X. Yang, M. Sun, R. Tapia, J.A. Koudelka, C.J. Bastidas Pacheco, and K.G. Cafferty
Idaho National Laboratory, Idaho, United States
Keywords: Critical Water Infrastructure, irrigation canals, semantic segmentation, geospatial AI, cyber-physical security
Geospatial understanding of water infrastructure is vital for managing resources, supporting hydropower, and protecting assets from natural and cyber-physical threats. Increasing risks highlight the need for automated methods to detect and monitor critical infrastructure such as stormwater channels and irrigation canals, which are often undocumented. Traditional survey and digitization are costly, and even national infrastructure assessments exclude canals. This work presents an AI-powered approach for canal mapping using semantic segmentation of high-resolution satellite imagery and Normalized Difference Water Index (NDWI) data. We apply transfer learning with the DeepLabV3+ architecture, originally trained for urban street mapping, and adapt it for hydrological feature detection. Optimized mask labeling improves Intersection over Union (IoU) performance, while canal-focused heatmaps enable accurate delineation of conveyance pathways under canopy cover, terrain camouflage, or image misalignment. The developed framework provides reliable baseline maps and supports automated change detection, enabling rapid anomaly identification and incident response. By integrating with SCADA and IoT data, this approach enhances situational awareness and resilience of critical water infrastructure across defense and civilian applications.