Northern Minnesota Ecological Conservation​: Exploring Classification of Forest Cover Types in Northern Minnesota Using Earth Observations​

The Superior National Forest encompasses the northeastern part of Minnesota and contains critical ecological and economic resources. Responsible for assessing the forest’s resources, the Minnesota Department of Natural Resources (MNDNR) requires data about the spatial distribution of tree species in order to inform their management decisions. Remote sensing can provide spectral and topographic data across large spatial extents, and machine learning models can leverage these data to generate spatial distribution maps of forest cover types needed to inform these decisions. The MNDNR partnered with NASA DEVELOP to explore a machine-learning method to classify the Superior National Forest into a cover type map. The team developed a supervised machine learning model that combined the MNDNR’s plot-based forest inventory data with multispectral Harmonized Landsat Sentinel-2 imagery and topographic data to classify the study area into 12 different land cover classes. With an overall accuracy of 61.1%, the best performing model demonstrated some feasibility of our methods but was insufficiently reliable to produce a highly accurate land cover map containing classes with individual tree species. However, this project provided important insights on mapping methods. We found that topographic data had variable impact on accuracy, and the gradient boosting technique improved accuracy compared to the random forest technique. We also found a positive correlation between sample size and model accuracy, highlighting the need for additional and more proportionate distribution of training data across the 12 land cover classes. These findings will be used to support the MNDNR’s efforts to generate land cover classifications needed for assessing detecting changes in forest composition.