WetSAT-ML: A Machine Learning Framework for Mapping Wetland Flooding Dynamics using Sentinel-1 Observations

Wetlands contribute to human well-being in multiple ways. These contributions take various forms, including goods such as food and other raw materials, but also through services like carbon sequestration, flood and climate regulation. Despite their importance, wetlands are being lost at annual rates exceeding 0.5% of their global extent over the past sixty years, mainly due to conversion to other land uses. Several multilateral agendas emphasize the importance of wetland monitoring and protection; however, progress toward established targets remains limited by the lack of consistent national datasets and operational monitoring tools. Satellite observations provide globally consistent data that enable systematic wetland monitoring. In particular, synthetic aperture radar (SAR) has been successfully used for mapping wetland flooding dynamics due to their all-weather acquisition capability and the ability to identify below canopy inundation processes. The increasing availability from current and planned SAR missions poses an opportunity to advance space-based wetland monitoring, but their operational implementation requires the development of flexible and scalable frameworks.

This study aims to develop WetSAT-ML, a satellite-based machine learning (ML) framework for mapping wetland flooding dynamics and trends using Sentinel-1 SAR data. The approach combines radar features with supervised and unsupervised ML classification algorithms, to distinguish different inundation categories, including open water, vegetated water, and land. Random Forest and K-means models were trained using two training and validation areas in the South Florida Everglades (USA) and the lower Atrato River Basin (Colombia). These sites were selected considering their wetland variability and the availability of reference data. The first version of the models was trained using all available Sentinel-1 observations from 2024 over the selected regions, capturing the full hydrological seasonality. Reference training and validation datasets included gauge measurements and manually annotated data for both regions. The trained WetSAT-ML models are being evaluated through a proof of concept across five wetland systems in South Asia and South America, including the Meghna River wetlands in Bangladesh; the Atrato River, Ayapel, and Barbacoa wetlands in Colombia; and the Pantanal wetlands spanning the Brazil–Bolivia border. The test sites represent a wide range of hydroclimatic conditions, geomorphological settings, and vegetation cover.

Preliminary results indicate that WetSAT-ML consistently captures spatial patterns and intra-annual inundation dynamics that are coherent with the known hydrological regimes of the test regions. Cross-site comparisons reveal clear differences in key hydroperiod parameters related to inundation persistence, seasonal amplitude, and ecosystem connectivity. Overall, the results provide a foundation for operational wetland monitoring applications. WetSAT-ML is open access, and the first public version is available in a GitHub repository: https://github.com/sei-latam/WETSAT_v2. The next steps of the research will focus on cross-validation using independent datasets and expanding the training database across additional wetland areas.