TDAVM-UNet: task-driven attention VM-UNet for crop disease detection from UAV imagery.

Area of Science:

• Agricultural Science
• Computer Science
• Remote Sensing

Background:

• Crop diseases significantly threaten agricultural yield and global food security.
• Accurate crop disease detection using Unmanned Aerial Vehicle (UAV) remote sensing is crucial for precision agriculture.
• Challenges include complex backgrounds, diverse spectral-spatial features, irregular lesion boundaries, and variable textures.

Purpose of the Study:

• To propose a novel deep learning model, Task-Driven Attention VM-UNet (TDAVM-UNet), for accurate crop disease detection from UAV imagery.
• To address the limitations of existing methods in handling complex field conditions and disease variabilities.
• To provide a computationally efficient and robust solution for precision agriculture.

Main Methods:

• Developed TDAVM-UNet integrating Disease-Aware Dynamic Attention (DADA) and Channel-Spatial Visual State Space (CSVSS) modules.
• DADA enhances diseased region representation via feature enhancement, multi-scale channel attention, and texture-guided spatial attention.
• CSVSS enables efficient long-range dependency modeling and feature fusion with linear complexity.
• Employed a hybrid loss strategy (BCE, Dice, CE) to tackle class imbalance and boundary delineation.

Main Results:

• TDAVM-UNet achieved 82.22% mean Intersection over Union (mIoU) on a diverse UAV crop disease dataset.
• The model demonstrates O(N) linear computational complexity, significantly outperforming TransUNet (80% lower GFLOPs).
• Achieved 26.87M parameters and 31.45 GFLOPs for 256x256 inputs, indicating high efficiency.

Conclusions:

• TDAVM-UNet offers a high-accuracy, robust, and computationally efficient method for UAV-based crop disease detection.
• The proposed model provides significant technical support for precision agriculture applications.
• This research contributes to improving crop yield monitoring and safeguarding global food security.