Learning a lightweight multiscale bidirectional network for remote sensing image dehazing

Abstract

Remote sensing images acquired under hazy atmospheric conditions often exhibit reduced contrast, color distortion, and loss of fine structures, which adversely affect visual interpretation and subsequent tasks such as land cover classification and object detection. Existing dehazing approaches either rely on computationally intensive transformer architectures or convolutional models with limited global modeling capacity, making it difficult to balance performance and efficiency for high-resolution remote sensing imagery. We propose a Lightweight Multiscale Bidirectional Network (LMBNet) for remote sensing image dehazing. The framework adopts a hierarchical encoder–decoder architecture to capture degradation patterns across different spatial resolutions. A Sparsity-Compensated Dual-stream Transformer (SCDT) block is introduced as the core feature extractor. The Sparsity-Compensated Self-Attention (SCSA) mechanism employs rank-based masking to suppress irrelevant responses and enhance discriminative nonlocal aggregation. In parallel, a Dual-Stream Feed-forward Network (DSFN) leverages multiscale depth-wise convolutions and differential modulation to refine spatial details and reduce redundancy. A composite loss integrating Charbonnier, edge, and frequency terms further improves structural fidelity. Experimental results on benchmark datasets including StateHaze1k and RICE demonstrate that the proposed method achieves competitive performance with only 14.31 M parameters, yielding up to 29.58 dB in peak signal-to-noise ratio (PSNR) and 0.9411 in Structural Similarity Index (SSIM), indicating its suitability for practical remote sensing applications