Hybrid spatiotemporal modeling of nutrient cycling in wetland ecosystems using advanced mapping techniques and machine learning approaches.

Accurate spatiotemporal monitoring of nutrient cycling in wetlands is critical for conservation. However, traditional field-based methods are often inadequate for capturing the overall dynamics of wetlands. To address this challenge, we developed and validated a two-stage hybrid Random Forest regression framework that seamlessly integrated in-situ water quality data with wetland features and satellite imagery from Sentinel-1 and Sentinel-2 within the Google Earth Engine platform. The framework first models baseline nutrient concentrations using discrete wetland characteristics (stage 1) and then models the resulting spatial residual using continuous satellite-derived predictors (stage 2). We applied the model to predict quarterly nitrogen and phosphorus concentrations over four years (2021-2024) in the Beavercreek Wetlands Greenway (BWG), a mixed-use landscape. The two-stage model demonstrated exceptional predictive performance for both nitrogen (final [Formula: see text] = 0.90, RMSE = 0.129 mg/L) and phosphorus (final [Formula: see text] = 0.89, RMSE = 0.007 mg/L). The variable importance analysis revealed divergent predictive pathways: nitrogen concentration was driven by both landscape-level factors (e.g., land use classes, area and perimeter of wetlands, rainfall) and in-stream biophysical conditions captured by remote sensing (e.g., vegetation and SAR indices), whereas phosphorus was controlled by source-loading from developed land uses. We produced spatiotemporal maps to visualize these distinct patterns. The maps revealed that the BWG system is subject to seasonal nitrogen stress but has experienced significant recovery from prior phosphorus impairment. This study offers a diagnostic framework that could inform focused wetland management strategies and aid in monitoring the health and function of wetland ecosystems.