deSEO: Physics-Aware Dataset Creation for High-Resolution Satellite Image Shadow Removal
摘要
由地形和高大建筑物投射的阴影仍是高分辨率卫星图像分析的主要障碍，降低了 classification、detection 和 3D reconstruction performance。缺乏提供几何一致的 shadow/shadow-free 配对卫星影像的公开资源，大多数 Earth-observation 数据集设计用于 shadow detection 或 3D modeling，而非 removal。现有的 deep shadow-removal 数据集要么针对地面场景或航空场景，要么依赖 unpaired 和 weakly supervised 的 formulation，而非 explicit satellite pairs。我们通过 geometry-aware and physics-informed 方法填补了这一空白，deSEO is the first to derive paired supervision for satellite shadow removal from S-EO shadow detection 数据集，通过 fully replicable pipeline 实现。对于每个 tile，deSEO 选择 minimally shadowed acquisition 作为 weak reference，并通过 temporal 和几何 filtering、Jacobian-based orientation normalisation、LoFTR-RANSAC registration 配对 shadowed counterparts。per-pixel validity mask 限制学习到可靠对齐的区域，尽管存在 residual off-nadir parallax，仍可实现 supervision。除了该配对数据集外，我们开发了 DSM-aware deshadowing model，结合 residual translation、perceptual objectives 和 mask-constrained adversarial learning。与直接适配 UAV-based SRNet/pix2pix 架构不同，后者在 satellite viewpoint variability 下无法收敛。我们的模型在 diverse illumination and viewing conditions 下 consistently reduce the visual impact of cast shadows， achieving improved structural and perceptual fidelity on held-out scenes。deSEO 因此提供了第一个 reproducible、geometry-aware 配对数据集和基准，用于卫星 Earth observation 中的 shadow removal。
引用
@article{arxiv.2605.03610,
title = {deSEO: Physics-Aware Dataset Creation for High-Resolution Satellite Image Shadow Removal},
author = {Lorenzo Beltrame and Jules Salzinger and Filip Svoboda and Phillipp Fanta-Jende and Jasmin Lampert and Radu Timofte and Marco Körner},
journal= {arXiv preprint arXiv:2605.03610},
year = {2026}
}
备注
8 pages, 6 figures, 5 tables. Accepted in the annals track at the ISPRS 2026 Congress. Code and materials: https://github.com/AIT-Assistive-Autonomous-Systems/deSEO