Multi-Sensor Monitoring of Wetland Inundation Using a Machine Learning and Data Fusion Framework

Continuous, high-resolution inundation data are needed to understand how small-scale, short-term wetland flooding influences global methane emissions and carbon cycling. Small (less than 1,000 m²), variably inundated wetlands are significant methane sources, yet coarse satellite products often miss their dynamics. Integrating optical and radar imagery with resolutions less than 30 m offers a solution over single sensor approaches, but two needs remain: (1) understanding the relative strengths and limitations of different sensors for detecting inundation in vegetated wetlands, and (2) developing methods that leverage these sensor characteristics to transform sparse, irregular classification maps with different resolutions into complete, consistent coverage. To address this, we first evaluated sensors by classifying inundation from 2017 to 2022 across a wetland area in eastern North Carolina using Sentinel-1, Sentinel-2, and PlanetScope imagery. We used tree-based machine learning models to classify pixels as Dry Land, Inundated Vegetation, or Open Water, and compared maps of inundation frequency produced by each model. Sentinel-2 random forest achieved the highest accuracy (94.9%), followed by PlanetScope (92.6%) and Sentinel-1 (85.9%) models. Next, we introduce a fusion framework based on Fixed-Rank Kriging, a spatiotemporal statistical model, to fuse binary classification maps into daily inundation probabilities at 6 m resolution. The fused approach balanced individual sensor variability and detected short-term inundation fluctuations missed by the Landsat Dynamic Surface Water Extent, a widely used satellite-derived data product. By characterizing sensor performance and translating that information into a multi-sensor fusion approach, this work enables high-resolution, continuous monitoring of inundation dynamics vital to predicting wetland methane emissions.