Mapping Forest Cover Using Sentinel-2 Imagery and Machine Learning Techniques

Abstract

Accurate and up-to-date forest land cover information is essential for environmental monitoring, biodiversity conservation, and sustainable land management. The increasing availability of high-resolution satellite imagery combined with advances in machine learning (ML) techniques offers new opportunities for improving forest mapping accuracy. In this study, we evaluate and compare the potential of several machine learning algorithms for Mediterranean forest land cover mapping using Sentinel-2 multispectral imagery. A comprehensive set of predictor variables was derived from Sentinel-2 data, including, textural features based on gray-level co-occurrence matrices (GLCM), and topographic variables (elevation and slope). Reference samples were generated using Google Earth Pro and used to train and test multiple ML models, including KNN, Random Forest, Gradient Tree Boost. Model performance was assessed using standard accuracy metrics, including overall accuracy, precision, F1-score. The results reveal notable differences in classification performance among the tested algorithms, highlighting the influence of model structure and feature utilization on forest mapping accuracy. Tree-based ensemble methods generally outperformed simpler classifiers, particularly in heterogeneous landscapes. The findings demonstrate the added value of integrating multi-source features and advanced machine learning approaches for reliable forest land cover mapping. This comparative analysis provides valuable insights into the strengths and limitations of different ML algorithms and supports the selection of appropriate models for large-scale forest land cover mapping using Sentinel-2 imagery. Keywords: Forest land cover; Sentinel-2; Machine learning; Land cover classification; Textural features; GLCM; Topographic variables