FFGF-Net: A Forestry Fine-Grained Fusion Network for tree species classification from UAV imagery

Accurate classification of tree species is crucial for biodiversity conservation, forest resource management, and intelligent breeding. Traditional manual survey methods are inefficient and costly, while the AI-driven paradigm offers a new approach for intelligent forestry through computer vision technologies, enabling large-scale, high-precision tree species identification. However, fine-grained classification of tree species remains challenging due to high intra-class variation, low inter-class distinction, and complex environmental noise in real-world forest scenarios. To bridge the intra- and inter-class discrepancy, this study proposes a Forestry Fine-Grained Fusion Network (FFGF-Net), which integrates three key innovations: a CanopyPatchExtractor for local feature extraction, a CanopyTextureAnalyzer for multi-scale texture modeling, and a Cascaded Canopy Attention mechanism for noise suppression. For comprehensive evaluation, we introduce a newly collected UAV dataset, NJFUTreeData, and utilize established public datasets, including SZUTreeData and ETH dataset. Experimental results demonstrate that FFGF-Net achieves state-of-the-art performance, with Overall Accuracy (OA) of 75.00% on NJFUTreeData and 94.75% on the ETH dataset. Notably, our method consistently outperforms a wide range of existing model families, including representative CNN backbones (ConvNeXt-Tiny), lightweight architectures (e.g., EfficientNet-B0, MobileNet_V3-Small), and advanced Vision Transformers (e.g., Swin-Tiny). The proposed network provides a robust technical foundation for automated forest resource surveys, ecological monitoring, and smart forestry applications. • Proposes FFGF-Net for complex fine-grained tree classification. • Achieves SOTA accuracy on both custom and public tree datasets. • Validates generality and robustness on other fine-grained benchmarks. • Achieves high efficiency to support edge deployment.