Li et al. (2025) Transformer-based detection of abnormal rice growth using drone-based multispectral imaging

Identification

• Journal: Computers and Electronics in Agriculture
• Year: 2025
• Date: 2025-10-08
• Authors: Yanfen Li, Lujuan Dang, Hanxiang Wang, Muhammad Fayaz, Sufyan Danish, Junliang Shang, Hyoung‐Kyu Song, Hyeonjoon Moon
• DOI: 10.1016/j.compag.2025.111055

Research Groups

• School of Computer Science, Qufu Normal University, Rizhao, China
• Institute of Research and Development, Duy Tan University, Da Nang, Viet Nam
• Faculty of Information Technology, Duy Tan University, Da Nang, Viet Nam
• Department of Information and Communication Engineering and Convergence Engineering for Intelligent Drone, Sejong University, Seoul, South Korea
• Department of Computer Science and Engineering, Sejong University, Seoul, South Korea

Short Summary

This study proposes ARG-TR, a lightweight transformer-based semantic segmentation model, to accurately detect various abnormal rice growth patterns using drone-based multispectral imaging, demonstrating superior performance and computational efficiency compared to existing state-of-the-art methods.

Objective

• To develop a lightweight, generalizable, and computationally efficient transformer-based semantic segmentation model for early and accurate detection of complex abnormal rice growth patterns from drone-based multispectral imagery, suitable for on-board UAV deployment.

Study Configuration

• Spatial Scale: Large-scale rice fields, enabling field-scale assessment of individual rice plants.
• Temporal Scale: Focused on early and timely detection of anomalies to mitigate yield losses.

Methodology and Data

• Models used: ARG-TR (a lightweight transformer-based semantic segmentation framework built on SegFormer architecture), MaskFormer (baseline), KNet (baseline).
• Data sources: Large-scale, drone-captured multi-spectral dataset.

Main Results

• The proposed ARG-TR model achieves rapid convergence during training.
• It demonstrates strong generalization capabilities to unseen data.
• ARG-TR attains a robust Intersection over Union (IoU) of 64.8 on a challenging dataset of abnormal rice growth patterns.
• The model outperforms state-of-the-art baselines, such as MaskFormer and KNet, in both accuracy and computational efficiency.

Contributions

• Introduction of ARG-TR, a novel lightweight transformer-based semantic segmentation framework specifically designed for detecting complex abnormal rice growth patterns.
• Addresses the limitations of previous methods that either focused on single symptoms or lacked generalization across diverse field conditions.
• Provides a computationally efficient solution suitable for real-time inference and on-board deployment on unmanned aerial vehicles (UAVs), overcoming the high computational cost of many high-accuracy models.
• Leverages long-range dependencies through a hierarchical transformer encoder to identify intricate growth anomalies.

Funding

No specific funding projects, programs, or reference codes were mentioned in the provided text.

Citation

@article{Li2025Transformerbased,
  author = {Li, Yanfen and Dang, Lujuan and Wang, Hanxiang and Fayaz, Muhammad and Danish, Sufyan and Shang, Junliang and Song, Hyoung‐Kyu and Moon, Hyeonjoon},
  title = {Transformer-based detection of abnormal rice growth using drone-based multispectral imaging},
  journal = {Computers and Electronics in Agriculture},
  year = {2025},
  doi = {10.1016/j.compag.2025.111055},
  url = {https://doi.org/10.1016/j.compag.2025.111055}
}