PHENet: Toward Robust Fog-Resilient Change Detection in Remote Sensing Imagery With a Physics-Guided Height-Enhanced Network

Change detection from remote sensing imagery faces significant challenges under foggy conditions, where fog and haze severely degrade image quality and compromise detection accuracy. Current methods for foggy conditions rely on generic image enhancement without considering the physics-based correlation between fog distribution and terrain elevation. This paper presents PHENet, a physics-guided height-enhanced network that leverages the intrinsic correlation between atmospheric scattering and terrain elevation for robust change detection under foggy conditions. The proposed approach introduces three synergistic components: (1) Height-Aware Mixture Fusion module that leverages frequency-domain decomposition for elevation-conditioned spatial feature enhancement, selectively suppressing fog-related artifacts while preserving structural details; (2) Height-Enhanced Difference (HED)-Augmented Difference Fusion Module that incorporates height-derived change indicators invariant to atmospheric variations, significantly reducing false positives from weather-induced spectral changes; and (3) Physics-Guided Adversarial Refinement Module that enforces consistency with atmospheric scattering principles during training without requiring paired clear-foggy data. Comprehensive experiments on foggy variants of WHU-CD and LEVIR-CD datasets demonstrate that PHENet achieves superior performance across all fog densities. Furthermore, the proposed method maintains a compact architecture with 5.24M parameters, demonstrating that physics-informed design can deliver robust performance without excessive model complexity. This code can be available at https://github.com/sssy3/PHENet.