Аннотация
In Kazakhstan, most small and medium-sized dams lack geographic coordinates in official records, which hampers GIS-based monitoring. A machine learning framework has been developed by using neural networks trained on satellite images in order to detect small dams (100–400 meters) in remote areas. The freely available satellite images have a low resolution, 10m per pixel, meaning that each dam appears as only 10-40 pixels on the map, making reliable identification difficult. In order to address this, 15,000 fragments of the ESRI Satellite map have been automatically generated for dams at a resolution of 2m per pixel. The coordinates of the dams have been obtained from the database of dams registered in the Global Dam Watch project. These images have been semi-automatically annotated. By using standard augmentation techniques, the dataset has been expanded to 35,336 images. The YOLOv5 and the YOLOv8 model families have been trained to detect dams in the images. Results include detection examples and a description of model performance. Applied to 113,960 basemap fragments covering 67,000 square kilometers of Kazakhstan, the model has identified 467 previously ungeoreferenced structures.