Improving National Forest Mapping in Romania Using Machine Learning and Sentinel-2 Multispectral Imagery

Abstract

Forest attributes, such as standing stock, diameter at breast height (DBH), tree height, and basal area, are critical for effective forest management; yet, traditional estimation methods remain labor-intensive and often lack the spatial detail required for contemporary decision-making. This study addresses these challenges by integrating machine learning algorithms with high-resolution remotely sensed data and rigorously collected ground truth measurements to produce accurate, national-scale maps of forest attributes in Romania. To ensure the reliability of the model predictions, extensive field campaigns were conducted across representative Romanian forests. During these campaigns, detailed measurements were recorded for every tree within selected plots. For each tree, DBH was measured directly, and tree heights were obtained either by direct measurement—using hypsometers or clinometers—or, when direct measurements were not feasible, by applying well-established DBH—height allometric relationships that have been calibrated for the local forest types. This comprehensive approach to ground data collection, supplemented by an independent dataset from Brasov County collected using the same protocols, allowed for robust training and validation of the machine learning models. This study evaluates the performance of three machine learning algorithms—Random Forest (RF), Classification and Regression Trees (CART), and the Gradient Boosting Tree Algorithm (GBTA)—in predicting the forest attributes from Sentinel-2 satellite imagery. While Random Forest consistently delivered high R2 values and low root mean square errors (RMSE) across all attributes, GBTA showed particular strength in predicting standing stock, and CART excelled in basal area estimation but was less reliable for other attributes. A sensitivity analysis across multiple spatial resolutions revealed that the performance of all algorithms varied significantly with changes in resolution, emphasizing the importance of selecting an appropriate scale for accurate forest mapping. By focusing on both the methodological advancements in machine learning applications and the rigorous, detailed empirical forest data collection, this study provides a clear solution to the problem of obtaining reliable, spatially detailed forest attribute maps.