Mapping the spatiotemporal variation in wetland habitat using remote sensing-based machine learning models

Abstract:

This study develops an open-source process for mapping spatiotemporal variation of depressional wetland surface water dynamics, including surface water, water depth, and hydroperiods, using the Prairie Pothole Region (PPR) as a test case. The PPR is an extensive section of the U.S. Great Plains that contains numerous depressional wetlands (potholes), and is known for cycles of drought and deluge. Mapping these patterns has proved challenging and created difficulties for conservation managers trying to integrate dynamic processes into conservation plans for wetland obligates like waterfowl or shorebirds. Better estimates of the inter - and intra-annual surface water extent and pond depth across the PPR will provide needed information for identifying viable waterfowl habitat and forecasting conservation planning scenarios. For ground-truth data of surface water area, we used field studies and aerial photography. To estimate hydrodynamics, we created 10 x 10 m Landsat 8 and Sentinel-2 harmonized surface reflectance data, which we fused with Sentinel-1 SAR data and 1-m resolution digital elevation models. Remote sensing data underwent pre-processing for atmospheric correction. Pan-sharpening was also implemented to enhance the spatial resolution of Landsat 8 data and detect fine-scale wetland habitat features.

We used machine learning (random forest) (RF) to establish a mathematical relationship between target hydrologic features (surface water area, water depth) and their corresponding reflectance and backscatter coefficient values in remote sensing data. These relationships were used to classify the remote sensing data and to generate the temporal maps for surface water dynamics in high priority conservation areas of the PPR. The results were also compared to simple reflectance index based approaches (e.g., the normalized difference water index) (NDWI) and Landsat 8 CFMask Wet Flag. Machine learning techniques such as RF outperformed NDWI and Landsat 8 CFMask Wet Flag for classifying wet areas. Further analysis of remote sensing data is underway for mapping spatiotemporal variation in water depth and emergent vegetation through machine learning techniques (random forest, extreme gradient boosting, and deep learning). Results will be disseminated to waterfowl conservation managers in the PPR.