Comparing Stability in Random Forest Models to Map Northern Great Plains Plant Communities Using 2015 and 2016 Pleiades Imagery
Abstract. The use of high resolution imagery in remote sensing has the potential to improve understanding of patch level variability in plant structure and community composition that may be lost at coarser scales. Random forest (RF) is a machine learning technique that has gained considerable traction in remote sensing applications due to its ability to produce accurate classifications with highly dimensional data and relatively efficient computing times. The aim of this study was to test the ability of RF to classify five plant communities located both on and off prairie dog towns in mixed grass prairie landscapes of north central South Dakota, and assess the stability of RF models among different years. During 2015 and 2016, Pleiades satellites were tasked to image the study site for a total of five monthly collections each summer (June–October). Training polygons were mapped in 2016 for the five plant communities and used to train separate 2015 and 2016 RF models. The RF models for 2015 and 2016 were highly effective at predicting different vegetation types associated with, and remote from, prairie dog towns (misclassification rates