A research team from Wuhan University has developed a new artificial intelligence framework that can reconstruct partially hidden objects in satellite imagery with greater accuracy than existing methods. The approach, called Remote Sensing Amodal Completion (RSAC), moves beyond simple pixel filling to infer an object's full geometry, surface texture, and semantic identity from incomplete observations.
Satellite imagery is widely used in disaster response, urban planning, and environmental monitoring, but ground objects are often obscured by clouds, overlapping structures, or imaging angles. Traditional inpainting methods may produce visually plausible results but can distort object structure or generate incorrect content. The RSAC framework addresses these limitations by combining diffusion-based generation with remote-sensing-specific structural guidance.
The study, published in the Journal of Remote Sensing on April 7, 2026, proposes a Dual-Adaptive Diffusion-Based Framework. It adapts Stable Diffusion to the remote sensing domain using Low-Rank Adaptation (LoRA) and a four-channel ControlNet that uses image and mask information to guide completion. A prior-enhanced initialization strategy preserves low-frequency information from visible object parts, improving physical consistency.
In comparative experiments against methods like Stable Diffusion Inpainting and LaMa, the proposed framework achieved an Intersection over Union (IoU) of 0.853, an amodal completion IoU of 0.688, and a structural similarity index of 0.930. The method also produced 100% valid-output coverage and outperformed baselines in geometry and texture continuity.
The team built a dedicated dataset with 1,770 annotated instances across 10 categories, including planes, ships, and sports fields. The framework improved downstream object detection and supported layered 2.5D scene understanding. Researchers emphasized that the goal is to help machines infer what an object is and how it should be structured, not just make images look complete.
This technology could enhance geospatial intelligence in scenarios where objects are frequently obscured, such as post-disaster assessment and infrastructure mapping. Future work may extend the framework to more object categories, drone perspectives, and multimodal data. The study was supported by the National Natural Science Foundation of China under grants 42422109 and 42371366.
For more details, see the original study at https://doi.org/10.34133/remotesensing.1035.
