SA-BEF: Deep Fusion of Sentinel-1/2 for Forest Biomass Estimation with Occlusion-Aware Attention Mechanisms

The mitigation of climate change and sustainable forest management in the world needs proper techniques of forest biomass monitoring that aids in assessing carbon sequestration, as well as forest resource management of various ecosystems. Understanding vegetation dynamics heavily depends on biomass estimation through traditional methods, although these prove inefficient as they require on-field measurements, which are labour-intensive, highly time-consuming, and limited when extended across large-scale monitoring. Aerial remote sensing devices provide scalable alternatives, however, they are limited in handling data occlusion and missing data as well as the integration of multitemporal information. This article proposes a Spatiotemporal Attention-Biomass Estimation Framework (SA-BEF) that combines Sentinel-1 Synthetic Aperture Radar (SAR) and Sentinel-2 Multispectral Instrument (MSI) data to enhance biomass estimation. SA-BEF’s novel architecture includes specialized components for occlusion reconstruction and feature extraction, coupled with an attention mechanism that prioritizes critical temporal and spatial features. SA-BEF is trained and validated on the Biommaster’s Benchmark Dataset, demonstrating good biomass estimation performance. It showcases exceptional biomass estimation capabilities, achieving impressive results across various evaluation metrics with MAE of 9.966, RMSE of 13.14, rRMSE of 20.49, \(R^2\) of 0.751, MAPE of \(24.04\%\) and Pearson Correlation of \(86.70\%\) in comparison to state-of-the-art methods. Its ability to handle occlusion while keeping the relationship between spatial and temporal elements makes the framework best suited to support large-scale forest monitoring applications.