Chapter 4. Methodology and Modeling for Carbon Sequestration Pattern in Cashew Plantation

Transcription

1 Chapter 4 Methodology and Modeling for Carbon Sequestration Pattern in Cashew Plantation In this chapter, a new approach for carbon sequestration pattern of broad leaf vegetation is proposed using modified vegetation photosynthesis model (MVPM). This model utilises the features extracted from VGT and MODIS imagery data. The vegetation indices namely EVI, LSWI, TRVI along with climate data are used to estimate the carbon sequestration potential of cashew plantation. Section 4.2 briefly describes the MVPM. Section 4.3 explains the study area details and Section 4.4 describes the remote sensing data used for studying the carbon sequestration pattern of cashew plantation. Section 4.5 presents the experimental results. Section 4.6 deals with the classification of carbon sequestration pattern. 4.1 Introduction In tropical countries like India, forest carbon sinks are believed to offset a significant portion of carbon emission associated with fossil fuel combustion. But due to large scale industrialization and increased population, the forest area is slowly declining. Perennial fruit trees like cashew, mango and guava have similar potential like forest trees to sink atmospheric carbon. Cashew is an evergreen fruit tree; it occupies nearly 40,000 ha in Tamilnadu and has great potential for carbon sequestration. In India, 48

2 cashew is grown seasonally in moist tropical climate having a distinct dry and wet season. There is a strong seasonality of photo synthetically active radiation (PAR) usually being much larger in the summer season than in the winter season. The seasonally moist tropical fruit trees may have evolved two adaptive mechanisms to maximize carbon uptake in an environment with large seasonal variation of light and water. One adaptive mechanisms is that the perennial fruit trees have deep roots (1.0 m and deeper) for getting access to water in deep soil during dry season. In cashew plantation, the dry season evaporation is (4.2 mm/day), while the wet season evaporation is (3.4 mm/day). In cashew plantations the new leaf formation started during the end of the dry season is during month of September which coincides with the start of North east monsoon rain. In this study, we combined the analysis of satellite images with meteorological data and the objectives of this study are to estimate the seasonal dynamics of carbon sequestration in the tropical cashew plantation area of Cuddalore and Perambalur district of Tamilnadu, India using modified vegetation photosynthetic model (MVPM) and to explore the capability of SVM, RBFNN and AANN to classify the pattern of carbon sequestration in the cashew plantation on the basis of quantity of carbon sequestered in different periods. The MVPM model takes advantages of additional spectral bands (e.g. blue, and short wave infrared (SWIR)) that are available from advanced optical sensors - VEGETATION (VGT) sensor on board the SPOT-4 satellite, and the Moderate Resolution Imaging Spectroradiometer (MODIS) on board the terra and aqua satellites. The input data to the MVPM model are the enhanced vegetation index (EVI) [45], the land surface water index (LSWI) [12], total ratio vegetation index (TRVI), 49

3 air temperature (T), and photo synthetically active radiation (PAR). EV I = G ρ nir ρ red ρ nir + (C 1 ρ red C 2 ρ blue ) + L (4.1) where ρ blue, ρ red, ρ nir are surface reflectance of blue, red and near infrared regions respectively, G is the gain factor, L is the canopy background adjustment factor that addresses nonlinear, differential NIR and red radiant transfer through a canopy; and C 1 and C 2 are the coefficients of the aerosol resistance term, which uses the blue band to correct the aerosol influences in the red band. In the EVI algorithm, L=1, C 1 =6, C 2 = 7.5, and G = 2.5. EVI includes the blue band for atmospheric correction, which is one important feature for the study in the forest area where seasonal burning of pasture and forest takes place throughout the dry season, either for agricultural purpose (land clearing) or for natural fire events. The advanced optical sensors (VGT and MODIS) have additional spectral bands (e.g. blue and shortwave infrared), making it possible to develop time-series data of improved vegetation indices. As the short wave infrared (SWIR) spectral band is sensitive to vegetation water content and soil moisture, a combination of NIR and SWIR bands has been used to derive water sensitive vegetation indices [60] including the land surface water index (LSWI) [139]. LSWI = ρ nir ρ swir ρ nir + ρ swir (4.2) As leaf liquid water content increases or soil moisture increases, SWIR absorption increases and SWIR reflectance decreases, resulting in an increase of LSWI value. Recent work in evergreen tropical forests have shown that LSWI is sensitive to changes in leaf water content over time [12]. In arid and semi-arid regions, soil background has more reflectance in the near infrared (NIR) and red (RED) wavelengths of vegetation. Vegetation cover is usually 50

4 sparse compared to soil background, and soil and plant spectral signatures tend to mix non-linearly. Thus, arid plants tend to lack the strong red edge found in plants of humid regions due to ecological adaptations to the harsh desert environment. In order to nullify the effect of soil reflectance on vegetation studies, TRVI is introduced as a new parameter. The total ratio vegetation index (TRVI) is the ratio of NIR and the sum of visible and NIR wavelengths and is calculated using the following equation [ ] NIR RED TRV I = 4 NIR + RED + GR + B (4.3) where RED and NIR stand for spectral measurement acquired in the red and near infrared regions, respectively. The present study area lies in semiarid environment, and hence, it is decided to introduce the new vegetation index based on total wavelength (visible and NIR) in the GPP calculation. 4.2 Brief description of the modified vegetation photosynthesis model (MVPM) Overview of the MVPM: Leaves and canopy are composed of photosynthetically active vegetation (PAV; chloroplasts) and non-photosynthetic vegetation (NPV; e.g. stem, branch, cell wall, vein). Based on the conceptual partitioning of PAV and NPV, the VPM model was recently developed to estimate GPP of forests [12]. Here we have modified the VPM model by introducing another vegetation index called the total ratio vegetation index (TRVI), which is specifically developed for arid and semiarid region vegetation studies. The brief description of the MVPM model is GPP = α ε g FAPAR PAV PAR TRV I (4.4) 51

5 ε g = ε 0 T scalar W scalar P scalar (4.5) where α is a scalar, in this study it is set to be 5. PAR is the photosynthetically active radiation (µ mol/m 2 /s, photosynthetic photon flux density, PPFD), FAPAR PAV is the fraction of PAR absorbed by PAV (chloroplasts), and ε g is the light use efficiency (µ mol CO 2 /µ mol PAR). The parameter ε 0 is the apparent quantum yield or maximum light use efficiency (µ mol CO 2 /µ mol PAR), and T scalar, W scalar and P scalar are the down-regulation scalars for the effects of temperature, water and leaf phenology on the light use efficiency of vegetation, respectively. In MVPM model, FAPAR PAV is assumed to be a linear function of EVI, and the coefficient a in (4.6) is simply set to be 1.0 [12]. FAPAR PAV = a EV I (4.6) T scalar is estimated at each time step, using the equation developed for the Terrestrial Ecosystem Model [140]. T scalar = (T T min )(T T max ) [(T T min )(T T max )] (T T opt ) 2 (4.7) where T min, T max and T opt are minimum, maximum and optimal temperature for photosynthetic activities, respectively. If air temperature falls below T min, T scalar is set to be zero. W scalar, the effect of water on plant photosynthesis, has been estimated as a function of soil moisture and/or vapor pressure deficit (VPD) in a number of Production Efficiency Models [141]. W scalar = 1 + LSWI 1 + LSWI max (4.8) where LSWI max is the maximum LSWI within the plant growing season for individual pixels. P scalar is included to account for the effect of leaf phenology (leaf age) on photosynthesis at the canopy level. In MVPM model, calculation of P scalar is dependent 52

6 upon the longevity of leaves. For a canopy that is dominated by leaves with a life expectancy of 1 year, P scalar is calculated at two different phases as a linear function [12]: P scalar = 1 + LSWI 2 (4.9) During bud burst to leaf full expansion P scalar = 1 After leaf full expansion (4.10) LSWI values range from -1 to +1, and the simplest formulation of P scalar is a linear scalar with a value range of 0 to 1. Evergreen broad leaf vegetation in the tropical zone has a green canopy throughout the year, because foliage is retained for several growing seasons. Canopies of evergreen broad leaf vegetations are thus composed of green leaves of various ages. In this version of the MVPM model, a simple assumption of P scalar =1 is made for evergreen broad leaf vegetation [82]. Parameter estimation of the MVPM: The MVPM model has three sets of parameters to be estimated. In this study, α is the scalar and is set to be 5. We used an ε 0 value of µ mol CO 2 /µ mol PAR, by following [82], derived from the time-series data of NEE and incident PAR at the CO 2 flux tower sites [142]. The second parameter set is for calculation of T scalar (4.7). For tropical vegetation, we used a minimum temperature (T min ) of 18 o C, optimum temperature (T opt ) of 28 o C, and maximum temperature (T max ) of 48 o C, as implemented in the process-based Terrestrial Ecosystem Model [143]. The third parameter set is for calculation of W scalar. Estimation of site specific LSWI max is dependent upon the optical sensor and the time series of image data. The maximum LSWI value within the plant-growing season is selected as an estimate of LSWI max [12]. 53

7 Fig. 4.1: Map of India and Tamilnadu. 4.3 Study area This research was conducted in cashew plantation area of Cuddalore and Perambalur district of Tamilnadu, India. This area lies between 11 o.15 N to 11 o.43 N latitude and 79 o.16 E to 79 o.44 E longitude and covers a total area of 40,000 ha. The annual rainfall ranges from 750 mm to 1100 mm with an average of 900 mm. October to December is the peak of wet season with mean monthly rainfall of 300 mm. The study area has two rainfall regimes that are wet and dry seasons. The wet season stretches from October to January and the dry season ranges from February to September. The average mean temperature is 32 o C. The mean minimum temperature is 24.2 o C. Mean maximum temperature is 38 o C. The mean annual solar net radiation is 16.9 MJ m 2 day 1, whereas the relative air humidity ranges from 80 to 90% with an average of 85%. The soil in 54

8 Fig. 4.2: Cashew and casuarina plantation areas of Cuddalore and Perambalore district of Tamilnadu, India. the study area is characterized by neutral, deep, strongly weathered, and moderately well drained. Inceptisols and Alfisol are the dominant soil types in the study area. The elevation of the study area ranges from 48 to 55 MSL. Field measurement is conducted in order to collect data such as leaf area index, leaf moisture content. Photosynthetic photon flux density is measured using portable photosynthetic system Li Temporal variability analysis is conducted to assess the temporal dynamics of carbon sequestration in the plantation area. 4.4 Remotely sensed data 10-day composite images from the VEGETATION sensor: The VEGETATION (VGT) sensor on board the SPOT-4 satellite is one of a new 55

9 generation of space-borne optical sensors designed for the observation of vegetation and land surfaces. The VGT instrument has four spectral bands: blue ( nm), red ( nm), near-infrared (NIR, nm), and shortwave infrared (SWIR, nm). The blue band is primarily used for atmospheric correction. The SWIR band is sensitive to soil moisture, vegetation cover, and leaf moisture content. Unlike scanner sensors (e.g. AVHRR, MODIS), the VGT instrument uses the lineararray technology (push-broom), and thus produces high-quality images at moderate resolution (1-km) with greatly reduced distortion. Since its launch in March 1998, the VGT instrument has acquired daily images at 1-km spatial resolution for the globe. There are three 10-day composites for 1 month: day 1-10, day 11-20, and day 21 to the last day of the months. The VGT-S10 products are freely available to the public ( The VGT-S10 data was acquired over the period of January 2006 to December 2009 for the globe. The EVI, LSWI and TRVI were calculated using the surface reflectance of blue (ρ blue ), red (ρ red ), NIR (ρ nir ), and SWIR (ρ swir ) bands from the standard VGT- S10 data. Further 1km 1km pixels were selected (approximately 1 1 km 2 ) that covered Vridhachalam meteorological station for feature extraction, and calculated the GPP. 8-day composite images from MODIS sensor: Of the 36 spectral bands in the MODIS sensor, seven spectral bands are primarily designed for the study of vegetation and land surfaces: blue ( nm), green ( nm), red ( nm), NIR ( nm, nm), and SWIR ( nm, nm). The MODIS sensor acquires daily images of the globe at a spatial resolution of 250 m for red and NIR ( nm) bands, and at a spatial resolution of 500 m for blue, green, NIR ( nm), and SWIR bands. The MODIS Land 56

10 Science Team provides a suite of standard data products to the users ( including the 8-day surface reflectance product (MOD09A1) that has the above seven spectral bands at 500 m spatial resolution. The MOD09A1 dataset is provided to users in a tile fashion; each tile covers 108 o latitude by 108 o longitude. The downloaded the 8-day surface reflectance product (MOD09A1) was downloaded for the period of 1/ /2009 from the EROS Data Center, U.S. Geological Survey ( Surface reflectance values from these four spectral bands (blue, red, NIR ( nm) and SWIR ( nm)) were used to calculate the vegetation indices (EVI, LSWI and TRVI). Data from the MOD09A1 product were extracted from 3 3 MODIS pixels ( 1.5km 1.5km). Simulations of the MVPM model are driven by MODIS images in , temporally consistent with the available field data in Experimental Results for Carbon Sequestration pattern in Cashew plantation Biophysical performance of vegetation indices from VGT images The EVI value calculated from the VGT imagery (S10 data) shows that the highest EVI value (0.62) was recorded in the month of January, the value of EVI starts decreasing from March onwards due to decrease in soil moisture as well as increased air temperature as shown in Fig. 4.3(a). The lowest EVI (0.36)was noticed in the month of June. The EVI started increasing from the month of July onwards which coincides with the onset of Southwest monsoon rain. Similar trend is observed with LAI. The seasonal dynamics of EVI is likely driven by a change in the leaf area index as the canopy of seasonally moist tropical cashew plantation has varying LAI as shown in Fig. 4.3(b) over seasons. We hypothesize that the seasonal distribution of EVI in a 57

11 (a) (b) Fig. 4.3: Seasonal Dynamics of (a) Enahanced vegetation index (EVI). (b) Leaf area index (LAI) using VGT. year may be attributed to both fall of old leaves and emergence of new leaves resulting in dynamic changes in proportion of young and old leaves within a vegetation canopy over seasons. In general, the old leaves have less chlorophyll and water content, but have more structured material (e.g. lignin, cellulose) in comparison to young leaves which could lead to significant changes in absorbance, transmittance and reflectance 58

12 of leaves as the aging process of leaves progress. The EVI continued to maintain higher value even up to February which may be attributed to continued emergence of new flushes in the winter season. The peak EVI values had the time lag of 2 months (January-February). The observed decrease of EVI in the peak dry season (May-June) could be largely attributed to the aging process of leaves, including increase of both leaf thickness and the non-photosynthetic vegetation (NPV) component. This fact is supported by the field data. Field data collected at the experimental site showed that new leaf flush emergence starts during the end of September and continued up to January. The decrease in EVI from April to September may be attributed to both leaf age (older leaves) and epiphyll cover [144]. Young leaves have higher photosynthetic capacity than older leaves [145]. Time series data ( ) of LSWI was used to assess the status of leaf canopy water content of seasonally moist tropical cashew plantation. LSWI values were generally higher in the wet season than in the dry season as shown in Fig. 4.4(a). The seasonal dynamics of LSWI from was positively correlated with that of leaf moisture content. The observed evapotranspiration data from the meteorological research station were also higher in dry season than in the wet season as shown in Fig. 4.5(a). The high LSWI values in the month of December-February might be due to high proportion of young leaves and more leaf water content as indicated by leaf moisture data and seasonal dynamics of EVI. Usually young leaves have more water content than old leaves [146]. The seasonal dynamics of LSWI shows that water stress exists in the experimental area during the dry season from The TRVI time series data ( ) had distinct seasonal dynamics as that of EVI within the plant growing season as shown in Fig. 4.5(b). The maximum TRVI was 59

13 (a) (b) Fig. 4.4: Seasonal Dynamics of (a) Land surface water index (LSWI). (b) Leaf moisture content (LMC) using VGT. observed in the late winter season (January to March), and then declined gradually. The observed decrease of TRVI after reaching its peak in May might be caused by moisture stress in plants which reflect more red light and less NIR light. The TRVI starts increasing from July to December. The onset of South west monsoon in July might have increased leaf moisture content which resulted in increased TRVI from July 60

14 to December. (a) (b) Fig. 4.5: Seasonal Dynamics of (a) Evaoptranspiration (ET). (b) Total Ratio Vegetation Index (TRVI) using VGT Calculation of GPP by MVPM using 10 day VGT composites We used the MVPM [12] to estimate the GPP using LSWI, EVI, TRVI and site specific climate (air temperature). The MVPM model predicts high GPP in the late 61

15 wet season, as compared to dry season. For instance, the monthly GPP pred was g C/m 2 in January 2009 (wet season) but g C/m 2 in June 2009 (dry season). Fig. 4.6 shows that the relatively low GPP pred in the dry season can be attributed to a number of factors. First the moisture stress in soil and maturity of leaves, secondly the averaged EVI value was lower (0.42) in late dry season (July-September) than in wet season (0.55). EVI seasonal dynamics is related to leaf phenology at the canopy level. This suggests that leaf phenology could play an important role in the GPP calculation of seasonally moist tropical cashew plantation. The annual sum of predicted GPP in cashew plantation is ( g C/m 2 /year). Fig. 4.6: Seasonal Dynamics of Gross Primary Production using VGT for cashew plantation Seasonal dynamics of vegetation indices from MODIS sensor The EVI calculated using MODIS data (MOD09A1) clearly depicts that the EVI increased from January to March and reached the peak in the month of March which coincides with the maximum LAI (4.2) as observed in Fig. 4.7(a). Declining trend in 62

16 the EVI was noticed from April to June and thereafter increased due to the increase in soil moisture content by south west monsoon rain. Again the EVI decreased from November due to low PAR [144]. We hypothesize that the seasonal distribution of variation in EVI in a year may be attributed to formation of new leaves and twigs, and fall of senesence leaves, which resulted in changes of EVI over a period as shown in Fig. 4.7(a). It is more vivid that younger leaves are capable of accumulating more of carbon through photosynthesis than the older leaves [145]. Moreover, there is a positive correlation between LAI and photosynthetic activity. The sufficient amount of soil moisture and favourable temperature helps to increase the EVI from January to March. Decline in the EVI value from April to May might be largely attributed to the aging process of photosynthetically active portion of the plant, as well as decrease in leaf moisture content and increase in leaf thickness. The LSWI was calculated using the time series data ( ) of MODIS to assess the leaf moisture content of seasonally moist tropical cashew plantation. In general, the low LSWI was observed during summer and high LSWI value was noticed in winter as shown in Fig. 4.8(a). Although there is a decline in LSWI from January to June, a steep decline was noticed from April onwards. In the present study area, the soil moisture is sufficient to supply plant water need upto March, so that there is no much variation in leaf moisture content and hence LSWI is not much varied from January to March. The decrease in LSWI from April to June may be attributed to increased air temperature and decreased leaf moisture content. As the south west monsoon starts in July, the leaf moisture content as well as young leaves formation starts from July onwards due to increased soil moisture status. The maximum rainfall in the study area occurs during North east monsoon (October- 63

17 (a) (b) Fig. 4.7: Seasonal Dynamics of (a) Enahanced vegetation index (EVI). (b) Leaf area index (LAI) using MODIS. December) and hence LSWI increases steeply from October to December. The results of the study show the presence of water stress in the study area during summer from in Fig The seasonal variation in TRVI calculated using MODIS data is depicted in Fig. 4.8(b). The TRVI continued to maintain higher value till March, beyond that there was a 64

18 gradual decline up to June. TRVI increases again from July onwards up to May. The maximum TRVI during January to March might be due to the presence of sufficient water content in leaves and increased NIR reflectance by vegetation. As the soil moisture starts decreasing from April, the plants are subjected to moisture stress and hence the TRVI decreases up to June till the on set of south west monsoon rain. (a) (b) Fig. 4.8: Seasonal Dynamics of (a) Land surface water index (LSWI). (b) Total ratio vegetation index (TRVI) using MODIS. 65

19 4.5.4 Calculation of GPP by MVPM using 8-day MODIS composite images We used the MVPM [12] to estimate the GPP using LSWI, EVI, TRVI and site specific climate (air temperature). The predicted MODIS GPP decreases from March to June due to decrease in soil moisture and increased temperature. The GPP increased from July onwards due to the onset of southwest monsoon rain as shown in Fig Secondly the averaged EVI and TRVI value was lower (0.3 and 0.19) in late dry summer season than in winter season (0.56 and 0.35). As EVI and TRVI seasonal dynamics is related to leaf age and leaf moisture content at the canopy level, this suggests that leaf age and leaf moisture content could play an important role in the GPP calculation of seasonally moist tropical cashew plantation. The decrease in TRVI value in the summer season might have also contributed to low GPP in summer season. The annual sum of predicted GPP in cashew plantation was g C/m 2 /year. The seasonal dynamics of predicted GPP agreed seasonably well with observed evapotranspiration. As photosynthesis is closely coupled with water flux, we used observed water flux evapotranspiration from the nearest meteorological station in the experimental site to evaluate the performance of the MVPM model as shown in Fig Comparison of MVPM predicted GPP using VGT and MODIS with MODIS-GPP The seasonal dynamics of predicted GPP both from VGT and MODIS imagery derived data were compared with the MODIS GPP data derived from MOD-17A2 data (Fig. 4.10). The seasonal dynamics of predicted GPP over the plant growing season in , agreed reasonably well with those of MODIS GPP. The simple linear regression model also shows a good agreement between GPP predicted and GPP ob- 66

20 (a) (b) Fig. 4.9: Seasonal dynamics of (a) Evaprotranspiration (ET). (b) Gross Primary Production (GPP) using MODIS for cashew plantation. served in the cashew plantation area during When the VGT and MODIS predicted GPP are compared (Fig. 4.10), the MVPM GPP predicted using MODIS data was in best agreement (R 2 =0.95) with MODIS GPP (MOD 17A2), and hence, the predicted MVPM GPP data are taken for carbon sequestration pattern classification using SVM, RBFNN and AANN. 67

21 (a) (b) Fig. 4.10: Comparison of (a) VGT-MVPM GPP with MODIS GPP. (b) MODIS-MVPM GPP with MODIS GPP. 4.6 Carbon sequestration pattern classification of cashew plantation Classification of cashew plantation using SVM for MODIS data The 11 features (light use efficiency (ε), leaf phenology (P scalar ),land surface water index (LSWI), enhanced vegetation index (EVI), air temperature (T), minimum tem- 68

22 perature (T min ), maximum temperature (T max ), optimum temperature (T opt ), total ratio vegetation vegetation index (TRVI), global primary production (GPP) and scale factor (a)) derived from remote sensing image at the rate of four per month from 2006 to 2009 and climate data were used to train the model. Training data include the class attribute so a total of (11 features + 1 class attribute) 12 attributes were fed to the classifier while the test data had only 11 attributes excluding the class attribute. SVMTorch, a freely available C++ based object-oriented machine learning was used for training and testing the model [147]. The calculated GPP of different months in a year showed that the monthly GPP ranged from g C/m 2. The SVM is trained in a such a way to provide a value of 0, 1 and 2 classes for the periods carbon sequestration which ranged from g C/m 2, g C/m 2 and g C/m 2 respectively. In order to compare the performance of the four different kernels viz., linear, polynomial, Gaussian, sigmoid, each kernel is used in SVM and is trained with 184 samples and tested with 36 samples. Among the four kernels, Gaussian kernel showed highest accuracy as compared to other kernels as shown in Fig Table 4.1 shows the overall accuracy of SVM in terms of confusion matrix for Gaussian kernel. The data from three different classes were applied to calculate the accuracy of SVM. The accuracy obtained with linear, polynomial, Gaussian and Sigmoidal kernels were 45%, 73%, 98% and 78% respectively. The performance of SVM for carbon sequestration classification is 98% Classification of cashew plantation using RBFNN for MODIS data Eleven features were extracted using MODIS imagery and climate data at the maximum rate of four per month from So altogether 184 samples were obtained. 69

23 Fig. 4.11: Classification performance using in different kernels in SVM for cashew plantation (MODIS). Table 4.1: Confusion matrix for SVM with Gaussian kernel based classification for cashew(in %)(MODIS) Classes For RBFNN training, 164 samples were taken out of 184, each with 11 features are given as input to the RBFNN model. The RBF centers are located using k-means algorithm. The weights are determined using least squares algorithm. The value of k = 2, 3, 4, 5 and 6 has been used in our studies. The system gives optimal performance for k = 6. For training, the weight matrix is calculated using the least squares algorithm. For classification, the feature vectors are extracted and each of the feature vectors is given as input to the RBFNN model. The average output is calculated for 70

24 each of the output neuron. The class to which the each sample belongs is decided based on the highest output. The performance measures calculated for the three different classes such as class 0, class 1 and class 2 are shown in Table 4.2, Table 4.3 and Table 4.4. The classwise performance of RBFNN for carbon sequestration of cashew plantation using satellite imagery derived vegetation indices (MODIS) are expressed as sensitivity, specificity, precision, F-Score, and accuracy are given in Table 4.2 and Fig. 4.12(c). Sensitivity: The highest sensitivity of 82.35% in 0 class was noticed in cluster 3, where as in class 1 and 2 cluster 6 recorded the highest sensitivity of 95.83% and 92.59% respectively. Cluster 5 recorded the next highest sensitivity of 92.31% and 88.89% in class 1 and 2, where as in class 0 the cluster 5 recorded the lowest sensitivity value of 69.23%. Sepsificity: Specificity is also differ with cluster in all the three classes. As that of sencitivity, cluster 6 recorded the highest specificity of 96.67%, 97.78% in class 0 and 1, where as in class 2 cluster 5 recorded the highest sensitivity value of 95.45%. The lowest specificity of 84.62% was noticed in cluster 2 in class 1. The cluster 2 in class 1 and 2 recorded the lowest specificity value of 78.95% and 82.93% respectively. Precision: The precision value ranged from 77.78% to 91.67% in class 0, 69.23% to 97.80% in class 1 and 70.83% 92.59% in class 2. The lowest precision value of 77.78% was noticed in cluster 2 in class 0. In class 2 and 3, the lowest precision value of 69.23% and 70.83% respectively was observed in cluster in 1. The highest precision value of 91.67% in class 0, 97.80% in class 1 and 92.59% in class 2 was recorded in cluster 6. F-Score: The F-Score ranged from 75.00% to 84.62% in class 0, 63.16% to 97.87% in class 1 and 62.96% to 92.59%. The hand in F-Score is very similar to precision. 71

25 The cluster 6 recorded the largest F-Score of 84.62% in class 0, 97.87% in class 1 and 92.59% in class 2. Accuracy: Classwise accuracy shows that cluster 2 is registered the lowest accuracy of 79.55%, 69.57% and 71.83% in class 0, class 1, and class 2 respectively. Increase in the cluster from 2 to 6 increased the accuracy of RBFNN. The highest accuracy of 90.91%, and 94.37% in class 0, class 1, and class 2 repectively. The cluster performance of RBFNN for carbon sequestration pattern classification of cashew plantation using satellite imagery derived vegetation indices are given in Table 4.3 and Fig. 4.12(b). The results of the analysis clearly depicted that increase in the number of clusters from 2 to 6 increased the sensitivity, specificity, precision and F-Score. The sencitivity ranged from 61.45% to 90.77%, specificity ranged from 82.18% to 97.48%, precision ranged from 73.91% to 95.16% and F-Score ranged from 67.11% to 92.91%. The RBFNN with 6 clusters recorded the highest value in all the performance parameters studied. The clusterwise average class performance of RBFNN for carbon sequestration pattern classification of cashew plantation using satellite imagery derived vegetation indices (MODIS) are given in Table 4.4 and Fig. 4.12(a). As stated in the previous table, increase in the number of clusters increased the sensitivity, specificity, precision and Average accuracy. The lowest sensitivity, specificity, precision and average accuracy of 62.0%, 83.0%, 75.0% and 74% was noticed with cluster 2. However, this value increased progressively with increased number of clusters. The hights sncitivity (89%), specificity (97%), precision (95%) and accuracy (95%) was registered with 6 clusters RBFNN. In Fig shows that the performance of RBFNN for carbon sequestration classification is 95.2%. 72

26 Table 4.2: Class-wise performance of RBFNN for carbon sequestration pattern classification of cashew plantation using satellite imagery derived vegetation indices (MODIS). Class Clusters Sensitivity (%) Specificity (%) Precision (%) F-Score (%) Accuracy (%)

27 Table 4.3: Performance of RBFNN for carbon sequestration pattern classification of cashew plantation using satellite imagery derived vegetation indices (MODIS). Clusters Sensitivity (%) Specificity (%) Precision (%) F-Score (%) Table 4.4: Average class performance of RBFNN for carbon sequestration pattern classification of cashew plantation using satellite imagery derived vegetation indices (MODIS). Clusters Sensitivity (%) Specificity (%) Precision (%) Average Accuracy (%) Classification of cashew plantation using AANN for MODIS data The AANN was trained with 11 input features as described in Section Based on the overall GPP range, the GPP is classified into three classes as class 0 (30-75 g 74

28 (a) (b) Fig. 4.12: (a)average performance of RBFNN with different means for each class. (b) Performance of RBFNN for different means for all class. (c) Over all Performance of RBFNN for each mean in cashew plantation. C/m 2 ), class 1 ( g C/m 2 ) and class 2 ( g C/m 2 ). The AANN model is trained for 100, 500 and 1000 epochs for each class. Then the model was tested using the set of test data and then the input feature vector is compared with the output to compute the normalized squared error e k. The normalized squared error (e k ) for the (c) 75

29 feature vector x is given by e k = x o 2 x 2 (4.11) where o is the output vector given by the model. The error e k is transformed into a confidence score s using s = exp( e k ) (4.12) The data are classified based on highest confidence score. In order to test the performance of the AANN model, a set of test data is considered initially. A total data set of 184 samples is used in our studies. This includes 44 for class 0, 69 for class 1 and 71 for class 2. The eleven feature vectors of three classes (0, 1, 2) are given as input to the each AANN training and the network is trained for 100, 500 and 1000 epochs. One epoch of training is a single presentation of all the training vectors to the network. The performance of the AANNs increased with an increase in the epochs up to This is illustrated in Fig The result shows that there is no significant change in the performance of the AANN, beyond 1000 epochs. The AANN gives better performance for 1000 epochs and the final overall performance is calculated by taking the average of all three class runs with respect to 1000 epochs and is shown in Fig The performance measures calculated for the three different classes (class 0, class 1 and class 2) are shown in Table 4.5, Table 4.6 and Table 4.7 and Fig The classwise performance of AANN of carbon sequestrationpattern classification of cashew plantation using satellite imagery derived vegetation indices (MODIS) are given Table 4.5 and Fig. 4.14(b). In all the three classes tried, increase in the number of epochs from 100 to 1000 considerably increased the sensitivity, specificity, precision, F-Score and accuracy. In class 0, an increase in epochs from 100 to 1000 increased sensitivity from 85.71% 76

30 to 97.56%, specificity from 77.89% to 82.34%, precision from 83.33% to 95.24%, F-Score from 84.51% to 96.39% and accuracy from 80% to 93.18%. Similarly, class 1 and 2 also recorded the highest sensitivity value of 96.77% and at 1000 epochs respectively. Increase in epochs from 100 to 1000 in both class 1 and 2 recorded the increased specificity from 76.45% to 87.48% in class1 and 80.00% to 97.04% in class 2. Precision value also favorably increased in both class 1 and 2 as that of class 0. Increase in epochs from 100 to 1000 increased the precision value from 77.78% to 95.24% in class 1 and 83.33% to 97.10% in class 2. F-Score recorded for 100, 500 and 1000 epochs are 80.00%, 87.18% and 96.00% in class 1 and 86.21%, 93.85% and 97.81% in class 2. Similarly accuracy also increased with increased number of epochs in both class 1 and class 2. The results revealed that 1000 epochs recorded highest value which is close to 100 and hence 1000 epochs was identified as optimum number of epochs for AANN in carbon sequestration on pattern studies. Performance of AANN for carbon sequestration pattern classification of cashew plantation using satellite imagery derived vegetation indices are given in Table 4.6 and Fig. 4.14(a). The results of performance of AANN for different epochs for all classes showed that increase in number of epochs from 100 to 1000 increased the sensitivity, specificity, precision and F-Score values. Use of 100, 500 and 1000 epochs recorded the sensitivity of 85.92%, 92.59% and 97.66%, specificity of 82.47%, 91.78% and 97%, precision of 81.33%, 90.36% and 95.98% and F-Score of 83.56%, 91.46%, and 96.81% respectively. Average class performance of AANN for carbon sequestration pattern classification of cashew plantation using satellite imagery derived vegetation indices (MODIS) are 77

31 (a) (b) Fig. 4.13: (a)effect of epochs on the confidence score. (b) Average performance of AANN for different means for each class. in cashew plantation using MODIS. given in Table 4.7 and Fig. 4.13(b). The results of the average class performance of AANN clearly reveals that decrease in epoch number from 1000 to 500 or 100 considerably reduced the sensitivity, specificity, precision and average accuracy in percentage. Sensitivity, specificity, precision and average accuracy of 97.62%, 95.11% 95.86% and 93.9% was recorded which using 1000 epochs and the lowest sensitivity, specificity, 78

32 Table 4.5: Class-wise performance of AANN for carbon sequestration pattern classification of cashew plantation using satellite imagery derived vegetation indices (MODIS). Class Epochs Sensitivity (%) Specificity (%) Precision (%) F-Score (%) Accuracy (%) precision and average accuracy of 87.78%, 83.28%, 81.48% and 81.22% were observed with 100 epochs. 500 epoch recorded medium value between 100 and 1000 epochs in respect of sensitivity, specificity, precision and average accuracy. The structure of AANN model plays an important role in capturing the distribution of the feature vectors. The number of units in the third layer (compression layer) determines the number of components captured by the network. The AANN model projects the input vectors onto the subspace spanned by the number of units (n c ) in the compression layer. If there are n c units in the compression layer, then the feature vectors are projected onto the subspace spanned by n c components to realize them 79

33 Table 4.6: Performance of AANN for carbon sequestration pattern classification of cashew plantation using satellite imagery derived vegetation indices (MODIS). Epochs Sensitivity (%) Specificity (%) Precision (%) F-Score (%) Table 4.7: Average class performance of AANN for carbon sequestration pattern classification of cashew plantation using satellite imagery derived vegetation indices (MODIS). Epochs Sensitivity (%) Specificity (%) Precision (%) Avg.Accuracy (%) at the output layer. The effect of changing the value of n c on the performance of carbon sequestration pattern classification is studied. After some trial and error, the network structure 11L 33N 4N 33N 11L is obtained. This structure seems to give good performance in terms of classification accuracy. The results are shown in Table 4.8. Table 4.8: Performance of AANN for carbon sequestration pattern classification of cashew plantation in terms of number of units in the compression layer (MODIS). Number of units in the compression layer (n c ) Classification rate (in %)

34 (a) (b) Fig. 4.14: (a) Performance of AANN for different epochs for all class. (b) Over all Performance of AANN for each epochs in cashew plantation using MODIS. Similarly, the performance is obtained by varying the number of units in the second layer (expansion layer) keeping the number of units in the compression layer as 4. When the number of units in the expansion layer is changed from 22, 33 and 44, there is no favourable increase in the performance of AANN. The performance of carbon 81

35 sequestration pattern classification in terms of number of units in the expansion layer is shown in Table 4.9. Table 4.9: Performance of AANN for carbon sequestration pattern classification of cashew plantation in terms of number of units in the expansion layer (MODIS). Number of units in the expansion layer (n e ) Classification rate (in %) For testing, the feature vectors extracted from the MODIS imagery derived vegetation indices and climate data are given as input to the model, and the corresponding class is identified based on the maximum confidence score. From the results, it is observed that the overall classification accuracy is 96.38%. Table 4.10 and Fig show the average performance values of three classifiers. From Fig it is concluded that the accuracy of SVM is better (98%) than RBFNN (95.2%) and AANN (96.38%). SVM outperforms RBFNN and AANN in identifying carbon sequestration pattern of cashew plantation using satellite imagery derived vegetation indices. The Fig shows the snapshot of GPP estimation of cashew plantation from VGT using MVPM and Fig shows the snapshot of GPP estimation of cashew plantation from MODIS using MVPM. Fig. 4.18, Fig and Fig shows the snapshot of Classification of carbon sequestration pattern of cashew plantation using SVM, RBFNN and AANN respectively. 82

36 Table 4.10: Comparison of SVM, RBFNN and AANN for carbon sequestration pattern classification of cashew plantation using satellite imagery derived vegetation indices (MODIS). Measures SVM RBFNN AANN Accuracy (%) Fig. 4.15: Performance Comparison of SVM, RBFNN and AANN for carbon sequestration pattern classification of cashew plantation using satellite imagery derived vegetation indices (MODIS). 4.7 Summary In this chapter the existing light use efficiency model namely vegetation photosynthesis model (VPM) was modified by the introduction of another vegetation index called total ratio vegetation index (TRVI), which is specially developed for semearid vegetation studies. The GPP calculated using modified VPM was classified using SVM, RBFNN and ANN. Among the three techniques SVM classified the carbonsequestration pattern of cashew plantation with highest accuracy of (98.6%). 83

37 Fig. 4.16: GPP estimation of cashew plantation from VGT using MVPM. 84

38 Fig. 4.17: Snapshot of GPP estimation of cashew plantation from MODIS using MVPM. 85

39 Fig. 4.18: Classification of carbon sequestration pattern of cashew plantation using SVM. 86

40 Fig. 4.19: Classification of carbon sequestration pattern of cashew plantation using RBFNN. 87

41 Fig. 4.20: Classification of carbon sequestration pattern of cashew plantation using AANN. 88