Mrs. Mazal Adar for

here assistance, always with patience and smile.

The help provided by Ms. Silvia Asido and Mr. Saker Al-Atrash in the lab and fields of Gilat

Research Center was essential for making this research possible.

I would like to thank my spouse Eti for all her love, support and living in the middle of nowhere

that now is not easy to leave. Last but definitely not least to my parents and sister for all their

support, advises and being who they are.

v

Table of content:

List of Figures………………………………………………..………………………….……..vi

List of Tables …………………………………………………..……………..………….……vii

Abstract ……………………………………………….…………………………………….…ix

Abbreviations …………………………………………………………………………………xiv

Chapter 1

1. Introduction …………………………..…………...…………………………………........1

1.1. Hyperspectral remote sensing and precision agriculture ……………...………1

1.2. Dissertation outline ……………………………………………………………5

Chapter 2

2. SWIR-based spectral indices for assessing nitrogen content in potato fields…....……8

Chapter 3

3. LAI assessment of wheat and potato crops by VENμS and Sentinel-2 bands…......…26

Chapter 4

4. Field spectroscopy for weed detection in wheat and chickpea fields………………….38

Chapter 5

5. Ground-level hyperspectral imagery for detecting weeds in wheat fields…………….59

Chapter 6

6. Spectral monitoring of two-spotted spider mite damage to pepper leaves……….……91

Chapter 7

7. Summary and main conclusions…………………………………………………….……99

8. Reference of the introduction and summary (chapters 1 and 7)………………………104

vi

List of Figures:

Chapter 2

Figure 1 - N content predicted vs. observed ……….………………………..……………...…18

Figure 2 - Correlating VNIR-based and SWIR-based NIs to N cont.…………………….…...19

Chapter 3

Figure 1 - Band settings of VENμS and Sentinel-2 with respect to the atmospheric

transmittance……………….…………………………………………………………………...29

Figure 2 - Spectral variation of reflectance curves……………………………………………..30

Figure 3 - VIP values as function of wavelengths ……………………………………………..32

Figure 4 - LAI correlation with REIP and NDVI ………………………………………….…..33

Chapter 4

Figure 1 - Spectral reflectance of wheat, chickpea, BLW, GW, and soil in leaf and canopy

scales………………………………………………………………………………………..…..46

Chapter 5

Figure 1 - Setup of the hyperspectral camera in a wheat field…………………………………66

Figure 2 - Decision tree for separating five classes……………….……………………………71

Figure 3 - VIP values of model #3……………………………………………………………..78

Figure 4 - Two images and their classification…………………………………………………82

Figure 5 - The amount of cross validation pixels acquired related to accuracies……….……...84

Figure 6 - Relating relative coverage by field assessment to model # 3 results………………..85

Chapter 6

Figure 1 - Leaves from the four damage levels……………………………….…...……………94

Figure 2 - Averaged reflectance values of the four damage levels …………………………….95

Figure 3 - Averaged index values and standard deviations for the damage levels…………......96

vii

List of Tables:

Chapter 2

Table 1 - R values matrix of N content ……………………….……………………….………17

Table 2 - Relationships between N observed and N predicted …………………….…….……18

Table 3 - Sr values of coupled spectral indices with respect to N content ……………………20

Chapter 3

Table 1 - Sentinel-2 bands ……………………………………….………………….….….….29

Table 2 - VENμS bands………..…………………………………………………..…….…….29

Table 3 - Measurement distribution by: crops, growing seasons, DAE and DAS …………….30

Table 4 - LAI prediction by PLS models…………………………..………………….…….…31

Table 5 - Prediction quality of LAI by NDVI (linear and exponential) and REIP (linear)…….33

Table 6 - The t-test p values of coupled r values of the same index in different data

formations……………………………………………………………………………………....34

Table 7 - The t-test p values of coupled r values of NDVI and REIP for the same data

formation…………………………………………………………………………………….…34

Chapter 4

Table 1 - Classification of pure leaf spectra by vegetation category…………………...….…..47

Table 2 - Classification of pure leaf spectra by Genera………………………………..……....48

Table 3 - Classification of canopy spectra for wheat fields based on 11 bands…………...…..49

Table 4 - Classification of canopy spectra for chickpea fields, based on 8 bands………….....50

Table 5 - Classification of canopy spectra for wheat fields, based on VENμS and ALI bands

………………………………………………………………………………………………....51

Table 6 - Classification of canopy spectra for chickpea fields, based on VENμS and ALI

bands…………………………………………………………………………………………..52

Table 7 - Herbicide application for wheat fields with vegetation cover of 0-100%……….….53

viii

Table 8 - Herbicide application for wheat fields with vegetation cover >5%…………….......54

Chapter 5

Table 1 - Distribution of cross validation pixels amongst the classes and images………….…68

Table 2 - Model # 1 cross validation of PLS-DA classification of three classes……………...73

Table 3 - Model # 2 cross validation of PLSDA classification of two classes…………..…….73

Table 4 - Model # 3 cross validation of PLSDA classification of four classes…………..……74

Table 5 - Model # 4 cross validation of PLSDA classification odel of three classes.................74

Table 6 - Model # 5 cross validation of PLSDA classification of eight classes…..…..……….75

Table 7 - Model # 6 cross validation of PLSDA classification of six classes……………..…..76

Table 8 - Normal curve deviation (Z) values of coupled PLS-DA models……………..……..76

Table 9 - Prediction of each of the images by model # 3……………..……………………….80

Table 10 - Prediction of all the images together by model # 3……………...…………………81

Chapter 6

Table 1 - Spectral bands’ centres and widths……………………………...…………………..94

Table 2 - Equations of the vegetation indices………………………..………………………...94

Table 3 - Indices separation ability between damage levels by one-way ANOVA…..……….96

ix

Abstract

Enhancing crop productivity and reducing environmental impacts are primary goals of modern

farmers. Spectral wavelengths in specific regions of the electromagnetic spectrum can be related

to the state and condition of vegetation as well as to crop species or category (broadleaf or grass).

The ability to identify, within an agricultural field, differences that are significant and important

enough to be included in the site-specific crop management (termed also “precision

agriculture”), can satisfy the above-mentioned farmers’ goals, i.e., economical profit and

environmental benefits.

The current dissertation is aimed at using spectral data and respective techniques to explore four

potential field crop applications that could be used in framework of precision agriculture. These

applications are: (1) assessing nitrogen (N) content in potato plants; (2) assessing Leaf Area

Index (LAI) in wheat or potato fields; (3) detecting weeds in wheat and chickpea fields; and (4)

identifying two-spotted spider mite (TSSM) damage to greenhouse pepper leaves. These four

applications are presented in five studies as the weed detection topic is presented by two studies.

The main knowledge gaps that were explored are: vegetation indices that are not directly related

to N content in plants are the main assessors of it; potential ability to assess LAI by future

satellites; inadequate ability to spectrally separate between crops and weeds on ground level in

the critical period of weed control; spectral assessment of TSSM damage to leaves. Specific

objectives were formulated for each of the studies.

Assessing nitrogen (N) content in potato plants

N is an essential element in plant growth and productivity, and N fertilizer is therefore of prime

importance in cultivated crops. The amount and timing of N application has economic and

environmental implications and is consequently considered to be an important issue in precision

agriculture. Spectral indices derived from handheld, airborne and spaceborne spectrometers are

used for assessing N content. The majority of these indices are based on indirect indicators,

mostly chlorophyll content, which is proven to be physiologically linked to N content. The

current research aimed to explore the performance of new N spectral indices dependent upon the

shortwave infrared (SWIR) region (1200–2500 nm), and particularly the 1510 nm band because

it is related directly to N content. Traditional nitrogen indices (NIs) and four proposed new

SWIR-based indices were tested with canopy-level spectral data obtained during two growing

x

seasons in potato experimental plots in the northwest Negev, Israel. Above-ground biomass

samples were collected at the same location of the spectral sampling to provide in-situ N content

data. The performance of all indices was evaluated by three methods: (1) correlations between

the existing and proposed indices and N as well as correlations among the indices themselves;

(2) the root mean square error prediction (RMSEP) of the N content; and (3) the indices relative

sensitivity (Sr) to the N content. The results reveal a firm advantage for the proposed SWIR-

based indices in their ability to predict, and in their sensitivity to, N content. The best index is

one that combines information from the 1510 and 660 nm bands but no significant differences

were found among the new SWIR-based indices. Since these vegetation indices are combining

direct and indirect relation to N content they are related to the presence of N in the plant as well

as to the repercussions of it. Therefore the SWIR-based indices are concluded to be better

predictors and more sensitive to N content.

Assessing Leaf Area Index (LAI) in wheat and potato fields

Leaf Area Index (LAI) is an important variable that governs canopy processes and can be

monitored by satellites. The current study aims at exploring the potential and limitations of using

the red-edge spectral bands of the forthcoming superspectral satellites, namely – Vegetation and

Environmental New Micro Spacecraft (VENμS) and Sentinel-2, for assessing LAI in field crops.

The research was conducted in experimental plots of wheat and potato in the northwestern

Negev, Israel. Continuous spectral data (400-1000 nm) were collected by a field spectrometer

and LAI data were obtained by a ceptometer. The spectral data were resampled to the

superspectral VENμS and Sentinel-2 resolutions. The data were divided into seven datasets (four

seasons, two crops, and one including all data). The LAI prediction abilities by Partial Least

Squares (PLS) models for continuous spectra and the resampled spectra were compared and

evaluated. For wheat and potato of the continuous, VENμS, and Sentinel-2 data formations, the

PLS correlation coefficients (r) values were 0.93, 0.93, and 0.92, respectively. In most cases, the

red-edge region was found to be the most important spectral region for the three data formations,

according to the Variable Importance in Projection (VIP) analysis. Additionally, Normalized

Difference Vegetation Index (NDVI) and the Red-Edge Inflection Point (REIP) were computed

for the three data formations in order to observe relation to as well as prediction accuracy in

retrieving LAI values. The prediction abilities of the calculated indices by the data formations

were compared, peaking for wheat, with r values of 0.91 for the REIP for the three data

xi

formations. While the r values for potato were 0.73 or 0.72 for the three data formations. These

differences in r values are significant and can be related to the uniformity of the crops that

influences on the amount of soil measured in the plants readings. Therefore, it is concluded that

VENμS and Sentinel-2 can spectrally assess LAI as good as a hyperspectral sensor. The REIP

was found to be a significantly better predictor than NDVI for wheat data and therefore can

potentially be implemented for future LAI monitoring applications by superspectral sensors that

contain four red-edge bands.

Detecting weeds in wheat and chickpea fields by a point field spectrometer

Weed control is commonly performed by applying selective herbicides homogeneously over

entire agricultural fields. However, applying herbicide only where needed could have economic

and environmental benefits. The objective of this study was to apply remote sensing into

detection of grasses and broadleaf weeds among cereal and broadleaf crops. Spectral relative

reflectance values at both leaf and canopy scales were obtained by field spectroscopy for four

plant categories: wheat; chickpea; grass weeds; and broadleaf weeds. Total reflectance spectra

of leaf tissues for botanical genera were successfully classified by General Discriminant Analysis

(GDA) showing total classification higher than 99%. The total canopy spectral classification by

GDA for specific narrow bands was 95 ±4.19% for wheat and 94 ±5.13% for chickpea. The total

canopy spectral classification by GDA for future Vegetation and Environmental New

MicroSpacecraft (VENμS) bands was 77 ±8.09% for wheat and 88 ±6.94% for chickpea and for

the operative satellite Advanced Land Imager (ALI), bands was 78 ±7.97% for wheat and 82

±8.22% for chickpea. Within the critical period for weed control, overall classification accuracy

of 87 ±5.57% was achieved for >5% vegetation coverage in a wheat field thereby providing

potential for implementation of herbicide applications. Therefore it is concluded that the spectral

characteristics of the pure leaf spectra enable precise classification of the different plant

categories and genera. The red-edge is highly important for crops and weeds classification.

Spectral separation of crops and weeds is potentially useful for wheat fields with >5% vegetation

cover in the critical period for weed control. Qualitative models based on wheat, chickpea, grass

weed and broadleaf weed spectral properties have high quality classification and prediction

potential that can be used for site specific weed management.

xii

Ground-level hyperspectral imagery for detecting weeds in wheat fields

The study used ground-level image spectroscopy data, with high spectral and spatial resolutions,

for detecting annual grasses and broadleaf weeds in wheat fields. The image pixels were used to

cross validate Partial Least Squares Discriminant Analysis (PLS-DA) classification models. The

best model was chosen by comparing the cross validation confusion matrices in term of their

variances and Cohen’s Kappa values. This best model used four classes: broadleaf weeds, grass

weeds, soil, and wheat. Each of the classes contains sunlit and shaded data together. The

Variable Importance in Projection (VIP) method was applied in order to locate the most

important spectral regions for each of the classes. It was found that the red-edge is the most

important region for the vegetation classes. Ground truth pixels were randomly selected and their

confusion matrix resulted in total accuracy of 72%. The results obtained are reasonable although

the model used wheat and weeds from different growth stages, acquiring dates, and fields. It is

concluded that high spectral and spatial resolutions can provide separation between wheat and

weeds based on spectral data alone. Sunlit vegetation can be better classified than shaded

vegetation. Composition of four classes (broadleaf weeds, grass weeds, wheat, and soil) was the

best for weed detection in the current dataset. The red-edge is the most important region for

separation among wheat, BLW, and GW. The model’s cross validation and ground truth were

acquired for heterogenic data and the model obtained reasonable results. Therefore, there is

feasibility for up-scaling the spectral methods to air or spaceborne sensors as well as developing

ground-level application.

Identifying two-spotted spider mite (TSSM) damage to greenhouse pepper leaves

Two-spotted spider mites (TSSM; Tetranychus urticae Koch) cause significant damage to crops

and yields, in the field as well as in greenhouses. By feeding, TSSM destroy chloroplast-

containing cells; this damage can be spectrally detected in the reflectance of the visible and near-

infrared regions. This study focuses on hyperspectral reflectance data of greenhouse pepper

(Capsicum annuum) leaves, obtained by integrated sphere. The averaged reflectance curves of

each of the four damage levels were compared and spectral differences were identified. The

reflectance data were transformed into six vegetation indices allowing early TSSM damage

detection by separation between leaf damage levels. One-way analysis of variance of coupled

damage levels was applied to each of the vegetation indices. The results show that five out of six

xiii

indices can significantly separate between leaves with no damage and these with low damage.

Only the REIP and two other indices can significantly separate between the four damage levels.

Therefore, it was concluded that early identification of TSSM damage to greenhouse pepper leaf

can be obtained by multispectral means. REIP can provide consistent separation between

increasing damage levels of greenhouse pepper leaves in the laboratory. Furthermore, the

proposed methods may identify the damage on the upper side of the leaves although the TSSM

feed on the underside of leaves.

To conclude, several insights can be inferred from the wide perspective of the current

dissertation. These are: (1) Spectral data in the electromagnetic spectrum are capable to detect

and monitor different phenomena and properties of agricultural crops. These data can be of

continuous hyperspectral spectra, multi or super spectral spectra, and VIs based on both spectral

resolutions. Therefore, spectral data acquired by a variety of ground-level, airborne, and

spaceborne sensors can be applied for precision farming and site-specific management. (2) The

red-edge region is highly important spectral region for several potential field crop applications

e.g., LAI, weed detection, and monitoring TSSM damage to leaves. The importance of this

region has been recently recognized by the remote sensing community and appropriate spectral

bands are now designed in the forthcoming space systems (e.g., VENμS and Sentinel-2). (3) It

was found that multi or super spectral data perform as good as continuous hyperspectral data.

This might be beneficial for applying spectral data with reduced cost and increased area of

coverage.

xiv

Abbreviations

ALI – Advanced Land Imager

ANOVA – Analysis Of Variance

ASD – Analytical Spectral Devices

BLW – Broadleaf Weeds

cm – centimeter

DAE – Days After Emergence

DAS – Days After Seeding

DOY – Day Of Year

DT – Decision Tree

FOV – Field Of View

g – gram

GDA – General Discriminant Analysis

GW – Grass Weeds

h – hour

Kgha-1 – kilogram per hectare

km – kilometer

LAI – Leaf Area Index

m – meter

MCARI – Modified Chlorophyll Absorption in Reflectance Index

mm – millimeter

N – Nitrogen

NDVI – Normalized Difference Vegetation Index

NIR – Near Infrared

NIs – Nitrogen Indices

xv

nm – nanometer

NRI – Normalized Ratio Index

PLS – Partial Least Squares

PLS-DA – PLS Discriminant Analysis

r – Correlation coefficient

R2 – Coefficient of determination

REIP – Red-Edge Inflection Point

RMSEP – Root Mean Square Error Prediction

sec – seconds

Sr – Relative Sensitivity

SWIR – Shortwave Infrared

TCARI – Transformed Chlorophyll Absorption in Reflectance Index

TSSM – Two-Spotted Spider Mites

VENμS – Vegetation Environmental New Micro Satellite

VIP – Variable Importance in Projection

VIs – Vegetation Indices

VIS – Visible

VNIR – Visible Near Infrared

μm – micrometer

1

Chapter 1

1. Introduction

1.1. Hyperspectral remote sensing and precision agriculture

Hyperspectral remote sensing (RS) is based on the examination of hundreds of contiguous narrowly

defined spectral bands (Campbell 1996, Chang 2003). A range of sensors with a variety of spectral

and spatial resolutions are installed on different levels of platforms (i.e., ground, airborne, and

spaceborne level). Sensors that produce radiation and measure its reflectance, transmittance, or

absorbance are known as active, and those using the naturally available radiation are known as

passive. Sensors can acquire hyperspectral data by either point or image concept. The former

concept is a measurement of a certain target or an area that provides one spectrum that integrates all

the data available by the field of view. The image concept puts together many points of

measurements, i.e., pixels, to create a rectangular image. Such an image can contain many pixels,

each of these being composed of a specific spectrum that represents the detected feature. This

concept is known as a data cube in which the x and y dimensions create the spatial domain of pixels

and the z dimension the spectral domain (Campbell 1996).

Hyperspectral imaging improves the capability of multispectral imaging by identifying materials

that could not be resolved previously (Chang 2003). On the other hand, hyperspectral data create

relatively large amounts of data of which a significant part might be useless for specific application

(e.g., nitrogen content in plants; Herrmann et al. 2010). For specific applications (e.g., leaf area

index assessment), the hyperspectral data might not provide significantly better results than those

obtained by multi (up to 10 bands)- or super (10 to 50 bands)-spectral sensors (Herrmann et al.

2011). The latter systems are relatively more expensive, and analyzing the data can be time- and

computer-memory consuming. Hyperspectral data processing requires more complicated methods

than for multi- or super-spectral data. Therefore, in this dissertation in addition to analyzing

hyperspectral data as is, a study was conducted to explore the important spectral bands for specific

agricultural applications. In doing so, in some of the cases, the data were resampled to multi- or

super-spectral data as well as used for vegetation indices (VIs) computation.

The current dissertation presents an analysis of spectral data obtained by a point spectrometer for

nitrogen (N) content assessment in potato plants, leaf area index (LAI) estimation in wheat and

potato plants, weed detection in wheat and chickpea fields, and an assessment of spider mite

2

damage to greenhouse pepper leaves. In addition, data analyses were obtained by a hyperspectral

scanner for weed detection in wheat fields.

The spectral regions used in this dissertation are: visible , near infrared (NIR), and shortwave

infrared (SWIR); their boundaries are: 0.4-0.7 nm, 0.7-1.1 nm , and 1.1-2.5 nm, respectively (Ben-

Dor et al. 2003). These boundaries are not precise and can be affected by the intensity of the plant’s

biophysiological and structural properties. Vegetation reflectance values of each spectral region are

affected mainly by one property. With respect to vegetation monitored by remote sensing means,

pigments determine most spectral features in the visible region, structure controls the reflectance

intensity in the NIR region, and water content rules the SWIR region.

The blue and red wavelengths absorb 70 to 90 % of the incident radiation due to chlorophyll

involvement in the photosynthetic process, and therefore, the reflectance values are relatively low

(Gausman 1985, Verbyla 1995). The NIR radiation is scattered at the interfaces of cell walls and

intercellular air spaces in the spongy mesophyll, resulting in a relatively high reflectance of

radiation in the NIR region (Gausman 1985, Campbell 1996). A broadleaf leaf reflects relatively

higher in the NIR region than a grass leaf of the same thickness and age due to differences in leaf

structure (Gausman 1985, Raven et al. 2005). Canopy reflectance in the NIR radiation is usually

amplified since it penetrates leaves, in both directions, and is reflected from others (Campbell

1996). Hence, NIR reflectance is dependent on the canopy thickness and structure that can be

affected by several factors, e.g., species and stage of growth. In the SWIR region, there are

reflectance sinks at 1450 and 1950 nm due to absorption by the plant water content (Gausman 1985,

Tian et al. 2001). When radiation enters a plant canopy, multiple events of reflection, transmission,

absorption, scattering, and diffraction occur (Borregaard et al. 2000). The reflectance of a canopy,

in agricultural fields, will, in most cases, include shaded leaves and might include stems, flowers,

fruits, and some kind of background soil that might be partly shaded and of different humidity

levels. Spectral data analyzers should be aware of this.

An additional spectral region, specifically defined for vegetation, is the red-edge. The red-edge is

the steep increase of reflectance that marks the transition between the visible region of the red

chlorophyll absorption feature and the high reflectance values of the NIR plateau and, therefore,

occurs between 680 and 750 nm (Horler et al. 1983). It can be accurately defined as the position (in

terms of wavelengths) of the inflection point of this slope in the spectral reflectance of vegetation

(Pu et al. 2003, Mutanga and Skidmore 2007). Among other variables, e.g., chlorophyll content, the

3

red-edge position is influenced by the LAI (Moran et al. 2004) and, therefore, can be used for its

assessment. It is also an important element in plant species segregation (Vrindts et al. 2002, Smith

and Blackshaw 2003) and, therefore, potentially essential for weed – crop separation.

Spectral data of vegetation can be analyzed by the entire spectrum available, parts of the spectrum,

VIs, or specific wavelength. VIs, as listed by Bannari et al. (1995), are different combinations of red

and infrared channels’ reflectance values. A variety of VIs are widely employed as measures of

green vegetation density, vigor, and productivity (Bannari et al. 1995) as well as being highly

correlated to specific growth stages (Pimstein et al. 2007) and biophysiological variables of plants

(Hardinsky et al. 1983, Ceccato 2001, Fensholt and Sandholt 2003, Fitzgerald et al. 2006,

Rodriguez et al. 2006, Tilling et al. 2006, Tilling et al. 2007). Some VIs lose sensitivity in relatively

low values of the biophysiological variable they are correlated to (e.g., Buschmann and Nagel 1993,

Pimstein et al. 2007); some VIs’ relation to one biophysiological variable can be influenced by

another variable (El-Shikha et al. 2007), and many of them are influenced by the background

(Haboudane et al. 2002).

This dissertation presents analyses by the entire spectra as well as by VIs. N content was assessed

in potato plants by VIs in the visible-NIR (VNIR) and SWIR regions. The entire spectrum and VIs,

in the VNIR region, were also applied in order to assess LAI in wheat and potato plants. Weed

detection in wheat and chickpea fields were obtained by the entire spectrum in the VNIR region.

Separating the intensity of spider mite damage to greenhouse pepper leaves was analyzed by VIs in

the VNIR region.

Precision agriculture (PA; also known as site-specific management) was defined by Pierce and

Nowak (1999) as “the application of technologies and principles to manage spatial and temporal

variability associated with all aspects of agricultural production for the purpose of improving crop

performance and environmental quality.” PA is the management of agricultural crops at a spatial

scale smaller than the entire field (Moran et al. 1997, Plant 2001). To justify PA implementations,

there should be within-field variability (spatial or temporal) that is significant, identifiable, and

measurable, which exists in factors that are connected to or that influence the crop yield. Enhancing

crop productivity, monitoring fertilization, improving weed management, and reducing

environmental impact are the goals of modern farmers (Younan et al. 2004). To achieve these goals,

managers of agricultural lands are in need of an information-based crop management system or the

means for improving such an existing system. Such information concerning crops can be attained

4

by accurate remote sensing devices and processing techniques (Moran et al. 1997, Asner 1998,

Moran et al. 2004).

PA has been made possible and is becoming more and more practical by the invention,

improvement, and growing availability of modern technologies such as global positioning systems

(GPS), yield sensors, variable-rate chemical applications, geographical information systems, and

last but not least – RS through its various spatial, spectral, and temporal resolutions. In general, the

practical implementation of PA is a cycle including: measuring the field spatial variability and data

analysis, mapping the variability, applying the right treatment in the exact location at the right time,

and assessing the economic and environmental implications. As mentioned by Plant (2001), there

are three ways to measure field spatial variability: continuously (e.g., on-the-go monitoring),

discretely (e.g., point sampling of soil or plant properties), and remotely (e.g., hyperspectral or

multispectral images obtained by airborne or spaceborne sensor). It is important to mention that

spectral sensors and RS techniques can be at the service of each of these three ways to measure field

spatial variability. This dissertation analyzes data obtained on the ground level, while these results

are the first step towards applications on the air and space levels.

The hypothesis that RS is a valid tool for PA is well established, as mentioned above. In order to

provide potential RS applications for PA, there is a need to be more specific; therefore, the main

goal of exploring new RS methods for use in conjunction with PA is divided into several objectives:

• Improve the ability to evaluate total N content in plants based on direct and indirect relation

to it by spectral data from the SWIR and VNIR regions.

• Explore the potential of future super-spectral satellites to retrieve LAI values in wheat and

potato crops using the visible, red-edge, and NIR spectral regions by entire spectra as well

as by known VIs.

• Use point leaf spectral reflectance to distinguish between wheat, chickpea, grass weeds, and

broadleaf weeds as well as examine the potential of using point canopy spectral reflectance

data from the field to predict categories of crops and weeds.

• Examine ground-level image spectroscopy data, with high spectral and spatial resolutions,

ability to detect annual grass and broadleaf weeds in wheat fields by choosing the best class

determination for separating broadleaf weed, grass weed, and wheat, as well as discovering

the most important spectral bands for the classification.

5

• Explore the possibility to spectrally detect early damage of two-spotted spider mites (TSSM)

to greenhouse pepper leaves.

1.2. Dissertation outline

The current chapter includes a general introduction with objectives and dissertation outline. This

dissertation contains three published and one accepted peer-reviewed articles. In addition, the fifth

article has been submitted to a peer-reviewed journal. All these studies show methodology and

methods that can lead to RS applications for PA.

Chapter 2 – Herrmann, I., Karnieli, A., Bonfil, D., J., Cohen, Y., and Alchanatis, V., 2010, SWIR-

based spectral indices for assessing nitrogen content in potato fields. International Journal of

Remote Sensing, 31, pp. 5127-5143.

The article deals with N content assessment in potato plants by spectral means. Since N and

chlorophyll contents are proven to be physiologically linked, the majority of VIs for assessing N

content in plants are based on chlorophyll content. This article is aimed at exploring the

performances of new N content VIs that combine direct and indirect spectral relation to N content.

Aboveground biomass samples were collected at the same location of the spectral sampling, to

provide actual N content in-situ data. The indices’ performances were evaluated by three methods:

correlations between existing and proposed VIs and N as well as correlations among the indices

themselves; the Root Mean Square Error Prediction (RMSEP) of N content; and the VIs’ relative

sensitivity (Sr) to N content. The results revealed a firm advantage for the proposed indices in the

ability to predict N content as well as in showing sensitivity to it.

Chapter 3 – Herrmann, I., Pimstein, A., Karnieli, A., Cohen, Y., Alchanatis, V. and Bonfil, D., J.,

2011, LAI assessment of wheat and potato crops by VENμS and Sentinel-2 bands. Remote Sensing

of Environment, 115, pp. 2141-2151.

The article explores the potential and the limitations of LAI assessment, in wheat and potato plants,

by the spectral bands of the forthcoming super-spectral satellites, namely – Vegetation and

Environmental New Micro Spacecraft (VENμS) and Sentinel-2. LAI prediction abilities by

Hyperspectral data were evaluated and compared to the satellites spectral data by Partial Least

Squares (PLS) models. Based on the Variable Importance in Projection (VIP) analysis, in most of

the cases, the red-edge region was found to be the most important region for LAI assessment.

Additionally, the VIs’ Normalized Difference Vegetation Index (NDVI) and Red-Edge Inflection

6

Point (REIP) were computed in order to observe relation to, as well as prediction accuracy in,

retrieving LAI values. It was concluded that VENμS and Sentinel-2 can spectrally assess LAI as

effectively as a hyperspectral sensor and that the red-edge region is highly important for LAI

assessment by the entire spectra as well as by vegetation index.

Chapter 4 – Shapira, U., Herrmann, I.∗, Karnieli, A., and Bonfil, D., J., 2012, Field spectroscopy for

weed detection in wheat and chickpea fields. International Journal of Remote Sensing, in press.

The article examines the ability to detect weeds, grass as well as broadleaf, among wheat and

chickpea crops. General Discriminant Analysis (GDA) models were applied in order to spectrally

separate between the different classes on a leaf as well as a canopy level. Red-edge bands were

found to be highly important for classification by the entire spectra. When resampled to VENμS

and Advanced Land Imager (ALI) spectra, there was no significant advantage for either of the

satellites. It was concluded that qualitative models, based on wheat, chickpea, grass weed and

broadleaf weed spectral properties, have high quality classification and prediction potential that can

be used for site-specific weed management.

Chapter 5 – Herrmann, I., Shapira, U., Kinast, S., Karnieli, A., and Bonfil, D., J., 2012, Ground-

level hyperspectral imagery for detecting weeds in wheat fields, submitted to Precision Agriculture.

This manuscript also deals with weed detection in wheat fields but adds the spatial aspect. PLS -

Discriminant Analysis (PLS-DA) classification models were built based on data obtained from

ground-level hyperspectral images with high spatial resolution. The most suitable model was

chosen by comparison of cross-validation confusion matrices. The VIP analysis found that the red-

edge is the most important region for the vegetation class separation. The model was applied to the

images in order to allow ground truth. The results show feasibility for potential up-scaling of the

spectral methods to air or spaceborne sensors, as well as for developing ground-level application.

Chapter 6 – Herrmann, I., Berenstein, M., Sade, A., Karnieli, A., Bonfil, D.J. and Weintraub, P.G.,

2012, Spectral monitoring of two-spotted spider mite damage to pepper leaves. Remote Sensing

Letters, 3, pp. 277-283.

This study deals with spectral assessment of TSSM damage to greenhouse pepper leaves. There

were four levels of damage ranging from no damage to high damage. The spectral reflectance data

were transformed to VIs. The one way analysis of variance (ANOVA) was applied in order to check

∗ Equal contribution

7

if these indices’ values of the different damage levels were significantly different from one another.

It was concluded that early damage identification can be obtained by multispectral means.

Furthermore, the proposed methods may identify the damage on the upper-side of leaves even

though the TSSM feed on the under-side of leaves.

Chapter 7 presents the dissertation summary, conclusions, and contribution, as well as the study’s

significance and recommendations for future research.

8

Chapter 2

2. SWIR-based spectral indices for assessing nitrogen content in potato fields.

International Journal of Remote Sensing, 31, pp. 5127-5143.

26

Chapter 3

3. LAI assessment of wheat and potato crops by VENμS and Sentinel-2 bands. Remote

Sensing of Environment, 115, pp. 2141-2151.

38

Chapter 4

4. Field spectroscopy for weed detection in wheat and chickpea fields. International

Journal of Remote Sensing, in press.

39

Field Spectroscopy for Weed Detection in Wheat and Chickpea Fields*

URI SHAPIRA**†, ITTAI HERRMANN**†, ARNON KARNIELI†, DAVID J. BONFIL‡

† The Remote Sensing Laboratory, Jacob Blaustein Institutes for Desert Research, Ben-Gurion University of the Negev,

84990, Israel

‡ Field Crops and Natural Resources Department, Agricultural Research Organization, Gilat Research Center, 85280,

Israel

Weed control is commonly performed by applying selective herbicides homogeneously over entire

agricultural fields. However, applying herbicide only where needed could have economical and

environmental benefits. The objective of this study was to apply remote sensing into detection of

grasses and broadleaf weeds among cereal and broadleaf crops. Spectral relative reflectance values

at both leaf and canopy scales were obtained by field spectroscopy for four plant categories: wheat;

chickpea; grass weeds; and broadleaf weeds. Total reflectance spectra of leaf tissues for botanical

genera were successfully classified by General Discrimination Analysis (GDA). The total canopy

spectral classification by GDA for specific narrow bands was 95 ±4.19% for wheat and 94 ±5.13%

for chickpea. The total canopy spectral classification by GDA for future Vegetation and

Environmental monitoring on a New Micro-Satellite (VENμS) bands was 77 ±8.09% for wheat and

88 ±6.94% for chickpea and for the operative satellite Advanced Land Imager (ALI), bands was 78

±7.97% for wheat and 82 ±8.22% for chickpea. Within the critical period for weed control, overall

classification accuracy of 87 ±5.57% was achieved for >5% vegetation coverage in a wheat field

thereby providing potential for implementation of herbicide applications. Qualitative models based

on wheat, chickpea, grass weed and broadleaf weed spectral properties have high quality

classification and prediction potential that can be used for site specific weed management.

* Accepted for publication in the International Journal of Remote Sensing.

** U. Shapira and I. Herrmann contributed equally to this work.

40

1. Introduction

Weeds are the most acute pest in agriculture with estimated annual global damage of around 40

billion US dollars (USD) per year (Monaco et al. 2002). Weeds reduce crop yield and quality by

competing with crops for water, sunlight, and minerals (Pinter et al. 2003, Slaughter et al. 2008);

producing allelopathic substances (Moran et al. 2004); hosting diseases and insects (Pikart et al.

2011, Papayiannis et al. 2012); and disturbing tilling and harvesting (Monaco et al. 2002). One

increasing problem is weed resistance to herbicides (Mallory-Smith et al. 1990, Jones et al. 2005,

Marshall and Moss 2008). In Australia alone, herbicide resistance is estimated to impose an

additional annual cost of more than 1 billion USD (Gibson et al. 2008).

More than 60% of the pesticides developed worldwide are herbicides (Monaco et al. 2002). Thus, it

is not surprising that herbicides are also the most common pesticide found in ground water (Manh et

al. 2001). Herbicides can be an environmental hazard to fauna as well as humans (Dhawan et al.

2009, Brent and Schaeffer 2011). Consequently, the amount of herbicides that can be applied per

area unit is restricted in some countries (Biller 1998, Timmermann et al. 2003, Slaughter et al.

2008). Herbicide use regulations, consumer concerns, and growing interest in organically produced

foods, limit the long-term acceptability of herbicide applications (Slaughter et al. 2008).

Weed distribution in fields is non-uniform and confined to patches of varying size along field

borders (Gerhards et al. 1997, Lamb and Brown 2001, Vrindts et al. 2002, Gerhards and

Christensen 2003, Moran et al. 2004, Slaughter et al. 2008, Weis et al. 2008). Application of

herbicides on a field is often based on the previous year's weed problems and information obtained

from field scouting (Manh et al. 2001, Moran et al. 2004). By significantly reducing the quantity of

herbicide applied (Gerhards et al. 1997, Gerhards and Christensen 2003, Timmermann et al. 2003,

Eddy et al. 2006, Slaughter et al. 2008, Weis et al. 2008), site-specific weed control and

management could economically benefit farmers and consumers as well as the environment without

diminishing weed control efficacy (Pinter et al. 2003, Slaughter et al. 2008, Weis et al. 2008).

Reducing the amount of applied herbicides should reduce the probability of weeds building

resistance to herbicides and increase herbicide effectiveness.

According to Lindquist et al. (1998), it is possible to reduce the quantity of herbicide applied by

applying herbicides only where weeds are located. Site specific weed management has reduced

herbicide use by 11 to 90% without affecting crop yield (Brown et al. 1994, Brown and Steckler

1995, Johnson et al. 1995, Lindquist et al. 1998, Feyaerts and van Gool 2001, Gerhards and

41

Christensen 2003). Weed distribution and competition with crops are influenced by spatial

variability in topography, drainage, soil type, and microclimate. There is significant variation in

weeds within and between different fields (Moran et al. 2004) emphasizing the need for site-

specific weed management.

Real-time (on the go) non selective weed detection and control can be implemented by means of

tractor mounted, optical sensors that trigger a spray nozzle valve to open briefly upon detection of

green vegetation (Bennett and Pannell 1998, Biller 1998, Blackshaw et al. 1998). This approach can

be applied to entire fields before crop emergence or between the crop rows after emergence (Moran

et al. 1997, Alchanatis et al. 2005, Slaughter et al. 2008). Other ground-based, on the go sensing

methods are designed to detect the shape of weed leaves against a light-toned soil background and

thus, can be applied only in early growing stages (Gerhards and Christensen 2003, Weis et al.

2008). In addition, remote sensing from air or space has been used to identify and map weeds prior

to herbicide application (Gerhards et al. 1997, Weis et al. 2008). Remote sensing techniques can

provide fast and cost-effective mapping of weed populations over large areas, which otherwise

would be impractical to cover by manual ground survey methods (Zwiggelaar 1998, Hamouz et al.

2008). Remote sensing applications also allow early and late season, and spatial and spectral

methods for site specific weed detection and management (Zwiggelaar 1998, Moran et al. 2004,

Alchanatis et al. 2005).

Few studies have dealt with ground-level spectral classification of crops and weeds over multiple

growing seasons. Lopez-Granados et al. (2008) classified ground-level spectral reflectance of

wheat, four grass weeds, and soil, and concluded that one sampling date per growth season, when

phenological distinction is maximal, can provide high quality classification. However, relying upon

phenology for basing spectral differences will likely be ineffective if the optimal time for herbicide

application precedes the date of maximal phenological variability among crops and weeds. In their

extensive review, Gray et al. (2009) determined that short wave infrared (SWIR) bands are

important for classifying ground level spectral reflectance of soybean, six broadleaf weeds, and soil.

Likewise, Slaughter et al. (2008) found many reports in the literature in which studies were

conducted in ideal conditions with no spatial overlap of crops and weeds, which resulted in

classification accuracies of 65% to 95%. Zwiggelaar (1998) mentions in his review that using

selected wavelengths for the discrimination between row crops and weeds has not been shown so

far, and imaging using a limited number of wavelengths might not be sufficient. The first step

42

required to distinguish between crops and weeds is to obtain continuous spectra of pure plants of

each species, which can be done by means of data having high spatial and spectral resolutions.

Vrindts et al. (2002) employed such data and found that relative reflectance values were needed to

classify crops and weeds, and minimize the effect of different lighting conditions on spectral data.

In addition, their use of several wavebands resulted in high classification accuracy.

A dicotyledonous leaf has more air spaces within its spongy mesophyll tissue than a

monocotyledonous leaf (Raven et al. 2005) of the same thickness and age resulting in a higher

reflectance in the NIR region (Gausman 1985). The red-edge region is the slope connecting the red

(R) and Near Infra-Red (NIR) regions in the reflectance spectra of plants. It is an important element

in spectral separation of different plant species including weeds and crops (Vrindts et al. 2002,

Smith and Blackshaw 2003, Herrmann et al. 2011).

The Advanced Land Imager (ALI) is a multispectral sensor with nine bands, including two in the

red-edge region, onboard the Earth Observing-1 (EO1) satellite that provides spatial resolution of

30 m, swath of 37 km, and revisit frequency of 16 days (Chander et al. 2009). The Vegetation and

Environmental monitoring on a New Micro-Satellite (VENμS) is a future satellite with super-

spectral sensor (12 bands in visible, red edge, and NIR regions). VENμS will provide excellent

spatial resolution of 5.3 m, 27.5 km swath, and revisit frequency of 2 days (Herrmann et al. 2011).

These specifications of VENμS are very suitable for site specific weed management and other

precision agricultural applications.

In this research, remote sensing was used to detect annual grasses and broadleaf weeds amongst

broadleaf and cereal crops. Specific objectives were twofold: (1) use leaf spectral reflectance to

distinguish between wheat, chickpea, grass weeds (GW), and broadleaf weeds (BLW); and (2)

examine the potential of using canopy spectral reflectance from the field and band equivalent

reflectance of VENμS and ALI to predict categories of crops and weeds.

2. Methodology

2.1 Study sites

Field measurements and sampling were performed in rainfed as well as irrigated wheat and

chickpea experimental plots in winter 2007 and 2008 at the Gilat Research Center (31o20’N,

34o40’E) and Kibbutz Saad (31°28'N, 34°33'E) in the northwest Negev, Israel. The climate is semi-

43

arid with a short rainy season (November-April) that yields an average annual precipitation of 230

at Gilat and 385 mm at Kibbutz Saad (Har Gil et al. 2011). Soils are Calcic Xerosols with sandy

loam texture formed from alluvium and loess on shallow hills with average elevation of 80-150 m

above sea level (Kafkafi and Bonfil 2008).

2.2 Field measurements and sampling design

The Analytical Spectral Devices (ASD, Inc., Boulder, CO, USA) FieldSpec Pro FR spectrometer

was used to measure the spectral reflectance from plants at leaf and canopy scales in early growth

stages. The ASD instrument operates over a range of 400-2400 nm with a spectral sampling

resolution of 1.4 nm for 400–1000 nm and 2 nm for 1000–2400 nm. Spectra were resampled to 5

nm bands by means of linear interpolation. Atmospheric water absorption spectral regions (1350-

1420nm and 1800-1960nm) were then eliminated from the resampled spectra. This range of 400-

2400nm is defined hereafter as all wavebands. Spectral measurements were taken at leaf and canopy

scales. The high intensity contact probe of the ASD radiometer was used to obtain leaf scale spectra

as needed to determine the feasibility of spectrally separating between crops, and broadleaf weeds

or grasses. The total number of leaf spectral samples obtained by the contact probe was 608, with

the following distribution: wheat 63, chickpea 57, GW 136, and BLW 352. All leaf spectral samples

were acquired 30-40 Days After Emergence (DAE) of the crops.

The bare fiber adaptor of the ASD instrument was used to also collect canopy reflectance data at

solar noon ±1 h, under clear sky conditions with a bare fiber adaptor that was leveled in a nadir

angle. Reference measurements of spectral reflectance were obtained periodically using a standard

white reference panel (Spectralon Labsphere Inc.). At a viewing angle of 25°, the Field of View

(FOV) was a circle with radius of ~32 cm when the bare fiber optic adaptor was held 1.4 m above

ground. This radius was large enough to include a few plants in the FOV while being small enough

to select only one category of plants (e.g., wheat, chickpea, BLW, or GW). Since canopy spectral

measurements were obtained during the early growing stages of crops and weeds, and the height of

the probe was fixed, it was assumed that changes in the FOV between different measurements were

negligible. Wheat or chickpea crops, GW, and BLW were separately measured as sole targets in the

FOV against a soil background. Soil was also measured as sole target.

To explore the spectral feasibility and limitations of satellites, the ground spectral data were

resampled to VENμS bands by averaging these spectra in the range of each of the bands (Herrmann

et al. 2011) and to EO1-ALI bands by averaging with respect to the spectral response in the range of

44

each of the bands (Mendenhall et al. 2000). On ground level, a pixel or FOV could be obtained

containing one target (e.g., wheat, BLW, GW, and soil) versus airborne or spaceborne sensors

where it is likely that each pixel will contain number of targets. Mixed pixels will be a combination

of the spectra of the targets they contain (Biewer et al. 2009). Spectral measurements of wheat with

BLW, and wheat with GW, both with soil background, were acquired to examine the effect of

mixed pixel of crop and weeds on the classification quality for wheat fields.

The spectral data from each season were randomly divided into calibration (50%) and validation

(50%) data sets. Validation statistics were computed to assess the accuracy of the calibration.

Number of samples changed with crop, DAE, and relative coverage of vegetation. Canopy spectral

sampling sites were distributed in the crop fields based on presence or absence of weeds. The

canopy spectral samples were acquired 8-57 DAE in wheat fields and 10-79 DAE in chickpea

fields.

The vegetation coverage (Deardorff 1978) was assessed at every spectral measurement. To do this,

a 50 by 60 cm rectangular frame (the same area as the FOV) was placed in the center of the FOV.

The rectangular frame was divided to 20 equal size squares. The assessment was done for each of

the squares and accumulated, with 5% weight per square, to include the entire area surrounded by

the frame. All assessments were performed by the same person. Since herbicides are to be applied

before closure of the crop canopy (Thorp and Tian 2004), the spectral samples were obtained in

early growth stages of crops and weeds. The classification analysis was mainly limited to plots with

>30% vegetation coverage to reduce the negative effect of soil background on the crop canopy

reflectance while classification of data with 0-100% as well as >5% vegetation cover were applied

for specific cases as presented in the results.

2.3 General Discriminant Analysis.

Qualitative classification analysis was applied by the General Discriminant Analysis (GDA) method

(Wastell 1987, Baudat and Anouar 2000, Shen et al. 2007). The GDA applies the general linear

model to the discriminant function analysis problem. The general linear model is a generalization of

the linear regression model that tests for effects of categorical predictor variables and continuous

predictor variables; and accommodates experimental designs with either a single dependent variable

or multiple dependent variables. Discriminant function analysis involves the prediction of a

categorical dependent variable by one or more continuous or binary independent variables, and is

used to determine which variables discriminate between two or more naturally occurring groups.

45

There can be as many classification functions as there are groups. For each group it is possible to

determine the location of the centroid. A case would be classified as belonging to a group that the

Mahalanobis distance to the group’s centroid is the smallest (Mahalanobis et al. 1937). These

classifications are determined by the most influencing wavelengths as well as by the entire spectra.

Therefore, to separate between different crops and weeds as well as to determine the most important

wavelengths for the separation, GDA forward stepwise models were created and validated by

Statistica v.9 software (StatSoft®, Tulsa, OK, USA).

The quality of classification of the validation data sets was assessed by the Cohen’s Kappa

coefficient, overall accuracy, user’s accuracy, and producer’s accuracy for each confusion matrix.

Cohen’s Kappa as defined by Cohen (1960), is a unit less value ranging from 1 for perfect

agreement to -1 for complete disagreement. Cohen’s Kappa is presented in the following equation:

d qKappaN q−

=−

(1)

where d is the sum of ground truth pixels that were correctly classified, q is the multiplication of

total ground truth and total classification values summed and divided by the total number of

samples, and N is the total number of samples. The 95% confidence limits (CL) were calculated for

overall accuracy as shown by Foody (2008) and presented as follows:

( ), 1

11N d

p pCL t

N−

−= ±

− (2)

where p is the overall accuracy, tN,d-1 is the statistical value of 95% two tailed test for d samples, N

is the total number of samples, d is the sum of ground truth pixels that were correctly classified. The

CL of the total accuracy can allow comparison between models that is based on significance and by

that show if there is a model that is significantly better or worse than others (Foody 2008).

3. Results and discussion

The leaf spectra, obtained by contact probe, are pure, without any mixed category spectra (Figure

1(a)). Plant species differ in levels of reflectance, but otherwise have similar spectral features.

Differences in spectral reflectance can be observed in the NIR region (700-1200 nm) and may be

attributed to the different internal leaf structure of cereals (GW and wheat) and broadleaf plants

46

(BLW and chickpea). Conversely, the plant canopy spectra obtained by bare fiber probe (Figure

1(b)) are influenced by >60% vegetation cover with soil as background. The different plant

categories differ in level of reflectance and possess similar spectral features between 700 nm and

2400 nm. Besides soil background, the canopy spectra are influenced by the canopy’s structure and

thickness as well as by other external parameters (e.g., plant age, sun angle and wind). These

parameters can influence reflectance values in the visible (400-700 nm), NIR, and SWIR (1200-

2400 nm) regions. These differences in the level of reflectance and features for both leaf and

canopy scales form the basis for using GDA to classify the different categories.

Figure 1. (a) Leaf reflectance spectra of wheat, chickpea, broadleaf weeds (BLW) and grass weeds

(GW) obtained in the field with the ASD radiometer’s contact probe by one layer of leaves with

100% cover of the field of view. (b) Canopy reflectance spectra of wheat, chickpea, BLW, GW, and

soil obtained in the field with the ASD bare fiber adaptor. The vegetation spectra were obtained

when the vegetation cover is >60%.

47

Very high overall accuracy for classification was obtained from GDA model of pure leaf spectra by

general category (wheat, chickpea, BLW and GW, Table 1) thus indicating that GDA is capable of

noticing features that consistently appear in pure leaf spectra of general vegetation categories. The

overall accuracy was excellent for classification by plant species (Table 2) when 20 samples of

unknown genera that could not be related to any of the 13 weed species were excluded from the

dataset. These results indicate that classifying by genera would be as good as classifying by

category. Using hyperspectral data Smith and Blackshaw (2003) obtained perfect results when

classifying leaves but the quality of classification was less when species of crops and weeds were

classified. Gibson et al. (2004) applied multispectral (i.e., yellow, red and infrared wide bands)

aerial imagery to identify the presence of GW and BLW in soybean, but were unable to separate

between weed species. Therefore, this simple classification scheme is deemed warranted for

separating among these general plant categories.

Table 1. Confusion matrix of the classification of pure leaf spectra by vegetation category using all

wavebands.

* 95% confidence interval = ±0.6% for the overall classification accuracy and Kappa = 0.99

Ground Truth Classes

Wheat Chickpea BLW GW Total # of classified samples

User’s accuracy % correct

Map

Cla

sses

Wheat 32 0 0 1 33 97

Chickpea 0 28 0 0 28 100

Broad Leaf Weeds (BLW) 0 0 174 0 174 100

Grass Weeds (GW) 0 0 0 66 66 100

Total # of ground truth samples 32 28 174 67

Producer’s accuracy % correct 100 100 100 99 99.7*

48

Table 2. Classification by GDA of pure leaf spectra by Genera, all wavebands.

Genera % Correct Number of validation samples

Wheat 100 32

Chickpea 100 28

Hordeum 100 10

Hirschfeldia 100 20

Malva 100 40

Sinapis 96 24

Ipomoea 100 11

Avena 100 12

Solanum 100 11

Setaria 100 17

Silybum 100 11

Chrysanthemum 100 29

Sonchus 100 13

Lolium 100 9

Beta 100 14

Total 99.6 281

General DA model results for samples with more than 30% vegetation coverage are shown for

wheat in Table 3 and chickpea in Table 4. In both cases the user’s accuracy for each of the classes

is >91% and producer’s accuracy for each of the classes is >87%. The overall accuracies are 95%

for wheat and 94% for chickpea with 95% confidence intervals that reduce the total accuracies to

not less than 90% and 88%. In both cases, The BLW class has a perfect user’s accuracy and GW

has the same user’s accuracy. The producer’s accuracy of wheat is greater than chickpea, which

might be related to the biomass density of the crop (Thorp and Tian 2004), since the FOV of the

fiber optic adaptor can cover five or six rows of wheat compared to only one row of chickpea. In

both cases the soil is classified with high success. GDA-based classification results for wheat are

based on 11 narrow bands. Sorted in order of importance, they are 675, 715, 705, 745, 690, 875,

850, 1090, 750, 760, and 1070 nm. For chickpea, eight narrow bands are important: 675, 725, 705,

49

730, 690, 715, 685, and 680 nm. Included in each series are several highly ranked red-edge bands

(e.g., 675 nm, 715 nm, and 705 nm for wheat; and 675 nm, 725 nm, and 705 nm for chickpea).

Therefore, optical sensors with four or more red-edge bands might be required for implementing the

proposed GDA classification scheme. Another interesting result is that out of the five most

important bands, red edge bands occupy the first, third, and fifth places in both models narrow band

lists. These findings indicate that the NIR and red-edge regions contain information that is

important for detection of categories and species of vegetation in agreement with previous studies

(Vrindts et al. 2002, Jurado-Exposito et al. 2003, Smith and Blackshaw 2003). Thenkabail et al.

(2004) presented a list of 22 narrow wavebands in the range of 350-2500 nm best for discriminating

natural and agricultural vegetation and weeds. The six red-edge bands (i.e., 675, 680, 685, 690,705,

and 730 nm), found to be important for classification in the current study, agreed with this list. Gray

et al. (2009) reported that the most important bands for classification of soybean, soil, and six

BLWs are in the SWIR region. Reflectance in the SWIR region is influenced by plant water content

whereas reflectance in visible region is influenced by chlorophyll pigments (Gausman 1985).

Table 3. Canopy classification model for wheat fields based on 11 narrow bands (sorted by

importance: 675, 715, 705, 745, 690, 875, 850, 1090, 750, 760, and 1070nm) and homogeneous

sample with vegetation cover >30%.


Wheat BLW GW Soil

Total # of classified samples


Map

Cla

sses

Wheat 36 0 1 1 38 95


Grass Weeds (GW) 0 2 22 0 24 92

Soil 1 0 0 20 21 95


Producer’s accuracy % correct 97 92 96 95 95*

* 95% confidence interval = ±4.19% for the overall accuracy and Kappa = 0.94

50

Table 4. Canopy classification model for chickpea fields, based on 8 narrow bands (sorted by

importance: 675, 725, 705, 730, 690, 715, 685, and 680nm) and homogeneous sample with

vegetation cover >30%.


Chickpea BLW GW Soil



Map

Cla

sses

Chickpea 13 1 0 0 14 93


Grass Weeds (GW) 0 1 23 1 25 92

Soil 2 0 0 20 22 91


Producer’s accuracy % correct 87 92 100 95 94*


Classification results based on resampled VENμS and ALI bands, for samples with more than 30%

vegetation coverage, are shown in Table 5 for wheat and Table 6 for chickpea. The BLW class has

the highest user’s accuracy for wheat in VENμS data and 3% greater than soil in ALI data (Table

5). The BLW has the highest user’s accuracy together with soil for chickpea in VENμS data and the

second after soil inALI data (Table 6). Since chickpea is a broadleaf and wheat is a grass it might be

that BLW will be classified with higher accuracy against a crop that is not a broadleaf. By the same

logic the GW class obtained higher user’s accuracy values in chickpea than in wheat, but as

mentioned above the BLW class user’s accuracy values are still greater than the GW class even in

chickpea. This relatively high user’s accuracy of BLW agrees with Vrindts et al. (2002) who

classified BLW against GW using three to nine selected bands. Red-edge narrow bands were highly

important for classification, as mentioned above. There is no advantage of four red-edge bands (i.e.,

VENμS) over two red-edge bands (i.e., ALI) since the Cohen’s Kappa values as well as user’s

accuracies are similar and overall classification accuracies are overlapping when considering

confidence intervals. Nevertheless, VENμS would be better suited for site specific weed

management than ALI due to its greater spatial resolution (5.3 m vs. 30 m). The chickpea models

51

Table 5. Canopy classification models for wheat fields based on resampled VENμS and ALI broad bands, with vegetation cover >30%.

VENμS ALI

Ground Truth Classes Ground Truth Classes

Wheat BLW GW Soil



Wheat BLW GW Soil Total # of classified samples


Map

Cla

sses

Wheat 30 7 5 3 45 67 29 3 5 0 37 78

Broad Leaf Weeds (BLW) 0 16 0 0 16 100 1 18 2 0 21 86

Grass Weeds (GW) 4 3 18 0 25 72 3 5 16 1 25 64

Soil 3 0 0 18 21 86 4 0 0 20 24 83

Total # of ground truth samples 37 26 23 21 37 26 23 21

Producer’s accuracy % correct 81 62 78 86 77* 78 69 70 95 78**


** 95% confidence interval = ±7.97% for the overall classification accuracy and Kappa = 0.70

52

Table 6. Canopy classification models for chickpea fields based on resampled VENμS and ALI broad bands, with vegetation cover >30%.

VENμS ALI

Ground Truth Classes Ground Truth Classes

Chickpea BLW GW Soil Total # of classified samples


Chickpea BLW GW Soil Total # of classified samples


Map

Cla

sses

Chickpea 14 2 0 1 17 82 13 4 0 1 18 72

Broad Leaf Weeds (BLW) 0 19 1 0 20 95 0 16 2 0 18 89

Grass Weeds (GW) 0 5 22 0 27 81 0 6 21 0 27 78

Soil 1 0 0 20 21 95 2 0 0 20 22 91


Producer’s accuracy % correct 93 73 96 95 88* 87 62 91 95 82**


** 95% confidence interval = ±8.22% for the overall classification accuracy and Kappa = 0.76

53

provided better classification results than the wheat models, but this advantage is not significant

based on the 95% confidence intervals.

The total accuracy is 79 ±6.74% for all ASD wavebands, in wheat field, for six categories:

wheat, BLW, and GW that can have soil as background, and soil and two categories of mixed

vegetation in the FOV (Table 7). To simulate a situation where decision to spray or not spray is

to be made, the result of GDA classification for the following three options of herbicide

application: no spray, spray BLW, or spray GW are presented in Table 8 for all wavebands. No

herbicide would be applied if only wheat and soil were detected in the FOV of the sensor

whereas herbicide would be applied if BLW or GW with wheat, or only BLW or GW were

detected. This scenario is based on heterogeneous spectra with >5% vegetation cover 25 to 40

DAE of wheat, which is the optimal time for herbicide application. Spectral data obtained on

ground level by Lopez-Granados et al. (2008) resulted in high classification but relation to an

optimal time for herbicide application was

Table 7. Herbicide application model for wheat fields with vegetation cover of 0-100%,

restricted to 25-40 days after emergence, for all wavebands.


Wheat Wheat &

BLW Wheat & GW BLW GW Soil



Map

Cla

sses

Wheat 24 6 4 2 1 2 39 62

Wheat & Broad Leaf Weeds (BLW) 1 20 0 5 1 0 27 74

Wheat & Grass Weeds (GW) 0 1 13 0 1 0 15 87

BLW 0 3 0 18 0 0 21 86

GW 0 0 1 0 11 0 12 92

Soil 0 0 0 1 0 26 27 96

Total # of ground truth samples 25 30 18 26 14 28

Producer’s accuracy % correct 96 67 72 69 79 93 79*


54

not mentioned. For the application window in wheat field the overall classification accuracy was

87 ±5.57% for all ASD wavebands, (Table 8). The user’s accuracy for no herbicide application

indicates that 79% of the decisions not to spray are indeed correct. Basing site-specific herbicide

application on a map derived from remote sensing and the GDA method would be highly

effective as indicated by user’s accuracy of 96% for both spraying BLW and GW.

Table 8. Herbicide application model for wheat fields with vegetation cover >5%, restricted to

25-40 days after emergence, for all wavebands.


Herbicide application Wheat & soil

BLW & wheat or BLW

GW & wheat or GW



Map

Cla

sses

Wheat & soil No 52 9 5 66 79

Broad Leaf Weeds (BLW) & wheat or BLW

Yes (Spray BLW) 1 46 1 48 96

Grass Weeds (GW) & wheat or GW Yes (Spray GW) 0 1 26 27 96

Total # of ground truth samples 53 56 32

Producer’s accuracy % correct 98 82 81 87*


4. Summary and conclusions

Classification of crops and weeds was applied by GDA models for the leaf and canopy scales.

The leaf scale resulted in almost perfect classification by genera as well as by categories. The

canopy scale was applied for several spectral resolutions, hyperspectral and resample to current

and forthcoming satellites included, as well as different vegetation coverage percentage. The

GDA is negatively affected by non-uniform number of samples among classes (Fraley and

55

Raftery 2002) and it has difficulty in separating classes that are spectrally similar (Zhao and

Maclean 2000). In the current study the GDA results were influenced by an unequal number of

samples among classes. The classes were of different vegetation types that may have included

soil characteristics in many spectral samples as well as soil as a class of its own. Nevertheless,

the results of this study indicate that differentiation between crops and weeds is possible using

GDA thus potentially contributing to practical site-specific herbicide application. Specific

conclusions are:

• The spectral characteristics of the pure leaf spectra enable precise classification of the

different plant categories and genera.

• The red-edge region is highly important for crops and weeds classification.

• Spectral separation of crops and weeds is potentially useful for wheat fields with >5%

vegetation cover in the critical period for weed control.

Ground level sensors offer very high spatial resolution and the ability to apply the classification

by classes that are of one plant species each. In contrast, satellite sensors offer synoptic, map-

like views that cover large regions at less spatial resolution. If a space platform is chosen,

VENμS would be a better option than ALI because of greater spatial resolution and revisit

frequency. Ground sensors are less affected by the effects of atmosphere on vegetation

reflectance measurements. Further research is needed to determine how much ground truth data

are needed to adjust the GDA model to sensor spatial and spectral resolutions and determine the

effect of mixed pixels.

5. References

Alchanatis, V., Ridel, L., Hetzroni, A. and Yaroslavsky, L., 2005, Weed detection in multi-spectral images of cotton fields. Computers and Electronics in Agriculture, 47, pp. 243-260.

Baudat, G. and Anouar, F.E., 2000, Generalized discriminant analysis using a kernel approach. Neural Computation,

12, pp. 2385-2404. Bennett, A.L. and Pannell, D.J., 1998, Economic evaluation of a weed-activated sprayer for herbicide application to

patchy weed populations. Australian Journal of Agricultural and Resource Economics, 42, pp. 389-408. Biewer, S., Erasmi, S., Fricke, T. and Wachendorf, M., 2009, Prediction of yield and the contribution of legumes in

legume-grass mixtures using field spectrometry. Precision Agriculture, 10, pp. 128-144.

56

Biller, R.H., 1998, Reduced input of herbicides by use of optoelectronic sensors. Journal of Agricultural Engineering Research, 71, pp. 357-362.

Blackshaw, R.E., Molnar, L.J. and Lindwall, C.W., 1998, Merits of a weed-sensing sprayer to control weeds in

conservation fallow and cropping systems. Weed Science, 46, pp. 120-126. Brent, J. and Schaeffer, T.H., 2011, Systematic Review of Parkinsonian Syndromes in Short- and Long-Term

Survivors of Paraquat Poisoning. Journal of Occupational and Environmental Medicine, 53, pp. 1332-1336. Brown, R.B. and Steckler, J., 1995, Prescription maps for spatially variable herbicide application in no-till corn.

Transactions of the Asae, 38, pp. 1659-1666. Brown, R.B., Steckler, J. and Anderson, G.W., 1994, Remote-Sensing for Identification of Weeds in No-Till Corn.

Transactions of the Asae, 37, pp. 297-302. Chander, G., Markham, B.L. and Helder, D.L., 2009, Summary of current radiometric calibration coefficients for

Landsat MSS, TM, ETM+, and EO-1 ALI sensors. Remote Sensing of Environment, 113, pp. 893-903. Cohen, J., 1960, A coefficient of agreement for nominal scales Educational and psychological measurement, 20, pp.

37-46. Deardorff, J.W., 1978, Effcient prediction of ground surface-temperature and moistere, whitch inclusion of a layer

of vegetation. Journal of Geophysical Research-Oceans and Atmospheres, 83, pp. 1889-1903. Dhawan, A., Bajpayee, M. and Parmar, D., 2009, Comet assay: a reliable tool for the assessment of DNA damage in

different models. Cell Biology and Toxicology, 25, pp. 5-32. Eddy, P.R., Smith, A.M., Hill, B.D., Peddle, D.R., Coburn, C.A. and Blackshaw, R.E., 2006, Comparison of neural

network and maximum likelihood high resolution image classification for weed detection in crops: Applications in precision agriculture. 2006 Ieee International Geoscience and Remote Sensing Symposium, Vols 1-8, pp. 116-119.

Feyaerts, F. and van Gool, L., 2001, Multi-spectral vision system for weed detection. Pattern Recognition Letters,

22, pp. 667-674. Foody, G.M., 2008, Harshness in image classification accuracy assessment. International Journal of Remote

Sensing, 29, pp. 3137-3158. Fraley, C. and Raftery, A.E., 2002, Model-based clustering, discriminant analysis, and density estimation. Journal of

the American Statistical Association, 97, pp. 611-631. Gausman, H., 1985, Plant leaf optical properties in visible and Near Infrared light, pp. 9-60 (Lubbock TX. USA:

Texas Tech Press). Gerhards, R. and Christensen, S., 2003, Real-time weed detection, decision making and patch spraying in maize,

sugarbeet, winter wheat and winter barley. Weed Research, 43, pp. 385-392. Gerhards, R., Sokefeld, M., Schulze-Lohne, K., Mortensen, D.A. and Kuhbauch, W., 1997, Site specific weed

control in winter wheat. Journal of Agronomy and Crop Science, 178, pp. 219-225. Gibson, K.D., Dirks, R., Medlin, C.R. and Johnston, L., 2004, Detection of weed species in soybean using

multispectral digital images. Weed Technology, 18, pp. 742-749. Gibson, L., Kingwell, R. and Doole, G., 2008, The role and value of eastern star clover in managing herbicide-

resistant crop weeds: A whole-farm analysis. Agricultural Systems, 98, pp. 199-207.

57

Gray, C.J., Shaw, D.R. and Bruce, L.M., 2009, Utility of Hyperspectral Reflectance for Differentiating Soybean (Glycine max) and Six Weed Species. Weed Technology, 23, pp. 108-119.

Hamouz, P., Novakova, K., Soukup, J. and Holec, J., 2008, Detection of Cirsium arvense L. in winter wheat using a

multispectral imaging system. Journal of Plant Diseases and Protection, pp. 167-170. Har Gil, D., Bonfil, D.J. and Svoray, T., 2011, Multi scale analysis of the factors influencing wheat quality as

determined by Gluten Index. Field Crops Research, 123, pp. 1-9. Herrmann, I., Pimstein, A., Karnieli, A., Cohen, Y., Alchanatis, V. and bonfil, D., J., 2011, LAI assessment of wheat

and potato crops by VENμS and Sentinel-2 bands. Remote Sensing of Environment, 115, pp. 2141-2151. Johnson, G.A., Mortensen, D.A., Young, L.J. and Martin, A.R., 1995, The Stability of Weed Seedling Population-

Models and Parameters in Eastern Nebraska Corn (Zea-Mays) and Soybean (Glycine-Max) Fields. Weed Science, 43, pp. 604-611.

Jones, R.E., Vere, D.T., Alemseged, Y. and Medd, R.W., 2005, Estimating the economic cost of weeds in Australian

annual winter crops. Agricultural Economics, 32, pp. 253-265. Jurado-Exposito, M., Lopez-Granados, F., Atenciano, S., Garcia-Torres, L. and Gonzalez-Andujar, J.L., 2003,

Discrimination of weed seedlings, wheat (Triticum aestivum) stubble and sunflower (Helianthus annuus) by near-infrared reflectance spectroscopy (NIRS). Crop Protection, 22, pp. 1177-1180.

Kafkafi, U. and Bonfil, D., J., 2008, Integrated Nutrient Management: Experience and Concepts from the Middle

East. In Integrated Nutrient Management for Sustainable Crop Production, M. Aulakh and C.A. Grant (Eds.), pp. 523-565 (Binghamton NY. USA: The Haworth Press, Inc. ).

Lamb, D.W. and Brown, R.B., 2001, Remote-sensing and mapping of weeds in crops. Journal of Agricultural

Engineering Research, 78, pp. 117-125. Lindquist, J.L., Dieleman, J.A., Mortensen, D.A., Johnson, G.A. and Wyse-Pester, D.Y., 1998, Economic

importance of managing spatially heterogeneous weed populations. Weed Technology, 12, pp. 7-13. Lopez-Granados, F., Pena-Barragan, J.M., Jurado-Exposito, M., Francisco-Fernandez, M., Cao, R., Alonso-

Betanzos, A. and Fontenla-Romero, O., 2008, Multispectral classification of grass weeds and wheat (Triticum durum) using linear and nonparametric functional discriminant analysis and neural networks. Weed Research, 48, pp. 28-37.

Mahalanobis, P.C., Bose, R.C. and Roy, S.N., 1937, Normalization of statistical variaties and the use of rectangular

coordinates in theory of sampling distributions. Sankhya, 3, pp. 1-34. Mallory-Smith, C.A., Thill, D.C. and Dial, M.J., 1990, Identification of Sulfonylurea Herbicide-Resistant Prickly

Lettuce (Lactuca-Serriola). Weed Technology, 4, pp. 163-168. Manh, A.G., Rabatel, G., Assemat, L. and Aldon, M.J., 2001, Weed leaf image segmentation by deformable

templates. Journal of Agricultural Engineering Research, 80, pp. 139-146. Marshall, R. and Moss, S.R., 2008, Characterisation and molecular basis of ALS inhibitor resistance in the grass

weed Alopecurus myosuroides. Weed Research, 48, pp. 439-447. Mendenhall, J.A., Lencioni, D.E. and Evans, J.B., 2000, Earth Observing-1 Advanced Land Imager: radiometric

response calibration. Available online at: http://eo1.gsfc.nasa.gov/new/validationReport/Technology/Documents/EO-1-3.pdf (accessed 29 January 2012).

Monaco, T., J, Weller, S., C and Ashton, F., M, 2002, Weed science principles and practices pp. 3-126 (New York

NY. USA: John Wiley & sons, Inc.).

58

Moran, M., S., Inoue, Y. and Barnes, E., M., 1997, Opportunities and Limitations for Image-Based Remote Sensing

in Precision Crop Management. Remote Sensing of Environment, 61, pp. 319-346. Moran, M.S., Maas, S.J., Vanderbilt, V.C., Barnes, M., Miller, S.N. and Clarke, T.R., 2004, Application of image-

based remote sensing to irrigated agriculture. In Remote sensing for natural resource management and environmental monitoring S.L. Ustin (Ed.), pp. 648-650 (Hoboken: John Wiley & sons).

Papayiannis, L.C., Kokkinos, C.D. and Alfaro-Fernandez, A., 2012, Detection, characterization and host range

studies of Pepino mosaic virus in Cyprus. European Journal of Plant Pathology, 132, pp. 1-7. Pikart, T.G., Souza, G.K., Serrao, J.E. and Zanuncio, J.C., 2011, Leafcutter ants: a small dispersal agent of the

invasive plant Murraya paniculata. Weed Research, 51, pp. 548-551. Pinter, P., J., Hatfield, J., L., Schepers, J., S., Barnes, E., M., Moran, M., S., Daughtry, C., S, T. and Upchurch, D.,

R., 2003, Remote sensing for crop management. Photogrametric Engineering & Remote Sensing, 69, pp. 647-664.

Raven, P., H, Everet, R., F and Eichhorn, S., E, 2005, Biology of plants, pp. 35-88 (New-York: W. H. Freeman and

Company). Shen, L., Bai, L. and Fairhurst, M., 2007, Gabor wavelets and General Discriminant Analysis for face identification

and verification. Image and Vision Computing, 25, pp. 553-563. Slaughter, D.C., Giles, D.K. and Downey, D., 2008, Autonomous robotic weed control systems: A review.

Computers and Electronics in Agriculture, 61, pp. 63-78. Smith, A.M. and Blackshaw, R.E., 2003, Weed-crop discrimination using remote sensing: A detached leaf

experiment. Weed Technology, 17, pp. 811-820. Thenkabail, P.S., Enclona, E.A., Ashton, M.S. and Van der Meer, B., 2004, Accuracy assessments of hyperspectral

waveband performance for vegetation analysis applications. Remote Sensing of Environment, 91, pp. 354-376. Thorp, K.R. and Tian, L.F., 2004, A review on remote sensing of weeds in agriculture. Precision Agriculture, 5, pp.

477-508. Timmermann, C., Gerhards, R. and Kuehbauch, W., 2003, The economic impact of site-specific weed control.

Precision Agriculture, 4, pp. 249-260. Vrindts, E., De Baerdemaeker, J. and Ramon, H., 2002, Weed detection using canopy reflection. Precision

Agriculture, 3, pp. 63-80. Wastell, D.G., 1987, A simple randomization procedure for validating discriminant analysis- a methodological note

Biological Psychology, 24, pp. 123-127. Weis, M., Gutjahr, C., Rueda Ayala, V., Gerhards, R., Ritter, C. and Scholderle, F., 2008, Precision farming for

weed management: techniques. Gesunde Pflanzen, 60, pp. 171-181. Zhao, G. and Maclean, A.L., 2000, A comparison of canonical discriminant analysis and principal component

analysis for spectral transformation. Photogrammetric engineering and remote sensing, 66, pp. 841-847. Zwiggelaar, R., 1998, A review of spectral properties of plants and their potential use for crop/weed discrimination

in row-crops. Crop Protection, 17, pp. 189-206.

59

Chapter 5

5. Ground-level hyperspectral imagery for detecting weeds in wheat fields. Submitted

to precision agriculture.

60

Ground-Level Hyperspectral Imagery for Detecting Weeds in Wheat Field*

Herrmann, I(1), Shapira, U(1), Kinast, S(2), Karnieli, A(1) , Bonfil, D.J(3)

(1) The Remote Sensing Laboratory, Jacob Blaustein Institutes for Desert Research, Ben-Gurion University of the

Negev, 84990, Israel

(2) Department of Solar Energy and Environmental Physics, Jacob Blaustein Institutes for Desert Research, Ben-

Gurion University of the Negev, 84990, Israel

(3) Field Crops and Natural Resources Department, Agricultural Research Organization, Gilat Research Center,

85280, Israel

Abstract

Site-specific weed management can allow economically and environmentally more efficient

weed control. Spectral differences between plants species can lead to the ability to separate

wheat from weeds. The study used ground-level image spectroscopy data, with high spectral and

spatial resolutions, for detecting annual grasses and broadleaf weeds in wheat fields. The image

pixels were used to cross validate Partial Least Squares Discriminant Analysis (PLS-DA)

classification models. The best model was chosen by comparing the cross validation confusion

matrices in term of their variances and Cohen’s Kappa values. This best model used four classes:

broadleaf, grass weeds, soil, and wheat. Each of the classes contains sunlit and shaded data

together. The Variable Importance in Projection (VIP) method was applied in order to locate the

most important spectral regions for each of the classes. It was found that the red-edge is the most

important region for the vegetation classes. Ground truth pixels were randomly selected and their

confusion matrix resulted in total accuracy of 72%. The results obtained are reasonable although

the model used wheat and weeds from different growth stages, acquiring dates, and fields. It is

concluded that high spectral and spatial resolutions can provide separation between wheat and

weeds based on their spectral data. The results show feasibility for up-scaling the spectral

methods to air or spaceborne sensors as well as developing ground-level application.

Keywords: Weed detection; Hyperspectral imaging; Site specific; Plant spectroscopy; Wheat.

* Submitted to Precision Agriculture.

61

1. Introduction

Weeds are pest that can be generally defined as plants growing where it is not desired (Monaco

et al. 2002). From an agricultural point of view, weeds reduce crop yield and quality by

competing with crops for water, sunlight, and minerals (Pinter et al. 2003; Slaughter et al. 2008),

producing allelopathic substances (M. S. Moran et al. 2004), hosting diseases and insects

(Papayiannis et al. 2012; Pikart et al. 2011), and disturbing tilling and harvesting (Monaco et al.

2002). Weeds are the most acute pest in agriculture with estimated annual global damage of

around 40 billion US dollars per year (Monaco et al. 2002). In Australia alone, herbicide

resistance is estimated to impose an additional annual cost of more than 1 billion US dollars

(Gibson et al. 2008). An increasing problem is weed resistance to herbicides (Mallory-Smith et

al. 1990; Jones et al. 2005; Marshall and Moss 2008).

Weed distribution in fields is non-uniform and confined to patches of varying size or stripes

along field borders (Lamb and Brown 2001; Gerhards et al. 1997; Gerhards and Christensen

2003; Slaughter et al. 2008; Weis et al. 2008; Vrindts et al. 2002; M. S. Moran et al. 2004).

Application of herbicides at a field is often based on the previous year's weed problems and

information obtained from field scouting (Manh et al. 2001; M. S. Moran et al. 2004) and

therefore can result in applying herbicide were it is not needed. By significantly reducing the

quantity of herbicide applied (Gerhards et al. 1997; Gerhards and Christensen 2003;

Timmermann et al. 2003; Weis et al. 2008; Slaughter et al. 2008; Eddy et al. 2008), site-specific

weed control and management could economically benefit farmers and consumers as well as the

environment without diminishing weed control efficiency (Pinter et al. 2003; Weis et al. 2008;

Slaughter et al. 2008). Since an increasing problem is weed resistance to herbicides (Mallory-

Smith et al. 1990; Jones et al. 2005; Marshall and Moss 2008), another reason to reduce the

amount of applied herbicides is to lower the probability of having this effect. Since herbicides

can be an environmental hazard to fauna as well as humans (Brent and Schaeffer 2011; Dhawan

et al. 2009) several countries have restricted the amount of herbicides that can be applied per

area unit (Biller 1998; Timmermann et al. 2003; Slaughter et al. 2008). Herbicide use

regulations, consumer concerns, and growing interest in organically produced foods, limit the

long-term acceptability of herbicide applications (Slaughter et al. 2008). According to Lindquist

62

et al. (1998), it is possible to reduce the quantity of herbicide applied by applying herbicides only

where weeds are located. Site-specific weed management has reduced herbicide use by 11 to

90% without affecting crop yield (Feyaerts and van Gool 2001; Gerhards and Christensen 2003;

Johnson et al. 1995; Lindquist et al. 1998; Brown and Steckler 1995; Brown et al. 1994). Weed

distribution and competition with crops are influenced by spatial variability in topography,

drainage, soil type, and microclimate. There is significant variation in weeds within and between

different fields emphasizing the need for site-specific weed management (M. S. Moran et al.

2004).

Remote sensing techniques can provide fast and cost-effective mapping of weed populations, at a

chosen date, over large areas, which otherwise would be impractical to cover by manual ground

survey methods (Hamouz et al. 2008; Zwiggelaar 1998; M. S. Moran et al. 2004; Alchanatis et

al. 2005). Real-time (also termed “on-the-go”) non selective weed detection and control can be

implemented by means of tractor-mounted, optical sensors that trigger a spray nozzle valve to

open briefly upon detection of green vegetation (Bennett and Pannell 1998; Biller 1998;

Blackshaw et al. 1998). This approach can be applied to entire fields before crop emergence or

between the crop rows after emergence (Slaughter et al. 2008; M. Moran, S. et al. 1997;

Alchanatis et al. 2005). Other ground-based, on-the-go sensing methods are designed to detect

the shape of weed leaves against soil background and thus, can be applied only in early growing

stages when all leaves are not overlapped (Weis et al. 2008). Feyaerts et al. (2001) operated

ground-level, spectral classification experiment of crop and weeds for different plats that their

leaves were not shading or overlapping other leaves and resulted in better classification of weeds

than crops. Weeds were better classified than crop also by Nieuwenhuizen et al. (2010) that have

analyzed sugar beet (crop) and potato (weed) in early growth stages of one or two layers of

leaves. They applied adaptive classification with no need to choose training data in advance in

row crop. Combining both classification concepts (i.e., morphological and spectral methods) is

summoned by the morphological and spectral variety of plants. Eddy et al. (2008) have

combined morphological methods with spectral methods in order to classify between one crop

and one weed, per model, by ground level images. It was suggested that species separation can

be obtained at early plant growth stages, based on that earlier plant stage showed consistently,

better classification results than the later ones (seven days difference).

63

Spectral separation between species is commonly based on differences in the relative reflectance

values in the range of 400 to 2500 nm. Different leaf pigmentation (e.g., chlorophyll content)

will mainly affect the relative reflectance values in the visible region (400 to 700 nm) of plant

spectrum (Gausman 1985; Yoder and Pettigrew-crosby 1995). The near infrared (NIR) region

(700 to 1100 nm) is highly influenced by the leaf or canopy structure that can be affected by

phenology as well as species (Gausman 1985). A dicotyledonous leaf has more air spaces among

its spongy mesophyll tissue, than a monocotyledonous leaf (Raven et al. 2005) of the same

thickness and age resulting in a higher reflectance in the NIR region (Gausman 1985). The red-

edge region is the slope connecting the low red and high NIR values in the reflectance spectrum

of vegetation. The red-edge is an important element in spectral separation of different plant

species (Smith and Blackshaw 2003; Vrindts et al. 2002; Herrmann et al. 2011). The short wave

infrared (SWIR) region (1100 to 2500 nm) of plant reflectance spectrum is mainly influenced by

the plant’s water content (Gausman 1985; Tian et al. 2001). When radiation enters a plant

canopy, multiple events of reflection, transmission, absorption, scattering, and diffraction occur

(Borregaard et al. 2000). Reflectance of canopy will probably include shaded leaves and might

include stems, flowers, fruits, and background that is very likely to be soil that might be partly

shaded and of different humidity levels.

Several studies dealt with ground-level spectral classification of crop and weeds over multiple

growing seasons. Lopez-Granados et al. (2008) classified ground-level spectral reflectance of

wheat, four grass weeds, and soil, concluded that one sampling date, per growth season, when

phenological distinction is maximal can provide high quality classification. It is important to

mention that relaying on phenology for spectral separation will be less efficient in case the

optimal time for herbicide application precedes the date of maximal phenological variability

among crop and weeds. Gray et al. (2009) resulted in a three lists (one for each of the two

seasons and one of the combined data) of top 10 bands for classifying ground level spectral

reflectance of soybean, six broadleaf weeds, and soil. SWIR bands are in the top seven, five and

10 places in these lists, showing high importance for the SWIR region. Slaughter et al. (2008)

summarized in their review that the greatest extent of the studies were conducted in ideal

conditions, with no overlapping of crop and weeds, and resulted in classification accuracies of 65

to 95%. Zwiggelaar (1998) mentions in his review that using selected wavelengths for the

discrimination between crops and weeds in a row environment has not been shown so far, and

64

analyzing images with a limited number of wavelengths might not be sufficient. The first step

required in order to spectrally distinguish between crops and weeds is to obtain continuous

spectra of pure plant for each species. This can be done by high spatial and spectral resolutions,

as shown by Vrindts et al. (2002) concluding the need to employ relative reflectance values in

order to classify crops and weeds and to minimize the different lighting conditions effect on the

spectral data. It was also mentioned that several bands can result in high classification accuracy.

Okamoto et al. (2007) worked in the visible and NIR regions in order to separate between sugar

beet and four weeds (two BLW and two GW). The data were acquired in two successive dates,

cloudy and sunny, in field conditions. The growth stage is not mentioned but a figure presents

young seedlings apart from each other, with no more than two layers of leaves. To the best of our

understanding, Okamoto et al. (2007) analyzed digital numbers (rather physical units such as

radiance or reflectance), therefore, in order to allow gathering and comparing data from different

images the digital numbers were normalized . In their work, the preprocessing procedures did not

include radiometric correction (transforming from digital numbers to physical units), subtraction

of the sensor’s electronical noise, and atmospheric correction (reducing the atmospheric

influence on the reflectance values, this influence is changing between wavelengths). Therefore,

the results might be changed from the different weather conditions among the different images

and days. The validation results show 75 to 97 % success for the five classes for sampled pure

vegetation pixels. Prediction for an entire image and ground truth analyses were not mentioned.

Borregaard et al. (2000) have worked with a line scanner and artificial lighting on young crops

(sugar beet and potato) and three broadleaf weeds with soil back ground. When the weeds were

classified together the crops classification performance were improved significantly but the

weeds classification performance was not presented. In other studies, applying hyperspectral

cameras, the soil background was excluded and classification of crop and weeds were applied for

young plants with one or two layers of leaves (Okamoto et al. 2007; Borregaard et al. 2000;

Feyaerts and van Gool 2001; Nieuwenhuizen et al. 2010).

The current research used ground-level image spectroscopy data, with high spectral and spatial

resolutions, for detecting annual grasses and broadleaf weeds in wheat fields. Specific objectives

were threefold: (1) to choose the best classes determination, for this dataset, in order to separate

broadleaf weed (BLW), grass weed (GW), and wheat; (2) to find the most important spectral

bands needed for this separation; and (3) to examine the potential of using high spectral and

65

spatial resolutions ground level reflectance from the wheat fields to predict categories of wheat

and weeds.

2. Material and methods

2.1.Study area

Field measurements were performed in rainfed as well as irrigated wheat experimental plots in

winter 2009 at the Gilat Research Center in the northwest Negev, Israel (31o 20’ N, 34o 40’ E).

This region is defined as semi-arid with a short rainy season (November-April; Har Gil et al.

2011). It is characterized by soil type of sandy loam loess–Calcic Xerosol, shallow hills with an

average elevation of 80-150 m above mean sea level, other climatological information were

previously described (Kafkafi and Bonfil 2008).

2.2.Field work and pre-processing

Ground level images were obtained by the Spectral Camera HS (Specim), a push broom sensor,

with 1600 pixel per line and 849 spectral narrow bands (~ 0.67 nm wide) in the visible and NIR

regions. Since herbicides are usually applied before closer of the crop canopy (Thorp and Tian

2004), the images were acquired when the wheat was 10-54 days after emergence. The camera

was mounted on a tripod, 135 cm above the ground, pointing down to cover an area of 50 by 50

cm delimited by a metal frame at the canopy level (Figure 1). At this height, the spatial

resolution was approximately 0.5 mm. The square frame was divided to 16 equal size squares to

visually estimate the relative coverage of different features such as wheat, weeds, and soil. The

assessment was done for each of the squares and accumulated, with 6.25% weight per square, to

include the entire area surrounded by the frame. All assessments were performed by the same

person. Digital camera Coolpix S10 (Nikon), mounted at the same height as the hyperspectral

camera, was used to acquire true-color images. These photos were used as reference for the

cross validation classification as well as for obtaining ground truth. It should be noted that since

the integration time of each scene, resulting from the push broom instrument, is 28-35 sec

(depends on the frame rate and number of lines acquired), and since all images were acquired in

66

open field, gusts of wind can influence the relative location of leaves. Consequently, a slight

difference might exist between the hyperspectral image and the RGB photo.

The image preprocessing included subtraction of sensor’s electronical noise (dark current) and

radiometric correction by the AISA tools software. Then the images were converted to relative

reflectance values by the ENVI software environment. This process was based on the flat field

calibration method by white referencing to pressed and smoothed powder of barium sulfate

(BaSO4) positioned on the frame underneath the camera (Hatchell 1999) as presented in Figure

1. The flat field calibration method was performed for each image by its own white reference and

by that provided also atmospheric correction. The images were spectrally resampled to obtain 91

bands by averaging the original spectra every 5 nm in the range of 400 to 850 nm. The images

were rectangularly clipped to include only area within the frame boundaries. In cases where the

frame was not parallel to the image borders, the images were clipped to include maximal area

while the frame itself is not included. 21images were acquired this way for further image

processing and statistical analysis.

Figure 1. Setup of the hyperspectral camera in a wheat field. In the small frame: an example of a

sampled image including broadleaf and grass weeds, soil, and wheat.

2.3.PLSDA analysis

The Partial Least Squares Discriminant Analysis (PLS-DA) applies a Partial Least Squares

model to the discriminant function analysis problem in order to allow maximal separation among

classes (Musumarra et al. 2004). Since the Partial Least Squares was not initially designed for

67

classification, it has been rarely used for this purpose. Nevertheless, PLS-DA method can

produce plausible separation (Barker and Rayens 2003). In order to relate the PLS (numerical) to

the DA (categorical), in a two classes case, each sample is assigned with an arbitrary number that

indicates to which class it belongs (Xie et al. 2007). These two arbitrary numbers are the only

values acceptable for one artificial variable. In case more than two classes are to be separated,

there is a need for more (as many as the classes) binary artificial variables that will indicate to

which class each sample belongs (Musumarra et al. 2004).

A total number of 1857 spectra were picked out of the 21 raw images, as presented in Table 1.

Each of the spectra was obtained from one pixel as vector of reflectance values at all 91 bands.

Only pure pixels (i.e., containing one class) were selected. Approximately half of the spectra

were obtained from sunlit pixels and the rest from shaded pixels. In Table 1 the data is divided to

four classes: BLW, GW, soil, and wheat with 799, 364, 330, and 364 spectra, respectively. Six

models, all based on 1857 pixels, were examined:

• Model # 1 separates three classes: BLW, G (including GW and wheat), and soil;

• Model # 2 separates two classes: BLW and G (including GW and wheat);

• Model # 3 separates four classes: BLW, GW, soil, and wheat;

• Model # 4 separates three classes: BLW, GW, and wheat;

• Model # 5 separates the classes as in model # 3 that are divided to sunlit and shaded pixels

(i.e., eight classes);

• Model # 6 separates the classes as in model # 4 that are divided to sunlit and shaded pixels

(i.e., six classes).

The samples distribution to classes for each of the models can be obtained by Table 1.

In order to evaluate the relative importance of each band in the chosen PLS-DA model, the

Variable Importance in Projection (VIP) after Wold et al. (1993) was computed. The VIP is

defined as the summary of the importance for each predictor projections to find a number of

principal components of the PLS model (Chong and Jun 2005; Y. Cohen et al. 2010). The VIP

values are evaluated by “the higher the better” where the average VIP = 1 is considered to be the

putative threshold since it is the average value of the PLS model predictors’ VIP values.

68

Therefore in order to separate between wheat and weeds as well as to determine the most

important wavelengths for the separation, each PLS-DA model was cross validated (including

Table 1. Distribution of the 1857 picked pixels amongst the classes and images

Image #

Date acquired

Date of emergence

# of BLW spectra

sunlit; shaded

# of GW spectra

sunlit; shaded

# of soil spectra

sunlit; shaded

# of wheat spectra

sunlit; shaded

1 06-Jan-2009 14-Nov-2008 - ; - - ; 52 28 ; 23 40 ; 74

2 06-Jan-2009 14-Nov-2008 - ; - 35 ; 63 - ; - 11 ; 67

3 06-Jan-2009 14-Nov-2008 - ; - 18 ; - - ; - 28 ; 40

4 06-Jan-2009 14-Nov-2008 - ; 42 17 ; - - ; - 11 ; -

5 05-Feb-2009 27-Jan-2009 35 ; - - ; - 46 ; 45 17 ; -

6 05-Feb-2009 19-Dec-2008 24 ; - 39 ; - - ; - 31 ; -

7 05-Feb-2009 19-Dec-2008 - ; - 7 ; - - ; - 7 ; -

8 05-Feb-2009 19-Dec-2008 - ; - - ; - - ; - 10 ; -

9 05-Feb-2009 19-Dec-2008 - ; - 2 ; - - ; - 5 ; -

10 05-Feb-2009 19-Dec-2008 - ; - 2 ; - - ; - 12 ; -

11 05-Feb-2009 19-Dec-2008 - ; - 6 ; - - ; - 11 ; -

12 07-Jan-2009 07-Dec-2008 79 ; 59 - ; - - ; - - ; -

13 07-Jan-2009 07-Dec-2008 - ; 41 - ; - - ; - - ; -

14 07-Jan-2009 07-Dec-2008 71 ; 81 - ; - 47 ; - - ; -

15 07-Jan-2009 07-Dec-2008 70 ; - - ; - 42 ; 66 - ; -

16 06-Jan-2009 14-Nov-2008 - ; - - ; 35 - ; - - ; -

17 06-Jan-2009 14-Nov-2008 - ; 6 48 ; 40 - ; - - ; -

18 05-Jan-2009 14-Nov-2008 41 ; 101 - ; - - ; - - ; -

19 05-Feb-2009 19-Dec-2008 13 ; - - ; - - ; - - ; -

20 06-Jan-2009 14-Nov-2008 63 ; 73 - ; - - ; - - ; -

21 07-Jan-2009 07-Dec-2008 - ; - - ; - - ; 33 - ; -

Total 396 ; 403 174 ; 190 163 ; 167 183 ; 181

69

VIP analysis) and the best model performed prediction for each of the images. This process was

applied in Matlab environment by PLS-toolbox of Eigenvector. Building PLS-DA models

included preprocessing the X-block by mean centering the data. Mean centering is applied as

preprocessing for PLS models by reducing variation within the data (Navalon et al. 1999). It

operates by subtracting the mean reflectance value for each wavelength from each reflectance

value. Then the data were validated by cross validation as customary for empirical models

(Borregaard et al. 2000). The cross validation was by every 10th sample in order to set the

number of latent variables to be applied for the model (Wu et al. 1997; Nason 1996).

Classification quality assessment methods were applied for these confusion matrices in order to

obtain the most suitable model for prediction of weeds in wheat field. The chosen model was

applied for prediction of all images. The PLS-DA classification prediction intermediately

resulted in two-dimensional image per class, in which, the value of each pixel is the probability

of the pixel to belong to this class. A pixel was chosen to belong to a certain class if the

probability of this class was higher than the others. The threshold for classification was set to be

0.3, meaning pixels with probability values smaller than 0.3 were defined as unclassified. Pixel

that was unclassified for all classes was determined as unclassified in the final classification

result image. The PLS-DA classification prediction resulted in 21 two-dimensional images, in

which, each pixel is related to one of the classes or defined as unclassified.

2.4.Classification quality assessment

The quality of PLS-DA models was compared based on the cross validation confusion matrices.

Cohen’s Kappa was calculated as defined by Cohen (1960) a unitless value ranging from 1 to -1

that stands for perfect agreement to complete disagreement, respectively. Computation of

Cohen’s Kappa is presented in equation (1):

d qKappaN q−

=−

(1)

where d is the sum of ground truth pixels that were correctly classified, q is the multiplication of

total ground truth and total classification values summed and divided by the total number of

samples, and N is the total number of samples. The confidence limits (CL) units are percent and

were calculated for overall accuracy as shown by Foody (2008) and presented in equation (2):

70

( ), 1

11N d

p pCL t

N−

−= ±

− (2)

where p is the overall accuracy, tN,d-1 is the statistical value of 95% two tailed test for d samples,

N is the total number of samples, d is the sum of ground truth pixels that were correctly

classified. The CL of the total accuracy can allow comparison between models that is are based

on total accuracy and by that show if there is a model that is significantly better or worse than

others (Foody 2008). Comparing the quality of coupled PLS-DA classification models was

applied as mentioned by Congalton and Mead (1986) and presented in equation (3):

1 2

1 2

Kappa KappaZVar Var

−=

+ (3)

where Z is the normal curve deviation and if it is >1.96 or <-1.96 the difference between the

confusion matrices is significant in 95% (2.58 and -2.58 are the thresholds for significance in

99%), Kappa is defined in equation (1), and Var is the variance of the confusion matrix as show

by Hudson and Ramm (1987) and not presented in the current study. The comparison is done for

two confusion matrices at a time, their Cohen’s Kappa and variance are calculated in order to

determine if they are significantly different from one another.

Quality of prediction by the PLS-DA model was assessed by ground truth data that were

obtained by the digital photos taken in field. For each spectral image 50 pixels, not included in

the calibration data set, were randomly selected to be compared to ground truth data. The

classification results in comparison to the ground truth were analyzed as confusion matrices. The

quality of the classification was assessed by Cohen’s Kappa coefficient, overall accuracy, user’s

accuracy, and producer’s accuracy for each confusion matrix. The classification quality was

assessed for each image separately (i.e., 50 ground truth pixels) and for all the images together

(i.e., 1050 ground truth pixels).

2.5.Relative coverage assessment

Relative coverage assessment was performed by three methods: field estimation, counting pixels

in the PLS-DA classification results, and by a simple classification Decision Tree (DT). The field

estimation was described earlier and the classes were BLW, GW, soil, and wheat. The final

71

result of PLS-DA classification for each image was used to count the pixels related to each of the

four classes (i.e., BLW, GW, soil, and wheat) and divide it by the amount of pixels in the image

in order to obtain relative coverage. The DT classified each image to five classes: sunlit

vegetation, shaded vegetation, glinted vegetation, sunlit soil, and shaded soil. The DT

classification was applied in ENVI software environment for all the images and resulted in 21

two-dimensional images, in which, each pixel is related to one of the five classes.

The DT, presented in Figure 2, is based on conditions applied for relative reflectance values of

three narrow bands (i.e., 470, 555, and 670 nm) that reflect differences between the classes. The

first condition checks if reflectance values in bands 470 and 555 nm are both lower than 0.05. In

case the condition is fulfilled the pixel is assumed to be shaded and the next condition will

determine if it is soil or vegetation. In case the condition is not fulfilled the pixel is assumed to

be sunlit and the other two conditions will determine if it is sunlit soil, glinted vegetation, or

sunlit vegetation.

Figure 2. Decision tree for separating five classes: sunlit vegetation, shaded vegetation, glinted

vegetation, sunlit soil, and shaded soil. Each condition, written in rectangular, has two options:

yes or no. Each option is leading to another condition or to classification product. The ρ stands

for relative reflectance value at the specified wavelength (in nm).

72

3. Results and discussion

Tables 2-7 present the cross validation confusion matrices of six PLS-DA models. These six

models are divided to three couples, each with the same classes with and without soil. In order to

find the best model, a comparison of confusion matrices was computed by the normal curve

deviation (Z; Eq. 3) for all model couplings (Table 8). In model # 1 the classes are either

broadleaf or grass (including wheat data), therefore, it does not distinguish between crop and

weeds, for model # 2 a soil class is added. These models were analyzed and presented in order to

learn more about the importance of soil as a distinguishable class. The overall accuracy of model

# 1 is higher than of model # 2 (Tables 2 and 3) and the difference between these confusion

matrices is significant (Table 8). Similarly, in other couples of models, i.e., models # 3 and 4

(Tables 4 and 5) and models # 5 and 6 (Tables 6 and 7), the only difference between them is an

additional soil class. The overall accuracy is higher for each model containing soil class and the

difference between the coupled models is significant as presented in Table 8. These three

couplings show that the total accuracy is significantly better (more than 99%) when soil is added

as a class. Since the canopy in the current study is denser than mentioned in the literature for

operating similar sensors (Okamoto et al. 2007; Borregaard et al. 2000; Feyaerts and van Gool

2001; Nieuwenhuizen et al. 2010) and since the current study is expected to be a step towards

up-scaling (soil will be included in the larger pixel), the influence of soil on classification quality

was explored. It can be assumed that, for the coupled models (i.e., models # 1 and # 2, # 3 and #

4, and # 5 and # 6), the improvement in the overall accuracy for models including soil class is

mainly influenced by the accuracies of the soil classes. This assumption is incorrect based on the

user’s and producer’s accuracies in Tables 2-7.

Models # 5 and 6 were also analyzed in order to demonstrate the effect of sunlit and shaded

pixels. In most of the cases the sunlit class causes better user’s and producer’s accuracies than

the shaded class (Tables 6 and 7). The shaded soil, and in a lesser extent the shaded GW user’s

accuracy values of model # 5, produce values that are similar to the soil and GW classes in

model # 3. When comparing the user’s accuracy of sunlit classes from model # 5 to the classes of

model # 3, the values are similar. Since the images used for prediction include 21 to 65% shade

and shaded vegetation out of the total area of the image (obtained by the DT and not presented),

model # 3, including sunlit and shaded pixels together, is a more efficient classifier than model #

5. The distribution of sunlit and shaded pixels is presented in Table 1. Table 8 shows significant

73

superior of model # 3 over models # 5 and 6. Model # 3 comprises, beside wheat and weeds, a

class of soil, and each class combines data from sunlit and shaded pixels. Therefore it was

chosen to be the one for applying the prediction by images with the aim of identifying weeds in

wheat field.

Table 2. Model # 1 cross validation of PLS-DA classification model of three classes: broadleaf

weeds, grass weeds and wheat, and soil. The model is based on four latent variables resulting in

confidence interval of +/- 1.4% for the overall accuracy, and Kappa = 0.79.

Ground truth

BLW G Soil

Total # of

classified samples

User’s accuracy %

correct

Cla

ssifi

catio

n re

sults

BLW 618 46 0 664 93

G 177 656 0 833 79

Soil 4 26 330 360 92


Producer’s accuracy % correct 77 90 100 89

Table 3. Model # 2 cross validation of PLSDA classification model of two classes: broadleaf

weeds and grass weeds. The model is based on five latent variables resulting in confidence

interval of +/- 1.7% for the overall accuracy, and Kappa = 0.73.

Ground truth

BLW G

Total # of

classified samples

User’s accuracy %

correct

Cla

ssifi

cat

ion

resu

lts BLW 681 89 770 88

G 118 639 757 84

Total # of ground truth samples 799 728

Producer’s accuracy % correct 85 88 87

74

Table 4. Model # 3 cross validation of PLSDA classification model of four classes: broadleaf

weeds, grass weeds, soil and wheat. The model is based on seven latent variables resulting in


Ground truth

BLW GW Soil Wheat

Total # of

classified samples

User’s accuracy %

correct

Cla

ssifi

catio

n re

sults

BLW 695 15 0 28 738 94

GW 41 297 0 53 391 76

Soil 7 0 330 25 362 91

Wheat 56 52 0 258 366 70


Producer’s accuracy % correct 87 82 100 71 85

Table 5. Model # 4 cross validation of PLSDA classification model of three classes: broadleaf

weeds, grass weeds, and wheat. The model is based on four latent variables resulting in


Ground truth

BLW GW Wheat

Total # of

classified samples

User’s accuracy %

correct

Cla

ssifi

catio

n re

sults

BLW 611 15 90 716 85

GW 36 267 58 361 74

Wheat 152 82 216 450 48


Producer’s accuracy % correct 76 73 59 70

75

Table 6. Model # 5 cross validation of PLSDA classification model of eight classes: broadleaf weeds, grass weeds, soil and wheat, all

sunlit as well as shaded. The model is based on five latent variables resulting in confidence interval of +/- 2.1% for the overall

accuracy, and Kappa = 0.70.

Ground truth

BLW sunlit

BLW shade

GW sunlit

GW shade

Soil sunlit

Soil shade

Wheat sunlit

Wheat shade



Cla

ssifi

catio

n re

sults

BLW sunlit 347 6 2 0 0 0 13 0 368 94

BLW shade 0 371 0 126 0 0 0 164 661 56

GW sunlit 28 0 153 0 0 0 17 0 198 77

GW shade 3 7 7 47 0 0 1 3 68 69

Soil sunlit 0 0 0 0 163 11 2 0 176 93

Soil shade 0 6 0 0 0 156 3 2 169 92

Wheat sunlit 18 0 12 0 0 0 145 0 175 83

Wheat shade 0 13 0 17 0 0 0 12 42 29


Producer’s accuracy % correct 88 92 88 25 100 93 79 7 71

76

Table 7. Model # 6 cross validation of PLSDA classification model of six classes: broadleaf

weeds, grass weeds, and wheat, all sunlit as well as shaded. The model is based on four latent

variables resulting in confidence interval of +/- 2.4% for the overall accuracy, and Kappa = 0.60.

Ground truth

BLW sunlit

BLW shade

GW sunlit

GW shade

Wheat sunlit

Wheat shade



Cla

ssifi

catio

n re

sults

BLW sunlit 336 1 0 0 28 0 365 92

BLW shade 1 263 0 63 0 54 381 69

GW sunlit 26 0 160 0 20 0 206 78

GW shade 0 49 3 44 4 34 134 34

Wheat sunlit 33 0 11 0 131 0 175 75

Wheat shade 0 90 0 83 0 93 266 35

Total # of ground truth samples 396 403 174 190 183 181

Producer’s accuracy % correct 85 65 92 23 72 51 65

Table 8. Normal curve deviation (Z) values of coupled PLS-DA classification models. Z value

>1.96 shows significant (>95%) difference between the confusion matrices. High lightened are

the important couplings.

It is important to emphasize that although the cross validation was obtained by pure pixels, the

spectral samples of vegetation classes were acquired from the canopy and those of the soil class

were acquired next to canopy or in its shade. As mentioned above, leaves are semitransparent for

Model # (Kappa) 1 (0.79) 2 (0.73) 3 (0.79) 4 (0.55) 5 (0.70) 6 (0.60)

1 (0.79) *

2 (0.73) 2.63 *

3 (0.79) 0.32 2.98 *

4 (0.55) 10.81 7.2 11.41 *

5 (0.70) 4.81 1.24 5.39 7.21 *

6 (0.60) 9.84 5.82 10.54 2.08 5.71 *

77

NIR radiation. Therefore, spectrum of soil shaded by vegetation will most likely include

vegetation signals, hence, a spectrum from each of the classes also includes the spectral signals

of the neighboring objects (Borregaard et al. 2000) e.g., leaves and soil, GW and BLW, and

sunlit as well as shaded. These objects can be the same class of the sample, another class, or even

a target that is not included in any of the classes (e.g., stems, glint, and stones). Therefore, each

of the presented models includes some amount of error that is caused by field conditions.

Consequently, model # 3, that includes in each of its classes sunlit and shaded spectra, is

believed to provide prediction results that are based on real field conditions.

In order to find the important wavelengths for separating the classes in model # 3, the VIP

analysis was applied for each of the classes and presented in Figure 3. For the three vegetation

classes, the most important spectral region is the red-edge with the highest peak at 730 nm while

for the soil the highest peak is at 685 nm on the edge of red and red-edge spectral regions. This

makes sense, since, the soil does not absorb radiation in the red region as vegetation does due to

the photosynthetic process (Gausman 1985) and the 685 nm wavelength is the reflectance deep

for the three sunlit vegetation classes (data not presented). This goes in agreement with the

perfect producer’s accuracy for soil and sunlit soil classes in Tables 2, 4, and 6. Spectra of

shaded soil, in a vegetated area, can include elements of vegetation (e.g., absorption in the red

region and enhanced reflectance in the NIR region). This might be part of the reason that the soil

and soil shaded classes do not have perfect user’s accuracy in these three tables. For the wheat

and GW classes, beside the red-edge, the important regions are the blue and green regions. In

most of the cases, the GW plants, in the field as well as in the digital pictures, have a lighter

green hue than the wheat plants that seem to be more bluish. This combination of four

wavelengths (i.e., blue, green, red, and red-edge) resulted in being the most important for

separating the four classes of model # 3.

Hyperspectral satellites that are active today (e.g., Hyperion and CHRIS) as well as forthcoming

(e.g. HyspIRI and EnMAP) are rare and might not be able to meet the spatial requirements for

site specific weed managements. Currently there are only two operational multispectral satellites

WorldView-2, and RapidEye which can meet the four bands requirements (including a red-edge

band) along with high spatial resolution for site specific agricultural applications. Other two

forthcoming satellites are the Vegetation and Environmental New micro Spacecraft (VENμS)

78

and Sentinel-2 to be launched in 2014, the earliest. There is a need for further exploring the

influence of reduced spatial resolution on the quality of prediction, in other words, to deal with

the mixed pixel issue in order to find out what is the ideal pixel size for weed analysis. This

suggested to be a specific airborne (hyper or multi spectral) mission for calibrating several

common weeds and crops with relation to economic benefits of weed control. Although satellites

can obtain coverage of several fields in one image, it seems that high spatial resolution that

allows identifying early weed infestation, similar to ground level, is beyond the goals of current

or near future satellites. Therefore, air and ground level application based on similar sensors,

with higher spatial resolution, is also a course to be concentrated on in the near future.

Figure 3. Variable Importance in Projection (VIP) values of model #3 for the four classes in

model # 3 and the threshold for VIP values.

Table 9 presents the total accuracy and its confidence interval, Cohen’s Kappa, and user’s and

producer’s accuracies obtained from 21 confusion matrices computed in order to calculate the

quality of prediction by model # 3 for the 21 images. The total accuracies are ranging from 54 to

90 % and when considering the confidence intervals it can range from 40 to 98 %. The highest

total accuracies of 88 and 90 % were both obtained in images that do not contain all four classes

79

to begin with or the randomly picked ground truth pixels did not cover all available classes, as

can be seen by the 0 values in both user’s and producer’s accuracies. Correlating the total

accuracy to vegetation coverage by field assessment, PLS-DA and DT for images presented in

Table 9, resulted in very weak, non-significant, negative relations with R2 <0.025. Therefore, it

can be assumed that model # 3 predictions are not influenced by the vegetation cover range of

values of the current data. The vegetation cover was obtained by the three methods mentioned

above and resulted in 35-95% by the field assessment, 28-100% by the PLS-DA model, and 15-

99% by the DT.

The confusion matrix for all the ground truth point together is presented in Table 10. The 1050

ground truth pixels were distributed among the classes almost evenly and the amount of

unclassified pixels is neglect able. Soil is the class that its accuracy of prediction is the highest.

BLW can be predicted with 81% success but BLW is predicted as BLW only with 52% success.

For the wheat class it is the other way around as wheat can be predicted with 60% success but

wheat is actually predicted as wheat with 79% success. Pixels of BLW, GW, and soil that are

mistakenly classified are in most of the cases classified as wheat 68, 49, and 21, respectively.

Therefore, the user’s accuracy of wheat is the lowest. In Table 6, most of the miss-classifications

are between shaded vegetation classes and since in model # 3 the classes include sunlit and

shaded data in the same class, this might be the source of miss-classifications in the vegetation

classes presented in Table 10. The 1857 cross validation pixels were collected from homogenous

regions and were not picked from stones, stems, glinted, and upside down leaves as well as not in

adjacency to leaves edges. The 1050 ground truth pixels were randomly selected, therefore,

might be of target that its spectrum is similar to more than one class (e.g., pixel on the edge of a

leaf). In order to generalize the model, spectral data were collected in several plots and growing

stages. Tyystjarvi et al. (2011) obtained best results for species classification by training and test

leaf florescence measurements that were obtained on the same date. Therefore it is assumed that

the classification results could have been improved if the cross validation was applied for data

obtained from one date, one plot, or even one image.

80

Table 9. Vegetation coverage by three methods for each image and ground truth accuracy results (50 pixels per image) by model # 3.

Vegetation coverage % Users accuracy % Producers accuracy %

Image #

Total accuracy

%

Confidence interval %

Cohen’s Kappa

Field assessment PLSDA Decision

tree BLW GW Soil Wheat BLW GW Soil Wheat

1 78 11.9 0.67 52 61 77 33 75 94 73 20 75 94 80

2 70 13.3 0.49 75 96 99 0 89 40 62 0 57 100 94

3 78 11.9 0.67 67 81 93 100 50 85 85 17 56 92 96

4 82 11.1 0.67 90 96 100 33 100 0 73 50 77 0 94

5 90 8.6 0.51 40 15 28 60 0 100 0 100 0 89 0

6 84 10.6 0.76 80 94 98 100 86 75 81 44 90 100 94

7 62 14.1 0.46 70 54 71 56 23 95 50 45 33 76 80

8 76 12.3 0.64 85 86 93 75 50 57 100 33 100 100 79

9 54 14.6 0.37 73 86 92 100 56 27 67 14 63 80 76

10 78 11.9 0.66 78 79 87 100 67 89 79 14 73 89 96

11 70 13.3 0.54 80 76 91 100 45 64 81 33 56 78 84

12 66 13.7 0.51 65 52 71 100 29 90 9 52 67 83 50

13 58 14.4 0.39 45 52 69 100 0 96 9 21 0 82 100

14 70 13.3 0.54 55 57 77 80 0 94 29 67 0 88 50

15 74 12.7 0.53 50 40 60 100 0 97 14 36 0 91 100

16 74 12.7 0.44 90 97 98 0 82 0 86 0 94 0 50

17 64 14 0.29 90 96 100 75 75 0 22 43 79 0 22

18 66 13.7 0.29 90 95 98 76 33 67 25 85 29 50 17

19 54 14.6 0.37 90 86 92 100 64 38 47 6 78 75 75

20 88 9.3 0 95 99 100 88 0 0 0 100 0 0 0

21 76 12.3 0.62 35 41 61 80 0 92 53 36 0 100 83

81

Table 10. Prediction of all the images together by model # 3 with confidence interval

of +/- 2.7% for the overall accuracy, and Kappa = 0.63.

Ground truth

Unclassified BLW GW Soil Wheat


User’s accuracy %

correct

Cla

ssifi

catio

n re

sults

Unclassified 0 0 3 2 0 5 0

BLW 0 139 12 6 14 171 81

GW 0 39 175 8 27 249 69

Soil 0 20 9 232 16 277 83

Wheat 0 68 49 21 210 348 60

Total # of ground truth samples 0 266 248 269 267

Producer’s accuracy % correct 0 52 71 87 79 72

Figure 4 displays two images and their classification results by model # 3. Figure 4(a)

presents image with the lowest total accuracy and Figure 4(c) with above the average

values (Table 9). The relative coverage assessed in the field shows 30 and 25% of

GW, 40 and 50% of wheat, and 27 and 22% of soil, respectively, and each 3% of

BLW. The DT analysis resulted in similar relative coverage with differences that are

not higher than 4% for each of the five classes. The quality of prediction confusion

matrices comparison resulted in significant difference (more than 95%), based on Z=

2.27. The distribution of ground truth pixels in both images are similar with

differences of 1, 3, 6, and 4 pixels for the four classes BLW, GW, soil, and wheat,

respectively. Both images were acquired at the same field on the same day with less

than half an hour difference. Therefore, beside human assessment or other errors or

miss accuracies, the limitations of the model itself, the difference between the

confusion matrices is assumed to be related to targets not included in the cross

validation process (e.g., stones, stems, glint, and upside-down leaves). Another

possible option is that this can be also partly a result of model that was cross validated

from a variety of spectral, growth stages, and environmental conditions combined

with infield spectral variability.

82

Figure 4. (a) an example of hyperspectral ground-level image with the lowest total

accuracy and (b) its classification; (c) an example on an image with above the average

total accuracy and (d) its classification.

The total accuracy and Cohen’s Kappa of prediction can be a result of the number of

pixels obtained from the image by 11.3 and 0.2%, respectively, both not significant.

Therefore, it is assumed that the influence of the amount of cross validation pixels on

the Cohen’s Kappa as well as the total accuracy is neglect able and model # 3

classification prediction results for the 21 images, are not influenced by the cross

validation pixels distribution among the images. Figure 5 presents user’s and

producer’s accuracy of a class for each image related to the amount of cross validation

pixels, of the same class, acquired from each image. GW and soil classes correctly

identified among pixels classified (i.e., user’s accuracy) as GW and soil, respectively,

by model # 3 are significantly influenced by the amount of cross validation pixels

acquired in the image by 39 and 22 %, respectively. BLW class correctly classified

among pixels known to be BLW (i.e., producer’s accuracy) by model # 3 are

83

significantly influenced by the amount of cross validation pixels acquired in the image

by 47%. When including in the analysis only the images that cross validation pixels

were acquired from and the user’s accuracy, the R2 values of BLW, GW, soil, and

wheat were 0.01, 0.27, 0.83, and 0.1, respectively. When including in the analysis

only the images that pixels were acquired from and the producer’s accuracy, the R2

values of BLW, GW, soil, and wheat were 0.29, 0.06, 0.31, and 0.03, respectively.

Comparing these R2 values to the R2 values in Figure 5 it is resulting in reduction of

correlation between the amount of pixels and the user’s as well as producer’s

accuracies for the three vegetation classes. Therefore, the vegetation classification

quality is almost not influenced by the distribution of pixels among the images. In the

case of soil the tendency is opposite and therefore it seems that for soil it is important

to distribute the cross validation pixels among more images.

In order to decide whether to apply weed control there is a need to know the relative

coverage of weeds in the field (Slaughter et al. 2008). In Figure 4 (b) and (d) are

presented examples of classification results by model # 3, such images were used to

get the relative coverage of each of the four classes. There is positive, significant

correlation between the relative coverage obtained by PLS-DA classification and by

field assessment (Figure 6). BLW and soil that show relatively better users accuracy

results (Table 9) provide better correlation to field assessment. Therefore, it can be

assumed that relative coverage of the four classes can be assessed by model # 3 with

the limitations presented above. On top of human error in assessment and

misclassifications of the model, differences in relative coverage can be partly an

outcome of the clipping process. Since the frame that was used for field assessment

was not always parallel to image borders, narrow triangles, in the frame area, were

eliminated from the clipped images that were used to predict and ground truth for the

PLS-DA model.

84

Figure 5. The amount of cross validation pixels, of each class, acquired from each

image in relation to: (a) user’s accuracy; and (b) producer’s accuracy, of the class for

each image.

85

Figure 6. Relating relative coverage of BLW, GW, soil, and wheat by field

assessment and by model # 3 classification results.

4. Summary and conclusions

Classification of wheat and weeds was the aim of the current study. Pixels were

picked from ground level images and used to build six PLS-DA models. The best

model was found to be the one that includes soil as an additional class and combines

sunlit and shaded pixels in each of its four classes: BLW, GW, soil, and wheat. The

important wavelengths for each of the classes of the best model were obtained by the

VIP method. The best model was applied for the images and ground truth was applied

by RGB photos. The total ground truth has resulted in an overall accuracy of 72%.

These results indicate that differentiation between wheat and weeds is possible using

PLS-DA, therefore, potentially contributing to practical site-specific herbicide

application. Specific conclusions are:

• Composition of four classes: BLW, GW, soil, and wheat, was the best for

weed detection for the current data set.

• Sunlit vegetation can be better separated to classes than shaded vegetation.

• The red-edge is the most important region for separation among wheat, BLW,

and GW. For wheat and GW the blue and green regions are also important,

86

respectively. For soil separation among vegetation the edge between red and

red-edge is most important spectrally.

• Although the model cross validation and ground truth were acquired for

heterogenic data, the model obtained reasonable results and therefore is

potentially applicable.

• High spectral and spatial resolutions can provide separation between wheat

and weeds based on spectral data alone.

Future work that is more aimed at economical thresholds for applying weed control

should be concentrated in one of two directions: covering wider areas, with coarser

spatial resolution, by satellites (e.g., WorldView-2, RapidEye, VENμS, and Sentinel-

2) while exploring the mixed pixel issue; or covering relatively smaller areas, with

fine spatial resolution, by ground level sensors in order to deal with sunlit and shaded

parts in the canopy and soil separately or better understand their mutuality.

5. References

Alchanatis, V., Ridel, L., Hetzroni, A., & Yaroslavsky, L. (2005). Weed detection in multi-spectral images of cotton fields. Computers and Electronics in Agriculture, 47(3), 243-260.

Barker, M., & Rayens, W. (2003). Partial least squares for discrimination. Journal of Chemometrics, 17(3), 166-173, doi:10.1002/cem.785.

Bennett, A. L., & Pannell, D. J. (1998). Economic evaluation of a weed-activated sprayer for herbicide application to patchy weed populations. Australian Journal of Agricultural and Resource Economics, 42(4), 389-408, doi:10.1111/1467-8489.00059.

Biller, R. H. (1998). Reduced input of herbicides by use of optoelectronic sensors. [Article]. Journal of Agricultural Engineering Research, 71(4), 357-362.

Blackshaw, R. E., Molnar, L. J., & Lindwall, C. W. (1998). Merits of a weed-sensing sprayer to control weeds in conservation fallow and cropping systems. Weed Science, 46(1), 120-126.

Borregaard, T., Nielsen, H., Norgaard, L., & Have, H. (2000). Crop-weed discrimination by line imaging spectroscopy. Journal of Agricultural Engineering Research, 75(4), 389-400.

Brent, J., & Schaeffer, T. H. (2011). Systematic Review of Parkinsonian Syndromes in Short- and Long-Term Survivors of Paraquat Poisoning. Journal of Occupational and Environmental Medicine, 53(11), 1332-1336, doi:10.1097/JOM.0b013e318233775d.

Brown, R. B., & Steckler, J. (1995). Prescription maps for spatially variable herbicide application in no-till corn. Transactions of the Asae, 38(6), 1659-1666.

87

Brown, R. B., Steckler, J., & Anderson, G. W. (1994). Remote-Sensing for Identification of Weeds in No-Till Corn. [Note]. Transactions of the Asae, 37(1), 297-302.

Chong, I. G., & Jun, C. H. (2005). Performance of some variable selection methods when multicollinearity is present. Chemometrics and Intelligent Laboratory Systems, 78(1-2), 103-112.

Cohen, J. (1960). A coefficient of agreement for nominal scales Educational and psychological measurement, 20(1), 37-46.

Cohen, Y., Alchanatis, V., Zusman, Y., Dar, Z., Bonfil, D., J. , Karnieli, A., et al. (2010). Leaf nitrogen estimation in potato based on spectral data and on simulated bands of the VENuS satellite. Precision Agriculture, 11(5), 520-537, doi:10.1007/s11119-009-9147-8.

Congalton, R. G., & Mead, R. A. (1986). Areview of three discrete multivariate analysis techniques used in assessing the accuracy of remotely sensed data from error matrices. IEEE Transaction on Geoscience and Remote Sensing, 24(1), 169-174.

Dhawan, A., Bajpayee, M., & Parmar, D. (2009). Comet assay: a reliable tool for the assessment of DNA damage in different models. Cell Biology and Toxicology, 25(1), 5-32, doi:10.1007/s10565-008-9072-z.

Eddy, P. R., Smith, A. M., Hill, B. D., Peddle, D. R., Coburn, C. A., & Blackshaw, R. E. (2008). Hybrid segmentation - Artificial Neural Network classification of high resolution hyperspectral imagery for Site-Specific Herbicide Management in agriculture. [Article]. Photogrammetric engineering and remote sensing, 74(10), 1249-1257.

Feyaerts, F., & van Gool, L. (2001). Multi-spectral vision system for weed detection. Pattern Recognition Letters, 22(6-7), 667-674.

Foody, G. M. (2008). Harshness in image classification accuracy assessment. International Journal of Remote Sensing, 29(11), 3137-3158, doi:10.1080/01431160701442120.

Gausman, H. (1985). Plant leaf optical properties in visible and Near Infrared light. Lubbock TX. USA: Texas Tech Press.

Gerhards, R., & Christensen, S. (2003). Real-time weed detection, decision making and patch spraying in maize, sugarbeet, winter wheat and winter barley. [Article]. Weed Research, 43(6), 385-392.

Gerhards, R., Sokefeld, M., Schulze-Lohne, K., Mortensen, D. A., & Kuhbauch, W. (1997). Site specific weed control in winter wheat. Journal of Agronomy and Crop Science, 178, 219-225.

Gibson, L., Kingwell, R., & Doole, G. (2008). The role and value of eastern star clover in managing herbicide-resistant crop weeds: A whole-farm analysis. Agricultural Systems, 98(3), 199-207.

Gray, C. J., Shaw, D. R., & Bruce, L. M. (2009). Utility of Hyperspectral Reflectance for Differentiating Soybean (Glycine max) and Six Weed Species. Weed Technology, 23(1), 108-119.

Hamouz, P., Novakova, K., Soukup, J., & Holec, J. (2008). Detection of Cirsium arvense L. in winter wheat using a multispectral imaging system. Journal of Plant Diseases and Protection, 167-170.

Har Gil, D., Bonfil, D. J., & Svoray, T. (2011). Multi scale analysis of the factors influencing wheat quality as determined by Gluten Index. Field Crops Research, 123(1), 1-9, doi:10.1016/j.fcr.2011.04.001.

88

Hatchell, D. (1999). Reflectance. http://www.asdi.com/tg_rev4_web.pdf. Accessed 9, July, 2008.

Herrmann, I., Pimstein, A., Karnieli, A., Cohen, Y., Alchanatis, V., & bonfil, D., J. (2011). LAI assessment of wheat and potato crops by VENμS and Sentinel-2 bands. Remote Sensing of Environment, 115, 2141-2151.

Hudson, W. D., & Ramm, C. W. (1987). Correct formulation of the kappa-coefficient of agreement. Photogrammetric engineering and remote sensing, 53(4), 421-422.

Johnson, G. A., Mortensen, D. A., Young, L. J., & Martin, A. R. (1995). The Stability of Weed Seedling Population-Models and Parameters in Eastern Nebraska Corn (Zea-Mays) and Soybean (Glycine-Max) Fields. Weed Science, 43(4), 604-611.

Jones, R. E., Vere, D. T., Alemseged, Y., & Medd, R. W. (2005). Estimating the economic cost of weeds in Australian annual winter crops. [Article]. Agricultural Economics, 32(3), 253-265.

Kafkafi, U., & Bonfil, D., J. (2008). Integrated Nutrient Management: Experience and Concepts from the Middle East. In M. Aulakh, & C. A. Grant (Eds.), Integrated Nutrient Management for Sustainable Crop Production (pp. 523-565). Binghamton NY. USA: The Haworth Press, Inc. .

Lamb, D. W., & Brown, R. B. (2001). Remote-sensing and mapping of weeds in crops. Journal of Agricultural Engineering Research, 78(2), 117-125.

Lindquist, J. L., Dieleman, J. A., Mortensen, D. A., Johnson, G. A., & Wyse-Pester, D. Y. (1998). Economic importance of managing spatially heterogeneous weed populations. Weed Technology, 12(1), 7-13.

Lopez-Granados, F., Pena-Barragan, J. M., Jurado-Exposito, M., Francisco-Fernandez, M., Cao, R., Alonso-Betanzos, A., et al. (2008). Multispectral classification of grass weeds and wheat (Triticum durum) using linear and nonparametric functional discriminant analysis and neural networks. Weed Research, 48(1), 28-37.

Mallory-Smith, C. A., Thill, D. C., & Dial, M. J. (1990). Identification of Sulfonylurea Herbicide-Resistant Prickly Lettuce (Lactuca-Serriola). Weed Technology, 4(1), 163-168.

Manh, A. G., Rabatel, G., Assemat, L., & Aldon, M. J. (2001). Weed leaf image segmentation by deformable templates. [Article]. Journal of Agricultural Engineering Research, 80(2), 139-146.

Marshall, R., & Moss, S. R. (2008). Characterisation and molecular basis of ALS inhibitor resistance in the grass weed Alopecurus myosuroides. Weed Research, 48(5), 439-447.

Monaco, T., J, Weller, S., C, & Ashton, F., M (2002). Weed science principles and practices New York NY. USA: John Wiley & sons, Inc.

Moran, M., S., Inoue, Y., & Barnes, E., M. (1997). Opportunities and Limitations for Image-Based Remote Sensing in Precision Crop Management. Remote Sensing of Environment, 61(3), 319-346.

Moran, M. S., Maas, S. J., Vanderbilt, V. C., Barnes, M., Miller, S. N., & Clarke, T. R. (2004). Application of image-based remote sensing to irrigated agriculture. In S. L. Ustin (Ed.), Remote sensing for natural resource management and environmental monitoring (3 ed., Vol. 4, pp. 617-676, Manual of remote sensing). Hoboken NJ. USA: John Wiley & sons.

Musumarra, G., Barresi, V., Condorelli, D. F., Fortuna, C. G., & Scire, S. (2004). Potentialities of multivariate approaches in genome-based cancer research:

89

identification of candidate genes for new diagnostics by PLS discriminant analysis. Journal of Chemometrics, 18(3-4), 125-132, doi:10.1002/cem.846.

Nason, G. P. (1996). Wavelet shrinkage using cross-validation. Journal of the Royal Statistical Society Series B-Methodological, 58(2), 463-479.

Navalon, A., Blanc, R., del Olmo, M., & Vilchez, J. L. (1999). Simultaneous determination of naproxen, salicylic acid and acetylsalicylic acid by spectrofluorimetry using partial least-squares (PLS) multivariate calibration. Talanta, 48(2), 469-475, doi:10.1016/s0039-9140(98)00268-9.

Nieuwenhuizen, A. T., Hofstee, J. W., & Van Henten, E. J. (2010). Adaptive detection of volunteer potato plants in sugar beet fields. Precision Agriculture, 11(5), 433-447, doi:10.1007/s11119-009-9138-9.

Okamoto, H., Murata, T., Kataoka, T., & Hata, S. I. (2007). Plant classification for weed detection using hyperspectral imaging with wavelet analysis. Weed Biology and Management, 7(1), 31-37.

Papayiannis, L. C., Kokkinos, C. D., & Alfaro-Fernandez, A. (2012). Detection, characterization and host range studies of Pepino mosaic virus in Cyprus. European Journal of Plant Pathology, 132(1), 1-7, doi:10.1007/s10658-011-9854-7.

Pikart, T. G., Souza, G. K., Serrao, J. E., & Zanuncio, J. C. (2011). Leafcutter ants: a small dispersal agent of the invasive plant Murraya paniculata. Weed Research, 51(6), 548-551, doi:10.1111/j.1365-3180.2011.00881.x.

Pinter, P., J., Hatfield, J., L., Schepers, J., S., Barnes, E., M., Moran, M., S., Daughtry, C., S, T., et al. (2003). Remote sensing for crop management. Photogrametric Engineering & Remote Sensing, 69(6), 647-664.

Raven, P., H, Everet, R., F, & Eichhorn, S., E (2005). Biology of plants (7ed.). New-York: W. H. Freeman and Company.

Slaughter, D. C., Giles, D. K., & Downey, D. (2008). Autonomous robotic weed control systems: A review. Computers and Electronics in Agriculture, 61(1), 63-78.

Smith, A. M., & Blackshaw, R. E. (2003). Weed-crop discrimination using remote sensing: A detached leaf experiment. [Article]. Weed Technology, 17(4), 811-820.

Thorp, K. R., & Tian, L. F. (2004). A review on remote sensing of weeds in agriculture. Precision Agriculture, 5(5), 477-508.

Tian, Q., Tong, Q., Pu, R., Gou, X., & Zhao, C. (2001). Spectroscopic determination of wheat water status using 1650-1850 nm spectral absorption features. International Journal of Remote Sensing, 22(11), 2329-2338.

Timmermann, C., Gerhards, R., & Kuehbauch, W. (2003). The economic impact of site-specific weed control. Precision Agriculture, 4(3), 249-260.

Tyystjarvi, E., Norremark, M., Mattila, H., Keranen, M., Hakala-Yatkin, M., Ottosen, C. O., et al. (2011). Automatic identification of crop and weed species with chlorophyll fluorescence induction curves. Precision Agriculture, 12(4), 546-563, doi:10.1007/s11119-010-9201-6.

Vrindts, E., De Baerdemaeker, J., & Ramon, H. (2002). Weed detection using canopy reflection. Precision Agriculture, 3(1), 63-80.

Weis, M., Gutjahr, C., Rueda Ayala, V., Gerhards, R., Ritter, C., & Scholderle, F. (2008). Precision farming for weed management: techniques. Gesunde Pflanzen, 60(4), 171-181.

90

Wold, S., Johansson, E., & Cocchi, M. (1993). PLS- partial least squars projections to latent structures. In H. Kubinyi (Ed.), 3D QSAR in drug design: theory, methods, and applications (pp. 523-550). Leiden: ESCOM.

Wu, W., Massart, D. L., & deJong, S. (1997). Kernel-PCA algorithms for wide data, part 2- Fast cross-validation and application in classification of NIR data. Chemometrics and Intelligent Laboratory Systems, 37(2), 271-280, doi:10.1016/s0169-7439(97)00027-0.

Xie, L., Ying, Y., & Ying, T. (2007). Quantification of chlorophyll content and classification of nontransgenic and transgenic tomato leaves using visible/near-infrared diffuse reflectance spectroscopy. Journal of Agricultural and Food Chemistry, 55(12), 4645-4650, doi:10.1021/jf063664m.

Yoder, B. J., & Pettigrew-crosby, R. E. (1995). Predicting nitrogen and chlorophyll content and concentrations from reflectance spectra (400-2500nm) at leaf and canopy scales. Remote Sensing of Environment, 53(3), 199-211.

Zwiggelaar, R. (1998). A review of spectral properties of plants and their potential use for crop/weed discrimination in row-crops. Crop Protection, 17(3), 189-206.

91

Chapter 6

6. Spectral monitoring of two-spotted spider mite damage to pepper leaves.

Remote Sensing Letters, 3, pp. 277-283.

99

Capter 7

7. Summary and main conclusions

This dissertation combines four different potential RS applications for PA of field

crops under one roof, namely – N content assessment, LAI assessment, weed

detection and early detection of TSSM damage. These applications were presented in

five studies and the summary, main conclusions and innovative contributions as well

as hypotheses evaluation of each of the studies are described below.

Chapter 2 – For assessing the total N content of plants, several known VIs are based

on spectral wavelengths that are indirectly related to N content (i.e., via chlorophyll

content) but none combines wavelengths that are directly and indirectly related to N.

It is hypothesized that applying indices based on wavelengths that are directly and

indirectly related to total N content in plants will improve the ability to evaluate it.

The new VIs for N content assessment in plants (e.g., NRI1510) combine direct and

indirect relations to N content, meaning combining the actual presence of N in the

plant and the repercussions of it. Three methods for evaluating and comparing the

indices' performances demonstrate the unequivocal advantages of the four proposed

new VIs. These indices are higher correlated, better predictors of N content, and more

sensitive to it than the other VIs examined in this study. These findings support the

hypothesis of amplifying the N predicting ability by combining direct and indirect

relations to N content as well as reinforcing the sensitivity of the four new indices to

N content. In addition, the NRI1510 also presents the advantage of combining N and

chlorophyll absorption features. Among the four new SWIR-based N indices, there is

not one with an apparent absolute advantage over the others. It is important to

mention, however, that although there are advantages of applying the SWIR region

for PA research, the SWIR sensors are not as common and are more costly than the

VNIR sensors (Karnieli et al. 2001, Ben-Ze’ev et al. 2006, Herrmann et al. 2010).

Therefore, the following studies in this dissertation (i.e., Chapters 3-6) focused on the

VNIR region. When the SWIR sensors (ground, air, and space levels) are more

common, this study will be highly relevant for applications dealing with N content

assessment in vegetation for agriculture as well as in natural habitats.

Chapter 3 – Three spectral data formations (i.e., hyperspectral, VENμS, and Sentinel-

2) were studied, for wheat and potato crops, in order to explore the ability of super-

100

spectral resolutions to assess LAI values. It is hypothesized that the band formation in

the red-edge region of future super-spectral satellites can allow retrieving LAI values

in wheat and potato crops. LAI prediction was examined by the entire spectra by the

three data formations, as well as by two VIs (i.e., NDVI and REIP). The VIP over the

PLS analysis confirmed that the red-edge is a highly sensitive region to LAI

variability in the range of 400 – 1000 nm. The results, in agreement with the

hypothesis, show that super-spectral sensors can perform as well as the hyperspectral

sensor in the LAI prediction of wheat and potato. Therefore, analyses of data obtained

by the forthcoming super-spectral sensors, VENμS and Sentinel-2, might take

advantage of these results for agricultural (and related) applications. When

considering the designed spatial resolutions, VENμS might potentially provide better

practical applications for precision agriculture. For wheat, and, to a lesser extent, for

potato, and in agreement with the hypothesis, it is demonstrated that the REIP is more

sensitive to LAI variability than the NDVI. Therefore, the REIP can be implemented

by the super-spectral sensors for this application. The main conclusions can be

summarized as:

1. LAI prediction by continuous data, in the range of 400-1000 nm, does not

provide any significant advantage over the VENμS and Sentinel-2 resampled

data. Therefore, these satellites’ band formations are as good as hyperspectral

sensors for LAI prediction by the entire spectra as well as by NDVI and REIP.

2. In the range of 400-1000 nm, the red-edge region is the most or second most

sensitive spectral region to changes in LAI values. The degree of importance

is determined by the specific band formation as well as by the crop.

3. While NDVI loses sensitivity for LAI values exceeding 2 (Gitelson 2004,

Coyne et al. 2009, Herrmann et al. 2011), the REIP can be used for predicting

LAI at all growth stages together, even for high biomass, and it is a better LAI

ground-level predictor for wheat than for potato.

To the best of my knowledge, there are no precision agricultural specific

implementations based on spectral data to be obtained by the future satellite VENμS.

Canisius and Fernandes (2012) applied data obtained by the super-spectral satellite

Medium Resolution Imaging Spectrometer (MERIS) with 300 m spatial resolution for

large area LAI monitoring. It was concluded that two red-edge indices are more stable

than others over time, and it was recommended to further study the LAI assessment

101

by red-edge VIs for Sentinel-2. Delegido et al. (2011) presented green LAI retrieval

by simulated Compact High Resolution Imaging Spectroscopy (CHRIS) data to

Sentinel-2 spectral resolution. It resulted in an LAI assessment by a red-edge VI with

similar quality to the results presented in this dissertation, and the importance of the

red-edge region for LAI assessment was cited from Herrmann et al. (2011).

Chapter 4 – RS data obtained by ground-level point spectrometer, from leaf and

canopy levels, were applied in order to detect grass and broadleaf weeds among cereal

and broadleaf crops. It is hypothesized that point spectral reflectance of leaf and

canopy can be used to identify categories of wheat, chickpea, grass weeds, and

broadleaf weeds. The leaf spectra were classified by GDA and resulted in almost

perfect overall accuracies, approving the hypothesis. The canopy spectral

classification by GDA was 95% successful for wheat and 94% for chickpea. The

results indicate that differentiation between weeds and crop is possible using GDA as

was hypothesized. Therefore, it is concluded that qualitative models based on wheat,

chickpea, grass weed and broadleaf weed spectral properties have high quality

classification and prediction potential that can be used for site-specific weed

management. Specific conclusions are:

4. Wavelengths in the red-edge region were found to be highly important for

crop and weed classification.

5. Site-specific herbicide use, based on spectral separation, was proven to be

potentially applicable for wheat fields with >5% vegetation cover in the

critical period for weed control (e.g., before the canopy closes).

In order to make this an operational application, there is a need to decide on the

platform: satellite, aerial, or ground-level; each has its benefits and shortcomings.

Applying a spaceborne sensor, for a site-specific weed management application, such

as VENμS, should be done while or after resolving the mixed pixel issue as

mentioned by Biewer et al. (2009). If the ground-level platform is chosen, using a

sensor with high spatial resolution is recommended.

Chapter 5 – RS data were obtained by ground-level imagery in order to detect grass

and broadleaf weeds in a wheat field. It is hypothesized that high spectral and spatial

ground-level image spectroscopy data, can allow spatial separation of broadleaf weed,

grass weed, and wheat. Pixels from images were used to cross validate PLS-DA

102

classification models. The best model used four classes: broadleaf weeds, grass

weeds, soil, and wheat. Each of the classes contained sunlit and shaded data together.

The VIP method was applied in order to locate the most important spectral regions for

each of the classes. Randomly selected ground truth pixels resulted in a total accuracy

of 72%. Since the model used wheat and weeds from different growth stages,

acquisition dates, and fields, the results obtained are reasonable, therefore supporting

the hypothesis. The main conclusions were that:

• The red-edge is the most important region for separation among wheat,

broadleaf weeds, and grass weeds. For wheat and grass weeds, the blue and

green regions are second in importance, respectively. For soil separation

among vegetation, the edge between the red and red-edge is most important

spectrally.

• High spectral and spatial resolutions can provide reasonable separation, with

spatial attributes, between wheat, weeds and soil, based on heterogenic

spectral data.

To the best of my knowledge, this is the first time that the spatial ability to spectrally

separate by the visible and NIR regions, on the ground-level, between wheat and both

grass and broadleaf weeds in the critical period of weed control, has resulted in such

affirmative quality. Fitzgerald et al. (2005) obtained ground-level spectral data and

evaluated plant coverage by separating between four endmembers (plant, shaded

plant, soil and shaded soil). The inclusion of economical thresholds for weed control

should be concentrated in one of two directions: 1) covering wider areas, with coarser

spatial resolution, by satellites while exploring the mixed pixel issue as an ad hoc

mission for specific weeds and crops or as a broader mission for several weeds and

crops; and 2) covering small areas, with fine spatial resolution, by airborne or ground-

level sensors in order to deal with sunlit and shaded parts in the canopy and soil

separately or at least to better understand their mutuality.

The combined contribution of the studies presented in chapters 4 and 5 is the in situ

spectral separation between crops and weeds, grass as well as broadleaf, by ground-

level hyperspectral relative reflectance of a point spectrometer and by scanned image

data as a step forward in a real-time system of weed control and, possibly, as part of

an operating wheat decision support system (Bonfil et al. 2004).

103

Chapter 6 – RS data obtained by a point spectrometer and an active light source for

leaves in the lab were used to compute VIs. It is hypothesized that early damage of

TSSM to greenhouse pepper leaves can be detected. The abilities of several VIs to

separate different damage levels of TSSM to greenhouse pepper leaves were

compared. The results, supporting the hypothesis, showed the ability of five out of six

VIs to separate between no damage and low damage classes, and three out of six VIs

could consistently separate between the four damage levels. The main conclusions

are:

• Although the TSSM activity is on the underside of leaves, early damage to

greenhouse pepper leaves can be identified by RS means in the laboratory.

• One green NIR and two red-edge VIs can provide consistent separation

between increasing damage levels to greenhouse pepper leaves in the

laboratory.

To the best of my knowledge, assessment of TSSM damage to leaves is a new

measure, and other studies looked at the number of arthropods living on a leaf or

young plant. This concept of damage location is innovative and should be beneficial

for growers to base their control on the damage itself and not on the number of pests.

Some of the aforementioned conclusions, from chapters 2-6, have common ground.

This common ground is an added value to the wide definition of this dissertation titled

“Hyperspectral applications of precision agriculture for field crops in drylands”

including four different potential applications. The common ground allows us to

connect the conclusions dealing with the red-edge spectral region to the

comprehensive understanding that the red-edge is important to LAI assessment, weed

detection, and monitoring TSSM damage to leaves. Another comprehensive

understanding is that using a sensor (ground, air, and space level) capable of obtaining

spectral data in the blue, green, red, red-edge, NIR and SWIR regions might be

practical for N content assessment, LAI assessment, weed detection, and monitoring

TSSM damage to leaves.

Each of the four potential applications can be further developed and refined. For the N

content assessment in potato plants, an analysis based on spatial dimension could be

added by ground-level imaging. The LAI assessment by the red-edge region could be

up-scaled to a bigger field of view or larger pixels as will be available by VENμS or

104

Sentinel-2. Since the REIP is not sensitive to soil cover, as long as the crop would be

uniform it is assumed that LAI can be assessed. In order to create a real-time system

for site-specific weed detection, there is a need to study further the weed detection

abilities by several recommended specific bands. In order to add the spatial dimension

for early detection of TSSM damage, ground-level images should be obtained in the

greenhouse. Although previous knowledge might be helpful, it is not common to

spectrally separate between different vegetation stresses (Pinter et al. 2003). The

concept of using the same data for several applications is neither common nor easy to

implement. Although the conclusions of this dissertation can be used to support such a

study, it is not concluded here. A study that will combine water and N content

assessments with weed detection is a worthy task for the future.

105

8. Reference for the introduction and summary (chapters 1 and 7).

Asner, G.P., 1998, Biophysical and biochemical sources of variability in canopy reflectance. Remote Sensing of Environment, 64, pp. 234-253.

Bannari, A., Morin, D., Bonn, F. and Huete, A.R., 1995, A review of vegetation indices. Remote Sensing Reviews, 13, pp. 95-120.

Ben-Dor, E., Goldlshleger, N., Benyamini, Y., Agasi, M. and Blumberg, D., 2003, The spectral reflectance properties of soil structural crusts in the 1.2 to 2.5 μm spectral region. Soil Science Society of America Journal, 67, pp. 289-299.

Ben-Ze’ev, E., Karnieli, A., Agam, N., Kaufman, Y. and Holben, B., 2006, Assessing vegetation condition in the presences of biomass burning smoke by applying the Aerosol-free Vegetation Index (AFRI) on MODIS images. International Journal of Remote Sensing, 27, pp. 3203-3221.

Biewer, S., Erasmi, S., Fricke, T. and Wachendorf, M., 2009, Prediction of yield and the contribution of legumes in legume-grass mixtures using field spectrometry. Precision Agriculture, 10, pp. 128-144.

Bonfil, D.J., Karnieli, A., Raz, M., Mufradi, I., Asido, S., Egozi, H., Hoffman, A. and Schmilovitch, Z., 2004, Decision support system for improving wheat grain quality in the Mediterranean area of Israel. Field Crops Research, 89, pp. 153-163.

Borregaard, T., Nielsen, H., Norgaard, L. and Have, H., 2000, Crop-weed discrimination by line imaging spectroscopy. Journal of Agricultural Engineering Research, 75, pp. 389-400.

Buschmann, C. and Nagel, E., 1993, In vivo spectroscopy and internal optics of leaves as basis for remote-sensing of vegetation. International Journal of Remote Sensing, 14, pp. 711-722.

Campbell, J., B, 1996, Introduction to remote sensing pp. (London: Taylor & Francis).

Canisius, F. and Fernandes, R., 2012, Evaluation of the information content of Medium Resolution Imaging Spectrometer (MERIS) data for regional leaf area index assessment. Remote Sensing of Environment, 119, pp. 301-314.

Ceccato, P., 2001, Estimation of vegetation water content using remote sensing for the assessment of fire risk occurrence and burning efficiency. Greenwich

Chang, C., I, 2003, Hyperspectral imaging techniques for spectral detection and classification, pp. (New York: Kluwer Academic \ Plenum Publishers).

Coyne, P.I., Aiken, R.M., Maas, S.J. and Lamm, F.R., 2009, Evaluating yieldtracker forecasts for maize in western Kansas. Agronomy Journal, 101, pp. 671-680.

106

Delegido, J., Verrelst, J., Alonso, L. and Moreno, J., 2011, Evaluation of Sentinel-2 Red-Edge Bands for Empirical Estimation of Green LAI and Chlorophyll Content. Sensors, 11, pp. 7063-7081.

El-Shikha, D.M., Waller, P., Hunsaker, D., Clarke, T. and Barnes, E., 2007, Ground-based remote sensing for assessing water and nitrogen status of broccoli. Agricultural Water Management, 92, pp. 183-193.

Fensholt, R. and Sandholt, I., 2003, Derivation of a shortwave infrared water stress index from MODIS near- and shortwave infrared data in semiarid environment. Remote Sensing of Environment, 87, pp. 111-121.

Fitzgerald, G.J., Pinter, P.J., Hunsaker, D.J. and Clarke, T.R., 2005, Multiple shadow fractions in spectral mixture analysis of a cotton canopy. Remote Sensing of Environment, 97, pp. 526-539.

Fitzgerald, G.J., Rodriguez, D., Christensen, L.K., Belford, R., Sadras, V.O. and Clarke, T.R., 2006, Spectral and thermal sensing for nitrogen and water status in rainfed and irrigated wheat environments. Precision Agriculture, 7, pp. 233-248.

Gausman, H., 1985, Plant leaf optical properties in visible and Near Infrared light, pp. 9-60 (Lubbock TX. USA: Texas Tech Press).

Gitelson, A.A., 2004, Wide dynamic range vegetation index for remote quantification of biophysical characteristics of vegetation. Journal of Plant Physiology, 161, pp. 165-173.

Haboudane, D., Miller, J.R., Tremblay, N., Zarco-Tejada, P.J. and Dextraze, L., 2002, Integrated narrow-band vegetation indices for prediction of crop chlorophyll content for application to precision agriculture. Remote Sensing of Environment, 81, pp. 416-426.

Hardinsky, M.A., Lemas, V. and Smart, R.M., 1983, The influence of soil salinity, growth form, and leaf moisture on the spectral reflectance of Spartina alternifolia canopies. Photogrammetric engineering and remote sensing, 49, pp. 77-83.

Herrmann, I., Karnieli, A., Bonfil, D., J., Cohen, Y. and Alchanatis, V., 2010, SWIR-based spectral indices for assessing nitrogen content in potato fields. International Journal of Remote Sensing, 31, pp. 5127-5143.

Herrmann, I., Pimstein, A., Karnieli, A., Cohen, Y., Alchanatis, V. and bonfil, D., J., 2011, LAI assessment of wheat and potato crops by VENμS and Sentinel-2 bands. Remote Sensing of Environment, 115, pp. 2141-2151.

Horler, D.N.H., Dockray, M. and Barber, J., 1983, The Red Edge of Plant Leaf Reflectance. International Journal of Remote Sensing, 4, pp. 273-288.

107

Karnieli, A., Kaufman, Y., Remer, L. and Wald, A., 2001, AFRI – aerosol free vegetation index. Remote Sensing of Environment, 77, pp. 10-21.

Moran, M., S., Inoue, Y. and Barnes, E., M., 1997, Opportunities and Limitations for Image-Based Remote Sensing in Precision Crop Management. Remote Sensing of Environment, 61, pp. 319-346.

Moran, M.S., Maas, S.J., Vanderbilt, V.C., Barnes, M., Miller, S.N. and Clarke, T.R., 2004, Application of image-based remote sensing to irrigated agriculture. In Remote sensing for natural resource management and environmental monitoring, S.L. Ustin (Ed.), pp. 617-676 (Hoboken NJ. USA: John Wiley & sons).

Mutanga, O. and Skidmore, A.K., 2007, Red edge shift and biochemical content in grass canopies. Isprs Journal of Photogrammetry and Remote Sensing, 62, pp. 34-42.

Pierce, F.J. and Nowak, P., 1999, Aspects of precision agriculture. Advances in Agronomy, Vol 67, 67, pp. 1-85.

Pimstein, A., Bonfil, D.J., Mufradi, I. and Karnieli, A., 2007, Spectral index for assessing heading timing of spring wheat grown under semi-arid conditions, In J.V. Stafford (Ed.), (Skiathos, Greece: 6th European Conference on Precision Agriculture, 663-669.

Pinter, P., J., Hatfield, J., L., Schepers, J., S., Barnes, E., M., Moran, M., S., Daughtry, C., S, T. and Upchurch, D., R., 2003, Remote sensing for crop management. Photogrametric Engineering & Remote Sensing, 69, pp. 647-664.

Plant, R.E., 2001, Site-specific management: the application of information technology to crop production. Computers and Electronics in Agriculture, 30, pp. 9-29.

Pu, R.L., Gong, P., Biging, G.S. and Larrieu, M.R., 2003, Extraction of red edge optical parameters from Hyperion data for estimation of forest leaf area index. IEEE Transactions on Geoscience and Remote Sensing, 41, pp. 916-921.

Raven, P., H, Everet, R., F and Eichhorn, S., E, 2005, Biology of plants, pp. 35-88 (New-York: W. H. Freeman and Company).

Rodriguez, D., Fitzgerald, G.J., Belford, R. and Christensen, L.K., 2006, Detection of nitrogen deficiency in wheat from spectral reflectance indices and basic crop eco-physiological concepts. Australian Journal of Agricultural Research, 57, pp. 781-789.

Smith, A.M. and Blackshaw, R.E., 2003, Weed-crop discrimination using remote sensing: A detached leaf experiment. Weed Technology, 17, pp. 811-820.

108

Tian, Q., Tong, Q., Pu, R., Gou, X. and Zhao, C., 2001, Spectroscopic determination of wheat water status using 1650-1850 nm spectral absorption features. International Journal of Remote Sensing, 22, pp. 2329-2338.

Tilling, A.K., O'Leary, G.J., Ferwerda, J., Jones, S.D., Fitzgerald, G.J. and Belford, R., 2006, Remote Sensing to Detect Nitrogen and Water Stress in Wheat, In 10-14 September 2006, (Perth, Western Australia: the 13th Australian Agronomy Conference,

Tilling, A.K., O'Leary, G.J., Ferwerda, J.G., Jones, S.D., Fitzgerald, G.J., Rodriguez, D. and Belford, R., 2007, Remote sensing of nitrogen and water stress in wheat. Field Crops Research, 104, pp. 77-85.

Verbyla, D., L. , 1995, Satellite remote sensing of natural resources, pp. (New-York: Lewis).

Vrindts, E., De Baerdemaeker, J. and Ramon, H., 2002, Weed detection using canopy reflection. Precision Agriculture, 3, pp. 63-80.

Younan, N.H., King, R.L. and Bennett, H.H., Jr., 2004, Classification of hyperspectral data: A comparative study. Precision Agriculture, 5, pp. 41-53.

Herrmann, I., Shapira, U., Kinast, S., Karnieli, A., and Bonfil, D., J., 2012, Ground-level hyperspectral imagery for

detecting weeds in wheat fields, submitted to Precision Agriculture.

זיהוי נזק אקריות קורים לעלי פלפל חממה

גורמות נזק משמעותי לגידולים ולייבולים בשדה כמו גם ) Tetranychus urticae Koch(אקריות קורים

נזק זה ניתן לגלות באמצעים . תוך כדי אכילה האקריות משמידות תאים המכילים כלורופלסטים. בחממה

מחקר זה מתמקד במידע ). NIR(ספקטראליים בעזרת החזרת האור בתחום הנראה והאינפרה אדום הקרוב

גרפי . שנרכש בעזרת כדור כיול, חממה) Capsicum annuum(עלי פלפל רוצים של ספקטראלי מרובה ע

המידע הספקטראלי הומר . זוהו םההחזרה הממוצעת של כל אחת מרמות הנזק הושוו והבדלים ספקטראליי

הערכת . מדדי צמחייה המאפשרים זיהוי מוקדם של נזק האקריות באמצעות הפרדת רמות נזק שונותשישה ל

התוצאות מראות כי .של צמדי דרגות נזק בוצעה עבור כל אחד ממדדי הצמחייה) ANOVA(כיוונית שונות חד

רק ה . מועטחמישה מתוך ששת המדדים יכולים להפריד באופן מובהק בין עלים ללא נזק לאלו שניזוקו באופן

REIP ניתן להסיק שזיהוי ,לכן. רמות הנזק להפריד באופן מובהק בין כל ארבע ושני מדדים נוספים יכלו

יכול לספק הפרדה REIP ה .מוקדם של נזק אקריות לעלי פלפל חממה ניתן לביצוע באמצעים מעוטי ערוצים

השיטות המוצעות יכולות לזהות , בנוסף לכך .בתנאי מעבדה, עקבית בין רמות נזק מתגברות בעלי פלפל חממה

. צידו התחתוןאת הנזק לעלה בצידו העליון למרות שהאקריות ניזונות מ

Herrmann, I., Berenstein, M., Sade, A., Karnieli, A., Bonfil, D.J. and Weintraub, P.G., 2012, Spectral monitoring of

two-spotted spider mite damage to pepper leaves. Remote Sensing Letters, 3, pp. 277-283.

מידע ספקטראלי בתחום ) 1: (ר מאפשרת להסיק מספר תובנותזווית מבט רחבה על עבודת המחק, לסיכום

מידע זה יכול להיות . האלקטרו מגנטי מסוגל לזהות ולנטר תופעות ותכונות שונות של גידולים חקלאיים

, לכן. כמו גם מבוסס מדדי צמחייה עבור שתי הרזולוציות, מולטי או סופר ספקטראלי, מרובה ערוצים ורציף

האווירית והחללית יכול להיות מיושם עבור , כש בעזרת מגוון חיישנים ברמה הקרקעיתמידע ספקטראלי שנר

אזור גבול האדום תוחם טווח ספקטראלי שהינו חשוב ביותר עבור ) 2. (חקלאות מדייקת וניהול מדויק מקום

. יות לעליםוניטור נזק אקר, איתור עשבים, מדד שטח עלה, ביניהם, מספר יישומים פוטנציאליים לגידולי שדה

י קהילת החישה מרחוק וערוצים ספקטראליים מתאימים מתוכננים "חשיבות איזור זה הוכרה לאחרונה ע

התברר כי מידע מולטי וסופר ספקטראלי מניב תוצאות שאינן ) Sentinel-2) .3ו VENμSהעתיד ללווייני

ם ליצירת יישומים פחותי עלות תובנה זו עשויה לתרו. נופלות באיכותן מזו של מידע רציף מרובה ערוצים

. שיכסו שטח גדול יותר

ערכיכלל . 99%והציגו מיון כולל הגבוה מ )GDA(מוינו בהצלחה בעזרת שיטת ניתוח האבחנה הכללית

דיוק עבור חיטה ±4.19% 95מסוימים הניבה םבאותה שיטה ועבור ערוצים צרי וההחזרה של העלווה מוינ

77הייתה VENμSההחזרה של העלווה עבור ערוצי לווין העתיד ערכיכלל . עבור חימצה ±5.13% 94ו

עבור חיטה ו ±7.97% 78עמדה על ALIעבור חימצה ועבור ערוצי לווין ±6.94% 88עבור חיטה ו ±8.09%

87דיוק המיון הכולל עמד על , במהלך התקופה הגורלית להדברת עשבים. עבור חימצה ±8.22% 82

ניתן , לפיכך .כיסוי צומח בשדה חיטה ולכן התגלה פוטנציאל ליישום קוטל עשבים 5%תר מ יו עבור ±5.57%

. להסיק שהמאפיינים הספקטראליים של רקמת עלה מאפשרים מיון מדויק לקטגוריות כמו גם מין בוטאני

שדה הפרדה ספקטראלית של גידולי . גבול האדום הינו בעל חשיבות רבה למיון גידולי שדה ועשבים רעים

כיסוי צומח בזמן 5%ועשבים רעים הינה שימושית באופן פוטנציאלי עבור שדות חיטה בהם יש מעל ל

, למודלים כמותיים המבוססים על תכונות ספקטראליות של חיטה .המשמעותי עבור טיפול בעשבים רעים

ש עבור ניהול עשבים עשבים דגניים ועשבים רחבי עלים יש איכות מיון ופוטנציאל חיזוי שעשוי לשמ, חימצה

.מדויק מיקום

Shapira, U., Herrmann, I.∗, Karnieli, A., and Bonfil, D., J., 2012, Field spectroscopy for weed detection in wheat and

chickpea fields. International Journal of Remote Sensing, in press.

הדמאות קרקעיות מרובות ערוצים עבור איתור עשבים בשדות חיטה

לשם , ברזולוציות גבוהות ספקטראלית ומרחבית, מרובות ערוצים, המחקר השתמש בהדמאות קרקעיות

הפיקסלים בהדמאה שימשו להצלבת אימות של . דגניים ורחבי עלים בשדות חיטה, חד שנתיים, איתור עשבים

המודל הטוב ביותר נבחר לאחר ). DA-PLS(ועים הפחותים והאבחנה הכללית הריבמודלי מיון בשיטת

מודל זה השתמש בארבע . של הצלבת האימות העירובשל טבלאות ) Cohen’s Kappa(השוואת ערכי הקאפה

ספקטראלי כל אחת מהקבוצות כללה מידע . וחיטה; קרקע; עשבים דגניים; עשבים רחבי עלים: קבוצות

שימשה לזיהוי האזורים הספקטראליים VIPשיטת ה . יחדמפיקסלים מוצלים ומידע יםמוארמפיקסלים

. נמצא כי אזור גבול האדום הוא החשוב ביותר עבור קבוצות הצמחייה. החשובים להגדרת כל אחת מהקבוצות

התוצאות מספקות . 72%פיקסלי אימות הקרקע נבחרו רנדומלית וסך הדיוק של טבלת הבלבול שלהם עמד על

ניתן . וחלקות ניסוי שונות, תאריכי רכישה שונים, יות שהמודל כולל חיטה ועשבים במגוון שלבי גידולה

בין בלבד יכולות להניב הפרדה המבוססת על ספקטרה, ספקטראלית ומרחבית, להסיק שרזולוציות גבוהות

ילוב של ארבע ש. ניתן למיין באיכות גבוהה יותר צמחייה מוארת מאשר צמחיה מוצלת .חיטה לעשבים

הניב את תוצאות המיון הטובות ביותר עבור מאגר ) קרקעו ,חיטה, עשבים דגניים, עשבים רחבי עלים(קבוצות

ועשבים , עשבים רחבי עלים, גבול האדום הינו התחום החשוב ביותר עבור הפרדה בין חיטה. הנתונים הקיים

ור מידע הטרוגני והמודל הפיק תוצאות הצלבת האימות של המודל והאימות הקרקעי נאספו עב. דגניים

היתכנות לאיסוף נתונים ספקטראליים ברמה אווירית או חללית כמו גם על פיתוח יישום ישנה, לכן. סבירות

. קרקעי

∗ Equal contribution

Herrmann, I., Karnieli, A., Bonfil, D., J., Cohen, Y., and Alchanatis, V., 2010, SWIR-based spectral indices for

assessing nitrogen content in potato fields. International Journal of Remote Sensing, 31, pp. 5127-5143.

אדמה - הערכת מדד שטח העלה בשדות חיטה ותפוח

המחקר הנוכחי מכוון אל . ניםמשתנה השולט בתהליכים בעלווה ואפשר לנטרו מלווייהמדד שטח העלה הינו

חקר הפוטנציאל והמגבלות של שימוש בערוצים ספקטראליים מתחום גבול האדום של לווייני העתיד

המחקר בוצע . לשם הערכת מדד שטח העלה בגידולי שדה, Sentinel-2ו VENμSהסופרספקטראליים

נאסף בעזרת ) ננומטר 400-1000(ף מידע ספקטראלי רצי. אדמה בצפון הנגב- בחלקות ניסוי של חיטה ותפוח

המידע הספקטראלי הומר לרזולוציות . ספקטרומטר שדה וערכי מדד שטח העלה נרכשו בעזרת ספטומטר

שני , ארבע עונות גידול(המידע חולק לשבעה מערכי מידע . Sentinel-2ו VENμSהספקטראליות של

בעזרת מודלים של שיטת הריבועים הפחותים חיזוי מדד שטח העלה ). ואחד הכולל את כל המידע, גידולים

)PLS (אדמה של רזולוציית - עבור חיטה ותפוח. עבור הספקטרה הרציפות כמו גם המומרות הושווה והוערך

, 0.92ו , 0.93, 0.93היו PLSשל ה ) r(ערכי מקדם הקורלציה , Sentinel-2ו VENμS , המידע הרציף

תחום גבול האדום כתחום נמצא ברוב המקרים) VIP(חשיבות המשתנה בהטלה הבשיטת . בהתאמה

מדד הצמחייה של ההפרש , כמו כן. הספקטראלי החשוב ביותר עבור שלוש הרזולוציות הספקטראליות

. חושבו עבור שלוש הרזולוציות הספקטראליות) REIP(ונקודת הפיתול של גבול האדום ) NDVI(המנורמל

יכולות החיזוי של המדדים . המדדים והוערכה יכולת חיזויו נבחן הקשר של מדד שטח העלה לשני, בנוסף

וערך מקדם הקורלציה , השיא היה עבור חיטה, המחושבים מתוך שלוש הרזולוציות הספקטראליות הושוו

בזמן שערכי מקדם .שהופק מתוך שלוש הרזולוציות הספקטקראליות REIPעבור ה 0.91הגבוה ביותר היה

עבור שלוש הרזולוציות הספקטראליות הפרשים אלו בערכי , 0.72או 0.73ו אדמה הי- הקורלציה של תפוח

מקדם הקורלציה הם מובהקים וניתן לשייכם לאחידות הגידולים המשפיעה על כמות הקרקע הנכללת בשדה

יכולים להעריך ספקטראלית מדד Sentinel-2ו VENμSניתן להסיק ש , לכן .הראיה במדידות הצמחייה

בעל יכולת חיזוי REIPעבור חיטה נמצא כי ה .ת שאינה נופלת מזו של חיישן מרובה ערוציםשטח עלה באיכו

טור מדד שטח עלה בעזרת חיישנים יולכן יכול פוטנציאלית לשמש לנ NDVIהטובה באופן מובהק מזו של ה

. ול האדוםבבעלי ארבעה ערוצים בג

Herrmann, I., Pimstein, A., Karnieli, A., Cohen, Y., Alchanatis, V. and Bonfil, D., J., 2011, LAI assessment of

wheat and potato crops by VENμS and Sentinel-2 bands. Remote Sensing of Environment, 115, pp. 2141-2151.

איתור עשבים רעים בשדות חיטה וחימצה בעזרת ספקטרומטר נקודתי

, בכל אופן. יישום קוטל עשבים באופן אחיד בכלל השדה החקלאי י"הדברת עשבים רעים מבוצעת לרוב ע

המטרה של מחקר זה . ליישום קוטל עשבים רק היכן שהוא דרוש עשויים להיות יתרונות כלכליים וסביבתיים

ערכים . הייתה להשתמש בחישה מרחוק לשם זיהוי עשבים דגניים ורחבי עלים בין גידולים דגניים ורחבי עלים

של החזרה יחסית ברמת העלה והעלווה נאספו בעזרת ספקטרומטר שדה עבור ארבע קטגוריות ספקטראליים

ההחזרה של רקמת עלה לפי מין בוטאני ערכיכלל .ועשבים רחבי עלים; עשבים דגניים; חימצה; חיטה: צומח

תקציר

אורכי גל . הסביבתי הינן מטרות עיקריות של החקלאים המודרניים הזיהוםהגדלת תפוקת הגידול והפחתת

קשר למצב הצמחייה כמו גם למיני גידולים יבתחומים מסוימים של הספקטרום האלקטרו מגנטי יכולים לה

הבדלים מובהקים החשובים מספיק , בתחומי שדה חקלאי, היכולת לזהות). רחב עלים או דגן(או לקטגוריות

, ל"יכולה לענות על המטרות הנ, )מוכר גם כחקלאות מדייקת(הול גידולים מדויק מיקום כלל בנייכדי לה

. רווח כלכלי ויתרונות סביבתייםלספק כלומר

עבודת המחקר הנוכחית מכוונת אל שימוש במידע ספקטראלי ושיטות המתאימות לו לשם חקר ארבעה

הערכת תכולת חנקן ) 1: (היישומים הינם. יישומים אפשריים בסביבת חקלאות מדייקת עבור גידולי שדה

איתור עשבים רעים בשדות ) 3( .אדמה-הערכת מדד שטח עלה בשדות חיטה ותפוח) 2( .אדמה-בצמחי תפוח

ארבעת יישומים אלו מוצגים בחמישה . זיהוי נזק של אקריות קורים לעלי פלפל חממה) 4( .חיטה וחימצה

מדדי : פערי המידע העיקריים שנבחנו הם. י שני פרסומים"פרסומים כאשר נושא איתור העשבים מוצג ע

יכולת פוטנציאלית של ; צמחייה הקשורים באופן עקיף לתכולת החנקן בצמח הינם המעריכים החשובים שלו

יכולת בלתי מספקת להפרדה ספקטראלית בין גידולי שדה לעשבים ; הערכת מדד שטח עלה בעזרת לוויני עתיד

הערכה ספקטראלית של נזק אקריות ; ק הזמן המשמעותי לטיפול בעשבים רעיםברמת הקרקע ובפר, רעים

. קורים לעלים

אדמה -הערכת תכולת חנקן בצמחי תפוח

לכמות . לכן לדישון בחנקן חשיבות רבה בגידולים מעובדים, חנקן הוא גורם חשוב להתפתחות הצמח ותפוקתו

בתיות וכתוצאה מכך הוא נושא חשוב בתחום תזמון של יישום דשן חנקן יש השלכות כלכליות וסבילו

וחלליים , אוויריים, מדדים ספקטראליים המופקים בעזרת ספקטרומטרים קרקעיים. החקלאות המדייקת

, בעיקר תכולת כלורופיל, רוב המדדים הללו מבוססים על סמנים עקיפים. משמשים להערכת תכולת החנקן

המחקר הנוכחי מכוון אל חקר הביצועים של מדדי חנקן . קןאשר הוכח שהיא קשורה פיזיולוגית לתכולת החנ

ובמיוחד , )ננומטר SWIR ,1200-2500(ספקטראליים חדשים התלויים בתחום האינפרה אדום קצר הגל

מדדי חנקן ידועים וארבעת המדדים . הוא קשור ישירות לתכולת החנקןשננומטר היות 1510באורך הגל

בחנו בעזרת מידע ספקטראלי ברמת העלווה שנרכש במהלך שתי עונות מבוססי האינפרה אדום קצר הגל נ

, דגימות של כלל הביומאסה מעל הקרקע נאספו. אדמה בצפון מערב הנגב- גידול בחלקות ניסוי של תפוחי

ביצועי כל המדדים . כדי לספק נתוני תכולת חנקן ממקומה המקורי, ממקום איסוף הנתונים הספקטראליים

קורלציה בין מדדים ידועים וחדשים לבין תכולת החנקן כמו גם קורלציה בין ) 1: (ש שיטותהוערכו בעזרת שלו

רגישותם היחסית של ) 3( .של תכולת החנקן) RMSEP(השורש הריבועי של שגיאת החיזוי ) 2( .המדדים עצמם

ם קצר התוצאות חשפו יתרון מוצק למדדים החדשים מבוססי תחום האינפרה אדו. המדדים לתכולת החנקן

1510המדד הטוב ביותר הוא כזה המשלב מידע מאורכי הגל . הגל ביכולתם לחזות וברגישותם לתכולת החנקן

היות ומדדי . ננומטר אך לא נמצא הבדל מובהק בין ארבעת המדדים מבוססי האינפרה אדום קצר הגל 660ו

חנקן בצמח כמו גם להשלכות צמחייה אלו משלבים קשר ישיר ועקיף לתכולת חנקן הם קשורים לנוכחות של

מאפשרים חיזוי משופר ורגישים יותר לתכולת SWIRלכן ניתן להסיק שהמדדים מבוססי תחום ה . שלה

. החנקן

:המחקר בוצע בהדרכת

ארנון קרניאלי' פרופ

המעבדה לחישה מרחוק ש יעקב בלאושטיין"המכונים לחקר המדבר ע

אוניברסיטת בן גוריון בנגב

בונפיל. ר דוד י"ד

חקלאות חרבה ושטחים פתוחים

המחלקה למדעי הצמח מרכז מחקר גילת, מינהל המחקר החקלאי

חקלאות מדייקת עבור גידולי שדה באזורים צחיחיםיישומים מרובי ערוצים של

מחקר לשם מילוי חלקי של הדרישות"דוקטור לפילוסופיה"לקבלת תואר

מאת

איתי הרמן

הוגש לסנאט אוניברסיטת בן גוריון בנגב

:אישור המנחים

__________ארנון קרניאלי ' פרופ

____________בונפיל . ר דוד י"ד

: ש קרייטמן"אישור דיקן בית הספר ללימודים מתקדמים ע

___________מיכל שפירא ' פרופ

2012, ב מאי"התשע, סיון

שבע- באר

יישומים מרובי ערוצים של חקלאות מדייקת עבור גידולי שדה באזורים צחיחים

מחקר לשם מילוי חלקי של הדרישות"לפילוסופיהדוקטור "לקבלת תואר

מאת

איתי הרמן

הוגש לסנאט אוניברסיטת בן גוריון בנגב

2012, ב מאי"התשע, סיון שבע- באר