Machine Learning Based Modeling for Forest Aboveground Biomass Retrieval

This study aims at evaluating the potential of Support Vector Machine (SVM) and Random Forest (RF) algorithm for the prediction of forest aboveground biomass. In this study, mean forest aboveground biomass density (AGBD) data obtained from GEDI L4B product, Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EVI) from Moderate Resolution Imaging Spectroradiometer (MODIS) at 1km spatial resolution for Uttarakhand region. Uttarakhand falls under biodiversity rich region and consists of enormous resources such as forests, glaciers, river etc. The machine learning algorithm SVM and RF was trained using AGBD GEDI, NDVI and EVI data. RF achieved an absolute training accuracy having root mean square error (RMSE) of 38.95 Mg/ha, Pearson correlation coefficient (r) of 0.89 and bias as -0.01 whereas testing conveyed RMSE, r and bias as 71.09Mg/ha,0.45 and -0.02 respectively. On the basis of correlation derived from RF model between biomass, NDVI and EVI, a potential biomass map was generated using raster prediction in R software package. The predicted biomass map has biomass ranging from 67.05Mg/ha to 348.65Mg/ha as compared to the GEDI L4B product biomass map provided which ranges from 52.64 to 400.43Mg/h. These biomass datasets show a moderate correlation (r=0.41), but, show a similar pattern of distribution across the study area. RF outperformed the SVM (polynomial) and SVM (RBF) but further need to explore the biomass estimation using some deep learning (i.e., CNN) and ensemble models.