Bare ground classification using a spectral index ensemble and machine learning models optimized across 12 international study sites

This research investigates a global approach to map bare ground across diverse geographies with an ensemble of spectral indices using optimal thresholds identified in testing to train and evaluate machine learning models to extract bare ground pixels from Sentinel-2 imagery. Twelve locations in four Köppen climate zones with data from two seasons were evaluated. Accuracy assessment showed a mean F1 score of 80% and a mean Overall Accuracy (OA) of 81% for random forest and an F1 score of 78% and OA of 79% for support vector machine. Higher accuracies were observed in climate region-based models with mean F1 > = 84% in three of four climate zones. Low accuracies occurred in winter imagery with leaf-off tree cover or building materials similar to bare ground. This framework provides a global approach to map bare ground without need for high-density time-series or deep learning models and moves beyond locally effective methods.