DiffShadow-AI: A Novel Differentiable Geometry-Aware Framework for Joint Cloud Shadow Detection and Cloud Height Estimation From Satellite Imagery
Year: 2026
Type: article
Abstract: Cloud shadow detection in moderate-resolution satellite imagery remains an open problem, as coarser pixel spacing blurs shadow boundaries and increases spectral confusion with dark surfaces such as open water and bare land. This study introduces DiffShadow-AI, the first differentiable, geometry-aware deep learning framework to jointly detect cloud shadows and retrieve cloud height from a single end-to-end trainable pipeline applied to Visible Infrared Imaging Radiometer Suite (VIIRS) imagery. Unlike existing approaches that treat shadow detection and height retrieval as independent tasks, DiffShadow-AI embeds a differentiable geometric projection module based on soft rasterization directly into the learning pipeline, enabling cloud height to emerge as a physically constrained by-product of shadow delineation rather than requiring a separate retrieval step. The framework combines a Lightweight ResNet18 encoder for multiscale spectral and geometric feature extraction with a joint regression and segmentation objective, ensuring that predicted shadow extents and cloud heights remain mutually consistent throughout training. Evaluated over Alaska under diverse seasonal and illumination conditions, DiffShadow-AI achieves a relaxed intersection-over-union of 85.3%, with Grad-CAM visualizations confirming that the model attends to geometrically meaningful regions at cloud bases and edges. Cloud-top height estimates are physically consistent with both high-resolution rapid refresh (HRRR) numerical weather prediction outputs and Community Satellite Processing Package (CSPP) operational retrievals, with discrepancies in deep convective scenes attributed to differences between the shadow-casting layer and the cloud top. These results demonstrate that embedding differentiable physics into deep learning pipelines is an effective strategy for simultaneous cloud shadow delineation and height inference from moderate-resolution optical imagery.
Authors: Baya Cherif, Mahmoud Ayyad, Marouane Temimi
Institution: Stevens Institute of Technology
Language: English
Subfield: Global and Planetary Change
Field: Environmental Science
Domain: Physical Sciences
SDG: Climate action
Open Access status: gold
Estimated APC paid: $1,800
OpenAlex ID: W7214207331