Johnson, LucindaBartsch, WillKovalenko, KatyaKloiber, SteveNixon, Kristi2024-01-112024-01-112024-01-11https://hdl.handle.net/11299/259969Out-of-bag testing with the randomForest package indicated that uplands were classified with an error rate of 12% and wetlands and restorable wetlands were classified with an error rate of 15%. Input Data: -MN DNR National Wetland Inventory (vector) -MN DNR Hydrography (vector) -MN DNR Ecological Classification System (vector) -MN DNR Digital Elevation Model (3m) -MN DNR Topographical Position Indexes (3m) -MN DNR Hydrologic Position Indices (3m) -MN DNR Compound Topographic Index (3m) -MN DNR Slope (3m) -NRRI Normalized Difference Vegetation Index (10m) -NRRI Normalized Difference Water Index (10m) -USDA Hydric soil status (10m) Modeling and Processing: Random Forest modeling was used to predict the probability of a pixel being either an upland wetland or an existing wetland or restorable wetland. The model had 15 predictive variables: -Distance to nearest wetland -Distance to nearest stream -Ecological Classification System Provinces and Sections -Elevation -Topographical Position Indices at radii of 100m, 150m, 200m, and 240m -Hydrologic Position Index -Compound Topographic Index -Slope -Normalized Difference Vegetation Index -Normalized Difference Water Index -Hydric soil status Vector data were converted to 3m grid data. 10m grid data were sampled to align with the 3m grid. The randomForest and raster packages in R (version 3.4.4) were used to develop and apply the two-way classification Random Forest model. Approximately 6000 points with field-verified wetland and upland status and 300 remote-verified points of restorable wetlands were provided by the Minnesota Department of Natural Resources and were used for training and verification of the model. Out-of-bag testing with the randomForest package indicated that uplands were classified with an error rate of 12% and wetlands and restorable wetlands were classified with an error rate of 15%. The model was then applied to the entire state using the Mesabi High-Performance Computing System of Minnesota Supercomputing Institute. Post-processing of the model output was conducted using ArcMap 10.6 to highlight pixels that are most likely to represent restorable wetlands. The first step was masking out existing National Wetland Inventory wetlands (except for those that are identified as farmed or partially drained or ditched). The second step was classifying pixels based on the probability of being a wetland/restorable wetland. A five-class system was used. Classes were assigned based on the probability of being a wetland/restorable wetland: 0.0 - 0.2 = 1 0.2 - 0.4 = 2 0.4 - 0.6 = 3 0.6 - 0.8 = 4 0.8 - 1.0 = 5 The final step was to smooth the results by aggregating individual or small groups of pixels into common classes using a majority pass filter. The RWI development process and final product were reviewed by regional wetland experts.The Minnesota Restorable Wetland Index (RWI) was developed by the Natural Resources Research Institute (NRRI) in collaboration with the Minnesota Department of Natural Resources (MN DNR). The RWI was developed statewide on a 3m grid by applying machine learning models to predict the location of existing and restorable wetlands based on hydrological, geomorphological, and geological variables. Post-processing was done to remove existing wetlands and smooth the results. This data layer replaces the original NRRI Restorable Wetland Inventory that was developed statewide on a 30m grid using a different methodology.Attribution 3.0 United Stateshttp://creativecommons.org/licenses/by/3.0/us/wetlandRWIrestorable wetland indexrestorable wetland inventorywaterrestorationDatasethttps://doi.org/10.13020/vmmd-rs84