TSSRes2Former: A Spatio-Temporal–Spectral Residual Transformer Network for Satellite Image Time Series Classification

The joint temporal, spatial and spectral information integrated by satellite image time series (SITS) not only reflects the comprehensive attributes and intrinsic physical properties of land cover types, but also captures their dynamic evolution, which is an important basis for high-precision land cover classification. Current classification approaches based on SITS data often input stacked multi-temporal images into models without distinguishing between spatial, spectral, and temporal characteristics, relying solely on model-driven feature extraction. However, the intrinsic heterogeneity of these dimensions means that undifferentiated processing can obscure deep semantic features, thereby limiting performance in complex scenarios. Recent studies have shown that separately modeling features from different dimensions can enhance classification accuracy. Additionally, extracting global-local multi-scale features across time remains a crucial challenge in time-series classification. We argue that explicitly modeling domain knowledge and classification experience in remote sensing, through targeted enhancement of spatio-temporal-spectral representations prior to model input, presents a promising paradigm for understanding the unique information structures of SITS. So, we propose the Temporal-Spectral-Spatial Residual Transformer (TSSRes2Former) network, which integrates both data-level representation enhancements and model-level architectural innovations to fully exploit and synergize temporal, spatial, and spectral features for high-precision classification. Specifically, we design dedicated enhancement strategies for each dimension, followed by a hybrid neural architecture that combines Res2Net and Vision Transformer, enabling multi-scale local feature extraction and both short- and long-range dependency modeling. Extensive experiments conducted on multiple publicly available datasets demonstrate the effectiveness and superiority of our approach.