A spectral-spatial-temporal deep learning model for large area aboveground biomass mapping of semiarid forest using Sentinel-1 and 2 time series

Aboveground biomass (AGB) serves as a crucial metric for evaluating carbon sequestration within terrestrial forest ecosystems. Notably, semiarid forests cover approximately 15% of the Earth’s land surface, playing a considerable role in global primary productivity. However, owing to the high spatiotemporal heterogeneity of these ecosystems, reliable biomass estimation remains a substantial challenge. To address this gap, we developed a spectral-spatial-temporal features enhancement transformer (SST-TF) model, using diverse dense time series data configurations from optical and SAR sources and their derivatives. By integrating multi-modal spatio-temporal features and introducing a multi-head attention mechanism, the traditional decoder layer and position encoding module were replaced. This substitution overcomed the limitations of traditional architectures in effectively handling complex spatio-temporal data and enhanced the model's feature extraction capability. The GEDI Level 4A product provides reference AGB values for model training and validation. The multivariate methods with unbiased variable selection in R language (MUVR) are used to identify optimal features at different scales (min, mid and max). Results showed that high temporal resolution improved performance. The SST-TF outperformed other methods (R2 increased by up to 25.6%). Best results used mid-scale features with 5-day resolution combining optical and SAR variables (R2: 0.765; RMSE: 43.960; RMSEr: 31.211 Mg/ha).