Integrating Google Earth Engine and machine learning for urban land use and land cover dynamics analysis

The accurate land use and land cover (LULC) classification in the data-scarce urbanized region of Peshawar remains challenging due to computational limitations, accuracy assessment, and traditional techniques. This study, for the first time, addresses this research gap by introducing different robust machine learning (ML) algorithms in Google Earth Engine (GEE). The crux of this study is to analyze the comparative performances of four classifiers, namely, classification and regression tree (CART), minimum distance (MiD), random forest (RF), and support vector machine (SVM) within GEE using Sentinel data for reliable LULC classification from 2020 to 2024. The performance of each classifier was evaluated by validation and accuracy assessment. The composed points of each class were run in a scripted code and assigned 70% data for training and 30% for testing. The overall accuracy of RF and CART classifiers was 95% followed by the same values of Kappa coefficients. In contrast, MiD shows the weakest performance. The CART and RF classifiers maintain high producer accuracy (PA, > 90) and user accuracy (UA, > 90) for each class. The classification consistency was confirmed with mean Mathew correlation coefficient (MCC) values of 0.98 (for CART) and 0.99 (for RF), with an average F1 score of over 95%. The McNemar test showed no significant difference between CART, RF, and SVM classifiers; however, the confidence interval (CI $=$ 95%) confirmed the superior performance of CART and RF. This study confirms that the selected classifiers are transferable for a complex urban environment.