F-CSFT: FiLM-driven cross-scale factorized spatio-temporal transformer for robust tree species classification in subtropical forests

Accurate tree species identification and mapping are crucial for forest management. While satellite image time series (SITS) are widely used for this purpose, subtropical forests present challenges due to limited dense temporal optical data. Leveraging multi-scale spatial features offers a promising solution, but integrating spatial and temporal features remains difficult. In this study, we propose F-CSFT, a FiLM-driven cross-scale factorized spatio-temporal Transformer, to address this issue. F-CSFT fuses multi-scale spatial features from high-resolution Jilin-1 images with temporal features from Sentinel-2, employing cross-scale multi-head attention and FiLM-based conditional modulation for enhanced spatio-temporal fusion. We evaluated F-CSFT on nine subtropical tree species across two regions of the Ta-pieh Mountains, comparing its performance with five state-of-the-art (SOTA) models. Results show that incorporating multi-scale spatial features significantly improves classification accuracy, increasing overall accuracy (OA) and kappa from 86.61% and 83.34% using temporal data alone to 91.10% and 88.93%. F-CSFT consistently outperformed other methods, with OA (kappa) improvements ranging from 2.88% (3.63%) to 6.07% (7.32%). Ablation experiments further demonstrate the critical contribution of FiLM modulation and factorized Transformer architecture to robust spatio-temporal fusion. Our findings highlight the potential of combining multi-scale spatial and temporal features to enhance tree species classification in subtropical forests.