Full Terms & Conditions of access and use can be found athttps://www.tandfonline.com/action/journalInformation?journalCode=tres20

International Journal of Remote Sensing

ISSN: 0143-1161 (Print) 1366-5901 (Online) Journal homepage: https://www.tandfonline.com/loi/tres20

Mapping oak decline through long-term analysisof time series of satellite images in the forests ofMalekshahi, Iran

Sadra Imanyfar, Mahdi Hasanlou & Vahid Mirzaei Zadeh

To cite this article: Sadra Imanyfar, Mahdi Hasanlou & Vahid Mirzaei Zadeh (2019): Mapping oakdecline through long-term analysis of time series of satellite images in the forests of Malekshahi,Iran, International Journal of Remote Sensing

To link to this article: https://doi.org/10.1080/01431161.2019.1620375

Published online: 27 May 2019.

Submit your article to this journal

View Crossmark data

Mapping oak decline through long-term analysis of timeseries of satellite images in the forests of Malekshahi, IranSadra Imanyfar a, Mahdi Hasanlou a and Vahid Mirzaei Zadeh b

aSchool of Surveying and Geospatial Engineering, College of Engineering, University of Tehran, Tehran,Iran; bSchool of Forest Sciences, Faculty of Agriculture and Natural Resources, University of Ilam, Ilam, Iran

ABSTRACTThe Zagros Mountains forests extend across 11 provinces in Iranand constitute approximately 40.0% of the country’s woodlands.These forests have important soil conservation and water regula-tion functions. Over the last decade, these forests have beendeclining in oak populations in many places, triggered by factorssuch as drought, pathogens like the fungus Biscogniauxia mediter-ranea, and pests such as borer beetles. Mapping the regions thatshow such a decline is the first step to addressing and managingthe risks posed by this environmental calamity. In this research, wefocus on the forests surrounding Malekshahi city in the Ilamprovince of Iran. Using Landsat data from the years 2000 to2016, we determined the spatial distribution of oak decline inthe region. After applying a forest/non-forest classification, appro-priate spectral indices including Enhanced Vegetation Index (EVI),Normalized Difference Vegetation Index (NDVI) and NormalizedDifference Water Index (NDWI) were selected. Together withground truth data, two regression methods (linear regressionand support vector regression (SVR)) were used to model thedecline score of each pixel based on the slope of variation ofselected spectral indices during the observed 17 years. The oakforests were then classified into four categories: healthy forests,low-severity-declined forests, mid-severity declined forests, andhigh-severity declined forests, based on the respective estimateddecline scores. SVR mapped different severities of oak decline withan overall accuracy of 51%, which appears to be due to thedependency of the method on the time of decline during the 17-year timeframe. However, in a binary classification mode – mean-ing classifying decline score to be either ‘Healthy’ or ‘decline’ –both regression methods were able to detect declined pixels witha producer’s accuracy of 100%.

ARTICLE HISTORYReceived 8 June 2018Accepted 26 March 2019

1. Introduction

Forests cover 31.0% of the global land area while only 7.40% of Iran is covered by forests(Keenan et al. 2015; Gooshbor et al. 2016). The per capita forest area in Iran is only 0.15ha (FAO 2015), against a world average of 0.64 ha, which means that maintaining forestresources in the country is especially important. The Zagros Forest, with an area of about

CONTACT Mahdi Hasanlou hasanlou@ut.ac.ir School of Surveying and Geospatial Engineering, College ofEngineering, University of Tehran, Tehran 1439957131, Iran

INTERNATIONAL JOURNAL OF REMOTE SENSINGhttps://doi.org/10.1080/01431161.2019.1620375

© 2019 Informa UK Limited, trading as Taylor & Francis Group

five million hectares, constitutes about 40.0% of all of Iran’s forests (Heshmati 2007;Salehi, Wilhelmsson, and Söderberg 2008; Valipour et al. 2014), and spans 11 provinces(Badian and Azim Nejad 2016).

Available data indicate that the global forest area declined by 3% from 1990 to 2015(Keenan et al. 2015). Also in Iran, there have been alarming reports about the health ofoak trees, which are the dominant tree species in the Zagros Forest. The first worryingreports came from Ilam province in 2009, where the status of the forest health was morealarming than in other provinces of Iran (Imanyfar and Hasanlou 2017). Symptoms ofa serious disease called ‘oak decline’, including unusual leaf colour change, leaf loss andsap secretion from infected parts of the trees could be observed (Mirabolfathy 2013).The disease is triggered by factors such as drought, dust, pathogens like the fungusBiscogniauxia mediterranea and pests such as borer beetles (Mattson and Haack 1987;Vannini and Valentini 1994; Vannini, Paganini, and Anselmi 1996; Hoseiny 2011; Pirouziand Tavakoli 2015) and now extends across about 1.5 million hectares – constitutinga major ecological disaster (Zakeri, Hojjati, and Kiadaliri 2013).

Mapping the declining regions is the first step towards addressing and managing therisk posed by this environmental calamity (Imanyfar and Hasanlou 2017). To that end,the importance of mapping is already well recognized as part of the Food andAgriculture Organization (FAO) of the United Nations experts’ recommendations, aswell as relevant instructions from the Forests, Range and Watershed ManagementOrganization (Forests, Range and Watershed Management Organization of Iran 2011).There are different ways to produce such a map, but considering the large extent of theZagros Forest, remote sensing (RS) methods should be particularly suitable and cost-effective provided that relevant phenomena can be detected with these. In RS research,numerous studies have been carried out using satellite imagery to monitor forest healthand growth (J. Wang et al. 2010; Lambert et al. 2013; Meng et al. 2016; Lewińska et al.2016; Housman et al. 2018). For example, Lambert et al. (2013) extracted vegetationactivity variations using a 12-year time series of Moderate Resolution ImagingSpectroradiometer (MODIS) data and, as a result, remarkable relationships were foundbetween the indicator of spring vitality derived from remote sensing images and theobserved status of forest stands (Lambert et al. 2013).

Diseases affecting Oak species are nothing new and have occurred in several coun-tries before, such as Japanese Oak Wilt (JOW) in Japan and Sudden Oak Death (SOD) inthe United States of America (USA) (Imamura et al. 2017; Kozanitas et al. 2017). It shouldbe noted that these diseases differ from those affecting the Zagros trees, at least interms of the fungus species involved. Many researchers have investigated the phenom-ena using different kinds of RS data, including medium resolution optical imagery, e.g.,Landsat data, high-resolution optical imagery, e.g., IKONOS data, and hyperspectral data(Komura et al. 2005; Weissling, Xie, and Jurena 2005; Wang, Lu, and Haithcoat 2007;Wang, He, and Kabrick 2008; Uto et al. 2008; Gillis 2014; Mahdavi et al. 2015; Rostamniaand Akhoondzadeh Hanzaei 2016). In the following paragraphs, selected research effortsrelated to the study of oak disease using each of these satellite data types will bediscussed.

Wang, Zhenqian, and Haithcoat (2007) investigated the detection of an oak decline inthe Mark Twain National Forest, in the Ozark Highlands, Missouri, USA. They applied theNormalized Difference Water Index (NDWI) to map the continuous forest dynamics

2 S. IMANYFAR ET AL.

associated with oak decline. Landsat 5 Thematic Mapper (TM) imagery from 1992 andEnhanced Thematic Mapper plus (ETM+) imagery from 2000 were processed to computethe differential NDWI (DNDWI). A simple thresholding method was used to map oakdieback, recovery, and stable areas. Although the overall accuracy was found to berelatively high (76%), this approach was not fully validated because of limitations inground data collection (Wang, Zhenqian, and Haithcoat 2007). In another study, a riskrating system based on a series of biophysical variables was developed for the samegeographical region (Wang, He, and Kabrick 2008).

In another study (Gillis 2014), a methodwas proposed for mapping SOD in the Santa CruzMountains in central California. In that research, TM imagery was used. It should be notedthat regions, where trees were destroyed for certain other reasons (such as fire), weredetermined using auxiliary data. The researchers calculated some vegetation indices andexamined them to determine which of them to apply, and what threshold value to use inorder to separate declined regions from healthy ones. The ratio of the Short Wave Infrared(SWIR) to the Near Infrared (NIR) proved most useful for this task. An accuracy evaluationshows that the method identified pixels containing dead trees with 24% accuracy, which, asthe author pointed out, makes the results rather unreliable (Gillis 2014).

Landsat data were used in research to study the part of the Zagros Forest located inIlam province (Rostamnia and Akhoondzadeh Hanzaei 2016). In this study, the effect ofdust and rainfall on the oak decline was investigated using linear regression. Resultsindicate that the effects of dust and rainfall are 38% and 62%, respectively. This studydid not, however, consider mapping the oak decline (Rostamnia and AkhoondzadehHanzaei 2016).

In another study (Mahdavi et al. 2015), the effects of physiographic factors (terrain relief),forest cover density, and soil on the severity and distribution of oak decline were evaluatedusing a logistic regression model. The results showed that with increasing altitude in thewestern and southern aspects, and with increasing forest cover density and in areas withshallow soils and steep terrain, the severity and distribution of the decline in oak treesincreased. Although this research did produce a prediction map for the oak decline, onlyone of the input parameters was influenced by the spectral behaviour of the forest, whilethe soil data were supplied by an external source (Mahdavi et al. 2015).

Other researchers have used high-resolution satellite imagery (HRSI) to detect dis-eases affecting oak trees. Using high-resolution imagery can be useful in detectingphenomena at the individual tree scale. Komura et al. (2005) used high spatial resolutionimages acquired by IKONOS and proposed a method to identify individual dead treeskilled by JOW. In this study, orthophoto imagery was transformed into both HueIntensity Saturation (HSI) and Normalized Difference Vegetation Index (NDVI) datasets.Individual dead tree crowns were recognized by cluster analysis, using datasets of H, S,and NDVI. Comparing the final map with training data indicated that the recognitionrate using HRSI was overestimated (Komura et al. 2005).

In another reasearch (Karami et al. 2017), oak decline was mapped into four levels ofseverity for some parts of Ilam forests using Worldview-2 satellite data. Several methodsincluding; Maximum likelihood, naive Bayes, K-nearest neighbours, and artificial neuralnetwork classification algorithm were applied to the image and the results showed thatthe artificial neural network classification algorithm provided the highest overall accu-racy of 73% (Karami et al. 2017).

INTERNATIONAL JOURNAL OF REMOTE SENSING 3

Early detection of oak diseases is very important from a managerial point of view. Forthis reason, there has been a tendency to use hyperspectral imagery in some studiesbecause of their greater spectral resolution. In a study by Weissling, Xie, and Jurena(2005), a great amount of in-situ hyperspectral data from different parts of a tree fordiscriminatory signatures of the disease in different stages of pathology were collected,processed, and analysed. The output led to the development of classification methodsthat can be applied to hyperspectral imagery from the Hyperion satellite obtained overthe study area in a scheduled acquizition for early disease detection (Weissling, Xie, andJurena 2005).

Other researchers have investigated whether or not symptoms of JOW in early stagescould be detected using hyperspectral sensor data. Model trees suffering from JOWwere created and observed until their death using multi-spectral cameras as well asspectrometers (Analytical Spectral Devices (ASD) FieldSpec HandHeld). The resultsshowed that the difference spectrum changed after beetle infestation, which isa possible warning symptom of JOW (Komura et al. 2005).

The research of Gillis (2014) did not produce reliable results, and although themethod proposed by Wang, Zhenqian, and Haithcoat (2007) achieved relatively highoverall accuracy, it was not fully validated. Mahdavi et al. (2015) modelled oak decline,but the modelling inputs were not completely based on RS products. The research ofRostamnia and Akhoondzadeh Hanzaei (2016) was not focused on oak decline mapproduction. In this paper, a method is proposed which uses time series of medium-resolution Landsat data to map oak decline and which is validated using ground truthdata (field data and Google-Earth-based data). Figure 1 illustrates the workflow of theproposed method. The method involves three steps. First, a binary classification isapplied to extract forest regions from non-forest ones. In the second step, three spectralindices are then selected according to ground truth data, rainfall data, and the results ofprevious research including Enhanced Vegetation Index (EVI) (Huete, Justice, andLeeuwen 1999), NDVI (Rouse 1974), and NDWI (Gao 1996). A time series of the indicesis calculated for each pixel, and a sequence of values is accessible from each index point.The trend of each index is examined by fitting a linear equation. During the third step,two different regression models use slope values of fitted lines to estimate a declinescore for each pixel. Decline scores are classified into four categories: healthy, low-severity, mid-severity, and high severity. The support vector regression (SVR)-basedmethod is found to be able to detect different severities of decline with an overallaccuracy of more than 50%.

2. Study area

Malekshahi county, with an area of 180,000 ha, lies at the geographic centre of Ilamprovince in Iran and is located between 46.2708° – 46.8800° East and 32.0844° – 32.5150°North. Malekshahi county is considered part of the arid and semi-arid regions of theZagros Mountains. Its altitude ranges from 330 m to 2,737 m above sea level. In additionto forest, there are several other land covers including urban area, waterbody, farmlandand rangeland. The location of the case study is shown in Figure 2.

The dominant oak species same as whole Zagros ecological region is Persian oak orQuercus persica (Salehi, Wilhelmsson, and Söderberg 2008). One of the forest regions of

4 S. IMANYFAR ET AL.

Malekshahi, which is denser than others, is called Bivareh. It extends from the centre ofMalekshahi to the southwest for more than 10 km and is indicated in green colour inFigure 2(d).

3. Data and pre-processing

3.1. Satellite data

This research used Landsat data of medium (30 m) resolution. Ordinarily, it would not bepossible to detect individual declined oak trees at this resolution (as canbe seen in Figure 7(b));however, if a method could be developed to map oak decline using these data, that wouldhave two important advantages. Firstly, these data are available free of charge, which meansthe method would be more practical for routine application. Secondly, the results could beused as a guide for other, more expensive, methods using HRSI.

Landsat satellite imagery from the years 2000 to 2016 was used (Table 1). The best time ofthe year for annual time series production was selected based on phenological studies. Thecriterion was rangeland dryness and forest greenness. To this aim, 16-year time series ofMODIS 16 day composite NDVI data of rangeland and forest pixels were analysed using TIME-series of SATellite (TIMESAT) software (Zhang et al. 2003; Hmimina et al. 2013). The output of

Figure 1. Workflow of the proposed method. Different steps are separated.

INTERNATIONAL JOURNAL OF REMOTE SENSING 5

TIMESAT was several phonological parameters of each land cover, including the beginningand end of the growth season (Jönsson and Eklundh 2004; Eklundh and Jönsson 2017).

The date of the higher peak of the NDVI values of the forest (which is expected to bemore useful) is strongly correlated with the peak of rangeland (as can be visuallyobserved by the red line in Figure 3). This causes inaccurate results if it is selected asthe best date. According to phenological parameters of forest and rangeland, the besttime of the year is determined based on the lower peak of forest cover which isthe second half of September and the first half of October (or the 18th 16-day period

Table 1. Date of used satellite imagery.Sensor Date Sensor Date

ETM+(Single Line Corrector(SLC) -on)

11 October 2000 ETM+(SLC-off)

4 October 200914 October 2001 23 October 20101 October 2002 10 October 2011

ETM+(SLC-off)

20 October 2003 26 September 201220 September 2004 OLI 7 October 20139 October 2005 24 September 201426 September 2006 13 October 201515 October 2007 14 October 201617 October 2008

Figure 2. Geographical location of the study area; (a) Iran is highlighted in the world map with redcolour, (b) Ilam province is highlighted in the map of Iran’s provinces, (c) Malekshahi county ishighlighted in the map of Ilam’s counties, and (d) Malekshahi county. Bivareh forest is highlightedwith green colour.

6 S. IMANYFAR ET AL.

of the year which is indicated by the green line in Figure 3(b); green line). Table 1 showsall dates for when the images were acquired.

Products and services provided by the United States Geological Survey (USGS) EarthResourcesObservation andScience (EROS) Centre’s ScienceProcessingArchitecture (ESPA) on-demand interface were used in this research (Jenkerson 2013). Landsat ETM+ andOperationalLand Imager (OLI) images were atmospherically corrected using the Landsat EcosystemDisturbance Adaptive Processing System (LEDAPS) (Masek et al. 2013) and Landsat SurfaceReflectance Code (LaSRC) (Vermote et al. 2016), respectively, and theywere clipped to the areaof Malekshahi county. The pixels were evaluated in terms of quality based on the ‘Pixel_qa’data layer. In this step, all pixels covered by snow, ice, clouds, cloud shadows, and water weremasked out.

3.2. Ground truth data

Calibrating oak decline models as well as evaluating the accuracy of the results requiresground truth data. The ground truth of this research is a combination of field data andGoogle-Earth-based data. We selected two small parts of Bivareh forest, which based onpreliminary studies, were known to have suffered more decline. Fieldwork was performed on9 September 2016 and on 12, 19, 20, and 21 October 2016. The health condition of all trees inboth parts of Bivareh forest wasmeasured visually and classified into four categories (Table 2).Their location was determined using smartphone Global Positioning System (GPS) data with

Figure 3. Seasonal decomposition of 16-year time series of MODIS 16 day composite NDVI data(23 periods per year). Redline proves high correlation between first growing season of rangelandand forest. (a) Forest pixel which presents to have two growing season per year, (b) rangeland pixelwhich has only one growing season per year, and (c) Timing of the second growing season of forestpixel which occurs in 18th 16 day period.

INTERNATIONAL JOURNAL OF REMOTE SENSING 7

a positional accuracy of about 3 m. Figure 4 shows an example of the collected field groundtruth data. Figure 5 shows different parts of declined oak trees observed in the field.

Field data alone, however, is not sufficient as ground truth data for our time seriesmethod, because a tree may simply have declined or have been felled before thefieldwork took place. Operators consequently would not be able to observe and mea-sure its health condition, and would instead consider its former location as a spot thatnever held a tree. From the point of view of satellite imagery acquired as early asthe year 2000, however, it is clear that there was indeed a tree in this spot that hadsimply declined in the meantime. This discrepancy causes a problem because it leads toinconsistencies between images and ground truth data. The problem can be solvedusing historical images of Google Earth. By comparing images from 2006 and 2017,vanished trees are easily detected. These missing trees are then classified as the highestlevel of decline, because they are already dead.

There still remain, however, some other inconsistency problems which should be solvedwith a few pre-processing steps. The first problem is that ground truth data (aside from felledtrees) are independent of the area of tree canopies. The more area a tree canopy covers, themore it will affect pixel values, but the area of canopy does not affect ground truth data at all,because collected data are not considered as a polygon layer but a point layer. To solve this

Table 2. Assigned code of trees based on the level of decline severity.Tree status Assigned code

Healthy forest(declines less than 25.00%) 1Low-severity declined (declines between 25.00% and 50.00%) 2Mid-severity declined (declines between 50.00% and 75.00%) 3High-severity declined (declines more than 75.00%) 4

Figure 4. Field collected data.

8 S. IMANYFAR ET AL.

problem, Google Earth images were used once again. First, the canopy layer of those twoparts of the Bivareh forest was drawn manually, and then a matching between this polygonlayer and the field-data-based point layer was performed. Next, the decline information of thecollected points was assigned to polygons. The output of this step can be seen in Figure 6.

The second pre-processing step is co-registration between the Google Earth image(canopy layer) and the satellite image. The final problem is that our algorithm producesa raster map of oak decline, and cannot detect individual trees, but the ground truth datadoes include individual trees. For this stage of pre-processing, it was decided to assign eachpixel a score that describes the severity of its decline. This solution can be considered a kindof weighted averaging. Each tree in a pixel will affect the score of that pixel according to twofactors: (1) the area of its canopy, and (2) the severity of its decline (Table 3).

Two kinds of ground truth data are needed to implement themethod. Ground truth datarelated to oak decline are needed for the third step (of ground truth data pre-processing)and also for effective spectral index selection (which is the second step of the proposedmethod). The second necessary ground truth data are related to canopy cover, which is usedin the forest/non-forest classification (first step of the proposed method).

The forest canopy cover, also known as crown cover, is defined as the ratio of theforest floor covered by the vertical projection of the tree crowns (Azizi, Najafi, andSohrabi 2008). Figure 7(a,b) show the true colour of a part of Bivareh forest and its

Figure 5. Examples of different parts of declined oak trees.

INTERNATIONAL JOURNAL OF REMOTE SENSING 9

equivalent forest/non-forest classified data (based on Google Earth data), respectively.As shown in Figure 7(b), Landsat data are also superimposed over the forest/non-forestclassified data. Each pixel of the canopy cover data layer – coinciding with the Landsatdata pixel – indicates how many percent of it is covered by the forest canopy.

Ground truth data of canopy cover assigns a value between 0.00 (empty of treecover) and 1.00 (fully covered by the tree) to a pixel, applying unsupervized classificationmethods (k-means and Iterative Self-Organizing Data Analysis (ISODATA)) to GoogleEarth images. For example, pixels of green and red boundaries have values of about0.50 and 0.00, respectively. According to the Forests, Ranges, and WatershedManagement Organization of Iran, for the Zagros ecological region, the canopy covershould be higher than 0.05 in order for it to be classified as forest. This kind of groundtruth data was prepared for 307 pixels.

Figure 6. Canopy layer of field-observed trees, drawn manually using Google Earth high-resolutionimagery, presented with olive colour. Also, cut trees are drawn based on comparison with GoogleEarth historical imagery, presented with orange colour.

Table 3. Scoring of different levels of decline severity.Decline severity Assigned score

1 02 13 34 5Cut tree (based on Google Earth images comparison) 5

10 S. IMANYFAR ET AL.

4. Method

To achieve objectives of this study, a method is proposed which includes three steps; (1)forest/non-forest classification, (2) effective spectral indices selection, and (3) oak declinemodelling.

4.1. Forest/non-forest classification

Forest pixels should be determined via a binary classification. In addition to forest, thereare several other land covers including urban area, waterbody, farmland and rangeland.The existence of land covers that are spectrographically similar to forest may be detri-mental to the quality of the final results, and so it seems useful to apply a binaryclassification to reduce probable inaccuracies.

A decision tree classifier containing two internal decision nodes was used (Lu, Di, and Ye2014). First, the NDVI value of each pixel is calculated at the right time of the year, when treesshow high photosynthetic activity. All pixels with an NDVI value less than T0 are considerednon-forest. Then, if the difference between the NDVI values of early autumn and winter is

Figure 7. Canopy cover ground truth data generation. (a) True colour of a region of Bivareh forestand (b) forest/non-forest classified data of the region superimposed by Landsat data with a spatialresolution of 30 m. Pixels of green and red boundaries have canopy cover values of about 0.5 and0.0, respectively.

INTERNATIONAL JOURNAL OF REMOTE SENSING 11

higher thanT1, they are classified as forest. For the purpose of determining T0 andT1, groundtruth data related to canopy cover was used.

4.2. Effective spectral indices selection

Selection of indices was done based on three criteria. The first criterion, according toexperts on the oak decline, relates to the fact that the less it rains, the more trees areprone to decline. This means that an index should be chosen that contains sufficientinformation on rainfall data. Precipitation data measured at meteorological stations isneeded to evaluate correlations with different indices. As mentioned previously, one ofthe symptoms of oak decline is an irregular loss of leaves, which affects the density ofcanopy layer. Another criterion for index selection, therefore, is related to canopy layerdensity. The third criterion is the extent of sensitivity to oak decline. Oak decline groundtruth data were prepared and used to determine the degree of sensitivity to oak declineacross different indices.

4.3. Oak decline modelling

The oak decline phenomenon was modelled based on the slope values of the fitted lineof the selected indices. Two different regression methods, linear regression and SVR, areused and compared with each other. Conventional regression techniques like linearregression are more appropriate for data that are clearly linear or exponential, whereasmachine-learning techniques are usually better suited to modelling strong nonlinearitybetween biophysical and biochemical parameters and reflection spectra (Liang et al.2015; Yue et al. 2018). Input features of regression methods are the general trend ofselected indices during the period, including NDVIs, NDWIsand EVIs.

In RS studies, the linear regression method has been used for different purposes, fromspectral unmixing (Theys et al. 2009) to determining relationships between primary andauxiliary variables (Haack and Rafter 2010). A mathematical model of linear regressioncan be seen in Equation (1). In this equation, a, b, c, and d are regression coefficients andS is the output parameter which is to be estimated, meaning in our case, the declinescore.

S ¼ aðNDVIÞs þ b ðNDWIÞs þ c ðEVIÞs þ d (1)

On the other hand, it has been proven that SVR can be used for several applications(Kazuhito et al. 2017; Okujeni et al. 2017; Rosentreter et al. 2017). Parameters of SVRinclude kernel parameter and penalty coefficient. Equation (2) shows a mathematicalmodel of the radial basis function (RBF) kernel used in this research. xi � xj

�� ���� ��2may berecognized as the squared Euclidean distance between the two feature vectors and isthe γ kernel parameter.

k xi; xj� � ¼ exp �γ xi � xj

�� ���� ��2� �; γ > 0 (2)

Regression methods estimate the decline score, which is a continuous value. For thesake of classification, however, score values should be converted into four discretecategories. The boundary values of these categories are determined by simulation.

12 S. IMANYFAR ET AL.

5. Results and discussion

The year selected for applying the classification was 2001. The parameters of thedecision tree classifier (T0 andT1) should be determined using canopy cover groundtruth. Seventy percent of ground truth data are then selected for training the classifier,while the remaining 30% is set aside for test purposes. After training, 0.145 and 0.020were assigned to T0 andT1, respectively. Using these two parameters, each pixel ofinterest was classified. Figure 8 shows the results of the classification. Using test data,classification accuracy was then evaluated. Table 4 shows the confusion matrix.According to the Table 4, an acceptable overall accuracy of 88% was obtained forforest/non-forest classification map.

For the first part of the second step, monthly precipitation data were obtained from Ilammeteorological station for the years 2000 to 2014. By normalizing precipitation values ata given timescale, the Standard Precipitation Index (SPI) (McKee, Doesken, and Kleist 1993)was obtained at the same timescale, which is one of the most useful indices for droughtmonitoring. Five indices, namely EVI, NDVI, NDWI, Soil-Adjusted Vegetation Index (SAVI)(Hongrui Ren, Zhou, and Zhang 2018) and Adjusted Transformed Soil-Adjusted VegetationIndex (ATSAVI) (H. Ren and Feng 2015) were investigated to find the one that mostcorrelated to precipitation data using simple linear regression. The average value of these

Figure 8. Forest map of study area. Black regions indicate the forest.

Table 4. Confusion matrix of forest/non-forest classification.Non-forest Forest Producer’s accuracy

Non-forest 4 10 0.28Forest 1 76 0.99User’s accuracy 0.80 0.88 Overall accuracy = 0.88

INTERNATIONAL JOURNAL OF REMOTE SENSING 13

indices over the forest pixels of Malekshahi was computed for each year. The rainfall valuefor each year was the average of monthly precipitation data from the start of autumn ofthe year to the end of spring of the next year. The greater coefficient of determination (R2) is,the more representative the index is of rainfall data. R2 values can be seen in Table 5.According to the listed R2 values, EVI is the best index on the list.

For the second criterion of step two, we need to find the index most correlated withcanopy cover density. Results of previous studies indicate that NDVI is a suitable index for thispurpose (Larsson 2002; La et al. 2013). Considering the final criteria, five indices werecomputed for each image, namely EVI, NDVI, SAVI, ATSAVI and NDWI. For each pixel, 17values (or less, if the Digital Number (DN) of a pixel is invalid for some years because of cloudsor other factors) are then available for each index. The general trend of each index wasexamined by fitting a linear equation. Figure 9 is the visual explanation of the process. Tomake it clearer, the slope image of the NDVI is shown in Figure 10, which also can give a visualimpression of decline. Although using NDVI or even false colour composite (FCC) images forsome years of research time frame can give a better visual impression, in our view thismethodmay be even misleading for the readers, because of two reasons; the case study is a sparseforest and used satellite data is of medium resolution.

Such an image (Figure 10) was produced for each selected index. A possible result ofnon-random patterns of Figure 10 is the possibility of oak decline detection using

Table 5. The coefficient of determination values betweenrainfall time series and spectral indices time series.Spectral index R2

NDVI 0.50NDWI 0.21SAVI 0.42EVI 0.59ATSAVI 0.47

Figure 9. Time series of NDVI, (a) the plot of how NDVI values change for an arbitrary pixel during17-year time series, and (b) the red line is the fitted line which its slope is used in the followingsteps.

14 S. IMANYFAR ET AL.

30 m satellite imagery. For example, the red-bounded region experienced deteriorationduring research time frame.

According to the ground truth data on oak decline, the number of pixels with a declinescore higher than 1.00 is 104. The correlation between these two sets of values (decline scoresand linear equation slope) was measured for each of the indices. The results are shown inTable 6.

Low correlation values were unexpected. One of the probable reasons was due tothe time frame of the study. If the time frame is not selected appropriately then forestspectral changes caused by the disease are less effective on the slope value andconsequently a lower correlation will be obtained. The probability of that assumptionincreased in accordance with the first years of oak decline outbreak, which were in themiddle of the time frame of our research. This hypothesis was examined, and theresult showed that the most appropriate time frame is a 9-year period: from 2008 to2016. Correlation values – for ground truth pixels with a declining score higher than5.00 – were computed and indicated the correctness of the assumption, are shown inTable 6. It becomes clear that NDWI is more sensitive to oak decline.

Figure 10. Slope image of NDVI index. The redder colour is, the higher the severity of declineoccurred in NDVI’s point of view.

Table 6. Correlation values between decline score and fitted line slope for 16 year and 9 year timeframes.

Spectral index

Time frame NDVI EVI NDWI SAVI ATSAVI

2000–2016 −0.19 −0.13 −0.04 −0.06 −0.062008–2016 −0.57 −0.56 −0.63 −0.56 −0.60

INTERNATIONAL JOURNAL OF REMOTE SENSING 15

In total, NDWI, EVI, and NDVI were considered as effective indices in relation to theoak decline phenomenon. For the purpose of modelling, approximately one-third of theground truth pixels – related to oak decline – was used to test the results (30 pixels outof 104 pixels), while the remainder was used for the modelling process. The linearregression method was trained first. The results of solving the system of equations areshown in Table 7.

Themodel is then validated using test data. The results indicate that the root-mean-squareerror (RMSE) for estimating the decline score is 3.41. It should be noted that the score (outputof regression) is a continuous quantity, but after the regression, the estimation (output) can beclassified into several categories. Boundary values for those categories were computed usinga simulation of some hypothetical pixels (to consider all possible levels of decline severitywhich may not happen in the ground truth data). The results are shown in Table 8.

According to the boundary values presented in Table 8, the Landsat data wereclassified for the oak decline phenomenon (Figure 11).

Table 7. Coefficient parameters for linear regression.Model parameter a b c d

Value −0.0220 −0.0007 −0.0270 5.2500

Table 8. Separating score values of different classes of oakdecline severity using simulated declined pixels.Decline severity Low Mid High

Decline score range 0.3–4.0 4.0–7.0 > 7.0

Figure 11. Oak decline map based on linear regression method.

16 S. IMANYFAR ET AL.

In the process of validation, the accuracy of the classifications should also be evaluated.This is achieved through twomodes. In the first mode, the model’s ability to identify declinedpixels against healthy ones is evaluated. In other words, in this step, the accuracy of a binaryclassification (declined/healthy) is evaluated using 85 pixels of the ground truth data. Theresults show that themethod can identify a declined pixel with a producer’s accuracy of 100%(but with lower user’s and overall accuracy). Of course, since healthy forest pixels were notavailable in the ground truth data, this would be ‘producer’s accuracy’, this means that alldeclined regions, in reality, are classified as declined in the final map. In the secondmode, thesuccess rate of the method to examine and distinguish different severities of decline wasevaluated. The error matrix can be seen in Table 9.

A second regression method for modelling oak decline is SVR. First, all three indices werescaled. In this research, the RBF kernel function was used. Then RBF parameter (γ) and penaltyfactor (C) were optimized using grid search. C and γ were equal to 104 and 2�13, respectively.Then the model was validated using test data. The RMSE was equal to 3.39. As in the previousregressionmethod, the oak declinewas classified and the resultingmap is shown in Figure 12.

Table 9. Confusion matrix of oak decline classification, based on linear method.Method

Low-severity Mid-severity High-severity Total Producer’s accuracy

Ground truth Low-severity 8 9 6 23 0.35Mid-severity 12 17 5 34 0.50High-severity 2 9 17 28 0.60Total 22 35 28 Overall

accuracy = 0.49User’s accuracy 0.36 0.48 0.60

Figure 12. Oak decline map based on SVR regression method.

INTERNATIONAL JOURNAL OF REMOTE SENSING 17

Next, the accuracy of the classification was also evaluated. SVR, like linear regression, detecteddeclined pixels (in the first mode) with a producer’s accuracy of 100% (but with lower user’sand overall accuracy). In the second mode, Table 10 shows the confusion matrix of theclassification.

Before proceeding with different regression methods, it was expected that SVR wouldprovide better results because it is less dependent on training data. Validation outcomesindicate the correctness of the expectation. It is worth noting thatmodelling oak decline usingthe SVRmethod results in amore pessimistic outcome thanmodellingwith the linearmethod.

Irrespective of which regression method is used, there might be a defect in the time-series method considering the overall accuracy of about 50%. According to the analysisof the authors, the output of the method depends on the time of decline in the 17-yeartime frame of the research, but the ground truth data are independent of that. In otherwords, if trees are declined in a late section of the time frame, they are less likely to bedetected by the method.

The study tried tomake a step forward in the domain of oak decline related researches. Thestudy of Gillis (2014) did not produce reliable results, and although the method proposed byWang, Zhenqian, and Haithcoat (2007) achieved relatively high overall accuracy, it was notfully validated. Mahdavi et al. (2015) modelled oak decline, but themodelling inputs were notcompletely based on RS products. The research of Rostamnia and Akhoondzadeh Hanzaei(2016) was not focused on oak decline map production. The method proposed here in thisstudy maps oak decline based on only remote sensing data and validated using ground truthdata.

In spite of the probable defect of themethod, the results still seem to be useful for relevantorganizations, because of the high producer’s accuracy in detecting declined regions. So theproducts can be used as a guide for optimizing other methods (like forest inventory or aerialimagery), which are time-consuming and costly. Another advantage of themethod is that it isbased only on the satellite imagery and it does not need any other auxiliary data. Also, itshould be noted that the used satellite imagery is provided free of charge. Another point ofthe study was the approach used to select effective spectral indices, which is based onphysical parameters (rainfall, etc.) instead of only mathematical approaches.

6. Conclusion

Over the last few decades, RS techniques have been used to monitor forest health. ZagrosForest, as the most extensive ecological region of Iran, has faced harsh conditions duringrecent years. In this research, a method was proposed to map oak decline using Landsat data.In thismethod, after applying a forest/non-forest classification, three optimum spectral indices

Table 10. Confusion matrix of oak decline classification, based on SVR.Method

Low-severity Mid-severity High-severity Total Producer accuracy

Ground truth Low-severity 10 7 6 23 0.43Mid-severity 12 17 5 34 0.50High-severity 4 8 16 28 0.57Total 26 31 27 Overall accuracy = 0.51User accuracy 0.38 0.55 0.59

18 S. IMANYFAR ET AL.

including EVI, NDVI, and NDWI were selected. These indices were selected considering rainfalldata, canopy cover, and oak decline, respectively. Then two regression methods were used tomodel the phenomena: SVR and linear regression. The regressionmethodsmodelled a declinescore for each pixel. The oak forest was classified into four categories: healthy forest, low-severity declined forest, mid-severity declined forest, and high-severity declined forest. Theclassification was validated using ground truth data. SVR was more successful than linearregression. SVR mapped oak decline with an overall accuracy of more than 51%. However,both methods were able to detect declined regions with a producer’s accuracy of 100% if thedecline score was classified as binary healthy/declined categories (but with lower user’s andoverall accuracy). It seems that the method suffers a defect because of a dependence on thetime of decline during the study period. Considering the objectives of the study, mapping oakdecline was performed using free satellite data andwithout using any other auxiliary data andalso the products seem to be useful for time and cost reduction of other complementary oakdecline monitoring methods. However, the overall accuracy of the products – especially forclassifying declined region to different decline severities – means that there is still a need toresearch for the development of higher performance algorithms.

Acknowledgements

We would like to thank the staff and directorate of the Natural Resources and WatershedManagement Organization of Malekshahi county, who were of great help to us during the in-situdata collection phase.

Disclosure statement

No potential conflict of interest was reported by the authors.

ORCID

Sadra Imanyfar http://orcid.org/0000-0002-8347-947XMahdi Hasanlou http://orcid.org/0000-0002-7254-4475Vahid Mirzaei Zadeh http://orcid.org/0000-0002-2857-6728

References

Azizi, Z., A. Najafi, and H. Sohrabi. 2008. “Forest Canopy Density Estimating Using Satellite Images.”The International Archives of the Photogrammetry, Remote Sensing and Spatial InformationSciences, XXXVII, Part B8. Beijing, China.

Badian, Z., and Z. Azim Nejad. 2016. “Preliminary Investigation of the Causes of Oak Decline inZagros Forests.” In World Conference on Agri-Environment and Agenda for SustainableDevelopment and Safe Production. Shiraz, Iran.

Eklundh, L., and J. Per 2017. “TIMESAT 3.3 With Seasonal Trend Decomposition and ParallelProcessing Software Manual.”

FAO. 2015. Iran (Islamic Republic Of) - Global Forest Resources Assessment 2015 – Country Report.Rome, Italy.

Forests, Range and Watershed Management Organization of Iran. 2011. “Sustainable ForestManagement Instructions of Zagros Forest Ecosystems for the Prevention and Control of OakDecline.” Forest, range and watershed management organization.I.R. of Iran.

INTERNATIONAL JOURNAL OF REMOTE SENSING 19

Gao, B.-C. 1996. “NDWI—A Normalized Difference Water Index for Remote Sensing of VegetationLiquid Water from Space.” Remote Sensing of Environment 58 (3): 257–266. doi:10.1016/S0034-4257(96)00067-3.

Gillis, T. 2014. “Use of Remotely Sensed Imagery to Map Sudden Oak Death (PhytophthoraRamorum) in the Santa Cruz Mountains.” Master of science, university of southerncalifornia.

Gooshbor, L., M. Pirbavaghar, J. Amanollahi, and H. Ghobari. 2016. “Monitoring Infestations of OakForests by Tortrix Viridana (Lepidoptera: Tortricidae) Using Remote Sensing.” Plant ProtectionScience 52 (4): 270–276. doi:10.17221/185/2015-PPS.

Haack, B., and A. Rafter. 2010. “Regression Estimation Techniques with Remote Sensing: A Reviewand Case Study.” Geocarto International 25 (1): 71–82. doi:10.1080/10106040802711679.

Heshmati, G. A. 2007. “Vegetation Characteristics of Four Ecological Zones of Iran.” InternationalJournal of Plant Production 2(Sept). doi:10.22069/ijpp.2012.538

Hmimina, G., E. Dufrêne, J.-Y. Pontailler, N. Delpierre, M. Aubinet, B. Caquet, A. de Grandcourt, et al.2013. “Evaluation of the Potential of MODIS Satellite Data to Predict Vegetation Phenology inDifferent Biomes: An Investigation Using Ground-Based NDVI Measurements.” Remote Sensing ofEnvironment 132: 145–158. May. doi:10.1016/j.rse.2013.01.010.

Hoseiny, A. 2011. “Infestation of Forest Trees to the Borer Beetle and Its Relation to HabitatConditions in the Persian Oak (Quercus Brantii) in Ilam Province.” Forest and Range ProtectionResearch 9 (1): 53–66.

Housman, I., R. Chastain, M. Finco, I. W. Housman, R. A. Chastain, and M. V. Finco. 2018. “AnEvaluation of Forest Health Insect and Disease Survey Data and Satellite-Based Remote SensingForest Change Detection Methods: Case Studies in the United States.” Remote Sensing 10 (8):1184. doi:10.3390/rs10081184.

Huete, A., C. Justice, and V. L. Wim 1999. “MODIS VEGETATION INDEX (MOD 13) ALGORITHMTHEORETICAL BASIS DOCUMENT-Version 3.”

Imamura, K., S. Managi, S. Saito, and T. Nakashizuka. 2017. “Abandoned Forest Ecosystem:Implications for Japan’s Oak Wilt Disease.” Journal of Forest Economics, Land Use, ForestPreservation and Biodiversity in Asia 29 (Dec): 56–61. doi:10.1016/j.jfe.2017.08.005.

Imanyfar, S., and M. Hasanlou. 2017. “Remote Sensing Analysis of the Extent and Severity of OakDecline in Malekshahi City, Ilam, Iran.” Engineering Journal of Geospatial Information Technology4 (4): 1–19. doi:10.29252/jgit.4.4.1.

Jenkerson, C. 2013. “User Guide: Earth Resources Observation and Science (EROS) Center ScienceProcessing Architecture (ESPA) on Demand Interface.” U.S. Geological Survey.

Jönsson, P., and L. Eklundh. 2004. “TIMESAT—A Program for Analyzing Time-Series of SatelliteSensor Data.” Computers & Geosciences 30 (8): 833–845. doi:10.1016/j.cageo.2004.05.006.

Karami, O., A. Fallah, S. Shataei, and H. Latifi. 2017. “Investigation on the Feasibility of Mapping ofOak Forest Dieback Severity Using Worldview-2 Satellite Data (Case Study: Ilam Forests).” IranianJournal of Forest and Poplar Research 3(Sept). doi:10.22092/ijfpr.2017.112879

Kazuhito, I., U. Masahito, K. Masayuki, S. Nobuko, K. Joon, A. M. Carmelita, A. Jonas et al. 2017. “NewData-Driven Estimation of Terrestrial CO2 Fluxes in Asia Using a Standardized Database of EddyCovariance Measurements, Remote Sensing Data, and Support Vector Regression.” Journal ofGeophysical Research: Biogeosciences 122 (4): 767–795.

Keenan, R. J., G. A. Reams, F. Achard, J. V. de Freitas, A. Grainger, and E. Lindquist. 2015. “Dynamicsof Global Forest Area: Results from the FAO Global Forest Resources Assessment 2015.” ForestEcology and Management 352(Sept): 9–20. Changes in Global Forest Resources from 1990 to2015. doi:10.1016/j.foreco.2015.06.014

Komura, R., N. Kamata, M. Kubo, and K.-I. Muramoto. 2005. “Identification of Dead Tree of JapaneseOak Wilt (JOW) Using High Spatial Resolution Satellite Imagery.”Paper presented at theGeoscience and Remote Sensing Symposium (IGARSS), 2005 IEEE International: 4351–4354.doi:10.1109/IGARSS.2005.1525882

Kozanitas, M., T. W. Osmundson, R. Linzer, and M. Garbelotto. 2017. “Interspecific Interactionsbetween the Sudden Oak Death Pathogen Phytophthora Ramorum and Two Sympatric

20 S. IMANYFAR ET AL.

Phytophthora Species in Varying Ecological Conditions.” Fungal Ecology 28(Aug): 86–96.doi:10.1016/j.funeco.2017.04.006.

La, H. P., Y. D. Eo, J. H. Kim, C. Kim, M. W. Pyeon, and H. S. Song. 2013. “Analysis of Correlationbetween Canopy Cover and Vegetation Indices.” International Journal of Digital ContentTechnology and Its Applications 7 (11): 10–17. doi:10.4156/jdcta.vol7.issue11.2.

Lambert, J., C. Drenou, J.-P. Denux, G. Balent, and V. Cheret. 2013. “Monitoring Forest Declinethrough Remote Sensing Time Series Analysis.” GIScience & Remote Sensing 50 (4): 437–457.doi:10.1080/15481603.2013.820070.

Larsson, H. 2002. “Acacia Canopy Cover Changes in Rawashda Forest Reserve, Kassala Province,Eastern Sudan, Using Linear Regression NDVI Models.” IEEE Journal of Selected Topics in AppliedEarth Observations and Remote Sensing 23 (2): 335–339. doi:10.1080/01431160010014279.

Lewińska, K., E. Ivits, M. Schardt, M. Zebisch, K. E. Lewińska, E. Ivits, M. Schardt, and M. Zebisch.2016. “Alpine Forest Drought Monitoring in South Tyrol: PCA Based Synergy between scPDSIData and MODIS Derived NDVI and NDII7 Time Series.” Remote Sensing 8 (8): 639. doi:10.3390/rs8080639.

Liang, L., D. Liping, L. Zhang, M. Deng, Z. Qin, S. Zhao, and H. Lin. 2015. “Estimation of Crop LAIUsing Hyperspectral Vegetation Indices and a Hybrid Inversion Method.” Remote Sensing ofEnvironment 165(Aug): 123–134. doi:10.1016/j.rse.2015.04.032.

Lu, L., L. Di, and Y. Ye. 2014. “A Decision-Tree Classifier for Extracting Transparent Plastic-MulchedLandcover from Landsat-5 TM Images.” International Journal of Remote Sensing 7 (11):4548–4558. doi:10.1109/JSTARS.2014.2327226.

Mahdavi, A., V. M. Zadeh, M. Niknezhad, and O. Karami. 2015. “Assessment and Prediction of OakTrees Decline Using Logistic Regression Model (Case Study: Bivareh Forest, Malekshahi-Ilam).”Forest and Range Protection Research 13 (1): 20–33.

Masek, J. G., E. F. Vermote, N. Saleous, R. Wolfe, F. G. Hall, K. F. Huemmrich, F. Gao, J. Kutler, andT. K. Lim. 2013. “LEDAPS Calibration, Reflectance, Atmospheric Correction Preprocessing Code,Version 2.” ORNL Distributed Active Archive Center. doi:10.3334/ORNLDAAC/1146.

Mattson, W. J., and R. A. Haack. 1987. “The Role of Drought in Outbreaks of Plant-Eating Insects.”BioScience 37 (2): 110–118. doi:10.2307/1310365.

McKee, T. B., N. J. Doesken, and J. Kleist. 1993. “The Relationship of Drought Frequency andDuration to Time Scales.” In 8th Conference on Applied Climatology, 17, Boston, 179–183.

Meng, J., L. Shiming, W. Wang, Q. Liu, S. Xie, M. Wu, J. Meng, et al. 2016. “Mapping Forest HealthUsing Spectral and Textural Information Extracted from SPOT-5 Satellite Images.” RemoteSensing 8 (9): 719. doi:10.3390/rs8090719.

Mirabolfathy, M. 2013. “Outbreak of Charcoal Disease on Quercus Spp and ZelkovaCarpinifolia Trees in Forests of Zagros and Alborz Mountains in Iran (Short Article).”Iranian Journal of Plant Pathology 49 (2194): 257–263.

Okujeni, A., S. van der Linden, S. Suess, and P. Hostert. 2017. “Ensemble Learning fromSynthetically Mixed Training Data for Quantifying Urban Land Cover with Support VectorRegression.” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing10 (4): 1640–1650. doi:10.1109/JSTARS.2016.2634859.

Pirouzi, F., and M. Tavakoli. 2015. “Investigating Effects of Dust as One of the Effective Factors onDecline of Oak Forest of Lorestan Province.” tehran.

Ren, H., and G. Feng. 2015. “Are Soil-Adjusted Vegetation Indices Better than Soil-UnadjustedVegetation Indices for Above-Ground Green Biomass Estimation in Arid and Semi-AridGrasslands?” Grass and Forage Science 70 (4): 611–619. doi:10.1111/gfs.12152.

Ren, H., G. Zhou, and F. Zhang. 2018. “Using Negative Soil Adjustment Factor in Soil-AdjustedVegetation Index (SAVI) for Aboveground Living Biomass Estimation in Arid Grasslands.” RemoteSensing of Environment 209(May): 439–445. doi:10.1016/j.rse.2018.02.068.

Rosentreter, J., R. Hagensieker, A. Okujeni, R. Roscher, P. D. Wagner, and B. Waske. 2017. “SubpixelMapping of Urban Areas Using EnMAP Data and Multioutput Support Vector Regression.” IEEEJournal of Selected Topics in Applied Earth Observations and Remote Sensing 10 (5): 1938–1948.doi:10.1109/JSTARS.2017.2652726.

INTERNATIONAL JOURNAL OF REMOTE SENSING 21

Rostamnia, M., and M. Akhoondzadeh Hanzaei. 2016. “Assessment of Hazardous Drought of IlamProvince Forests Using Landsat Satellite Images.” Journal of Geomatics Science and Technology 6(2): 131–144.

Rouse, J. W. 1974. Monitoring the Vernal Advancement and Retrogradation (Green Wave Effect) ofNatural Vegetation. Texas A & M University, Remote Sensing Center.

Salehi, A., E. Wilhelmsson, and U. Söderberg. 2008. “Land Cover Changes in a Forested Watershed,Southern Zagros, Iran.” Land Degradation & Development 19 (5): 542–553. doi:10.1002/ldr.860.

Theys, C., N. Dobigeon, J. Y. Tourneret, and H. Lanteri. 2009. “Linear Unmixing of HyperspectralImages Using a Scaled Gradient Method.” IEEE International Geoscience & Remote SensingSymposium, IGARSS 2008, July 8-11, 2008, Boston, Massachusetts, USA, Proceedings. IEEE 2008,ISBN 978-1-4244-2808-3.

Uto, K., Y. Takabayashi, Y. Kosugi, and T. Ogata. 2008. “Hyperspectral Analysis of Japanese Oak Wiltto Determine Normalized Wilt Index.” II-295-II-298. IEEE. doi:10.1109/IGARSS.2008.4778986.

Valipour, A., T. Plieninger, Z. Shakeri, H. Ghazanfari, M. Namiranian, and M. J. Lexer. 2014. “TraditionalSilvopastoral Management and Its Effects on Forest Stand Structure in Northern Zagros, Iran.” ForestEcology and Management 327(Sept): 221–230. doi:10.1016/j.foreco.2014.05.004.

Vannini, A., R. Paganini, and N. Anselmi. 1996. “Factors Affecting Discharge and Germination ofAscospores of Hypoxylon Mediterraneum (De Not.) Mill.” European Journal of Forest Pathology26 (1): 12–24. doi:10.1111/j.1439-0329.1996.tb00706.x.

Vannini, A., and R. Valentini. 1994. “Influence of Water Relations on Quercus Cerris-HypoxylonMediterraneum Interaction: A Model of Drought-Induced Susceptibility to A Weakness Parasite.”Tree Physiology 14 (2): 129–139. doi:10.1093/treephys/14.2.129.

Vermote, E., C. Justice, M. Claverie, and B. Franch. 2016. “Preliminary Analysis of the Performance ofthe Landsat 8/OLI Land Surface Reflectance Product.” Remote Sensing of Environment 185(Nov):46–56. Landsat 8 Science Results. doi:10.1016/j.rse.2016.04.008

Wang, C., H. S. He, and J. M. Kabrick. 2008. “A Remote Sensing-Assisted Risk Rating Study to PredictOak Decline and Recovery in the Missouri Ozark Highlands, USA.” GIScience & Remote Sensing 45(4): 406–425. doi:10.2747/1548-1603.45.4.406.

Wang, C., L. Zhenqian, and T. L. Haithcoat. 2007. “Using Landsat Images to Detect Oak Decline inthe Mark Twain National Forest, Ozark Highlands.” Forest Ecology and Management 240 (1):70–78. doi:10.1016/j.foreco.2006.12.007.

Wang, J., T. W. Sammis, V. P. Gutschick, M. Gebremichael, S. O. Dennis, and R. E. Harrison. 2010.“Review of Satellite Remote Sensing Use in Forest Health Studies.” The Open Geography Journal3 (1): 28–42. doi:10.2174/1874923201003010028.

Weissling, B., H. Xie, and P. Jurena. 2005. “Early Detection of Oak Wilt Disease in Quercus Ssp .:A Hyperspectral Approach.” Sioux Falls,south dakota.

Yue, J., H. Feng, G. Yang, and L. Zhenhai. 2018. “A Comparison of Regression Techniques forEstimation of Above-Ground Winter Wheat Biomass Using Near-Surface Spectroscopy.” RemoteSensing 10 (1): 66. doi:10.3390/rs10010066.

Zakeri, F., S. M. Hojjati, and H. Kiadaliri. 2013. “Analysis of the Trend of Zagros Forests Dieback andDecline.” Tehran.

Zhang, X., M. A. Friedl, C. B. Schaaf, A. H. Strahler, C. F. John, F. G. Hodges, B. C. Reed, and A. Huete.2003. “Monitoring Vegetation Phenology Using MODIS.” Remote Sensing of Environment 84 (3):471–475. doi:10.1016/S0034-4257(02)00135-9.

22 S. IMANYFAR ET AL.