Vision Transformer-Based Image Dehazing for Climate-Resilient Maritime Navigation

Abstract

As climate change intensifies, maritime transportation systems face critical challenges from frequent fog, high humidity, and adverse weather, which degrade visual perception and threaten navigation safety. To address these climate-induced disruptions, we propose TWRM-Net, a Transformer-based dehazing network designed for climate-resilient maritime vision. Firstly, the network employs a hierarchical encoder–decoder backbone that integrates convolutional priors with window-based self-attention, enabling joint modeling of local textures and global haze structures for maritime haze removal. Secondly, we introduce a Dual-Residual Attention Block (DRAB), which enhances structural awareness and haze localization by combining spatial and channel attention with residual learning, thus improving robustness in fog-dense environments. Lastly, the adaptive fusion module (AFM) adaptively balances low-frequency global context from the encoder and high-frequency details from the decoder, ensuring frequency-selective feature integration while suppressing aliasing artifacts. We construct a realistic maritime haze dataset using monocular depth estimation and the atmospheric scattering model, simulating climate-induced haze conditions. Extensive experiments on synthetic and real-world maritime datasets demonstrate that TWRM-Net achieves state-of-the-art performance, with PSNR of 35.07dB, SSIM of 0.9735, LPIPS of 0.0880, and FID of 23.65, significantly outperforming existing methods. These results highlight the effectiveness of our approach for climate-adaptive intelligent transportation, providing reliable perception for safe and sustainable maritime navigation in the era of climate change.