Object-Based Classification of Unmanned Aerial Vehicle (UAV)/Drone Imagery to monitor H2Ohio Wetlands

Wetlands are referred to as the kidney of the catchment due to their ability to reduce nutrient loads adjoining water bodies, mitigating eutrophication. The Ohio government has employed this beneficial mechanism as part of the H2Ohio program to abate the immediate release of nutrients in the water body by reconstructing wetlands at a number of locations in the Maumee watershed. We are using a combination of Unmanned Aerial Vehicles (UAVs), machine learning and field mapping to generate maps of wetland vegetation communities which assist in establishing the effectiveness of these restored sites in nutrient removal. The vegetation in most of these wetlands can take in and recycles most of the nutrients (especially Nitrogen and Phosphorus) from the incoming runoff, thereby reducing the nutrient loads into the waterbodies. At one of wetland restoration projects in northwest Ohio, high-resolution near-infra-red (NIR) and visible images acquired with UAV were corrected geometrically and radiometrically to generate Orthomosaic, Digital Elevation Model (DEM), and vegetation indices (VIs). Using eCognition (commercial software) and QGIS (open-source software), these products are combined and then segmented into homogenous units (objects) based on the similarities in shape, scale, color, smoothness, texture, etc. Then, object-based classifiers, including Support Vector Machine (SVM) and Random Forest (RF), were used to classify the landcover into desired classes based on field sampling. Each class represents a vegetation type or other landcover (like waterbody, bare ground, etc.) present on the wetland. The accuracy of the classification model was tested with held-back field validation data using a confusion matrix. The Kappa coefficient for SVM is 0.82 and RF is 0.75 in eCognition, while they are 0.76 and 0.69, respectively, in QGIS. The result shows that SVM was to be the best model to distinguish between vegetation in these wetland communities. Although the eCognition implementation classifies the community better, QGIS still offers a reliable classification, enabling this monitoring procedure to be adopted by a broader range of researchers. This process has been iterated for two seasonal years and will be repeated on a temporal basis to observe the effectiveness of the plant community in taking up nutrients.