| Abstract [eng] |
Forest health monitoring from unmanned aerial vehicle (UAV) imagery is essential for the timely detection of pine wilt disease (PWD), yet accurate pixel-wise delineation remains challenging because diseased crown regions are often small, fragmented, and visually confounded by surrounding canopy texture, shadows, and background clutter. This paper presents SilvaFormer, a transformer-based semantic segmentation framework for diseased-crown mapping in high-resolution RGB aerial imagery. The proposed model combines a hierarchical MiT/SegFormer-style encoder with a lightweight multi-scale fusion decoder, enabling the network to capture long-range contextual dependencies while preserving fine spatial detail. The final formulation uses a segmentation head for dense PWD prediction and an auxiliary boundary head during training to improve crown-edge delineation, together with differential fine-tuning of the pretrained encoder and task-specific decoding layers. We evaluate SilvaFormer on the public PWD-master UAV benchmark using a fixed train/validation/test split and compare it with representative segmentation baselines, including U-Net, PSPNet, DeepLabV3 +, ADE-Net, and two recent transformer-based baselines, SegFormer and UNetFormer. Against the classical CNN and attention-enhanced baselines, SilvaFormer achieves the strongest result, with an IoU of 0.7701, Dice/F1 of 0.8474, precision of 0.8449, and recall of 0.8730. In the additional transformer-focused comparison, SegFormer and UNetFormer achieve higher IoU values of 0.7842 and 0.8053, respectively, while SilvaFormer remains substantially more compact in parameter count (6.11M versus 28.15M and 30.21M). These results show that SilvaFormer provides a competitive and compact transformer-based solution for UAV-based forest health monitoring, while also clarifying its trade-off relative to larger recent transformer segmentation models. |
