Frequency-conditioned spatially adaptive decoding for bitemporal remote sensing segmentation

Semantic segmentation of bitemporal remote sensing images is challenging due to scene heterogeneity, fine structural boundaries, temporal variations, and high computational cost. To address these issues, this paper proposes ASGS-UNet (Adaptive Shearlet-Gated Skip-Exit UNet), a unified framework that integrates frequency-aware feature learning with spatially adaptive decoding. The proposed model introduces a learnable shearlet front-end to capture multi-scale directional features, cross-frequency gated skip connections to regulate encoder-decoder information flow, and per-pixel exit modules that dynamically adjust decoding depth based on local semantic confidence. This design enables effective representation of both spatial and frequency characteristics while reducing redundant computation in homogeneous regions. Experimental results on the SECOND dataset demonstrate strong performance with 91.6% mIoU and 95.3% F1-score. The model also generalizes well to fused SEN12MS imagery, achieving 88.2% mIoU without retraining. Furthermore, adaptive decoding reduces computational cost to 34.2 GFLOPs with an inference latency of 41 ms. These results indicate that combining frequency-aware representation learning with adaptive inference improves segmentation accuracy, boundary preservation, and computational efficiency for large-scale remote sensing applications.