Ecological Systems Classification: Integrating Machine Learning, Ancillary Modeling, and Sentinel-2 Satellite Imagery

Michael Sunde, David Diamond, Lee Elliott

Spatial land cover depictions are essential for ecological and environmental management. The thematic resolution of land cover and vegetation maps is also a significant factor affecting the ability to effectively develop policy and land management decisions based on spatial data. Natural resource and conservation planners often seek to develop strategies at broad scales; however, high-quality spatial data depicting current vegetation and ecosystem types over large areas are often unavailable. Since widely available land cover and vegetation datasets are generally lacking in either thematic resolution or spatial coverage, there is a need to integrate modeling approaches and ancillary data with traditional satellite image classifications to produce more detailed ecosystem maps for large areas. In this study, we present a comprehensive approach using satellite imagery, machine learning, and ancillary modeling approaches to develop high-resolution ecological system type maps statewide for Arkansas, USA. A RandomForest land cover classification of Sentinel-2 imagery was generated and further articulated into ecological types using a comprehensive set of secondary modeling approaches. A total of 123 types were mapped in Arkansas, including common cultural and ruderal land cover and vegetation such as pine plantations and developed types. Ozark–Ouachita Dry–Mesic Forest covered the most area, 17.51% of the state. Row Crops covered 17.16%. Twenty-five pine or pine plantation types covered 19.73% of the state, with Ozark–Ouachita pine woodland or mature pine plantation covering 6.15%. Field survey points were used to assess the quality of the mapped ecological systems. The approaches presented here provide a framework for finer resolution mapping of ecological systems at broad scales in other regions.