Exploring the inclusion of Sentinel-2 MSI texture metrics in above-ground biomass estimation in the community forest of Nepal

Abstract

The potential of the improved resolution Sentinel-2 MSI data was explored through texture metrics, vegetation indices (VIs) and pooled dataset using the Random Forest (RF) machine learning algorithm to estimate Above-ground Biomass (AGB) in a sub-tropical forest of Nepal. Texture metrics were derived based on different working window sizes (3 × 3, 5 × 5, 7 × 7 and 9 × 9), and the results were compared with those obtained, using raw traditional bands (Analysis set 1: 2, 3, 4, 8, 11 and 12), raw traditional and red edge bands (Analysis set 2: Set 1 + Band 5, 6, 7 and 8A), and red edge bands (Analysis set 3) only. Comparatively, the use of pooled data (texture and VIs) yielded higher biomass estimates. The results from pooled data based on the 7 × 7 window size resulted in models with better model fitting parameters. For instance, pooled data produced an R2 = 0.99 and a RMSE = 4.51 t ha−1 (relRMSE = 2.82). Further, the RF model selected dissimilarity, variance and mean from Band 2 and SAVI (Soil adjusted vegetation index) as the most important AGB predictor variables. The results demonstrated that like the red-edge bands, traditional bands were equally important in AGB estimation.