Comparison of Supervised Machine Learning Algorithms for Extracting Tree Canopy Cover using High-Resolution Imagery and Google Earth Engine

Accurate tree canopy cover (TCC) mapping is essential for managing forest ecosystems. Although machine learning (ML) methods have advanced remote sensing (RS) based image classification, there has been limited research on their effectiveness on TCC mapping. This study aims to compare different ML algorithms for TCC mapping under various conditions: (i) training sample sizes, (ii) parameter settings, and (iii) feature variable combinations. Additionally, Shapley Additive Explanation (SHAP) was employed to gain insights into the importance of features across classes. The accuracy of each algorithm was assessed using various performance metrics, and statistical comparisons were conducted using McNemar's chi-squared test for class-wise performance and the two-proportion Z-test for overall accuracy (OA). The results indicated that tuning parameters, using moderate training sample sizes, and combining feature variables significantly enhanced the overall performance of all algorithms. Among them, the Random Forest (RF) algorithm consistently outperformed the others, achieving an accuracy of 81.19% before optimization and 89.52% after optimization. The SHAP analysis further revealed that the RF algorithm provided a balanced profile of feature importance across variables. These findings serve as a strong foundation for selecting algorithms to extract TCC information accurately.