Unexploded Ordnance Mapping Across the Ukrainian Front-Lines using Commercial VHR Optical Satellite Imagery and Semantic Segmentation Convolutional Neural Networks

The difficulty in estimating unexploded ordnance (UXO) dangers in conflict zones is significant. However, utilizing very-high-resolution (VHR, <5 meter) commercial multispectral satellite imagery, and a supervised U-Net semantic segmentation convolutional neural network (CNN), artillery and rocket craters may be detected and mapped on an individual basis. The detection of exploded ordnance craters indicates the high likelihood of unexploded ordnance in the near vicinity, and thus artillery crater detection acts as a proxy for unexploded ordnance likelihood on a fine scale. In this study, we used a CNN model based upon the U-Net architecture to segment artillery craters in pansharpened multispectral images from the Maxar WorldView-3 satellite. The U-Net CNN was trained on rocket and artillery craters from the 2014 conflict in the Eastern Ukrainian provinces of Donetsk and Luhansk, as well as craters from Kherson and Mykolaiv Oblasts in the continued hostilities in 2022. We calculated sensitivity, precision, and F1 score on a stratified random sample to evaluate the models performance in detecting artillery and rocket craters. The trained CNN model was able to accurately detect 89% of craters compared to human marking, indicating its proficiency at the task of crater detection. The UXO detection methodology has been applied to regional studies in Kherson, Donetsk, and Mykolaiv Oblasts with contemporary conflict imagery. The resulting detection maps show UXO risk areas on a large scale, which provides valuable information for the post-conflict rehabilitation of Ukraine. The detections also aid our understanding of the wide-reaching consequences of armed conflict on agriculture, health, and the economy of front-line areas.