ACG-Net: an attention context-guided network for fine segmentation of Fagus canopies in high-resolution UAV imagery

The genus Fagus, a key community-forming taxon in northern temperate forests, plays a vital role in maintaining biodiversity and ecosystem functions. However, natural Fagus forests are threatened by poor regeneration and community degradation, underscoring the need for precise and efficient monitoring techniques. Existing methods are constrained by the lack of public datasets and the limitations of standard architectures like U-Net in capturing fine-grained features within complex forest scenes.

To address these challenges, we constructed a high-resolution UAV image segmentation dataset for Fagus and proposed the Environment Simulation-based Robust Data Augmentation Framework (ES-REF). By actively simulating realistic disturbances such as fog, local overexposure, and motion blur, ES-REF significantly enhances model generalization under complex conditions. Additionally, we developed ACG-Net, which uses VGG as the encoder backbone and incorporates SPConv, Criss-Cross Attention, and context-guided downsampling to improve multi-scale feature extraction, global context awareness, and spatial detail preservation.

Experimental results demonstrate that ES-REF improves model robustness, increasing the mIoU of U-Net and ACG-Net by 0.71 and 2.18 percentage points, respectively. On the test set, ACG-Net achieved an mIoU of 89.55%, outperforming the U-Net baseline by 4.31 percentage points and surpassing models such as DeepLabv3+.

This work establishes a data and methodological foundation for automated Fagus community mapping and provides a reliable framework for forest resource monitoring and smart forestry management.