About the Project

This transfer project addressed a need identified by the St. Louis River Habitat Workgroup to support the identification and prioritization of areas for future restoration and conservation. The multi-phase project approach included the transfer of a repeatable habitat mapping process developed by the Lake Superior Reserve to a larger area encompassing 57,000 acres of wetlands and adjacent uplands spanning the lower twenty-one miles of the St. Louis River below the Fond du Lac dam. The team applied accessible image classification methods–including use of common machine learning classifiers and freely available, non-proprietary data–to create a reproducible approach that can easily be adopted in other locations and redeployed at regular intervals to illuminate change over time.

About this Resource

These workflow resources walk users through the methodology employed in this project, with the goal of facilitating updates to future habitat maps or transferring the mapping methodology to other geographic locations.

Project Phase 1: Land Cover Mapping with U-Net for the St. Louis River Estuary

The primary objective of this project was to develop tools and methods using open-source geospatial technologies for land cover mapping of the St. Louis River Estuary, along the border between Wisconsin (WI) and Minnesota (MN) in the Duluth-Superior metro area. The main data product from this phase is a classified land cover map of the area of interest produced by segmentation of high-resolution aerial imagery with a deep learning approach (U-Net).

Project Phase 2: St. Louis River Estuary Habitat Mapping Post-processing of initial UNET land cover classes

The objective of UNET "Phase 2" post-processing was to refine UNET-derived land cover classes using GIS-based rules and ancillary datasets. Input to the post-processing steps was a completed UNET land cover map in raster format, derived from NAIP aerial imagery. The study area is a one-mile buffer around the St. Louis River estuary in the Duluth-Superior area.

The post-processing steps took the form of a series of 8 rulesets. Each ruleset was applied to the output from the previous step.