A Multi-Stage Deep Learning Framework for Shadow Detection in Aerial Orthophotos

Abstract. Shadow correction is an important preprocessing step not only for visual enhancement but also for improving object recognition performance in remote sensing imagery. Although many datasets and deep learning models have been proposed for shadow detection and removal, most of them focus on natural images. In contrast, high-resolution aerial orthophotos contain large continuous shadows caused by tall buildings, especially in urban areas, and existing models often fail to handle such large-scale structures effectively.
In this study, we construct a new shadow annotation dataset specifically designed for aerial orthophotos with spatial resolutions of 20 cm/pixel and 5 cm/pixel. Furthermore, we propose a three-stage multi-resolution segmentation framework that progressively refines shadow predictions from low to high resolution. Predictions from lower-resolution stages are used as auxiliary information to guide higher-resolution prediction. Experimental results demonstrate that the proposed approach improves fuzzy Intersection over Union (IoU) by approximately 0.05 compared with a previously published shadow detection model, and also outperforms a single-stage baseline, particularly for large continuous shadow regions. The framework is also applicable to other large-scale segmentation tasks requiring extensive receptive fields.