Automated Forest Loss Detection Using Machine Learning Algorithms and Sentinel-2 Imagery on Google Earth Engine: A Study in Ca Mau Province, Vietnam

Abstract

This study presents the development and testing of machine learning models for automated forest loss detection in Ca Mau Province. The entire workflow, from preprocessing Sentinel-2 satellite imagery (Surface Reflectance - L2A) to model training and evaluation, was implemented on the Google Earth Engine (GEE) cloud computing platform. The study applied cloud filtering techniques to obtain images with cloud cover below 40%, while extracting 10 spectral bands and 8 vegetation indices, with delta indices (ΔNDVI, ΔEVI, ΔNBR, ΔSAVI) serving for model training and evaluation. Three machine learning algorithms were employed including Random Forest, Support Vector Machine (SVM), and Gradient Boosting Trees (GBT). The study utilized 2,500 sample points collected with a balanced ratio of 50% forest loss and 50% non-forest loss, and after extracting features for these sample points, the data was divided into 80% for training (train and validation) and 20% for testing, maintaining the 50:50 ratio in each subset. Evaluation results showed that the Gradient Boosting Trees model achieved the highest performance with an overall accuracy of 85.40%, recall of 72.00%, precision of 98.36%, and F1-score of 83.14%. With considerable recall and very high precision, the system demonstrates high reliability when alerting forest loss while detecting the majority of actual forest loss cases. The study successfully developed an automated forest loss detection tool on GEE, providing effective information to support local forest management efforts. However, further research on integrating Sentinel-1 radar data is needed to overcome cloud cover limitations during the rainy season and further enhance detection capabilities.