An automatic classification method with weak supervision for large-scale wetland mapping in transboundary (Irtysh...

Wetlands are essential ecosystems that play a significant role in biodiversity conservation and environmental stability. Monitoring their changes is crucial for understanding ecological dynamics and informing conservation strategies, particularly those in transboundary basins. This study introduces a novel automatic classification method for mapping and detecting wetland changes in the Irtysh River Basin. Utilizing Google Earth Engine (GEE) as the primary platform, this method integrates unsupervised classification, sample transfer techniques, and object-oriented random forest (OORF) algorithms to generate accurate training samples and delineate wetlands. Using Sentinel-1 and Sentinel-2 satellite data, we created high-resolution wetland distribution maps. The process begins with unsupervised classification to identify wetland inundation zones, followed by overlaying permanent water bodies and surface depressions to refine the sample set. Sample transfer, using spectral similarity metrics with the GWL_FCS30 product, further enhances the robustness of the training data. The selected features from Sentinel-1 and Sentinel-2 data, including spectral indices, phenological parameters, and textural features, were optimized, resulting in 18 optimal features for the OORF classification. The classification achieved a high overall accuracy of 96.96 %, with a sample accuracy of 98.1 %, and both User's and Producer's Accuracies consistently above 88 %. Spatiotemporal analysis of wetland changes from 2017 to 2023 revealed significant fluctuations, including a net loss of approximately 1,743.92 km² of wetlands in the Irtysh River Basin. This study provides an effective and innovative method for large-scale wetland monitoring, offering valuable insights to support conservation and management efforts.