Integrating spectral, textural, and topographic features for ecological forest type classification with XGBoost

Abstract

Forests play a critical role in ecological stability and resource management, and their sustainable management increasingly relies on advanced remote sensing and data-driven approaches. In this study, we evaluate the potential of multimodal satellite data for detailed ecological forest site type classification within a heterogeneous forested region of the Czech Republic. We compiled a comprehensive dataset of 156 features integrating Sentinel-2 multispectral bands, Sentinel-1 SAR backscatter, vegetation indices, seasonal median composites, texture metrics, topographic variables, and a high-resolution canopy height model. Using this feature set, an Extreme Gradient Boosting (XGBoost) classifier was trained to map 16 ecological forest site type groups, dominated by rich/enriched beech-oak forests, beech-oak forests, and loamy oak-beech forests. The model achieved an overall accuracy of 88.4% and a macro-averaged F1-score of 0.89, with the highest classification performance obtained for fir forests, other beech forests, and calcareous forest types. Feature importance analysis shows that radar texture features, seasonal vegetation indices, and topographic variables are among the most influential predictors, while canopy height texture metrics further enhance class separability. These results demonstrate that combining diverse remote sensing modalities with advanced machine learning enables reliable, fine-scale mapping of complex ecological forest site types, extending beyond traditional tree-species classification and offering a scalable and cost-effective tool for ecological research and forest management planning.