Satellite fusion tracks wetland flooding daily

Small, short-lived wetland flooding is a major but poorly tracked methane source. This study compares radar and optical satellite classifications of dry land, inundated vegetation and open water, then combines them statistically. The result is a daily, high-resolution probability map that fills gaps and captures brief floods for carbon and methane research.

Wetlands sit at the junction of water and carbon cycles. When inundation creates low-oxygen conditions, wetlands can shift from carbon sinks to methane sources. Yet methane estimates remain uncertain because small, temporary flooded areas are easily missed by moderate-resolution satellite products. Cloud cover also blocks optical imagery, while radar alone can confuse wet soil with standing water. Single-sensor maps are sparse and inconsistent. Based on these challenges, in-depth research is needed on how different sensors detect vegetated wetland inundation and how their classifications can be fused into consistent, high-resolution coverage.

Researchers at North Carolina State University’s Center for Geospatial Analytics, with forestry and statistics colleagues, published (DOI: 10.34133/remotesensing.1077) the study in Journal of Remote Sensing on 15 October 2026. They tackle a monitoring gap: operational surface-water products miss small, dynamic wetlands, while individual sensors trade resolution, cloud penetration and spectral sensitivity. The team asks whether machine-learning classifications from three satellites can be fused into a complete daily record and whether that improves detection of ephemeral flooding.

The comparison found sensor strengths. Sentinel-2, an optical satellite, was most accurate overall (93.9%), followed by PlanetScope (88.4%) and Sentinel-1 radar (84.6%). Sentinel-2 also detected inundated vegetation best, while PlanetScope captured fine open-water features. Sentinel-1 was valuable under clouds but overpredicted inundation in fields. Fusion balanced these biases, filled gaps and produced daily 6 m inundation probabilities. It captured short flood pulses missed by individual sensors and mapped more wetland inundation than the most conservative Landsat Dynamic Surface Water Extent (DSWE) class at a moderate threshold, though the gap narrowed at stricter thresholds.

The team classified dry land, inundated vegetation and open water across eastern North Carolina’s Albemarle-Pamlico Peninsula, including Alligator National Wildlife Refuge, using five years of Sentinel-1, Sentinel-2 and PlanetScope imagery. Training and validation points came from selected dates and were checked against high-resolution reference imagery. Random forest and extreme gradient boosting models used spectral indices, radar backscatter, terrain, canopy cover and recent precipitation. The best model for each sensor was applied over time. For fusion, fixed-rank kriging (FRK) treated each classification as a binary observation, assigned error based on sensor accuracy and predicted a latent inundation process. A simulation first confirmed FRK could recover a known process from irregular observations. The method was tested in three focal wetland areas at 6 m resolution, producing daily probabilities and 90% credible intervals. Maps were compared with water-table, precipitation and lake-elevation records, as well as DSWE.

The authors stress that sensors are not interchangeable. Optical sensors excel in vegetated wetlands, radar adds cloud-free observations, and fusion turns disagreement into probability and uncertainty. They describe the framework as transferable to new sensors and existing products, while cautioning that dense canopies and cloud-mask errors remain limitations. This offers a path toward more continuous wetland monitoring.

Imagery was processed with Google Earth Engine and atmospheric correction tools, with strict cloud and cloud-shadow masking. Predictors included spectral water and vegetation indices, radar backscatter combinations, slope, topographic wetness, canopy cover and 7- and 30-day precipitation. Models were trained and validated with stratified sampling and accuracy assessment. Fusion used the FRK R package, accuracy-based measurement-error variances and slope as a covariate. Benchmarks included Landsat DSWE and last-observation-carried-forward gap filling. The team also ran threshold sensitivity tests from 30% to 90%.

The approach could feed methane and carbon models by carrying inundation uncertainty into ensembles, and may support monitoring of sea-level rise, saltwater intrusion and hydrologic regime shifts. Because it is largely sensor- and region-agnostic, forthcoming L-band radar or water-level observations could be added. Computational cost and subcanopy detection remain barriers. Multiresolution designs, near-real-time filtering, stricter cloud screening and denser field networks could improve operational reliability. For wetland science, the larger promise is a more continuous record of brief floods that shape greenhouse-gas exchange.

###

References

DOI

10.34133/remotesensing.1077

Original Source URL

https://spj.science.org/doi/10.34133/remotesensing.1077

Funding Information

This work was supported by NASA through Future Investigators in NASA Earth and Space Science and Technology (FINESST) Program Grant No. 80NSSC24K0016; author Jenna N. Abrahamson was also supported by National Science Foundation Graduate Research Fellowship Grant No. DGE-2137100. Any opinions, findings, and conclusions or recommenda-tions expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation.

About Journal of Remote Sensing 

The Journal of Remote Sensing , an online-only Open Access journal published in association with AIR-CAS, promotes the theory, science, and technology of remote sensing, as well as interdisciplinary research within earth and information science.