PGCNet: a Transformer-CNN hybrid segmentation model for pine wilt disease identification

Area of Science:

• Forestry
• Remote Sensing
• Computer Vision

Background:

• Pine wilt disease is a severe threat to forest ecosystems due to its rapid spread and high mortality.
• Current UAV-based identification methods using single CNN or Transformer architectures have limitations in capturing global context and local details.
• Existing feature fusion strategies lack effective hierarchical interaction, leading to poor integration of semantic and detailed information and high computational costs.

Purpose of the Study:

• To develop an efficient semantic segmentation model for accurate pine wilt disease identification from UAV remote sensing imagery.
• To address limitations of existing methods by effectively fusing CNN and Transformer representations.
• To create a computationally efficient model suitable for edge deployment in real-time forestry monitoring.

Main Methods:

• Proposed PGCNet model, a semantic segmentation network that fuses Convolutional Neural Network (CNN) and Transformer features.
• Utilized CSWin Transformer as the backbone for comprehensive global contextual information.
• Introduced a Progressive Guidance Fusion Module (PGFM) with spatial-channel attention for cross-layer feature fusion.
• Incorporated a lightweight Context-Aware Residual Atrous Spatial Pyramid Pooling (CAR-ASPP) module for multi-scale feature enhancement and reduced complexity.

Main Results:

• PGCNet demonstrated superior performance over mainstream semantic segmentation models across various metrics.
• The model showed strong accuracy in complex backgrounds and for identifying small-scale disease targets.
• Achieved high accuracy with excellent computational efficiency, outperforming existing methods.

Conclusions:

• PGCNet offers a practical and efficient solution for real-time pine wilt disease detection using UAV imagery.
• The model's computational efficiency makes it suitable for edge computing environments in forestry.
• The approach shows potential for broader application in agricultural remote sensing for disease identification.