Comparative analysis of deep learning approaches for forest stand type classification: insights from the new VHRTreeSpecies benchmark dataset

Remotely sensed data are widely used for forest stand type classification; however, traditional classification methods are time consuming and mostly implemented for spatially limited areas and species, restricting the transferability of these models to other locations and diverse species. With the increasing availability of high-resolution satellite imagery, deep learning techniques have become widely used in forest management and species classification research. In this study, we generated a new benchmark dataset, VHRTreeSpecies, designed for forest stand type classification. We used Very High Resolution satellite images and corresponding labels derived from forest stand maps provided by the General Directorate of Forestry, covering 15 dominant tree species from Türkiye’s diverse forest ecosystems. Various deep learning models including CNN-based architectures (ResNet50, ResNet101, VGG16, ResNeXt, InceptionV3, ConvNeXt, EfficientNet), transformer models (DeepViT, Swin Transformer, CaiT, SegFormer-b0-b1-b2), and YOLO models (YOLOv8-l and YOLOv11-l) were utilized to assess the performance of these models for the forest tree species classification. Our results demonstrate the effectiveness of SegFormer, which achieved the highest classification accuracy of 96.25%, outperforming traditional CNN models. Additionally, YOLOv8-l proved to be a highly efficient model for both precision and real-time classification tasks, demonstrating its robustness across different tree species.