Nonuniform Haze Removal via Dual-Domain Complementary Strategy and Cross-Subband Transformer for Remote Sensing Images

Abstract

Remote sensing images (RSIs) are frequently contaminated by haze interference, which causes spectral distortion and degradation, hindering subsequent interpretation tasks. Current haze removal restoration faces two key challenges: 1) homogeneous processing strategies applied across frequency components fail to account for their differential sensitivity to haze interference and 2) nonuniform haze interference leads to severe loss of local information in heavily obscured patches. To address this issue, we propose a novel dual-domain complementary strategy (DDCS) for nonuniform haze removal, which decomposes the restoration process into two pathways. In the wavelet domain pathway, we innovatively integrate a transformer with wavelet analysis into a cross-subband transformer (CST). It explicitly separates frequencies, reconstructing structural coherence from low-frequency (LL) bands and restoring fine details via interband attention across high-frequency components. In the spatial domain, we propose haze-aware self-learning (HASL) that performs in-image feature learning and transfer. Without relying on external exemplars, HASL estimates haze concentration, identifies clear regions within the same scene as reference sources, and adaptively transfers their features to heavily degraded patches via a learnable compensation mechanism. These two branches are adaptively fused via a dual-stream aggregation gate (DSAG), enabling complementary restoration. Extensive experiments demonstrate that the proposed method achieves state-of-the-art (SOTA) performance in terms of both quantitative metrics and visual quality, particularly in recovering spectral fidelity and spatial details.