Neural network restores color and clarity in sandstorm images

Sandstorms do more than darken the sky; they warp the images captured through them. Photographs taken during these events suffer from a combination of problems at once: a strong color cast that can make scenes look unnaturally yellow or red, dim lighting, and a veil of haze caused by suspended sand particles. A study published in Scientific Reports proposes a new deep learning approach, called RSTNet, designed to correct all of these distortions in a single pipeline.

The core insight behind the method is that color distortion in sandstorm images stems from imbalanced color channels, and that this imbalance is closely tied to an image’s saturation. Rather than adjusting color channels directly, which the authors note can introduce artificial-looking color components when the correction overshoots, RSTNet uses a convolutional neural network to estimate what they call residual saturation. By working with saturation, which strongly reflects a scene’s color information, the network aims to rebalance the channels in a way that keeps the enhanced result looking natural.

Color correction alone, however, does not solve the second major problem. Even after the channels are rebalanced, the image can remain hazy because fine sand particles in the atmosphere scatter light between the scene and the camera. To address this, the framework incorporates a second network that estimates transmission, a quantity used in image dehazing that describes how much light from the scene reaches the camera unscattered. With an estimated transmission map, the system can reverse the hazy effect and recover scene detail hidden behind the sand-laden air.

The full architecture, named a residual saturation and transmission estimation network, therefore chains two learned stages: one that estimates residual saturation for color balancing, and one that estimates transmission for dehazing. According to the paper, the enhanced images produced by RSTNet appear natural, and the authors report that the method was verified to be suitable for enhancing degraded sandstorm images. The article was published open access on 24 October 2026, with Hosang Lee as the author, and is being shared early as a citable, peer-reviewed accepted version that will be replaced by the final Version of Record after further edits.

The study addresses a genuine gap in the image enhancement literature. Existing sandstorm enhancement methods typically improve color components and mitigate haze separately, but the authors point out that enhancing color channels without accounting for balance can create artificial color components, an artifact that undermines the usefulness of the restored image. By coupling saturation-based color balancing with transmission-based dehazing inside one network framework, RSTNet is presented as an integrated answer to the diverse degradation that sandstorm images exhibit, from color casting and dim conditions to haze and channel imbalance.

Subject of Research: Deep learning enhancement of sandstorm-degraded images using residual saturation estimation and transmission-based dehazing

Article Title: RSTNet: sandstorm image enhancement using a residual saturation and transmission estimation network

Article References: Lee, H. (2026). RSTNet: sandstorm image enhancement using a residual saturation and transmission estimation network. Scientific Reports. https://doi.org/10.1038/s41598-026-76471-3

Image Credits: AI Generated

DOI: 10.1038/s41598-026-76471-3

Keywords: RSTNet, sandstorm images, image enhancement, convolutional neural network, color balancing, image saturation, transmission estimation, dehazing, color cast, Scientific Reports, computer vision, haze removal