MT-GSR4B: Multi-Temporally Guided Super-Resolution for Above-Ground Biomass Estimation

Abstract Above-ground biomass (AGB) is a critical indicator of various ecological parameters, such as forest degradation or carbon storage. However, obtaining ground-truth AGB measurements is costly and time-intensive, requires expert knowledge, and is limited to small areas. Current methods estimate AGB using observable parameters like canopy height, often relying on spaceborne platforms. Despite advances in satellite sensor technology, the spatial resolution of AGB maps remains limited (e.g., $$100 \times 100$$ meters in ESA CCI), compared to the resolution of some openly available satellite imagery, such as ESA Sentinel data. To increase the spatial resolution of AGB maps, Karaman et al. (2025) proposed a guided super-resolution approach in which a low-resolution biomass map is upsampled to a resolution of $$10 \times 10$$ meters using a Sentinel-1/2 image as a guide. Here, we show that incorporating multi-temporal satellite data, rather than a single time step, further enhances the accuracy of high-resolution AGB maps by 185 Mg/ha MAE and 266 Mg/ha RMSE on the BioMassters dataset. In addition, we analyze the impact of different input features and design modifications on model performance and provide a rationale for their effectiveness.