Mapping Proportions of Eastern Redcedar in Kansas Using Sentinel-2 Imagery and Random Forest Model

Abstract

Eastern redcedar (Juniperus virginiana L.) encroachment is transforming grasslands across the central United States, with significant ecological and hydrological consequences. Effective management requires maps quantifying proportional cover to detect early invasion and prioritize interventions. We developed a workflow producing annual 10 m fractional cover maps at the University of Kansas Field Station (2016–2023). A 1 m reference dataset was generated by intersecting light detection and ranging (LiDAR)-derived canopy-height stability with leaf-off aerial imagery and aggregated to the Sentinel-2 grid. Random forest models were trained on seasonal Sentinel-2 composites using proportion-aware sampling, with LandTrendr temporal segmentation applied to reduce interannual spectral noise. Classification accuracy reached 0.84 (κ = 0.67), comparable to models incorporating National Ecological Observatory Network (NEON) hyperspectral data (0.85); however, high spatial autocorrelation among training samples suggests these represent upper-bound estimates. Winter normalized difference vegetation index (NDVI) and Simple Ratio Index were the most influential predictors, confirming dormant-season imagery value for evergreen discrimination. Regression models achieved low error (RMSE = 0.04–0.05), though accuracy decreased in stands exceeding 70% cover. LandTrendr-stabilized maps closely matched seasonal composites (R = 0.95–0.99) while producing smoother trajectories. The workflow demonstrates particular sensitivity to low-density invasion (<10% cover), enabling early detection when interventions remain cost-effective. Because independent validation was unavailable after 2016, temporal predictions warrant caution. Operational deployment elsewhere requires local reference data and model recalibration; transferability depends on species composition, phenological contrast, and landscape configuration.