Estimating mean groundwater levels in peatlands using a Bayesian belief network approach with remote sensing data

Large-scale management, protection, and restoration of wetlands require knowledge of their hydrology, i.e., the status and dynamics of the groundwater table, which determine the evolution of the wetland ecosystem, its conservation value, and possible economic use. Unfortunately, in many cases, hydrological monitoring data are unavailable, resulting in the search for a proxy for the average annual depth of the groundwater level (GWL). This study presents an approach to estimating the mean GWL in peatlands using a Bayesian belief network (BBN) model, leveraging long-term hydrological and remote sensing data in the Biebrza National Park in Poland. The remote sensing data employed includes the synthetic aperture radar (SAR) backscatter coefficient, peat subsidence, rate and distance to watercourses. The BBN model achieved a predictive accuracy of 83.3% and 73.1%, depending on the validation approach used. Among the remote sensing variables considered, the SAR backscatter coefficient was the most sensitive in predicting the GWL in the peatlands. However, the model presents multiple uncertainties resulting from limitations of the available remote sensing data, low variability of class combinations in the conditional probability table, and lack of upscaling to other regions performed. Despite these uncertainties, the developed BBN model remains a valuable next step in reaching the goal of efficient peatland monitoring and management.