WaDiT: A Water-Aware Diffusion Transformer for Spectral-Consistent Cloud Removal in Multispectral Remote Sensing

Abstract

Cloud contamination severely limits the availability of optical remote sensing data for water quality estimation, which demands strict spectral fidelity rather than mere visual restoration. To address this challenge, we propose the water-aware diffusion transformer (WaDiT), a novel conditional diffusion framework that explicitly preserves water-specific spectral characteristics during cloud removal. WaDiT introduces two key innovations: 1) a water spectral feature extraction (WSFE) module that leverages contrastive learning with band-ratio constraints to encode physically meaningful spectral priors from cloud-free water regions; and 2) a spectral and auxiliary fusion (SAF) module that employs distance-weighted cross-attention to dynamically inject spectral features and geographically selected auxiliary patches into the denoising process. This hierarchical fusion strategy ensures that water spectral signatures near cloud boundaries guide the reconstruction of obscured regions. We construct a dedicated multispectral dataset over Lake Ontario with paired cloudy-clear images and bathymetric metadata. Extensive experiments demonstrate that WaDiT achieves state-of-the-art performance, notably reducing the spectral angle mapper (SAM) by 68% compared to existing methods, validating its effectiveness for spectral-critical water quality applications.