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

Destruction caused by violent conflicts play a big role in understanding the dynamics and consequences of conflicts, which is now the focus of a large body of ongoing literature in economics and political science. However, existing data on conflict largely come from news or eyewitness reports, which makes it incomplete, potentially unreliable, and biased for ongoing conflicts. Using satellite images and deep learning techniques, we can automatically extract objective information on violent events. To automate this process, we created a dataset of high-resolution satellite images of Syria and manually annotated the destroyed areas pixel-wise. Then, we used this dataset to train and test semantic segmentation networks to detect building damage of various size. We specifically utilized a U-Net model for this task due to its promising performance on small and imbalanced datasets. However, the raw U-Net architecture does not fully exploit multi-scale feature maps, which are among the important factors for generating fine-grained segmentation maps, especially for high-resolution images. To address this deficiency, we propose a multi-scale feature fusion approach and design a multi-scale skip-connected Hybrid U-Net for segmenting high-resolution satellite images. In our experiments, U-Net and its variants demonstrated promising segmentation results to detect various war-related building destruction. In addition, Hybrid U-Net resulted in significant improvement in segmentation performance compared to U-Net and other baselines. In particular, the mean intersection over union and mean dice score improved by 7.05% and 8.09%, respectively, compared to those in the raw U-Net.