Developing an open-source remote sensing framework for long-term wetland restoration monitoring

Abstract

Wetland restorations can take decades to show measurable success, yet long-term monitoring remains uncommon. Field-based monitoring is often costly and labor-intensive, making remote sensing an appealing alternative. In this project, we developed an open-source, reproducible remote sensing workflow calibrated with field data from a wetland restoration site in Galveston, Texas. The restoration was initiated in 2004, with additional efforts in 2011, with goals to increase wetland habitat, buffer wave energy and slow the erosion of remaining natural marsh. Our objective was to build an automated monitoring framework to assess restoration trajectory over time using biophysical metrics including elevation, land cover, and aboveground biomass. We trained models using Sentinel-2 satellite imagery and digital elevation model (DEM) data to estimate these metrics. To enhance elevation estimates, we developed a correction model that reduced root mean squared error (RMSE) from 0.29-m to 0.12-m. Our land cover model achieved an overall testing accuracy of 90% for classifying water, vegetation, and bare soil, while the aboveground biomass model performed with an RMSE of 82.4 g m-2 (normalized RMSE of 14%). These models were applied in time series analyses to evaluate site-wide landscape level restoration progress. We found that restoration mounds of 20+ years old had lower elevations than reference sites, but were similar to mid-aged sites, suggesting they are relatively stable. Mounds also appeared to be buffering remaining natural areas from lateral marsh erosion. However, mounds created in 2004 showed increased rates of vegetation transitioning to water and should be monitored closely. By applying this scalable and cost-effective framework, managers can more readily detect emerging restoration challenges and make timely, data-driven decisions, such as planting more vegetation, implementing more erosion control structures or increasing mound elevation through thin layer soil placement. Through leveraging high-resolution and freely available Sentinel-2 imagery and elevation data, the models achieved high classification accuracy and effectively captured key trends in biophysical parameters such as elevation, land cover and AGB, offering insight into restoration trajectories over time. While expert interpretation is still necessary for management decisions, this framework offers a powerful tool for improving long-term restoration monitoring and adaptive management.