A COMPREHENSIVE APPROACH TO ABOVE-GROUND BIOMASS ESTIMATION USING MULTI-SOURCE DATA INTEGRATION AND ADVANCED LEARNING TECHNIQUES

The measurement of biomass across multiple crops is critical for optimizing resource utilization, projecting yields, and soil fertility. As part of this research on the development of remote sensing systems for biomass estimation, we have conducted multiple studies related to modelling and methods for the analysis of multisource data. The research objective presented in this paper is the estimation of above-ground biomass using multisource remote sensing data. This study examines the effect of integrating data from several sources, comprising spectral reflectance from Sentinel 2 and spectral vegetation indices (NDVI, GDVI, SIPI, NDRE, SAVI, and RECL) sensitive to canopy structure, chlorophyll content, and soil variables (nitrogen content, organic carbon, and water content). We have used various machine-learning and deep-learning approaches to estimate the biomass. The results of our investigation showed that all the models significantly improved their training accuracy when used with soil data. The results show slight improvement in Random Forest, where the R2 score improved from 0.75 to 0.78 and RMSE decreased from 42.26 Mg ha-1 to 38.75 Mg ha-1 followed by PNN, BNN, and XGBoost. However, in the test dataset, the BNN showed a significant improvement in the R2 value from 0.54 to 0.66. Interestingly, the BNN model achieved an average RMSE test accuracy of 59 60 after multiple runs. While, the XG Boost model showed only a slight improvement in its performance. This depicts the importance of using a complex multimodal dataset and its relevance in the context of precision farming. Also, the ability of the BNN model to uncover complicated correlations and generalizability in varied agroecosystems.