A research team from Wuhan University has developed a new artificial intelligence framework that can reconstruct complete ground objects from partially visible satellite imagery, addressing a longstanding challenge in geospatial analysis. The method, detailed in the Journal of Remote Sensing (DOI: 10.34133/remotesensing.1035), goes beyond traditional image inpainting by inferring an object's full shape, surface texture, and semantic identity from incomplete observations.
Satellite imagery is critical for disaster response, urban planning, and environmental monitoring, but objects are often obscured by clouds, overlapping structures, or imaging angles. Existing inpainting methods often produce visually plausible results that distort object structure or introduce incorrect content. The proposed framework, called Remote Sensing Amodal Completion (RSAC), shifts from scene-level inpainting to object-level reasoning.
The framework adapts Stable Diffusion using Low-Rank Adaptation (LoRA) for the remote sensing domain, while a four-channel ControlNet uses image and mask data to guide structural completion. A prior-enhanced initialization strategy preserves low-frequency information from visible parts, improving physical consistency. In tests against methods like Stable Diffusion Inpainting and LaMa, RSAC achieved superior geometry, clearer boundaries, and realistic texture continuity.
The researchers built a dedicated dataset of 1,770 annotated instances across 10 object categories, including planes, ships, and sports fields. The method achieved an Intersection over Union (IoU) of 0.853, an amodal completion IoU of 0.688, and a structural similarity index of 0.930, outperforming baselines that showed distorted geometry or unrealistic backgrounds.
The technology could enhance geospatial intelligence in scenarios like post-disaster assessment and automated mapping. By restoring complete object morphology, it may also improve training data for detection models and help AI interpret satellite imagery more like human analysts. The team plans to extend the framework to more object categories, drone perspectives, and 3D reconstruction in future work.
This research was supported by the National Natural Science Foundation of China under grants 42422109 and 42371366. For more information, visit Chuanlink Innovations.
