Shadows from tall structures are a persistent problem for remote sensing segmentation: they change how surfaces appear in the imagery and blur the edges of objects, which can confuse models. The new paper from arXiv introduces ShadowCast, a method that addresses this by augmenting training data with realistic shadows.
ShadowCast performs on-the-fly, context-aware shadow augmentation. Rather than adding generic shadow shapes, it takes the surrounding scene context into account when generating shadows, so the synthetic shadows resemble those that would actually be cast by elevated objects. The goal is to make segmentation models less sensitive to shadow effects.
The abstract notes that most existing approaches handle shadows as a separate or post-hoc concern, whereas ShadowCast builds them directly into the training process. The paper is available on arXiv under identifier 2610.23440; the provided abstract is truncated, so full evaluation details are not available here.