Mapping the first global long time series wetland multitype sample dataset via the Google Earth Engine: a hybrid method of automated generation—index thresholding—spectral matching

Wetlands are characterized by high diversity and complexity and present formidable challenges for global-scale remote sensing mapping. A high-quality, robust global wetland sample dataset (GWSD) is essential for overcoming these challenges. However, yet the absence of a reliable methodology for generating long-term, consistent wetland samples has persisted as a critical gap. Herein, we propose a novel hybrid approach that combines automated generation – index thresholding – spectral matching (AG – IT – SM) to produce the first multicategory global wetland sample dataset from 1985 to 2020. Using the full Landsat 5/7/8 archive within the Google Earth Engine (GEE), we generated 349,952 training samples and 67,952 validation samples. Globally, wetland samples are distributed predominantly in the Northern Hemisphere, with a relatively sparse representation in the Southern Hemisphere. Independent expert validation through crosschecking confirmed that all wetland-type samples achieved an accuracy exceeding 90%. A comparative analysis with the GLC_FCS30D dataset demonstrated strong temporal consistency across all evaluated years. Classification experiments demonstrated that the refined wetland samples achieved accuracies exceeding 80%. The proposed method was validated as an effective approach for producing reliable wetland samples, resulting in the first global wetland reference dataset that may serve as a fundamental data resource for large-scale wetland mapping applications.