Identification of some tree species from sentinel-2 imagery using different machine learning algorithms: A case study of the Aladağ Forest Enterprise

The sustainable management of forest ecosystems and the conservation of biological diversity, particularly in the contemporary context, necessitate the accurate analysis of environmental changes and the development of detection and monitoring methods that provide reliable results at the species level from both scientific and practical perspectives. This doctoral study aims to classify five different tree species located in a forested area of Bolu-Aladağ using remote sensing techniques. Sentinel-2 satellite imagery served as the primary data source, and a comprehensive dataset was constructed by deriving 12 distinct vegetation, water, and soil indices, supplemented by topographic variables. The resulting dataset comprises 29 independent variables. To perform the classification, several machine learning algorithms were employed, including Random Forest (RF), Support Vector Machines (SVM) with various kernel functions, Artificial Neural Network (ANN) and Ensemble learning (EL) methods. The classification performance of these algorithms was comparatively evaluated, revealing that the RF algorithm outperformed the others in terms of overall accuracy. Additionally, the impact of seasonal variation on classification success was assessed, showing notably higher accuracy levels during the summer season. According to the results, the highest accuracy values for each species were obtained with different algorithms: for Black pine 88.06% with ANN (April), for Scotch pine 96.91% with linear SVM (November), for Fir 85.44% with RBF-based SVM (August), for Beech 85.66% with EL (August), and for Sessile oak 83.70% with RBF-based SVM (April). The findings indicate that species separability is directly associated with spectral band selection, data resolution, and the interaction of seasonal phenology. The study demonstrates that enhancing the performance of classification models requires not only algorithm optimization but also the adoption of an appropriate data strategy.