Abstract
Wetlands serve critical functions for protecting ecosystems but are declining globally due to anthropogenic activities and environmental change. This dissertation presents four manuscripts that focus on using multi-source imagery to support frequent and accurate wetland mapping. The first manuscript conducted a meta-analysis of 172 research papers to evaluate the impact of synthetic aperture radar configurations on wetland classification accuracy. The study found increasing multi-sensor data use and integration of C- and L- bands (19 studies), and higher classification accuracy with multi-frequency (29 studies) and multi-polarized (115 studies) data. We also found the highest median overall accuracy with high spatial resolution data. The second manuscript investigated the accuracy of different machine learning models for wetland mapping using features extracted from multi-polarimetry radar data. We found that a random forest model (81.9%) had higher overall accuracy compared to a support vector machine model (74.3%) with volume scattering features highly ranked in variable importance analysis. The third manuscript evaluated multi-source spatiotemporal features in classifying wetlands across New York State. The Gini impurity index found no variables universally resulted in the best classification results across all ecoregions. However, some seasonal textural features generally enhanced separation of wetland classes. The study also found that using a 30 m canopy height model (CHM) derived from Global Ecosystem Dynamics Investigation (GEDI) lidar and Landsat data reduced confusion caused by high backscatter and spectral similarity of wetland types with a 6% increase in overall accuracy compared to a baseline classification using optical and radar features. The fourth manuscript integrated GEDI height footprints with Sentinel-2 data to create a 10 m CHM to support wetland classification. We found the CHM had highest accuracy using GEDI 95th percentile relative height as the dependent variable. The CHM was more accurate when GEDI data was integrated with Sentinel-2 compared to Landsat though underestimated height across forested wetlands and overestimated height in areas with slopes above 15°. We used the 10 m CHM in a wetland classification and found the higher resolution generated a spatially defined output. This dissertation developed approaches to enhance the management and monitoring of wetlands across large extents.