DEMANet: A dehazing enhanced multi-branch attention network for remote sensing images

Abstract

Haze represents a key constraint on the application of optical remote sensing imagery. It not only impairs visual quality but also lowers the accuracy of remote sensing interpretation tasks such as classification and change detection. To tackle this issue, we present a Dehazing Enhanced Multi-branch Attention Network (DEMANet) for effective remote sensing image dehazing. The network adopts a U-Net-like structure with three hierarchical downsampling stages to implement progressive feature extraction from shallow to deep layers. Shallow and middle layers use a residual dual-path module to enhance local detailed features via a main-auxiliary dual-branch structure. Deep layers employ a dual-attention module with a two-layer attention mechanism to break the limitation of local receptive fields in traditional convolutions and accurately capture global semantics and haze features. A cross-stage feature interaction module embeds a haze-guided mechanism to locate haze regions based on edge and color differences, enabling cross-stage feature alignment and interaction between encoding and decoding. This reduces information loss during upsampling and improves detail preservation and dehazing performance. Experiments conducted on widely used public remote sensing datasets demonstrate that our innovative approach outperforms existing algorithms in haze removal, while simultaneously preserving intricate image details and color fidelity.