Evaluating Classification Accuracy, Area Estimation, and Area Deviations of Five Algorithms in Google Earth Engine: Policy Implications

Land use land cover (LULC) classification accuracy is essential in ensuring reliable evidenced-based environmental monitoring and policy making, especially within protected transboundary areas. This study evaluates the performance of five machine learning algorithms and their area estimation differences in the Suam region (40,000 ha) located near Mt. Elgon forest located along the Kenyan Ugandan border. The study utilized the ESA WorldCover and the Dynamic World consensus map to derive 863 training samples using an automated stratified random sampling technique. A medium-composite, cloud-free 2025 Landsat 9 image, SRTM DEM, and MERIT Hydro datasets were used to generate an 82-band feature stack composed of spectral metrics, spectral indices, texture bands, and terrain and hydro features to improve classifier accuracy. The final stack was used to generate LULC maps in Google Earth Engine (GEE) for each classifier using six LULC classes. Accuracy metrics such as overall accuracy (OA), kappa coefficient (K), mean F-1 scores (mF1), quantity disagreement (QD) and allocation disagreement (AD) were calculated for each classifier. Results showed that random forest (RF) had the highest OA (93.4%), K (0.91), and mF1 (0.922), followed by gradient tree boosting (GTB), classification and regression trees (CART), minimum distance (MD) and support vector machines (SVM), thereby achieving a balanced overall area estimation. Contrary to this, SVM experienced a complete model collapse, classifying all pixels as cropland while failing to accurately predict other land cover types. This research highlights the role of classifier choices in creating credible LULC maps for planning and management in heterogeneous tropical landscapes.