Use of Hyperion for Mangrove Forest Carbon Stock Assessment in Bhitarkanika Forest Reserve: A Contribution Towards Blue Carbon Initiative

Mangrove forest coastal ecosystems contain significant amount of carbon stocks and contribute to approximately 15% of the total carbon sequestered in ocean sediments. The present study aims at exploring the ability of Earth Observation EO-1 Hyperion hyperspectral sensor in estimating aboveground carbon stocks in mangrove forests. Bhitarkanika mangrove forest has been used as case study, where field measurements of the biomass and carbon were acquired simultaneously with the satellite data. The spatial distribution of most dominant mangrove species was identified using the Spectral Angle Mapper (SAM) classifier, which was implemented using the spectral profiles extracted from the hyperspectral data. SAM performed well, identifying the total area that each of the major species covers (overall kappa = 0.81). From the hyperspectral images, the NDVI (Normalized Difference Vegetation Index) and EVI (Enhanced Vegetation Index) were applied to assess the carbon stocks of the various species using machine learning (Linear, Polynomial, Logarithmic, Radial Basis Function (RBF), and Sigmoidal Function) models. NDVI and EVI is generated using covariance matrix based band selection algorithm. All the five machine learning models were tested between the carbon measured in the field sampling and the carbon estimated by the vegetation indices NDVI and EVI was satisfactory (Pearson correlation coefficient, R, of 86.98% for EVI and of 84.1% for NDVI), with the RBF model showing the best results in comparison to other models. As such, the aboveground carbon stocks for species-wise mangrove for the study area was estimated. Our study findings confirm that hyperspectral images such as those from Hyperion can be used to perform species-wise mangrove analysis and assess the carbon stocks with satisfactory accuracy.


Introduction
Mangrove forest coastal ecosystems provide several beneficial functions, both to terrestrial and marine resources [1,2]. Mangrove forests contain significant amount of carbon stocks and are one of Sensing is crucial as a tool for assessing and monitoring mangrove forests, primarily because many mangrove swamps are inaccessible or difficult to field survey [33].
Previous work by the authors as well as other researchers has allowed assessing the biomass of the several mangrove plant species and has provided the biomass of species individually. Chaube et al. [34] employed AVIRIS-NG (Airborne Visible InfraRed Imaging Spectrometer Next Generation) hyperspectral data to map mangrove species using a SAM (Spectral Angle Mapper) classifier. Authors identified 15 mangrove species over Bhitarkanika mangrove forest, reporting an overall accuracy (OA) of 0.78 (R 2 ). They also concluded that the hyperspectral images are very useful in discriminating mangrove wetlands, and having a finer spectral and spatial resolution can be crucial in investigating fine details of ground features. Kumar et al. [35] used the five most dominant classes of mangrove species present in Bhitarkanika as training sets to classify using SAM on Hyperion hyperspectral images, and archived an OA of 0.64. Ashokkumar and Shanmugam [36] demonstrated the influence of band selection in data fusion technique; they performed classification using support vector machine and observed that factor based ranking approach shown better results (R 2 of 0. 85) in discriminating mangrove species than other statistical approaches. In another study, Padma and Sanjeevi [37] used an identical algorithm by integrating Jeffries-Matusita distance and SAM to map the mangrove species within the Bhitarkanika using Hyperion Image with an OA of 0.86 (R 2 value).
Presently, the spatial distribution maps of mangroves are generated using Earth Observation (EO) Hyperion datasets [26]. Table 1 illustrates the wetland research, which employed several algorithms for the assessment using various data types. Identifying different species in a mangrove forest is a fundamental yet difficult task, as it requires a high spatial and spectral resolution satellite images. To identify different species within the study area, EO-1 Hyperion hyperspectral data is currently acquired and field-sampling points are taken to generate the endmember spectra. This study demonstrated the use of vegetation indices (in this paper NDVI (Normalized Difference Vegetation Index) and EVI (Enhanced Vegetation Index)) for estimating carbon stock within an area with a significant accuracy. Presently, the field inventory data were incorporated with the hyperspectral image to derive the carbon stock. Three different NDVI and EVI based models were used to determine the total blue carbon sequestered by each species within the study area.
In purview of the above, this study aimed at evaluating the net above ground carbon stocks present at Bhitarkanika mangrove forest ecosystem, particularly with relevant field inventory and remote sensing approaches.

Study Area
Our study site is located in the Kendrapara district of Odisha, India, which lies between 20 • 41 36.70" and 24 • 45 28" N latitude and 86 • 54 17.29" and 86 • 92 8.96" E longitude (as shown in Figure 1). Geographically, it covers an area of around 41.05 Km 2 of which mostly low-lying (10-25 m above mean sea level) covered with dense mangrove forests. The Bhitarkanika Forest Reserve is a protected forest reserve with a unique habitat and ecosystem. About two-third of the Bhitarkanika Forest Reserve is covered by the Bay of Bengal, and this estuarial region (lies within Bramhani-Baitarni) is a predominant inter tidal zone. Bhitarkanika Forest Reserve is home to a diverse types flora and fauna including some endangered species; it is the second largest mangrove forest in India formed by the estuarial formation of Brahmani-Baitarni, Dhamra, and Mahanadi rivers [67].

Study Area
Our study site is located in the Kendrapara district of Odisha, India, which lies between 20°41′36.70′′ and 24°45′28′′ N latitude and 86°54′17.29′′ and 86°92′8.96′′ E longitude (as shown in Figure 1). Geographically, it covers an area of around 41.05 Km 2 of which mostly low-lying (10-25 m above mean sea level) covered with dense mangrove forests. The Bhitarkanika Forest Reserve is a protected forest reserve with a unique habitat and ecosystem. About two-third of the Bhitarkanika Forest Reserve is covered by the Bay of Bengal, and this estuarial region (lies within Bramhani-Baitarni) is a predominant inter tidal zone. Bhitarkanika Forest Reserve is home to a diverse types flora and fauna including some endangered species; it is the second largest mangrove forest in India formed by the estuarial formation of Brahmani-Baitarni, Dhamra, and Mahanadi rivers [67].
The study area comes under the humid sun-tropical climatic region broadly having three seasons namely, summer in which the temperature reaches up to 43 °C , winter in which the temperature goes down to as low as 10 °C, and the rainy season in which this region faces flash floods and frequent cyclones between the months of June to October. The Bhitarkanika Forest Reserve was chosen for the present study because it contains variety of heterogeneous species. In our work, the 10 most dominant mangrove species (as shown in Table 2) were identified and used for further analysis.  The study area comes under the humid sun-tropical climatic region broadly having three seasons namely, summer in which the temperature reaches up to 43 • C, winter in which the temperature goes down to as low as 10 • C, and the rainy season in which this region faces flash floods and frequent cyclones between the months of June to October. The Bhitarkanika Forest Reserve was chosen for the present study because it contains variety of heterogeneous species. In our work, the 10 most dominant mangrove species (as shown in Table 2) were identified and used for further analysis.

EO Data Acquisition
EO-Hyperion images (L1Gst) were obtained over the study area from the United States Geological Survey (USGS). The specifications of Hyperion sensor are illustrated in Table 3. Hyperion has a spatial resolution of 30 m and 242 spectral bands covering 356 nm to 2577 nm wavelengths. The Hyperion data strip passing over Bhitarkanika Forest Reserve is shown in Figure 2. Out of the 242 spectral bands, 46 bands are considered as bad bands (including 1-7, 58-78, 120-132, 165-182, 185-187, and 221-242 bands), and thus, these were not considered in further analysis. Bad bands have a high amount of noise caused by the water absorption in atmosphere, band overlaps, and lack of proper illumination. The performed image pre-processing includes noise removal and cross track illumination correction. In addition, atmospheric correction has been applied to remove atmospheric noises using the FLAASH (Fast Line-of-sight Atmospheric Analysis of Hyper Spectral-cubes) module in ENVI (v. 5.2) software [68]. After completing this step, endmember extraction was performed for each of the targeted species using the final Hyperion reflectance image and the in-situ GPS (Global Positioning System) locations.

EO Data Acquisition
EO-Hyperion images (L1Gst) were obtained over the study area from the United States Geological Survey (USGS). The specifications of Hyperion sensor are illustrated in Table 3. Hyperion has a spatial resolution of 30 m and 242 spectral bands covering 356 nm to 2577 nm wavelengths. The Hyperion data strip passing over Bhitarkanika Forest Reserve is shown in Figure 2. Out of the 242 spectral bands, 46 bands are considered as bad bands (including 1-7, 58-78, 120-132, 165-182, 185-187, and 221-242 bands), and thus, these were not considered in further analysis. Bad bands have a high amount of noise caused by the water absorption in atmosphere, band overlaps, and lack of proper illumination. The performed image pre-processing includes noise removal and cross track illumination correction. In addition, atmospheric correction has been applied to remove atmospheric noises using the FLAASH (Fast Line-of-sight Atmospheric Analysis of Hyper Spectral-cubes) module in ENVI (v. 5.2) software [68]. After completing this step, endmember extraction was performed for each of the targeted species using the final Hyperion reflectance image and the in-situ GPS (Global Positioning System) locations.

Field-Inventory Based Biomass Measurement
Field sampling was undertaken during 2015 for the study site. The foremost steps are the prior knowledge of the mangrove plant species; their location and its structure were essential for collecting the sample data for geospatial analysis. Random and the most homogenous patches within the Bhitarkanika Forest Reserve were selected for the field survey to measure tree height, number of samples (trees), Diameter at Breast Height (DBH), and total number of species within the plot.
As the study site selected is 36.42 km 2 falling within the range of Hyperion data strip ( Figure 2). Hyperion image has limited coverage over the Bhitarkanika forest range, and for this reason, a region was selected that falls within the area covered by the Hyperion field of view. The samples were collected by making a 90 × 90 m 2 grid and it is further divided into nine equal 30 × 30 m 2 sub-grids, i.e., 90 sub-grids were examined. The most homogenous grid was taken into consideration. This process was then repeated to identify the 10 most homogenous mangrove plant species within the study area and samples were collected using GPS and Clinometer. The field data records the vegetation parameters using GPS in multiple directions. The number of tree species was counted within the plot in random sampling design in the Bhitarkanika Forest Reserve [69]. An overview of the methodology implemented is available in Figure 3. These major species were identified for the study site and their spectral profile was extracted using EO-1 Hyperion dataset. Total area covered by these species was 36.42 km 2 (see Figure 2). Non-vegetative regions were masked out from the study region. developed in modified form. It is more general in nature ( [78,82,83]) and applicable in field. It is not possible to cut all the trees to estimate their biomass. Considering the mathematical terms, the models were developed by [76,77,83,84]. The model developed by [75] (1989) to estimate above ground biomass has been used in the present investigation. The literature revealed that this method is nondestructive and is the most suitable method. The biomass for each tree is calculated using the following allometric equation [76,83,85]: where Y is above ground biomass (t. ha 1 ), D is the diameter at breast height, H is the tree height, and S is the wood density. The average wood density (S) for each species is taken from the wood density database provided by the International Council for Research in Agroforestry (ICRAF). From the acquired wood density, it was found that the wood density of Cerbera odollam Gaertn. was lowest (0.3349 gcm 3 ), followed by Excoecaria agallocha L. (0.49 gcm 3 ) among all. Heritiera littoralis Dryland ex Ait. had the highest (0.848 gcm 3 ) wood density. The above ground carbon was calculated using the following formula to estimate biomass [83,85,86]: where Y is the above ground carbon stock (t. ha 1 ) and B is the above ground biomass per hectare (t. C ha 1 ). The precise location of the in-situ ground control points of each species were further used to generate the spectral profile using Hyperion hyperspectral data as shown in Figure 4. The generated spectra of each species were given as an input to the SAM classifier. It is observed that Intsia bijuga (Colebr.) Kuntze is showing the highest reflectance among other observed species, whereas Aegiceras corniculatum (L.) has the lowest reflectance. The Spectral Angle Mapper (SAM) supervised classification algorithm was used for the land use/cover classification using ENVI software [70,71]. SAM is a physically-based spectral classification algorithm, according to [72] that calculates the spectral similarity between a pixel spectrum and a reference spectrum as "the angle between their vectors in a space with dimensionality equal to the number of bands" [72]. SAM uses the calibrated reflectance data for classification and thus relatively insensitive to illumination and albedo effects. End-member reference spectra used in SAM were collected directly from acquired hyperspectral images. SAM compares the angle between reference spectrum and each pixel of an image in n-D space [72][73][74]. This 'spectral angle' (α) is calculated as: where α is the angle between reference spectra and endmember spectra, t is the endmember spectra, and r is the reference spectra. A thorough and detailed investigation was performed to develop a criterion to estimate different species and determine variety of communities present in that ecosystem. To perform the sampling, firstly, the area is sub-divided into homogeneous patches or units, and furthermore, the samples were taken within these homogenous patches. The total number of transect sampling units to determine the allowable error was calculated using (Chacko, 1965) as follows: where N is the total number of samples, t is the Student's (t-statistics) value at a 95% significance level, CV is the coefficient of variation (in %), and E is the confidence interval (in mean %). While performing the field sampling, a transect of 30 m × 30 m plot was laid on the most dominant patch for each species inside the protected area of Bhitarkanika forest reserve. The collected field sampling points were further distributed, and 2/3 of the samples were used for generating the models, whereas 1/3 of the samples were used for validation purpose. Table 2 has shown the field measurements of each species, e.g., scientific name, tree height, DBH, total number of trees within the sample plot, wood density of each species, biomass, and carbon stock. The trees whose girth height was below 1.32 m and DBH < 10 cm were not taken under consideration. The geographical location (latitude and longitude) was recorded using hand-held GPS. There were several mathematical equations developed and used by researchers for biomass estimation of trees [75][76][77][78][79][80][81]. These equations are species specific, particularly in the tropics. The general equation has been developed in modified form. It is more general in nature ( [78,82,83]) and applicable in field. It is not possible to cut all the trees to estimate their biomass. Considering the mathematical terms, the models were developed by [76,77,83,84]. The model developed by [75] (1989) to estimate above ground biomass has been used in the present investigation. The literature revealed that this method is non-destructive and is the most suitable method. The biomass for each tree is calculated using the following allometric equation [76,83,85]: where Y is above ground biomass (t. ha 1 ), D is the diameter at breast height, H is the tree height, and S is the wood density. The average wood density (S) for each species is taken from the wood density database provided by the International Council for Research in Agroforestry (ICRAF). From the acquired wood density, it was found that the wood density of Cerbera odollam Gaertn. was lowest (0.3349 gcm 3 ), followed by Excoecaria agallocha L. (0.49 gcm 3 ) among all. Heritiera littoralis Dryland ex Ait. had the highest (0.848 gcm 3 ) wood density. The above ground carbon was calculated using the following formula to estimate biomass [83,85,86]: where Y is the above ground carbon stock (t. ha 1 ) and B is the above ground biomass per hectare (t. C ha 1 ). The precise location of the in-situ ground control points of each species were further used to generate the spectral profile using Hyperion hyperspectral data as shown in Figure 4. The generated spectra of each species were given as an input to the SAM classifier. It is observed that Intsia bijuga (Colebr.) Kuntze is showing the highest reflectance among other observed species, whereas Aegiceras corniculatum (L.) has the lowest reflectance.

Covariance Matrix Based Band Selection
Hyperspectral data are a set of hundreds of narrow bands at different wavelengths posing problems related to computational complexity, high data volume, bad bands, etc. Therefore, dimensionality reduction of hyperspectral data is considered as one of the solutions for the aforementioned issue. The dimensionality reduction technique is further classified into two groups, namely, feature extraction and feature selection. In the present study, an approach has been made to select the best band for calculation of different vegetation indices. Band selection generally involves two major steps, which are selection of criterion function and optimum band searching. The selection criterion applied in this study is the one proposed by [87], which was named Maximum ellipsoid volume criterion (MEV).
Mathematically it can be formulated as: where M is the number of pixels and S is the selected bands with S = [x1, x2, …, xn] and S T is the column vector with S T = [x1, x2, …, xm] T . Here, n and m are the number of bands and m is the number of number of pixels. Additionally, for the band searching purpose, sequential forward search was implemented, which basically works on the principle of "down to top". Here, the first band is defined as the band

Covariance Matrix Based Band Selection
Hyperspectral data are a set of hundreds of narrow bands at different wavelengths posing problems related to computational complexity, high data volume, bad bands, etc. Therefore, dimensionality reduction of hyperspectral data is considered as one of the solutions for the aforementioned issue. The dimensionality reduction technique is further classified into two groups, namely, feature extraction and feature selection. In the present study, an approach has been made to select the best band for calculation of different vegetation indices. Band selection generally involves two major steps, which are selection of criterion function and optimum band searching. The selection criterion applied in this study is the one proposed by [87], which was named Maximum ellipsoid volume criterion (MEV).
Mathematically it can be formulated as: where M is the number of pixels and S is the selected bands with S = [x 1 , x 2 , . . . , x n ] and S T is the column vector with S T = [x 1 , x 2 , . . . , x m ] T . Here, n and m are the number of bands and m is the number of number of pixels. Additionally, for the band searching purpose, sequential forward search was implemented, which basically works on the principle of "down to top". Here, the first band is defined as the band with maximum variance and the remaining band is compared one by one. While selecting the optimum band, the constant value 1 M − 1 . is neglected. Thus, Equation (4) can also be written as: where B k is the covariance matrix and S k = [x 1 , x 2 , . . . , x k ]. Therefore, we have: According to the rule of determination, the relation between B k and B k+1 is described as: Equation (7) was further used for determining the optimum band; the band that maximizes the value of det(B k+1 ) was termed as the optimum band. This band selection method was applied at blue, red, and near infrared bands to further calculate the NDVI and EVI indices.

NDVI and EVI
In our study, the vegetation indices of NDVI and EVI were employed, which were computed from the Hyperion hyperspectral data to assess the total above ground carbon stock using different allometric regression models [26]. The covariance matrix based band selection algorithm as per described in Section 2.4 determines the specific band for the calculation of vegetation indices. It was observed that the optimum band in NIR (Near-Infrared) region is R 793.13 (surface reflectance at 793.13 nm), in Red region, it is R 691.37 (surface reflectance at 691.37 nm), and in Blue region the optimum band is observed at R 447.17 (surface reflectance at 447.17 nm). The NIR and Red bands were used to calculate the NDVI; as shown in Equation (5), its value ranges from −1 to +1. The negative NDVI values shows waterbody and bare soil, whereas positive values are the green vegetation. The higher the NDVI value, the higher will the density of forest or vegetation be because of the high NIR reflectance and low Red reflectance coming from dense vegetation [88,89]. NDVI has been widely used to monitor vegetation health, density, changes, amount and condition of vegetation: EVI (Enhanced Vegetation Index) was originally developed as an improvement over NDVI; EVI is basically an optimized vegetation index that is used to enhance the sensitivity of high biomass region and it decouples the background variables as well as the atmospheric influences [90,91]. EVI is calculated as follows: EVI = 2.5 * (R 793.13 − R 691.37 ) (R 793.13 + 6 * R 691.37 − 7.5 * R 447.17 + L) (9) where L is the adjustment factor, generally 1.
In the present study, both NDVI and EVI were employed to correlate the carbon stock of the Bhitarkanika mangrove forest. EVI is considered as more robust proxy of biomass and carbon stock estimation, as it has better resilience to saturation and resistant to atmospheric contamination and soil [90,92].
Five different models, linear, polynomial, logarithmic, Radial Basis Function (RBF), and sigmoidal function, were utilized for assessing carbon using hyperspectral data derived from NDVI and EVI indices. The relationship of field measured above ground carbon with the NDVI and EVI vegetation indices for all the five models were calculated. The field measured above ground carbon was trained with NDVI and EVI values retrieved from hyperspectral image in each of the five models. The 2/3 of the in-situ measurements were used for training the data, while 1/3 of the remaining data were used for testing the models.

Results
This section provides a concise and precise description of the experimental results for blue carbon for a mangrove forest.

Spatial Distribution of Species
This section demonstrates the species-wise carbon stock spatial distribution and overall carbon stock of the Bhitarkanika forest reserve and delivers a brief analysis on the overall results. SAM classification ( Figure 5) achieved an OA of 84% and a kappa coefficient (k) of 0.81. These results indicate that SAM classification algorithm performed very well in determining the major plant species. These outputs were further taken into account and were used to derive the estimated carbon stock for each species using NDVI and EVI models and illustrating the species-wise carbon stock.
As per Table 4, it has been observed that the total aboveground carbon from EVI and NDVI derived aboveground carbon are 459.82 kt. C and 514.47 kt. C, respectively. The NDVI derived carbon is showing higher value than the EVI derived carbon because NDVI values can be influenced by the atmospheric contaminants, topography, soil, and dense biomass. These can lead to the increase in the irradiance of the NIR band and result in bias. It should also be noted that NDVI saturates in dense vegetation so that the accuracy of NDVI values differ by land use, topography, and atmospheric conditions [90,[93][94][95]. Santin-Janin et al. [96] used non-linear model coupled with NDVI and EVI estimates to estimate the biomass and carbon stock. Wicaksono et al. [97] employed 13 vegetation indices to assess the above ground carbon of mangrove forest and concluded that the best fitted above ground carbon model for mangrove species derived from vegetation indices was EVI1 (R 2 =0.688), whereas for below ground carbon GEMI (R 2 =0.567) showed the best fit. Similarly, Adam et al. [95] utilized the narrow band vegetation indices with all possible band combinations using hyperspectral data for above ground biomass and concluded EVI is more robust for the assessment. Different band selections were used by them to enhance the predictive accuracy, the best three combinations for estimating EVI are (a) 445 nm, 682 nm, and 829 nm, (b) 497 nm, 676 nm, and 1091 nm, and (c) 495 nm, 678 nm, and 1120 nm.

Estimation of Carbon Stock Using Spectral Derived Indices
This section presents the carbon stock assessment for mangrove forest using different models namely, linear, logarithmic, polynomial (second degree), RBF, and sigmoidal function. All the models were trained with the EVI and NDVI generated relations with the ground measured data as well as tested with the modeled biomass and observed carbon stock as shown in Figure 6. The latter figure illustrates the performance of each model for EVI and NDVI based estimations; it can be observed that the RBF model performed better than the others.

Estimation of Carbon Stock Using Spectral Derived Indices
This section presents the carbon stock assessment for mangrove forest using different models namely, linear, logarithmic, polynomial (second degree), RBF, and sigmoidal function. All the models were trained with the EVI and NDVI generated relations with the ground measured data as well as tested with the modeled biomass and observed carbon stock as shown in Figure 6. The latter figure illustrates the performance of each model for EVI and NDVI based estimations; it can be observed that the RBF model performed better than the others.
According to the distributed EVI value, it has been concluded that a good amount of area is under dense coverage of forest species; moreover, it has shown higher estimation of carbon stock than NDVI. EVI varies from 0.35 to 6.9 and it is more sensitive to branches and other nonphotosynthetic parts of the vegetation (parts different from leaves). EVI is more sensitive to plant parameters, as it avoids the atmospheric effects as well as the soil background. The results illustrate that EVI derived carbon varies from 27.  In both cases, the index-derived carbon estimation shows good agreement between measured and estimated carbon stock and either index could provide a good estimation. From the results EVI (R 2 = 86.98%) seems to perform slightly better than NDVI (R 2 = 84.1%). However, since the sample size is small (10 observations) the results are too close to say with statistical confidence that this hypothesis is true. However, the literature (see Section 3.1) indicates that this is indeed the case. The EVI and NDVI based carbon stock for each species (identified in the present study) is shown in Table 4.
The carbon stock values from the satellite-derived indices fall within the expected ranges for mangrove carbon stocks. NDVI values range from 0.5 to 0.65; the latter shows a healthy, dense mangrove forest in Bhitarkanika. The final interpretation result reveals that the middle northern part of the study area is showing higher biomass values (~250 t. C ha −1 ). Thus, it is concluded that these regions are highly dense and stores an ample amount of blue carbon in it. The polynomial regression model using EVI is found to be suitable for the estimation of carbon stock at the study site, with an R 2 of 0.87. EVI has shown high amount of estimated carbon ranges as it is more sensitive to biomass, and ultimately affecting the carbon estimation as compared to the NDVI and can be seen from Figure 7 and Table 4 whereas, NDVI has shown more consistent outcomes in the case of minimum and maximum estimated carbon stocks. Figure 6. (a) Performance analysis of different models with EVI based carbon estimation and in-situ measurements (b) Performance analysis of different models with NDVI based carbon estimation and in-situ measurements. In both cases, the index-derived carbon estimation shows good agreement between measured and estimated carbon stock and either index could provide a good estimation. From the results EVI (R 2 = 86.98%) seems to perform slightly better than NDVI (R 2 = 84.1%). However, since the sample size is small (10 observations) the results are too close to say with statistical confidence that this hypothesis is true. However, the literature (see Section 3.1) indicates that this is indeed the case. The EVI and NDVI based carbon stock for each species (identified in the present study) is shown in Table 4.
According to the distributed EVI value, it has been concluded that a good amount of area is under dense coverage of forest species; moreover, it has shown higher estimation of carbon stock than NDVI. EVI varies from 0.35 to 6.9 and it is more sensitive to branches and other non-photosynthetic parts of the vegetation (parts different from leaves). EVI is more sensitive to plant parameters, as it avoids the atmospheric effects as well as the soil background. The results illustrate that EVI derived carbon varies from 27.  The carbon stock values from the satellite-derived indices fall within the expected ranges for mangrove carbon stocks. NDVI values range from 0.5 to 0.65; the latter shows a healthy, dense mangrove forest in Bhitarkanika. The final interpretation result reveals that the middle northern part of the study area is showing higher biomass values (~250 t. C ha −1 ). Thus, it is concluded that these regions are highly dense and stores an ample amount of blue carbon in it.
The polynomial regression model using EVI is found to be suitable for the estimation of carbon stock at the study site, with an R 2 of 0.87. EVI has shown high amount of estimated carbon ranges as it is more sensitive to biomass, and ultimately affecting the carbon estimation as compared to the NDVI and can be seen from Figure 7 and Table 4 whereas, NDVI has shown more consistent outcomes in the case of minimum and maximum estimated carbon stocks.

Species-Wise Carbon Stock Assessment
The classification results generated from SAM classifier and the covariance matrix based optimum band selection for generating vegetation indices were further used to extract the species-wise carbon stock as well as the area covered by each species in the Bhitarkanika forest reserve (see Figures 8 and 9). Figure 9 illustrates the NDVI derived carbon distribution map for each major species, while Figure 8 demonstrates the EVI derived carbon distribution map for each major species. It is also important to notice that the carbon stock of each species shows some variance, which is investigated and presented in Figures 10 and 11. Furthermore, the outcome of species-wise carbon stocks depends upon the species classification accuracies for species distribution classification maps.       Total area covered by the major mangrove species was around 36.42 km 2 . Cerbera odollam Gaertn covers the largest part of the forest, approximately 22.90% of the total area. Total estimated carbon for the EVI derived indices is 49.82 kt. C. and total carbon estimated for the Bhitarkanika forest derived from NDVI indices is 514.47 kt. C. Using EVI-derived carbon stocks, the highest contribution of carbon stock is the Intsia bijuga (Colebr.) Kuntze species with 53.10 kt. C (11.54%). From the NDVI derived carbon stocks, Cerbera odollam Gaertn seems to contribute the most with 56.36 kt. C (10.95%). Field measured carbon was recorded lowest for the species Xylocarpus mekongensis Pierre, which was 76.20 t. C ha −1 . Figure 8 shows the spatial distribution of carbon derived from EVI for each species. Intsia bijuga (Colebr.) Kuntze shows highest carbon content up to 253.4 t. C ha −1 . The highest carbon stocks as derived from NDVI were displayed for Xylocarpus mekongensis Pierre at 258.84 t. C ha −1 .
As such, while Cerbera odollam Gaertn covers most of the area (22.9%), differences in carbon per hectare (Carbon area density) promote Intsia bijuga (Colebr.) Kuntze as the highest contributing species in the Bhitarkanika forest with EVI-derived carbon stocks. This is due to the large difference between EVI and NDVI derived carbon area density for Cerbera odollam Gaertn (average 128.78 ± 15.702 t. C ha −1 and 150.498 ± 15.51 t. C ha −1 ). Cross-referencing with the measured values presented in Table 2 (165.03 ± 10.87167.02 t. C ha −1 ), leads to the conclusion that the NDVI derived carbon stocks for Cerbera odollam Gaertn are more accurate. This conclusion is not reflective of all the species. Out of the 10 species examined, the average Carbon area density of EVI is closer to the measured value in six of them, while NDVI derived Carbon area density is more accurate in the other four. The greatest divergence between EVI and NDVI estimated carbon area densities is for Cerbera odollam Gaertn. Significant differences are also shown for Intsia bijuga (Colebr.) Kuntze and Xylocarpus mekongensis Pierre.
A species-wise box-plot is generated to assess the variation in different species-wise carbon stock estimated using EVI and NDVI, which is shown in Figures 10 and 11, with the minima, maxima, median, 25% quartile, and 75% quartile. The average carbon stock measured from field sampling is 131.07 t. C ha −1 . Average EVI derived carbon stock ranges from 77.86 t. C ha −1 to 135.28 t. C ha −1 and for NDVI derived carbon stock 116.57 t. C ha −1 to 145.82 t. C ha −1 for the Bhitarkanika mangrove forest. As such, both EVI and NDVI estimated averages are in agreement with the average carbon stock measured from the field.

Conclusions
Mangrove forests store a large quantity of blue carbon in plants, both in the form of biomass and as sediment in the soil. Anthropogenic activities threaten these forests nowadays due to conversion to other land use types. Such transition of forest areas is a major source of carbon emissions to the atmosphere. As such, carbon stock assessment is essential to reduce the loss of biomass in such ecosystems. Species-wise blue carbon analysis can be used to assess the impact of global climate change on different mangrove species as well as to help policy makers to accurately evaluate the ecological and economical trade off associated with the management of mangroves ecosystem. The present study aimed at demonstrating the use of hyperspectral EO data for species identification in a highly diversified mangrove ecosystem and for calculating total carbon stored. The Bhitarkanika forest in India was chosen as a study site and Hyperion hyperspectral images were used.
There have been several studies on the blue carbon stored in mangroves, however, thus far, a species wide blue carbon analysis with significant accuracy was missing. This study attempts to mitigate that gap of knowledge by estimating the above-ground carbon stocks for each of the 10 major species that were identified and found dominant in the study area.
Hyperspectral data from EO-1 Hyperion were collected and processed to extract the biophysical parameters of interest. Near co-orbital field measurements of biomass and carbon measurements were acquired for validation. The in-situ locations of mangrove species were used to generate spectral profile. The spatial distribution of the major mangrove species was identified using the SAM classification algorithm, which performed reliably well (e.g., kappa coefficient κ = 0.81). NDVI and EVI radiometric indices were calculated from the optimum bands, obtained by covariance matrix based band selection algorithm. Several models were tested to relate NDVI and EVI with carbon stocks. The RBF model performed best (R 2 = 86.98% for EVI and R 2 = 84.1% for NDVI) and was subsequently used in this study to estimate carbon stocks for the 10 dominant species and the entire study area.
Despite the significance of mangrove ecosystem and blue carbon for local as well as global climate, the drastic transformation of mangrove forests into other land use types is directly affecting the livelihood around it, which can be seen through the shortage of firewood, regular soil erosion, and decrease in fishing zones. Therefore, there should be adequate digital information about the coverage, biomass, and carbon content of the mangrove forest for quick management and planning. The present study provides evidence that NDVI and EVI indices have a very promising potential to be applied in classifying the dominant species of mangrove forests and coastal ecosystems according to their carbon content. These indices can provide adequate estimates of maximum, minimum, and average carbon content for a large area and show the spatial distribution of carbon, and thus, biomass. The above-ground carbon stocks for each species were estimated and presented in this study. For the whole study area, the carbon stocks were estimated 459.82 kt. C. from EVI and 514.47 kt. C. from NDVI.
The only limitation faced in this study was the limited availability of Hyperion data and that too covering a part of Bhitarkanika as shown in Figure 2. Using the same methodology with spectral images from different satellites could provide better coverage, and thus carbon stock estimations of different areas. Future studies could focus on different ecosystems to assess the effectiveness for this method and estimate carbon stock for different areas and ecosystems in order to provide the tools for a better evaluation of biomass and global carbon stocks; this remains to be seen.