Aboveground biomass estimates of tropical mangrove forest using Sentinel-1 SAR coherence data - The superiority of deep learning over a semi-empirical model

Abstract

Abstract The availability of advanced Machine Learning algorithms has made the estimation process of biophysical parameters more efficient. However, the efficiency of those methods seldom compared with the efficiency of already established semi-empirical procedures. Aboveground biomass (AGB) of mangrove forests is a crucial biophysical parameter as it is positively correlated to the carbon stocks and fluxes. The free availability of Sentinel-1 C-band SAR data and machine learning algorithms hold promises in estimating AGB of tropical mangrove forests. We reported high AGB (70\u202ft/ha to 666\u202ft/ha) using 185 field quadrats of 0.04ha each from Bhitarkanika Wildlife Sanctuary, located on the eastern Indian coast that could be attributed to species composition. The AGB maps generated using Interferometric Water Cloud Model (IWCM) and Deep Learning models were different from each other as they rely on different variables. IWCM was more dependent, especially on ground and vegetation components of coherence, while canopy height acted as the most crucial variable in the Deep Learning model. However, the negligible variations in Deep Learning-based AGB maps can be attributed to interpreting the importance of coherence and VH backscatter. Due to low canopy penetration power of C-band SAR, high temporal decorrelation resulting from longer time gap between interferometric image pairs, and high spatial heterogeneity of mangrove forests, IWCM found as an unsuitable method for AGB estimation. Interestingly, a Deep Learning algorithm could translate the exact relationship between predictor variables and mangrove AGB in Bhitarkanika Wildlife Sanctuary. The AGB estimation studies in mangrove forests using Sentinel data should focus more on using machine learning algorithms like Deep Learning rather than semi-empirical models.