A Geophysically Constrained Deep Learning and Binary Encoding Framework for Long-Term River Ice Characterization Extraction

Based on a deep Convolutional Neural Network (CNN) model, this dataset reconstructs and classifies 37 years of long-term satellite imagery to generate monthly river ice maps with a spatial resolution of approximately 30 meters. To efficiently characterize the monthly spatiotemporal evolution of river ice during the ice season (November to the following March) at the pixel level, a binary weighting encoding scheme was implemented. This framework compresses the river ice status of five consecutive months into a single 8-bit integer value. The resulting product provides high-density spatiotemporal data with a high compression ratio, offering critical support for analyzing river ice dynamics under the context of global climate change.