RIce-Net: Integrating ground-based cameras and machine learning for automated river ice detection

River ice plays a critical role in controlling streamflow in cold regions. The U.S. Geological Survey (USGS) qualifies affected water-level measurements and inferred streamflow by ice conditions at a date later than the day of the actual measurements. This study introduces a novel computer vision-based framework, River Ice-Network (RIce-Net), that uses the USGS nationwide network of ground-based cameras whose images are published through the National Imagery Management System (NIMS). RIce-Net consists of a binary classifier to identify ice-affected images that are segmented to calculate the fraction of ice coverage, which is used to automatically generate a near real-time ice flag. RIce-Net was trained using images from selected NIMS stations collected in 2023 and tested using images collected in 2024. Also, the framework’s scalability and transferability were tested over another station that was not included in the training process. RIce-Net ice flags are well-aligned with those reported by USGS. • In cold regions, river ice affects the inferred discharge under ice conditions. • RIce-Net automatically generates near real-time ice flag to water level observations. • RIce-Net uses the published images through the National Imagery Management System (NIMS). • The generated ice flags are well-aligned with those reported by USGS. • RIce-Net framework showed potential scalability to new stations.