Geometry- and Physics-Aware Dataset Creation for Shadow Removal in High-Resolution Satellite Imagery

Shadows cast by terrain and tall structures are a persistent limitation in satellite imagery, since they degrade radiometric consistency and compromise downstream tasks such as classification, detection, and 3D reconstruction. In this context, machine learning methods for shadow removal provide a flexible and easy-to-deploy tool to assist satellite remote sensing tasks.  Nevertheless, one prominent issue for its development in Earth Observation (EO) is the scarcity of publicly available, geometry-consistent paired shadowed/shadow-free satellite data. Most EO resources support shadow detection or 3D modelling but not shadow removal, while existing shadow-removal datasets largely target ground-level or UAV imagery and do not reflect multi-date, multi-angle satellite acquisition.

To address this gap, we present deSEO, a physics-informed, geometry-aware methodology that converts anyinto paired training data for weakly supervised satellite shadow removal. We exemplified our procedure on the S-EO satellite dataset. Using the multi-temporal, multi-geometry S-EO dataset (WorldView-3 imagery with DSM priors, simulated shadow masks, and RPC camera models), deSEO selects a minimally shadowed acquisition per tile as a proxy reference and pairs it with more shadowed dates under explicit temporal and geometric constraints. Residual off-nadir parallax is mitigated through orientation normalisation and feature-based registration (LoFTR + RANSAC), yielding a per-pixel validity mask that can be used to restrict model supervision to reliably aligned regions.

To validate the usability of the shadow-removal dataset derived from S-EO, we first adapted UAV-oriented methods such as SRNet and pix2pix. However, these approaches fail to converge to a stable training regime under the viewpoint variability typical of satellite acquisitions. We therefore develop a more robust method and training strategy that mitigates this common failure mode of image-to-image translation on multi-date, multi-geometry satellite imagery. Our approach involves training a DSM-conditioned conditional GAN with a U-Net-based generator. The model incorporates perceptual reconstruction and mask-constrained adversarial objectives, with a soft shadow-mask attention prior that emphasises shadow-transition regions. These enhancements overcome the limitations of the classical GAN image translation setup that worked well for UAV data. We evaluate the model on a held-out test split, where the proposed approach achieves a PSNR of 18 ± 1 dB, SSIM of 0.49 ± 0.08, and LPIPS of 0.46 ± 0.05. Notably, improvements were most pronounced at cast-shadow boundaries, and ablation studies revealed that DSM conditioning was the dominant contributing factor, something absent in the SRNet model.

Overall, deSEO provides a reproducible approach to derive paired supervision for satellite shadow removal and establishes a geometry-aware baseline for robust deshadowing under realistic EO acquisition variability.