Bridging Temporal and Spatial–Spectral Features With Satellite Image Time Series: TAS2B-Net for Crop Semantic Segmentation

Abstract

Semantic segmentation based on satellite image time series (SITS) is fundamental to a wide range of geospatial applications, including land cover mapping and urban development analysis. By integrating crop phenological dynamics over time, SITS provides richer spatiotemporal information than static satellite imagery. However, existing models fail to effectively process the temporal and spatial–spectral dimensions of SITS independently, leading to reduced segmentation accuracy. In this letter, we propose a temporal aggregation spatial–spectral bridge network (TAS2B-Net), a novel architecture designed to extract fine-grained crop features from SITS. The network consists of two key components: the pixel-aware grouping temporal integrator (PGTI), which captures temporal dependencies within pixel groups, and the edge-aware contextual fusion head (ECFH), which enhances spatial boundary and global structural representation. Additionally, we introduce a lightweight multiscale spectral decoder (LMSD) to aggregate contextual information across multiple spectral scales, further improving feature learning for semantic segmentation. Extensive experiments on the panoptic agricultural satellite time series (PASTIS) and MTLCC datasets show that the proposed network achieves mIoU scores of 68.91% and 84.59%, respectively, outperforming eight state-of-the-art (SOTA) methods and setting new benchmarks for SITS-based semantic segmentation.