evaluating unsupervised and supervised image classification methods for mapping cotton root rot

Evaluating unsupervised and supervised imageclassification methods for mapping cotton root rot

Chenghai Yang • Gary N. Odvody • Carlos J. Fernandez •

Juan A. Landivar • Richard R. Minzenmayer • Robert L. Nichols

� Springer Science+Business Media New York 2014

Abstract Cotton root rot, caused by the soilborne fungus Phymatotrichopsis omnivora, is

one of the most destructive plant diseases occurring throughout the southwestern United

States. This disease has plagued the cotton industry for over a century, but effective

practices for its control are still lacking. Recent research has shown that a commercial

fungicide, flutriafol, has potential for the control of cotton root rot. To effectively and

economically control this disease, it is necessary to identify infected areas within fields so

that site-specific technology can be used to apply fungicide only to the infected areas. The

objectives of this study were to evaluate unsupervised classification applied to multi-

spectral imagery, unsupervised classification applied to the normalized difference vege-

tation index (NDVI)and six supervised classification techniques, including minimum

distance, Mahalanobis distance, maximum likelihood and spectral angle mapper (SAM),

neural net and support vector machine (SVM),for mapping cotton root rot from airborne

multispectral imagery. Two cotton fields with a history of root rot infection in Texas, USA

were selected for this study. Airborne imagery with blue, green, red and near-infrared

bands was taken from the fields shortly before harvest when infected areas were fully

expressed in 2011. The four-band images were classified into infected and non-infected

zones using the eight classification methods. Classification agreement index values for

infected area estimation between any two methods ranged from 0.90 to 1.00 for both fields,

C. Yang (&)USDA-ARS, Aerial Application Technology Research Unit, 3103 F and B Road, College Station,TX 77845, USAe-mail: [email protected]

G. N. Odvody � C. J. Fernandez � J. A. LandivarTexas AgriLife Research and Extension Center, 10345 State Highway 44, Corpus Christi, TX 78406,USA

R. R. MinzenmayerTexas AgriLife Extension Service, 613 Hutchins Avenue, Suite 302, Ballinger, TX 76821, USA

R. L. NicholsCotton Incorporated, 6399 Weston Parkway, Cary, NC 27513, USA

123

Precision AgricDOI 10.1007/s11119-014-9370-9

indicating a high degree of agreement among the eight methods. Accuracy assessment

showed that all eight methods accurately identified root rot-infected areas with overall

accuracy values from 94.0 to 96.5 % for Field 1 and 93.0 to 95.0 % for Field 2. All eight

methods appear to be equally effective and accurate for detection of cotton root rot for site-

specific management of this disease, though the NDVI-based classification, minimum

distance and SAM can be easily implemented without the need for complex image pro-

cessing capability. These methods can be used by cotton producers and crop consultants to

develop prescription maps for effective and economical control of cotton root rot.

Keywords Cotton root rot � Airborne multispectral imagery � Image classification �Vegetation index � Agreement index

Introduction

Cotton (Gossypiumhirsutum)is an economically important crop that is highly susceptible to

cotton root rot, a destructive plant disease that occurs throughout the southwestern United

States. Infected plants wilt and quickly die with the leaves attached to the plants (Smith

et al. 1962). The symptoms usually begin during the period of rapid vegetative growth, are

more visible during flowering and fruit development, and continue to increase through the

growing season. Plants infected earlier in the growing season will die before bearing fruit,

whereas infection occurring at later plant growth stages will reduce cotton yield and lower

lint quality (Ezekiel and Taubenhaus 1934; Yang et al. 2005).

Cotton root rot has plagued the cotton industry for more than 100 years. Despite dec-

ades of research efforts, effective practices for the control of this disease were still lacking.

Recently, several fungicides have been evaluated and a commercial formulation of flu-

triafol (Top guard� - Cheminova, Inc., Wayne, New Jersey, USA) was found to effectively

control cotton root rot (Isakeit et al. 2010, 2012). Consequently, temporary authorization

has been granted to cotton growers in Texas, USA to use flutriafolfor the control of this

disease since 2012. Flutriafol needs to be applied every year in order to suppress the

disease. Once plants are infected during the season, application of flutriafol will be too late

to help them recover. Therefore, flutriafol is recommended to be applied at planting. To

more economically control this disease, it is necessary to identify infected areas within the

field so that site-specific technology can be used to apply the fungicide only to infected

areas.Cotton root rot tends to occur in similar areas within fields over recurring year-

s.Infected areas identified near the end of the growing season when the fungus is fully

expressed can be used for site-specific treatment in subsequent years.

Remote sensing has been successfully used to map cotton root rot infections in cotton

fields (Nixon et al. 1987; Yang et al. 2005, 2010). Preliminary work has been conducted to

monitor the progression of the disease within a growing season or across different growing

seasons (Yang et al. 2012). In previous studies, ISODATA (Iterative Self-Organizing Data

Analysis) unsupervised classification applied to multispectral imagery has been used to

identify root rot-infected areas. With this method, the optimal number of spectral classes

was determined based on the average transformed divergence for each classification map

and the spectral classes were then grouped into root rot-infected and non-infected zones

(Yang et al. 2014). Although this method is effective, many other spectral measures and

classification techniques are available and may offer simpler or more accurate alternatives

Precision Agric

123

for this application. Therefore, the objectives of this study were to 1) evaluate unsupervised

classification applied to multispectral imagery, unsupervised classification applied to the

normalized difference vegetation index (NDVI), and six supervised classification tech-

niques (minimum distance, Mahalanobis distance, maximum likelihood, spectral angle

mapper (SAM), neural network and support vector machine (SVM)); and 2) compare the

performance among the eight classification methods for mapping cotton root rot from

airborne multispectral imagery.

Methods

Study sites

A circular cotton field (48.5 ha) with latitude and longitude co-ordinates (28�000500N,

97�3803300W) near Edroy, Texas, USA, designated as Field 1,and a semi-circular cotton

field (10.5 ha) with co-ordinates (31�2604200N, 100�1604900W) near San Angelo, Texas,

USA, designated as Field 2, were selected for this study. The fields were both irrigated by

center-pivots and had a history of cotton root rot. Cotton and grain sorghum (Sorghum

bicolor) have been cropped alternately and cotton was planted to Field 1 in early March

and to Field 2 in early May in 2011. The two fields were located in two different cotton

growing regions approximately 460 km apart.

Airborne multispectral image acquisition

Anairborne four-camera imaging system described by Yang (2012) was used to acquire

multispectral imagery. The system consisted of four high resolution charge-coupled device

(CCD) digital cameras and a ruggedized PC equipped with a frame grabber and image

acquisition software. The cameras were sensitive in the 400 to 1 000 nm spectral range and

provided 2 048 9 2 048 active pixels with 12-bit data depth. The four cameras were

equipped with blue (430–470 nm), green (530–570 nm), red (630–670 nm), and near-

infrared (NIR, 810–850 nm) bandpass interference filters, respectively. The image

acquisition software allowed the synchronized individual band images, the normal visible

color composite, or the color-infrared (CIR) composite to be viewed on the computer

monitor and then saved as a four-band Tiff image. A Cessna 206 single-engine aircraft

equipped with the imaging system was used to acquire imagery at 3 050 m above the

ground level between 1 130 h and 1 430 h local time under sunny conditions. The images

were taken for Field 1 on 13 July 2011 and for Field 2 on 4 October 2011 about 1 week

before defoliation. At the imaging time, the crop was at the maturity stage and root rot was

fully expressed in both fields for the season. The ground pixel size achieved for the images

was approximately 0.9 m.

Image alignment and rectification

An image-to-image registration procedure based on the first-order polynomial transfor-

mation model was used to align the four individual band images in the composite image.

The registered images were then georeferenced or rectified to the Universal Transverse

Mercator (UTM), World Geodetic Survey 1984 (WGS-84), Zone 14, co-ordinate system

based on a set of ground control points around the fields located with a Trimble GPS

Precision Agric

123

Pathfinder ProXRT receiver (Trimble Navigation Limited, Sunnyvale, California, USA).

The root mean square (RMS) errors for rectifying the images using first-order transfor-

mation were approximately 2 m. All images were resampled to 1 m resolution using the

nearest neighborhood technique. All procedures for image registration and rectification

were performed using ERDAS Imagine (Intergraph Corporation, Madison, Alabama,

USA).

Unsupervised classification of multispectral images

The rectified four-band multispectral images were classified into 2–20 spectral classes

using ISODATA unsupervised classification (Intergraph Corporation, 2013). To evaluate

the separability of the spectral classes for each classification map, the average transformed

divergence was calculated among all possible pairs of classes for each classification based

on the four bands. The transformed divergence value ranges from 0 to 2 000. If the

calculated divergence is equal to 2 000, then the signatures can be said to be completely

separable in the bands being studied. A calculated divergence of zero means that the

signatures are inseparable. Generally, the classes are well separated if the calculated

divergence is greater than a threshold value of 1 900 (Jensen, 1996). The spectral classes in

each classification map were then grouped into root rot-infected and non-infected zones by

comparing with the original image and based on ground observations. The root rot-infected

areas (zone 1) and non-infected areas (zone 2) were estimated from the best two-zone

classification maps.

Unsupervised classification of NDVI images

NDVI was calculated from the NIR and red bands in the multispectral images as follows

(Rouse et al. 1973):

NDVI =NIR � RED

NIR + REDð1Þ

The NDVI images were then classified into two spectral classes using unsupervised

classification. This method ensures that each pixel is assigned to the class that has the

minimum NDVI difference to the pixel. Thus pixels with lower NDVI values correspond to

the root rot-infected zone, while pixels with higher NDVI values belong to the non-infected

zone.

Supervised classification

Ground scouting within the two fields during the growing season confirmed that cotton root

rot was the dominant stressor affecting the crop, even though other biotic and abiotic

stressors may have been present. Since the fungus causes a devastating effect on cotton

plants, it has a unique signature as seen on the airborne images compared with other

stressors such as nutrient deficiencies and insect damage that appear minor in contrast to

the effects of root rot. Therefore, the images for these fields could be simply classified into

root rot-infected areas and non-infected areas.

Six supervised classification methods, including minimum distance, Mahalanobis dis-

tance, maximum likelihood, SAM, neural net and SVM,were applied to the four-band

multispectral image. The minimum distance classifier uses the class means derived from

Precision Agric

123

the training data and assigns each pixel to the class that has the closest Euclidean distance

from the pixel (Campbell 2002). The Mahalanobis distance method is similar to minimum

distance, except that the covariance matrix is used in the calculation (Intergraph Corpo-

ration 2013). Each pixel is assigned to the class for which Mahalanobis distance is the

smallest. Maximum likelihood classification assumes that the data for each class in each

band are normally distributed and calculates the probability that a given pixel belongs to a

specific class (Richards 1999). Each pixel is assigned to the class that has the highest

probability (i.e., the maximum likelihood). Spectral angle mapper or SAM is a spectral

classification technique that uses the n-dimensional angle to match pixels to end members

(Kruse et al. 1993). The algorithm determines the spectral similarity between a pixel

spectrum and an end member spectrum by calculating the angle between them, treating

them as vectors in a space with dimensionality equal to the number of bands. Each pixel is

assigned to the end member whose spectrum has the smallest spectral angle with the pixel

spectrum.The neural net technique uses standard backpropagation for supervised learning

(Richards 1999). Learning occurs by adjusting the weights in the node to minimize the

difference between the output node activation and the output. The SVM classifier is a

kernel-based machine learning technique that separates the classes with a decision surface

that maximizes the margin between the classes (Hsu et al. 2007).

For supervised training, a number of infected and non-infected areas were identified and

digitized on the multispectral images as the training samples to represent respective

classes. The numbers of digitized training pixels for Field 1 were 1 927 (0.40 % of the total

area) for the infected class and 2 265 (0.48 %) for the non-infected class, while those for

Field 2 were 709 (0.68 %) for the infected class and 1 045 (1.0 %) for the non-infected

class. A circular boundary encompassing 484 892 pixels for Field 1 and a semi-circular

boundary encompassing 104 632 pixels for Field 2 were defined to exclude the areas

outside the boundaries for image classification. Each classifier resulted in a two-class

classification map. ENVI (Research Systems, Inc., Boulder, Colorado, USA) was used for

supervised classification.

Agreement indices

Two agreement indices were proposed to evaluate the agreement between any two clas-

sifiers. One agreement index is defined as

aij ¼Aij

Aj

ð2Þ

where aij is the agreement index of classifier i with respect to classifier j for the estimation

of infected area; Aij is the common infected area between classifiers i and j; Aj is the total

infected area based on classifier j. The other agreement index is defined as.

tij ¼Tij

Tð3Þ

where tij is the agreement index of classifier i with respect to classifier j for the estimation

of both infected and non-infected areas; Tij is the total common infected and non-infected

area between classifiers i and j; T is the total area within the defined field boundary. It

should be noted that the index aij is different from the index aji, while the index tij is the

same as the index tji.

Precision Agric

123

Accuracy assessment

For accuracy assessment of the eight classification maps, 200 points for Field 1 and 100

points for Field 2 were generated and assigned to the two classes in a stratified random

pattern using ERDAS Imagine. The UTMco-ordinates of these points were loaded to the

Trimble GPS receiver, which was then used to navigate to these points for ground veri-

fication. For those points that were located in transitional areas or surrounded by a mixture

of healthy and infected plants, extra care was made to confirm their infection status. Error

matrices for each classification map were generated by comparing the classified classes

with the actual classes based on ground verification. Classification accuracy measures,

including overall accuracy, kappa coefficient, producer’s accuracy and user’s accuracy,

were calculated based on the error matrices (Congalton and Green 1999).

Results and discussion

Figures 1 and 2 show the normal visible color and CIR composite images for the two

fields. On the normal visible color images, non-infected plants had a green color, whereas

infected plants had a brownish or grayish tone. On the CIR images, non-infected plants

showed a reddish-magenta tone, while infected plants had a cyanishor greenish color. Root

rot-infected areas could be easily separated from the non-infected areas on both types of

images, especially on the CIR images. Cotton root rot progressed across the fields and

continued to develop along the north edge of the field toward the end of the growing season

in Field 1. Generally, plants with various infection stages exist during the season, even

though the progression slows down toward the end of the season. Once infected, plants will

die within several days. The multispectral imagery will be able to detect infected plants

with certain levels of damage, while newly infected plants without obvious symptoms may

not be detected.In practice, buffer zones around the image-derived infection areas can be

added to account for the omission and the possible further expansion of the fungus.

Figure 3 shows the NDVI images derived from the multispectral images for the fields.

Infected areas had lower NDVI values and exhibited a dark grayish color. Non-infected

areas had higher values and showed a light gray tone. The dark color on the images

represents areas with dead dry plants, whereas the slightly dark color in the north and

northeast portion of Field 1 indicates infected plants that were not completely dry.The

NDVI images reveal similar patterns as shown on the color and CIR images.

Table 1 gives the average transformed divergence values among all possible pairs of the

spectral classes when each image was classified into 2- to 20-class unsupervised classifi-

cation maps for the two fields. The transformed divergence values ranged from 1 677 to

1 910 for Field 1 and from 1 643 to 1 908 for Field 2. Divergence values generally

increased with the number of classes, but there were a few exceptions. When the trans-

formed divergence value reaches 1900, the classification is considered excellent.

According to this guideline, the optimal number of classes was 17 for Field 1 and 19 for

Field 2. These classification maps provided the optimal separation among the classes.

Some of the classes belonged to infected areas with different levels of severity, whereas the

others were non-infected areas with different levels of plant vigor. For flutriafol applica-

tion, it is necessary to treat only infected areas at a uniform rate. Therefore, the two-zone

classification maps were formed by grouping the optimal number of classes for each field

into the root rot-infected zone and the non-infected zone.

Precision Agric

123

Figures 4 and 5 show the two-zone classification maps based on the eight classification

methods for Fields 1 and 2, respectively. A visual comparison of the classification maps

and their respective color and CIR images indicated that all the classification maps

effectively identified apparent root rot areas within each field and that there were minimal

differences between them.

Table 2 gives the estimates of infected and non-infected areas in pixels, hectares and

percentage based on the eight methods for Field 1. Infected area estimates ranged from

40.5 % with the maximum likelihood classifier to 44.8 % with SAM for the field. Table 3

presents the agreement values (aij) for the infected areas between any two classification

methods based on a pixel-to-pixel comparison. For example, the unsupervised classifica-

tion method detected 213 757 pixels of infected areas and the minimum distance method

identified 214 958 pixels of infected areas. The total number of common pixels identified

by both methods was 202 723 (not shown in the table). Thus the agreement of the

Fig. 1 Airborne normal visible color and color-infrared (CIR) images acquired from a 48.5-ha cotton rootrot-infected cotton field (Field 1) near Edroy, Texas, USA in 2011

Fig. 2 Airborne normal visible color and color-infrared (CIR) images acquired from a 10.5-ha cotton rootrot-infected cotton field (Field 2) near San Angelo, Texas, USA in 2011

Precision Agric

123

minimum distance method with respect to the unsupervised classification method was

202 723/213 757 = 0.95, while the agreement of the unsupervised classification method

with respect to the minimum distance methodwas 202 723/214 958 = 0.94 (see Table 3).

Fig. 3 Normalized difference vegetation index (NDVI) images derived from multispectral images forFields 1 and 2

Table 1 Average transformeddivergence among all possiblepairs of the spectral classes (sig-natures) when each image wasclassified into 2 to 20 classesusing unsupervised classificationfor Fields 1 and 2

Divergence values range from 0to 2 000 and a value of 1 900 orhigher indicates that the classesare well separated

No. ofspectralclasses

Field 1 Field 2

2 1 835 1 759

3 1 752 1 708

4 1 677 1 643

5 1 751 1 706

6 1 740 1 706

7 1 815 1 753

8 1 781 1 789

9 1 834 1 799

10 1 854 1 822

11 1 856 1 840

12 1 868 1 847

13 1 872 1 859

14 1 879 1 868

15 1 889 1 876

16 1 893 1 886

17 1 904 1 893

18 1 906 1 897

19 1 908 1 902

20 1 910 1 908

Precision Agric

123

The agreement values for infected area estimation between any two methods ranged from

0.89 to 1.00, indicating a high degree of agreement among the eight classification methods.

Table 4 presents the agreement values (tij) for infected and non-infected area estimation

between any two classification methods based on a pixel-to-pixel comparison.For example,

the total of the common infected and non-infected areas between the unsupervised clas-

sification method and the minimum distance method was 461 623 pixels and the total

number of pixels within the field boundary was 484 892. Thus the agreement between the

two methods was 461 623/484 892 = 0.94 (see Table 4). The agreement values for

Fig. 4 Two-zone classification maps based on eight classification methods from a multispectral image forField 1. UC unsupervised classification, NDVI normalized difference vegetation index, MD minimumdistance, MAHD Mahalanobis distance, ML maximum likelihood, SAM spectral angle mapper, NN neuralnet and SVM support vector machine

Fig. 5 Two-zone classification maps based on eight classification methods from a multispectral image forField 2. UC unsupervised classification, NDVI normalized difference vegetation index, MD minimumdistance, MAHD Mahalanobis distance, ML maximum likelihood, SAM spectral angle mapper, NN neuralnet and SVM support vector machine

Precision Agric

123

infected and non-infected area estimation between any two classification methods ranged

from 0.94 to 0.99. In comparison, the index tij tends to have less variability than the index

aij. The index aij is sensitive to the common infected area and the infected area estimated

from either of the two methods, while the index tij is an indicator of overall classification

agreement and only sensitive to the total common infected and non-infected area between

the two methods.

Table 5 summarizes the accuracy assessment results for the eight classification maps.

Overall accuracy ranged from 94.0 % for the unsupervised classification to 96.5 % for the

minimum distance classifier, indicating that 94.0 to 96.5 % of the image pixels were

correctly identified in the classification maps based on the 200 ground points. These results

indicate that all eight methods were accurate for identifying cotton root rot. Producer’s and

user’s accuracy values ranged from 91.9 to 98.3 % for the infected and non-infected

classes among the eight methods. Producer’s accuracy, a measure of omission error,

indicates the probability of actual areas being correctly classified on the map, while user’s

Table 2 Estimates of infected versus non-infected areas in pixels, hectares and percentage based on eightclassification methods from a multispectral image for Field 1

Methoda Infected Non-infected

Pixels ha (%) Pixels ha (%)

UC 213 757 21.4 44.1 271 135 27.1 55.9

NDVI 204 239 20.4 42.1 280 653 28.1 57.9

MD 214 958 21.5 44.3 269 934 27.0 55.7

MAHD 204 087 20.4 42.1 280 805 28.1 57.9

ML 196 480 19.6 40.5 288 412 28.8 59.5

SAM 217 250 21.7 44.8 267 642 26.8 55.2

NN 197 158 19.7 40.7 287 734 28.8 59.3

SVM 213 360 21.3 44.0 271 532 27.2 56.0

a UC unsupervised classification, NDVI normalized difference vegetation index, MD minimum distance,MAHD Mahalanobis distance, ML maximum likelihood, SAM spectral angle mapper, NN neural net, SVMsupport vector machine

Table 3 Agreement (aij) between any pair of eight classification methods for identifying infected areasfrom a multispectral image for Field 1

Methoda UC NDVI MD MAHD ML SAM NN SVM

UC 1.00 0.92 0.95 0.92 0.89 0.95 0.91 0.95

NDVI 0.97 1.00 1.00 0.96 0.95 1.00 0.95 0.98

MD 0.94 0.95 1.00 0.94 0.91 0.99 0.92 0.97

MAHD 0.97 0.96 0.99 1.00 0.96 0.98 0.95 0.99

ML 0.97 0.98 1.00 0.99 1.00 1.00 0.97 0.99

SAM 0.93 0.94 0.98 0.92 0.90 1.00 0.90 0.95

NN 0.98 0.98 1.00 0.99 0.97 1.00 1.00 1.00

SVM 0.95 0.94 0.98 0.94 0.91 0.97 0.92 1.00


Precision Agric

123

accuracy, a measure of commission error, indicates the probability that a category clas-

sified on the map actually represents that category on the ground. Based on the minimum

distance method, the producer’s accuracy for the infected class was 97.7 %, while the

user’s accuracy for this class was 94.4 %. In other words, 97.7 % of the root rot areas on

the ground were correctly identified as root rot on the classification map, but only 94.4 %

of the areas called root rot on the classification map were actually root rot on the ground.

Table 6 gives the estimates of infected and non-infected areas in pixels, hectares and

percentage based on the eight methods for Field 2. Infected area estimates ranged from

38.2 % with the maximum likelihood classifier to 43.4 % with the minimum distance

classifier for the field. The infected areas for both fields in this study happened to be

similar. However, percentage infected areas within fields can be from 0 to over 70 % based

on our observations in the two study regions in Texas.

Tables 7 and 8 present the two agreement indices, aij and tij, respectively, for all pairs of

the eight classification methods for Field 2. The agreement values for infected area esti-

mation between any two methods ranged from 0.90 to 1.00, while those for infected and

Table 4 Agreement (tij) between any pair of eight classification methods for identifying infected and non-infected areas from a multispectral image for Field 1

Methoda NDVI MD MAHD ML SAM NN SVM

UC 0.95 0.95 0.95 0.94 0.95 0.95 0.96

NDVI 0.98 0.97 0.97 0.97 0.97 0.96

MD 0.97 0.96 0.99 0.96 0.98

MAHD 0.98 0.96 0.98 0.97

ML 0.95 0.97 0.96

SAM 0.95 0.97

NN 0.97


Table 5 Accuracy assessment results for eight classification maps generated from a multispectral image forField 1

Methoda Overallaccuracy(%)

Overall kappa Infected Non-infected

PA UA PA UA

UC 94.0 0.878 94.2 92.1 93.9 95.5

NDVI 94.5 0.887 91.9 95.2 96.5 94.0

MD 96.5 0.929 97.7 94.4 95.6 98.2

MAHD 96.0 0.919 96.5 94.3 95.6 97.3

ML 95.5 0.908 91.9 97.5 98.3 94.1

SAM 96.0 0.919 97.7 93.3 94.7 98.2

NN 95.5 0.908 91.9 97.5 98.3 94.1

SVM 95.0 0.899 96.5 92.2 93.9 97.3


Precision Agric

123

non-infected area estimation between any two methods ranged from 0.95 to 1.00. These

results were similar to those for Field 1, confirming a high degree of agreement among the

eight classification methods.

Table 9 summarizes the accuracy assessment results for the eight classification maps for

Field 2. Overall accuracy ranged from 93.0 to 95.0 % based on the 100 ground points.

Producer’s and user’s accuracy values ranged from 90.2 to 96.6 % for the infected and

non-infected classes among the eight methods. These results indicate that all eight methods

were accurate for identifying cotton root rot. These results are in agreement with those

from Field 1 and our previous results (Yang et al. 2005; 2010).

Although the eight classification methods provided similar classification results for this

application, each method has its advantages and disadvantages and may not necessarily

provide similar results in other applications. The unsupervised classification directly

applied to multispectral imagery doesn’t need the user to identify any training samples for

Table 6 Estimates of infected versus non-infected areas in pixels, hectares and percentage based on eightclassification methods from a multispectral image for Field 2

Methoda Infected Non-infected

Pixels ha (%) Pixels ha (%)

UC 43 879 4.4 41.9 60 753 6.1 58.1

NDVI 42 252 4.2 40.4 62 380 6.2 59.6

MD 45 394 4.5 43.4 59 238 5.9 56.6

MAHD 42 855 4.3 41.0 61 777 6.2 59.0

ML 39 997 4.0 38.2 64 635 6.5 61.8

SAM 43 120 4.3 41.2 61 512 6.2 58.8

NN 42 050 4.2 40.2 62 582 6.3 59.8

SVM 42 239 4.2 40.4 62 393 6.2 59.6


Table 7 Agreement (aij) between each pair of eight classification methods for identifying infected areasfrom a multispectral image for Field 2

Methoda UC NDVI MD MAHD ML SAM NN SVM

UC 1.00 0.93 0.98 0.94 0.90 0.94 0.93 0.93

NDVI 0.97 1.00 1.00 0.99 0.94 1.00 0.98 0.99

MD 0.94 0.93 1.00 0.94 0.88 0.94 0.93 0.93

MAHD 0.96 0.97 0.99 1.00 0.93 0.98 0.97 0.97

ML 0.99 0.99 1.00 1.00 1.00 0.99 1.00 0.99

SAM 0.96 0.98 0.99 0.97 0.92 1.00 0.97 0.98

NN 0.97 0.98 1.00 0.98 0.95 0.99 1.00 0.98

SVM 0.97 0.99 1.00 0.99 0.94 1.00 0.98 1.00


Precision Agric

123

classification, but about 20 classification maps may need to be generated before an optimal

classification map can be identified. The NDVI combined with unsupervised classification

requires the calculation of the NDVI image and a two-class unsupervised classification. No

training samples are needed for this method either. In contrast, the other six supervised

methods all require training samples for each class before classification. Nevertheless, once

the training samples are identified, they can be used for each of the six unsupervised

classifiers. Due to the unique spectral characteristics of infected and non-infected plants, all

the eight methods provided similar results. However, if there exist other stressors that can

cause plant wilting or death as root rot does, some methods may perform better for

distinguishing cotton root rot from other stressors. It is preferable that all the methods be

used if the imaging processing software has the capability. This will allow multiple

classification maps to be compared so that the optimal classification map can be identified.

If the user only has limited image processing capability, it appears that the NDVI-based

classification, minimum distance and SAM can be easily implemented without the need for

Table 8 Agreement (tij) between any pair of eight classification methods for identifying for both infectedand non-infected areas from a multispectral image for Field 2

Methoda NDVI MD MAHD ML SAM NN SVM

UC 0.96 0.96 0.96 0.95 0.96 0.96 0.99

NDVI 0.97 0.98 0.97 0.99 0.98 1.00

MD 0.97 0.95 0.97 0.97 0.99

MAHD 0.97 0.98 0.98 1.00

ML 0.97 0.98 0.99

SAM 0.98 1.00

NN 1.00


Table 9 Accuracy assessment results for eight classification maps generated from a multispectral image forField 2

Methoda Overall accuracy (%) Overall kappa Infected Non-infected

PA UA PA UA

UC 94.0 0.875 90.2 94.9 96.6 93.4

NDVI 95.0 0.897 95.1 92.8 94.9 96.6

MD 94.0 0.877 95.1 90.7 93.2 96.5

MAHD 95.0 0.897 95.1 92.9 94.9 96.6

ML 93.0 0.855 90.2 92.5 94.9 93.3

SAM 95.0 0.897 95.1 92.9 94.9 96.6

NN 93.0 0.855 90.2 92.5 94.9 93.3

SVM 95.0 0.897 95.1 92.9 94.9 96.6


Precision Agric

123

complex image processing software. These methods should be sufficient if cotton root rot

is the only dominant stressor.

Conclusions

Results from this study demonstrate that unsupervised classification applied to multi-

spectral imagery, NDVI combined with unsupervised classification, and the six supervised

classifiers (minimum distance, Mahalanobis distance, maximum likelihood, SAM, neural

net and SVM) are all effective tools for detecting cotton root rot from airborne multi-

spectral imagery.This study further confirms that unsupervised classification applied to

multispectral imagery is accurate and sufficient for mapping cotton root rot. The other

seven methods provide useful alternatives, depending on the user’s preferences and image

processing capability. Although some of the methods appeared to be slightly better than the

others, the differences among the classifiers were small.

For practical applications, several or all of the methods can be used for comparison if

the user has the imaging processing capability; otherwise, the NDVI-based classification,

minimum distance, or SAM, can be easily implemented. For site-specific fungicide

applications, buffer zones around classified root rot area scan be created to account for the

classification omission and the possible further expansion of the fungus. It should be noted

that these methods are accurate and effective for detecting and mapping cotton root rot-

infected areas if root rot is the dominant stressor within the field as in most cases. If

multiple stressors with similar symptoms co-exist within fields, all the methods need to be

evaluated to identify the optimal methods. More research is needed to evaluate the con-

sistency and reliability of these methods and other spectral techniques for identifying root

rot infection under diverse stress and environmental conditions.

Acknowledgments This project was partly funded by Texas State Support Committee and CottonIncorporated, Cary, North Carolina. The authors wish to thank Adam Garcia of Edinburg, Texas and FredGomez of USDA-ARS at College Station, Texas for taking the airborne imagery for this study and JimForward of U.S. Fish and Wildlife Service at Alamo, Texas for assistance in image registration and groundverification.

Disclaimer Mention of trade names or commercial products in this article is solely for the purpose ofproviding specific information and does not imply recommendation or endorsement by the U.S. Departmentof Agriculture (USDA). USDA is an equal opportunity provider and employer.

References

Campbell, J. B. (2002). Introduction to remote sensing (3rd ed.). New York, USA: The Guilford Press.Congalton, R. G., & Green, K. (1999). Assessing the accuracy of remotely sensed data: Principles and

practices. Boca Raton, Florida, USA: Lewis Publishers.Ezekiel, W. N., & Taubenhaus, J. J. (1934). Cotton crop losses from Phymatotrichum root rot. Journal of

Agricultural Research, 49(9), 843–858.Hsu, C.-W., Chang, C.-C., & Lin, C.-J. (2007). A practical guide to support vector classification. National

Taiwan University, Taipei. http://ntu.csie.org/*cjlin/papers/guide/guide.pdf. Accessed 12 May 2014.Intergraph Corporation. (2013). ERDAS Field Guide. Huntsville, Alabama, USA: Intergraph Corporation.Isakeit, T., Minzenmayer, R. R., Abrameit, A., Moore, G., & Scasta, J. D. (2010). Control of Phymato-

trichopsis root rot of cotton with flutriafol. In: Proceedings of Beltwide Cotton Conferences (pp.200–203). Memphis, Tennessee, USA: National Cotton Council of America.

Precision Agric

123

http://ntu.csie.org/~cjlin/papers/guide/guide.pdf

Isakeit, T., Minzenmayer, R. R., Drake, D. R., Morgan, G. D., Mott, D. A., Fromme, D. D., et al. (2012).Fungicide management of cotton root rot (Phymatotrichopsis omnivora): 2011 results. In: Proceedingsof Beltwide Cotton Conferences (pp. 235–238). Memphis, Tennessee, USA: National Cotton Councilof America.

Jensen, J. R. (1996). Introductory digital image processing: A remote sensing perspective (2nd ed.).Englewood Cliffs, New Jersey, USA: Prentice-Hall.

Kruse, F. A., Lefkoff, A. B., Boardman, J. W., Heidebrecht, K. B., Shapiro, A. T., Barloon, J. P., et al.(1993). The spectral image processing system (SIPS): Interactive visualization and analysis of imagingspectrometer data. Remote Sensing of Environment, 44, 145–163.

Nixon, P. R., Escobar, D. E., & Bowen, R. L. (1987). A multispectral false-color video imaging system forremote sensing applications. In: Proceedings of the 11th Biennial Workshop on Color Aerial Pho-tography and Videography in the Plant Sciences and Related Fields (Vol. 340, pp. 295–305).Bethesda, Maryland, USA: American Society for Photogrammetry and Remote Sensing.

Richards, J. A. (1999). Remote sensing digital image analysis (p. 240). Berlin, Germany: Springer-Verlag.Rouse, J. W., Haas, R. H., Shell, J. A., & D. W. Deering. (1973). Monitoring vegetation systems in the Great

Plains with ERTS. In: Proceedings of the 3rd ERTS Symposium, NASA SP-351 (pp. 309–317).Washington, DC: U.S. Government Printing Office.

Smith, H. E., Elliot, F. C., & Bird, L. S. (1962). Root rot losses of cotton can be reduced. Publication No.MP361. College Station, Texas, USA: Texas A&M Agricultural Extension Service.

Yang, C. (2012a). A high resolution airborne four-camera imaging system for agricultural applications.Computers and Electronics in Agriculture, 88, 13–24.

Yang, C., Fernandez, C. J., & Everitt, J. H. (2005). Mapping Phymatotrichum root rot of cotton usingairborne three-band digital imagery. Transactions of the ASAE, 48(4), 1619–1626.

Yang, C., Fernandez, C. J., & Everitt, J. H. (2010). Comparison of airborne multispectral and hyperspectralimagery for mapping cotton root rot. Biosystems Engineering, 107, 131–139.

Yang, C., Odvody, G. N., Fernandez, C. J., Landivar, J. A., Minzenmayer, R. R., Nichols, R. L., et al.(2012). Monitoring cotton root rot progression within and across growing seasons using remotesensing. In: Proceedings of Beltwide Cotton Conferences (pp. 475–480). Memphis, Tennessee, USA:National Cotton Council of America.

Yang, C., Odvody, G. N., Fernandez, C. J., Landivar, J. A., Minzenmayer, R. R., Nichols, R. L., et al.(2014). Monitoring cotton root rot progression within a growing season using airborne multispectralimagery. Journal of Cotton Science, 18(1), 85–93.

Precision Agric

123

evaluating unsupervised and supervised image classification methods for mapping cotton root rot

Documents