Detecting unknown dams from high-resolution remote sensing images: A deep learning and spatial analysis approach

The quality and integrity of available dam data are critical for broad efforts to produce fine-scale assessments of their basin cycle, water quality, and other environmental and ecological effects. This study proposes a dam identification method in broad areas, object identification based on remote sensing images, and geographical analysis. First, we extracted dam candidate regions from broad surface water raster data at a spatial resolution of 30 m. Second, we trained and adjusted the multi-target recognition models using the dam sample from Google images, scanning dam candidate regions and extracting highly confidential dam positions. Moreover, we analyzed the location characteristics of the dams and used three geographical constraints to reduce background region overestimation further. The proposed framework was tested across an area of 13 265 km2 (Aomori, Kanagawa, and Okinawa) and yielded promising results, which reduced the candidate areas to 13.43% of the total water area. We validate the framework results using the available high-resolution historical image series available on Google Earth. The framework recalled 112 dams at a rate of 91.06%, with a precision rate of 80%. We simultaneously identified 39 dams that were not recorded in the known datasets. Our results reveal that the overall framework is reliable for automatic and rapid dam detection with a foundation of open geographic products. The framework proposed in this paper is the new attempt to combine deep learning target detection technology and spatial analysis with dam identification in broad areas.