TDAVM-UNet: task-driven attention VM-UNet for crop disease detection from UAV imagery
Frontiers in Plant Science · 29.06.2026 · 17
DOI: 10.3389/fpls.2026.1855733 ↗Kurz gefasst
Fachartikel in der Zeitschrift Frontiers in Plant Science, online erschienen am 29.06.2026. Verfasst von Shanwen Zhang und 3 weiteren Autor:innen. Beteiligte Einrichtung: School of Electronic Information, XiJing University, Xi'an, China. Die Arbeit nennt 31 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Crop diseases pose a serious threat to agricultural yield and global food security. Accurate detection using Unmanned Aerial Vehicle (UAV) remote sensing imagery is of great significance for precision agriculture. However, this task remains challenging due to complex field backgrounds, diverse spectral-spatial characteristics of diseased leaf regions, irregular lesion boundaries, and variable texture patterns. To address these issues, this paper proposes a Task-Driven Attention VM-UNet (TDAVM-UNet), a novel deep learning model for crop disease detection from UAV imagery. The model integrates two task-driven attention modules: (1) Disease-Aware Dynamic Attention (DADA), which enhances the representation of diseased regions through disease feature enhancement, multi-scale dynamic channel attention, and texture-guided spatial attention; and (2) Channel-Spatial Visual State Space (CSVSS), which enables efficient long-range dependency modeling and local-global feature fusion while maintaining linear computational complexity. A hybrid loss strategy combining binary cross-entropy (BCE) loss, Dice loss, and cross-entropy (CE) loss with optimized coefficients is employed to address class imbalance and boundary delineation challenges. Extensive experiments are conducted on a self-constructed UAV crop disease unified mix dataset, comprising soybean disease images from Maharashtra, India, and rust disease images from wheat, corn, and other crops in Yangling, China, totaling 6,680 raw collected images, which after deduplication yields 5,000 images for experimentation. The results demonstrate that TDAVM-UNet achieves 26.87M parameters and 31.45 GFLOPs for 256×256 inputs, maintaining O(N) linear complexity (80% lower than TransUNet’s 156.78 GFLOPs), with 82.22% mIoU. This work provides a high-accuracy, robust, and computationally efficient method for UAV-based crop disease detection, offering significant technical support for precision agriculture applications.
Schlagwörter
Stichwörter der Autor:innen
Autor:innen und Institutionen
- School of Electronic Information, XiJing University, Xi'an, China.
- School of Network and Space Security, Xidian University, Xi'an, China.
Zitieren
APA
Zhang, S., Xu, C., Zhao, Y., & Zhang, T. (2026). TDAVM-UNet: task-driven attention VM-UNet for crop disease detection from UAV imagery. Frontiers in Plant Science, 17. https://doi.org/10.3389/fpls.2026.1855733
Vancouver
Zhang S, Xu C, Zhao Y, Zhang T. TDAVM-UNet: task-driven attention VM-UNet for crop disease detection from UAV imagery. Frontiers in Plant Science. 2026;17. doi:10.3389/fpls.2026.1855733
BibTeX
@article{Zhang2026,
author = {Zhang, Shanwen and Xu, Cong and Zhao, Yihang and Zhang, Ting},
title = {TDAVM-UNet: task-driven attention VM-UNet for crop disease detection from UAV imagery},
journal = {Frontiers in Plant Science},
year = {2026},
volume = {17},
doi = {10.3389/fpls.2026.1855733}
}
EUVIMED