Spatial quantification of biomass and carbon stock for different land use systems of Kallakurichi and Villupuram districts of Tamil Nadu, India

The transformations and sudden shift in the land use and land cover systems (LULC) greatly contributes to the human induced greenhouse gas emissions. With the carbon stock and biomass being quantified for each LULC systems, the sequestration potential and its associated parameters can be assessed aiding in the formulation of carbon related policy decisions. Fifteen different LULC classes including the crops cultivated in the study area were delineated by integrating optical (Sentinel 2A), microwave (Sentinel 1A), and its associated vegetation indices (26 Nos.) using several machine learning algorithms (i.e.) Random Forest (RF), Support Vector Machine (SVM), Multinomial Logistic Regression (MLR), Decision Tree (C5.0) and Extreme Gradient Boosting (XGB). The classification resulted with the random forest having the highest overall accuracy of 71.1% and a kappa coefficient of 0.69, which were enhanced through mask-based delineations. For biomass and stock quantification, a total of 105 observation samples have been collected from the agriculture and forest LULC systems randomly for analysing biomass, bulk density and soil organic carbon using standardized laboratory procedure. The vegetation indices (VI) from both the optical and SAR datasets were then used for the biomass modelling using Multiple Linear Regression (MLR). The regression was then performed with different combinations of the vegetation indices framed and their performance being validated using the test datasets partitioned. Though optical datasets had the evident highest correlation with the biomass values, when compared to the SAR datasets, the synergistic combination of both datasets (optical and SAR) increased the overall performance of the model for above ground biomass estimation. The efficiency of the quantifications was assessed based on the R² and RMSE to indicate the explained variance and the nature of the residuals in the derived model combinations. The integrated optical and the SAR dataset combinations resulted with the R² and RMSE highest for the training (0.84; 3.78 t/ha) and test (0.96; 2.38 t/ha) datasets for agricultural ecosystem. Similarly, for the forest ecosystem, the R² and RMSE metrics derived for the training (0.92; 11.25 t/ha) and the test datasets (0.73; 31.01 t/ha) had the highest measure among the combinations derived. The comprehensive results of the study reported that the random forest and MLR algorithm aided through optical and SAR datasets provided optimal classification and regression results, respectively. Further, the modeling framework resulted with sugarcane crop class having the highest total carbon stock values besides the evergreen forest sequestrating the maximum biomass and carbon stock. Thus, each of the agricultural and forest classes indicated their efficiency in accounting the carbon credit, which can be utilized by the policy makers in strategizing the regulations for carbon sequestration, sustainable land management, and climate change mitigation.