deSEO: Physics-Aware Dataset Creation for High-Resolution Satellite Image Shadow Removal

Shadows from terrain and tall buildings severely hinder high-resolution satellite image analysis, yet no public paired shadow/shadow-free datasets exist for training deep learning models. To close this gap, Beltrame et al. developed deSEO, a geometry-aware and physics-informed pipeline that derives paired supervision from the existing S-EO shadow detection dataset. For each tile, deSEO selects a minimally shadowed acquisition as a weak reference and pairs it with shadowed counterparts through temporal and geometric filtering, Jacobian-based orientation normalization, and LoFTR-RANSAC registration. A per-pixel validity mask restricts learning to reliably aligned regions, mitigating off-nadir parallax.

The team also built a DSM-aware deshadowing model combining residual translation, perceptual objectives, and mask-constrained adversarial learning. While a direct adaptation of UAV-based SRNet/pix2pix failed under satellite viewpoint variability, their model consistently reduces the visual impact of cast shadows across diverse illumination and viewing conditions. Experiments on held-out scenes demonstrated improved structural and perceptual fidelity. deSEO thus provides the first reproducible, geometry-aware paired dataset and baseline for satellite shadow removal, advancing Earth observation and downstream tasks like classification and 3D reconstruction.