2.1. Study Area
The study area covers 144 km2 within Tundi reserved forest in Dhanbad District, Jharkhand state, India. This Indian state is one of the richest in biodiversity due to diverse climatic and physiographic conditions. It spreads from a northwest to northeast direction with the Domunda hills having two heads (tops) on its eastern fringe.
There are a few isolated hills of varying dimensions scattered in the northern half of this district. The area is a typical deciduous forest that has a total forested area of 23,605 km2 constituting about 29.61% of the total land area of Jharkhand. In India, reserved and protected forests are declared by the respective state governments. In reserved forests, all activities are prohibited unless permitted; it is vice versa in protected forests, i.e., specific activities such as hunting, grazing, etc., are permitted unless prohibited. Out of the total forest land of Jharkhand, 81.28% is under the protected forest and 18.58% is under the reserved forest category.
Geographically, Jharkhand has two major forest categories, namely tropical zone dry forests and tropical zone wet forests [
43]. The test site falls under the tropical zone dry forest category and is covered with
Shorea robusta (local name Sal) as the most predominant tree species, lies within 23°45′40″ to 24°05′50″ N and 85°57′30″ to 86°35′55″ E (
Figure 1a). Tundi reserved forests under Dhanbad Forest Division fall under subgroup 5-B, namely northern tropical dry deciduous forest [
8]. The Tundi reserve forest cover contains an almost pure crop of
Shorea robusta saplings and poles. The test site is dominated by minor hillocks with undulating topography;
Shorea robusta is dominant and constitutes the top story of the forest, with an average height up to 12 m in the test site. Humus is absent except in remote areas where the forest cover is well protected. The area receives typical monsoon-type climate with three marked seasons: hot, rainy, and winter. Humidity is very high during the rainy season and very low during the hot weather. The maximum temperature rises above 45 °C and a hot wind known as ″Loo″ blows frequently from April to June, until the onset of monsoon. Thunderstorms usually occur in May, accompanied by a temporary fall of a few degrees. Due to heavy industrialization and mining activities, abundant suspended particles and specks of dust are found in the atmosphere during this season. The monsoon usually breaks in the middle of June and continues until the end of September. The average rainfall is between 1100 to 1200 mm.
Geologically, Tundi comprises Precambrian rock, the country rock being quartz-feldspathic schist intruded rather profusely by some igneous bodies, viz., metadolorites and amphibolites. The prominent hill and hillocks in the area comprise mostly dark color, hard, and compact intrusive rocks (mainly amphibolite and metadolorite). The soil and alluvium are derived largely from gneisses and quartz-feldspathic schist. The soil formation in the forest area is shallow to very shallow, and depth on the plain to undulating land rarely exceeds 60 cm (24 inches). It is generally red loam with pockets of clay. Erosion occurs and varies with slope. The topsoil is hard and compact [
44].
2.2. Datasets
Sentinel-2 multispectral image (
MSI)-level-1C product acquired on 4 December 2019, covering the study area has been chosen as test data of optical remote sensing.
Figure 1b shows the false-color composite (
FCC) image from Sentinel-2 (
L1C) that encompasses the selected test site within the study area. The spatial resolutions vary between 10 to 60 m and spectral resolutions vary between 15 to 180 nm for different bands of the collected Sentinel-2 image (
Table 1).
Sentinel-2 outperforms other available spaceborne datasets when using near-equivalent image bands when Sentinel-2 data are downsampled to 30 m pixel resolution (for example, Landsat 8). Additionally, Sentinel-2 includes high quality red edge band. Prediction of forest
AGB using Sentinel-2 confirmed better accuracy compared to Landsat 8 [
45]. The field survey was carried out to collect ground forest information for remote sensing-based method development and result validation. In order to cover the complete forest variability of the study area, ground truth locations were chosen randomly based on accessibility in the forest and irrespective to tree height categories. Moreover, ground samples were collected from different locations in such a pattern to cover the entire forest diversity of the study area.
The sample plots selected are a representation of the whole study area. The ground samples were collected from three different profiles separated by nearly 1 km. The distance between samples was also set at approximately 1 km. The ground samples were collected randomly from these profiles based on accessibility into the forest with independent directions. Field data collection was conducted using a global positioning system (
GPS), altimeter, and ancillary field equipment. The ground samples, including tree height, tree
DBH and the number of trees per sample plot, tree stem density, and
FVC from 22 different locations with plot size approximately 30 m × 30 m (approx. 0.09 hectare), were collected from the study site. The field samples from 22 locations comprised approximately 550 individual trees, and an average of 25 trees per plot were measured in different elevation strata. The average measured tree height for each field location was used as tree height of the location for further study. In the ground
FVC estimation at each sample location, initially, the total number of trees and the radius of each tree cover were calculated. The surface cover of individual trees was assumed circular. Therefore, surface area cover by individual trees was calculated using their covering radius. The total tree cover area at each sample location was estimated by summing individual tree cover area. Lastly, field
FVC was calculated by taking the ratio of area covered by trees to the total area of the sample location. Consequently, ground data and
FVC were utilized for ground
AGB calculation and validation of the modeled
AGB from Sentinel-2 data. Further details of the ground data and locations are provided in
Table 2. The ground samples sequence provided in
Table 2 is the order of data collection regardless of profile composition. Therefore, the sample sequence provided in
Table 2 shows random sample location irrespective to spatial pattern of the profile.
2.3. Methodology
The methodology is presented in two phases. In the first phase, a developed general model for forest AGB assessment from optical remote sensing imagery is presented, whereas the second phase shows the experimentation of the approach developed specific to the Sentinel-2 image, vegetation indices, and statistical parameters used in this study.
Phase I. Standard procedure for forest AGB estimation using optical remote sensing data.
The procedure for estimation of the forest
AGB was articulated into the following six steps and the flow diagram of the method is shown in
Figure 2.
Step 1: Preprocessing—In this step, the optical remote sensing image was preprocessed for noise reduction, radiometric correction and calibration, atmospheric correction, and spatial resampling. The noise persists in most of the satellite images due to sensor malfunctioning and heating, and environmental influences reduce the quality of the data. Hence, suppression of noise contents in an image is essential to enhance the quality of the data. The distributed remote sensing images usually contain the digital number, which is a calibrated radiance value by sensor-specific gain and offset. Thus, radiometric correction and calibration are required to convert digital numbers to radiance.
Further, atmospheric correction is performed to reduce the effects of atmospheric scattering and absorptions in the radiance data and convert the radiance into reflectance data. The Sentinel-2 image contains different spatial resolution in various spectral bands. Spatial resampling was used to achieve a single spatial resolution from different spatial resolutions in different spectral bands. In preprocessing, noise reduction and radiometric correction were performed using image statistics-based algorithms developed using the gain and offset of the sensor. A combined flat-field and dark object subtraction method was developed for atmospheric correction and conversion of a radiometric image into a reflectance image. The Gram–Schmidt pan sharpening algorithm was used to achieve the entire spectral bands into a single high spatial resolution.
Step 2: Forest
FVC calculation—In this step, the forest
FVC at each pixel level was calculated from optical remote sensing imagery using the
FVC model proposed by Zhang et al. [
46] as follows:
Suppose spectral information of pure vegetation pixel and pure soil pixel are represented as
Sv and
Ss. Then spectral information (
S) of an image at each pixel is composed of spectral information of vegetation cover (
Sveg) and nonvegetation cover (
So). If forest
FVC in a pixel is represented as
Fc, then a fraction of nonvegetation cover should be 1−
Fc. Thus, mathematically, forest FVC at a pixel level can be calculated as follows:
Using the Equation (4), the FVC is calculated from both the optical remote sensing image and field data.
Step 3: Field forest
AGB estimation—In this step, the field forest
AGB was calculated from the collected field data. The field forest
AGB was obtained at individual plot level in kilogram (kg) using the following simplified empirical model proposed by Chave et al. [
27] as follows:
where
μ is specific gravity (g cm
−3) of tree stem,
D is the
DBH (cm),
H is the tree height (m), and
κ is the constant for forest type.
Step 4: Model development between FVC and AGB—In this step, linear regression modeling is performed to find the best fit curve between the calculated ground forest AGB from field data using Step 3 and the image forest FVC obtained using Step 2.
The best fit linear model between the field forest
AGB and the corresponding image forest
FVC is represented as follows:
where
α and
β are the gain and offset, respectively, for image forest
FVC used to calculate the forest
AGB.
Step 5: Forest
AGB image generation using the forest
FVC image—Suppose the forest
FVC image is presented as
FVC(
i,
j) and the image of forest
AGB is represented as
AGB(
i,
j), where (
i,
j) is the pixel location. Then the pixelwise forest
AGB in kg is calculated as follows:
Step 6: Accuracy assessment—The accuracy of the forest
AGB calculated from the optical remote sensing image was evaluated by ground validation. The ground validation was performed by comparing the
AGB value between the field calculated and the image estimate at different ground locations. The accuracy of the result in percentage is represented as:
where
N is the total number of validation points, and
n is the number of validation points agreed with when equal or less than one standard deviation between image and field measurements.
Finally, the generalized forest AGB obtained by individual plot level in kg are standardized into tons per hectare (t ha−1).
Phase II. Forest AGB model evaluation using Sentinel-2 data.
To test and verify the effectiveness of the proposed method, experimentation in sequential procedure was applied on Sentinel-2 as an optical remote sensing image to calculate the forest AGB from the Tundi reserved forest test site, Jharkhand, India. The sequential steps in the forest AGB calculation are given below.
After preprocessing, the spatial resampling was initially performed to bring all of the spectral bands to 10 m spatial resolution (e.g., bands 2–4 with spatial resolution 10 m and bands 5–7, 8a with spatial resolution 20 m) and then resampled to 30 m spatial resolution, i.e., approximately equal to the field data spatial resolution.
The direct vegetation and nonvegetation (soil) spectral information extraction for the forest
FVC calculation is tricky. Thus, different vegetation indices were used to quantify the vegetation information within the pixel of an image. In our experiment, for quantifying vegetation amount in a pixel of Sentinel-2, four vegetation indices were assessed: normalized difference vegetation index (
NDVI), modified vegetation index (
MVI), soil-adjusted vegetation index (
SAVI), and modified soil-adjusted vegetation index (
MSAVI). The
NDVI,
MVI,
SAVI, and
MSAVI images were produced from the Sentinel-2 image using the following equations:
where
L is the canopy background adjustment factor (
L value as 0.5 in reflectance data) and
NIR and
RED are the reflectance values in near infrared and red bands of the multiband remote sensing image.
In our experiment,
NIR (band 8) and
RED (band 4) were carefully chosen from the Sentinel-2 reflectance image to calculate these vegetation indices. The highest vegetation index value in an image was considered as a pure vegetation pixel whereas the lowest value was decided as a pure soil pixel. Consequently, pure pixels for vegetation and nonvegetation (soil) information were identified to calculate the forest
FVC images from the vegetation indices calculated using Equation (4). The best vegetation index among vegetation indices for the forest
AGB calculation was selected based on the correlation coefficient (
R) and mean absolute error (
MAE) between the ground forest
FVC and image forest
FVC.
R is the degree of statistical similarity between a pair of variables, whereas
MAE is a measure of errors between paired observations.
R and
MAE were calculated between the ground
FVC and the
FVC from Sentinel-2 by
NDVI,
MVI,
SAVI, and
MSAVI corresponding to ground sample locations. Based on the highest correlation and lowest
MAE criteria, the
MSAVI-based image
FVC was chosen in
AGB calculation. Further, the trend line by linear regression between the ground and
MSAVI-based
FVC was fitted, which is shown in Equation (13):
In Tundi reserved forest, the predominant tree species is
Shorea robusta (Sal). Accordingly, based on [
27], the ground forest
AGB was calculated using the following equation from the field data (
Table 2):
where
AGB is the forest above ground biomass (kg),
μ (0.667) is specific gravity (tree stem density) (g cm
−3),
D is the
DBH (cm),
H is the tree height (m), and the constant 0.509 obtained from [
27]. The calculated ground forest
AGB from field data and the selected image forest
FVC are shown in Table 5.
A trend line is obtained between calculated ground forest
AGB and image-retrieved forest
FVC by linear regression. The following model (Equation (15)) was obtained from 15 ground training samples (Table 5) and the remaining 7 ground samples were used for validation of the results obtained from the developed model (Table 6).
The accuracy of the forest AGB based on Sentinel-2 was assessed by comparing the remaining seven field measurements to the corresponding locations. Kindly note that the samples selected for training were not chosen for the validation to avoid the biasness in the accuracy assessment. Thus, training and validation samples were separated before trend modeling. For data processing and implementation of the proposed procedure and experimentation in forest AGB assessment, mainly the Interactive Data Language (IDL) programing and Environment for Visualizing Images (ENVI) software were used.