Multi-Sensor Space-Time Data Fusion of Machine Learning Generated Time Series for Wetland Inundation Monitoring

The biogeochemistry of wetland ecosystems is driven by the presence and absence of water. Wetlands are known hotspots of methane (CH₄) emissions, particularly when inundated. Monitoring short-term, and possibly small-scale changes in inundation is therefore critical to quantifying both local and global CH₄ emissions. Despite their importance, these short-term changes have historically been under-reported in efforts to monitor CH₄. As sea levels rise and flood events increase, it’s imperative to account for these events to better project CH₄ cycle variation in a changing climate. Remote sensing is the only method capable of monitoring these changes over time at scale; however, no current remote sensing product has the spatial and temporal resolutions required to map ephemeral changes in inundation extents accurately. To address this, we developed a method to generate high spatiotemporal resolution inundation maps combining SAR and optical data from Sentinel-1 and Sentinel-2 imagery supplemented with commercial PlanetScope imagery from 2017–2022. This method was evaluated in the Albemarle-Pamlico Peninsula, a coastal wetland region in North Carolina, United States characterized by frequent and variable inundation.

Two decision-tree based machine learning algorithms were tested to map inundation extents: a random forest (RF) model and an extreme gradient boosted (XGBoost) model. The models were trained for each sensor based on a suite of spectral signals, terrain-derived features, and precipitation data for each image at the sensor’s native resolution. This work revealed minor differences between machine learning classifiers across the 5 years, with RF accuracies of 94.0%, 98.2%, and 98.6% and XGBoost accuracies of 89.1%, 98.3%, and 97.8% for PlanetScope, Sentinel-2, and Sentinel-1 respectively. The RF classified inundation maps from each sensor were then fused using a hierarchical spatiotemporal random effects model within a probit link function, to generate daily time series of inundation probabilities at 5 m resolution. This approach is unique in that we 1) address the differing sensor resolutions using a statistical change-of-support formulation with observations mapped to process locations, 2) fuse non-Gaussian (binary) responses from machine learning outputs, and 3) model spatial and temporal autocorrelation through spatial basis functions and a first-order autoregressive time series model. Overall, this work produced a novel 5-year inundation dataset, capturing both long-term and ephemeral changes in inundation extents that are critical for quantifying components of the water cycle and their interactions with biogeochemical cycles on Earth.