Refining global wetland characterization using an unsupervised, wetness-based dynamic framework

Existing global wetland datasets and monitoring approaches emphasizepersistent inundation, while intermittent inundation and waterlogged states—especially where vegetation is present—are underrepresented or of lower accuracy. This leads to inaccurate estimates of greenhouse gas emissions from carbon-rich systems (e.g., peatlands). Meanwhile, the predominance of annual mapping limits the capture of intra-annual variability, further reinforcing these inaccuracies and obscuring sub-seasonal disturbances from human activities (e.g., shifts in rice-cropping intensity). This study presents an unsupervised, wetness-driven framework for improving global wetland monitoring that leverages earth observation data streams. For framework development, the OPtical TRApezoid Model is applied to Harmonized Landsat-Sentinel imagery to retrieve surface wetness, followed by wetland delineation using a scene-adaptive grid-based thresholding algorithm. This framework is applied to 824 globally distributed 0.1° grid cells encompassing 9,781 land-cover-labeled sites and 134 sites with daily wet–dry labels across 28 Ramsar wetlands, and validated for spatial delineation, thematic, and temporal accuracy. Comparative analysis employs Dynamic World, the first global 30 m wetland map with a fine classification system (GWL_FCS30), and the modified Dynamic Surface Water Extent algorithm (DSWE). Our framework achieved moderate spatial delineation accuracy with F1 of 0.64 (recall 0.75, precision 0.56), comparable in F1 to Dynamic World and with higher recall than DSWE and GWL_FCS30. It delivered the highest temporal accuracy (F1 0.72; precision 0.81; recall 0.64) and improved thematic accuracy for vegetated wetland, reducing omission with modest commission. The proposed wetland monitoring framework enables more accurate targeted policy interventions.