Frequency-conditioned spatially adaptive decoding for bitemporal remote sensing segmentation.

Area of Science:

• Computer Vision
• Remote Sensing
• Image Analysis

Background:

• Semantic segmentation of bitemporal remote sensing images faces challenges like scene heterogeneity, fine boundaries, temporal changes, and high computational demands.
• Existing methods struggle to balance accuracy with computational efficiency for large-scale applications.

Purpose of the Study:

• To propose ASGS-UNet (Adaptive Shearlet-Gated Skip-Exit UNet), a unified framework to improve semantic segmentation of remote sensing images.
• To enhance feature representation by integrating multi-scale directional information and frequency characteristics.
• To reduce computational cost through spatially adaptive decoding.

Main Methods:

• Developed a learnable shearlet front-end for multi-scale directional feature extraction.
• Implemented cross-frequency gated skip connections to control encoder-decoder information flow.
• Introduced per-pixel exit modules for dynamic decoding depth adjustment based on semantic confidence.

Main Results:

• Achieved 91.6% mIoU and 95.3% F1-score on the SECOND dataset.
• Demonstrated generalization to fused SEN12MS imagery with 88.2% mIoU without retraining.
• Reduced computational cost to 34.2 GFLOPs with 41 ms inference latency.

Conclusions:

• Combining frequency-aware learning with adaptive inference significantly improves segmentation accuracy, boundary preservation, and computational efficiency.
• ASGS-UNet offers a promising solution for large-scale remote sensing semantic segmentation.
• The adaptive decoding approach effectively reduces computational load in homogeneous regions.