Remote sensing image dehazing based on frequency-domain feature-guided self-attention mechanism

Abstract

Existing remote sensing image dehazing methods are restricted by narrow receptive fields of spatial convolution operators, which fail to model long-range contextual dependencies and recover delicate high-frequency textures. To address this issue, this paper proposes a lightweight Frequency-Feature-Guided Stripe-Window Self-Attention Network (FFG-SANet) for remote sensing haze removal. We construct a bidirectional cross-stripe self-attention module to capture large-scale ground object context, paired with a multi-scale Fourier frequency branch to compensate missing fine textures with marginal computational burden. Comprehensive experiments across four synthetic and real hazy remote sensing benchmarks validate the independent contribution of the two core modules via split ablation studies. Compared with the competitive DehazeFormer, FFG-SANet lifts PSNR by 0.1dB while cutting parameter volume by nearly half. Supplementary inference efficiency tests further demonstrate that our network achieves lower latency and higher throughput than mainstream lightweight dehazing models. The proposed method substantially improves visual clarity and radiometric fidelity of degraded remote sensing imagery, providing reliable data support for downstream tasks including target detection, environmental monitoring, disaster assessment and refined urban planning, with promising practical deployment prospects.