Understanding changes in tropical wetlands with remote sensing, machine learning and land surface modleing

Tropical wetlands are among the Earth’s most critical ecosystems, playing a vital role in regulating global water and carbon cycles, buffering extreme weather events, and supporting biodiversity that sustains millions of people. Despite their importance, these ecosystems are highly vulnerable to climate change, and our understanding of their seasonal extent, their role in climate mitigation, and their response to changing climatic conditions remains limited. This lack of knowledge hinders the development of effective climate adaptation strategies and constrains projections of future carbon emissions. To address these gaps, this study and related ongoing work aim to develop an integrated framework that combines multiple methodologies, data sources, and analytical tools to improve the monitoring of surface water inundation in major floodplain systems.

A key component of this framework is understanding the availability, characteristics, and interpretative value of different remote sensing datasets. As a case study, this work focuses on two major wetland systems: the Sudd in South Sudan and the Pantanal in South America. Satellite observation datasets are compiled and categorized into static and dynamic products. The static datasets include GlobCover, GLWDv2 and SWAMP, while the dynamic datasets comprise WAD2M, JRC Global Surface Water (GSW), CYGNSS water mask, and GRACE Total Water Storage. Static datasets are used to assess long-term changes in wetland extent and classification, whereas dynamic datasets capture seasonal variability in wetland extent. Together, these datasets enable comparison between seasonal dynamics and long-term trends, providing improved insight into future projections of wetland extent. The dynamic datasets are projected onto a common grid to assess their consistency and agreement with one another. These datasets are also used to develop a Long Short Term Memory (LSTM) network, capable of capturing both seasonal variability and long-term trends, which is then applied to project future wetland dynamics.

This work represents an important first step towards reducing uncertainty in global wetland mapping. Building on this foundation, the study aims to use the Joint UK Land Environment Simulator (JULES) to simulate wetland extent and hydrological dynamics across selected tropical wetland regions (Sudd, Pantanal). Model simulations are driven by newly developed ancillary inputs, including land cover parameters, soil properties, and topographic information, to assess their influence on simulated wetland extent and seasonal flooding patterns. The resulting JULES outputs are systematically evaluated against EO-based wetland datasets such as GLWDv2, GlobCover, and WAD2M to identify areas of agreement, model sensitivities, and potential sources of bias. Through this comparative analysis, the study benchmarks the capability of the JULES land surface model to represent tropical wetland dynamics and provides insights into optimal data configurations for large-scale wetland modelling.

As the project develops, machine learning approaches are further applied to forecast wetland dynamics and to inform improvements in the representation of wetlands within climate models. Ultimately, this integrated modelling and data-driven framework aims to contribute to more reliable climate predictions and to provide decision-makers with clearer, evidence-based information for climate adaptation and mitigation planning.