LAPW-Net: learning adaptive physics-guided Wavelet Dehazing Network

High-resolution remote sensing imagery supports a wide range of Earth observation applications, but atmospheric haze can degrade radiometric quality and obscure structural details. Balancing restoration accuracy and computational cost remains a challenge in image dehazing. This paper presents the Learning Adaptive Physics-guided Wavelet Dehazing Network (LAPW-Net), which combines adaptive wavelet subband scaling with scattering-guided restoration in separate low- and high-frequency pathways. The Adaptive Scattering-Guided Low-Frequency Processing Block estimates spatially varying atmospheric parameters for low-frequency reconstruction, while the Multi-Scale High-Frequency Processing Block applies multi-scale feature extraction and learnable directional weighting to process structural details. Quantitative evaluations on synthetic benchmarks show competitive restoration performance and strong average ranks across the evaluated generalisation datasets. Qualitative comparisons on real-world imagery provide additional observations of haze removal and detail preservation. Ablation experiments indicate that the complete configuration outperforms the evaluated partial and replacement variants in restoration metrics, while a reference-informed proxy experiment assesses preservation of appearance features. The network variants exhibit different trade-offs between restoration accuracy and computational cost. Together, these experiments characterise the performance and limitations of combining adaptive frequency processing with scattering-guided restoration for remote sensing image dehazing.