Hybrid U-Net: Semantic segmentation of high-resolution satellite images to detect war destruction

Destruction caused by violent conflicts plays a significant role in understanding conflict dynamics, a growing area in economics and political science. However, existing data on conflict impacts typically come from news or eyewitness reports, which can be incomplete, unreliable, or biased. Using satellite imagery and deep learning techniques, the authors aim to extract objective, automated information on violent events.

The authors construct a dataset of high-resolution satellite images of Syria, manually annotating destroyed areas at the pixel level. Using this dataset, they train and test semantic segmentation models to detect building damage of various sizes. A U-Net model is chosen for its strong performance on small and imbalanced datasets, although the raw U-Net architecture does not fully exploit multi-scale feature maps—an important factor for fine-grained segmentation, especially with high-resolution images.

To address this limitation, the paper proposes a multi-scale feature fusion approach and designs a multi-scale skip-connected Hybrid U-Net for segmenting high-resolution satellite images. Experiments show that Hybrid U-Net and its variants produce strong segmentation performance in detecting various types of war-related building destruction.

Notably, the Hybrid U-Net significantly improves segmentation results over standard U-Net:

• Mean Intersection over Union (mIoU) improved by 7.05%
• Mean Dice Score improved by 8.09%

These findings demonstrate that the proposed Hybrid U-Net architecture is more effective at detecting war-related destruction from high-resolution satellite imagery compared to baseline models.