Multi-Spectral Gradient Residual Network for Haze Removal in Multi-Sensor Remote Sensing Imagery

Abstract

Remote sensing imagery is widely used for various Earth surface monitoring applications. However, the quality of these observations can be degraded by clouds during image acquisition. Haze, a type of thin cloud, commonly causes atmospheric absorption and scattering of visible light, resulting in partially obscured regions. Haze removal is an active research area with two main approaches: physics-driven computer vision and end-to-end data-driven machine learning. To leverage both approaches, we propose a deep neural network framework that utilizes large-scale multi-sensor data and geometric knowledge from image physics. This is achieved through a multi-spectral gradient residual network. This network transfers structural details from near-infrared (NIR) images, which have better haze penetration, to the visible (RGB) bands. During training, we incorporate a soft constraint using the partially available information under haze conditions. This constraint helps the model maintain atmospheric consistency, a concept commonly used in physical haze models. We validated our model’s performance on a multi-sensor benchmark dataset containing Landsat-8 and Sentinel-2 satellite images. Comparisons with state-of-the-art methods demonstrate significant improvements. Our model achieves a minimum of 18.71% improvement on MSE, 25.9% on SSIM, and 6.25% on MS-SSIM compared to the next best method. It also shows advancements in LPIPS (14.43%) and SAM (8.47%) measures.