Zhang et al. (2025) Automatic recognition of wheat growth stages with a lightweight multimodal data fusion network

Identification

• Journal: Computers and Electronics in Agriculture
• Year: 2025
• Date: 2025-12-26
• Authors: Lulu Zhang, Bo Zhang, Huanhuan Zhang, Libang Chen, Zhenpeng Zhang, Jianrong Cai, Chundu Wu, Xiaowen Wang
• DOI: 10.1016/j.compag.2025.111373

Research Groups

• School of Computer and Information, Anqing Normal University, China
• School of the Environment and Safety Engineering, Jiangsu University, China
• School of Agricultural Engineering, Jiangsu University, China
• School of Food and Biological Engineering, Jiangsu University, China
• Key Laboratory for Theory and Technology of Intelligent Agricultural Machinery and Equipment, Jiangsu University, China
• Jiangsu Province and Education Ministry Cosponsored Synergistic Innovation Centre of Modern Agricultural Equipment, Jiangsu University, China

Short Summary

This study proposes a lightweight multimodal data fusion network, leveraging UAV-acquired RGB, multispectral, and digital surface model data, to accurately recognize wheat growth stages. The developed MobileNetV3-Small-based model achieves 99.57% accuracy and high computational efficiency, effectively overcoming limitations of single-modal methods and enabling real-time deployment on edge devices.

Objective

• To develop an efficient and accurate method for automatic wheat growth stage recognition that overcomes the limitations of single-modal data (e.g., difficulty distinguishing adjacent stages) and the high computational demands of complex deep learning networks, enabling practical deployment on resource-constrained edge devices.

Study Configuration

• Spatial Scale: Field scale, focusing on individual wheat plant canopies.
• Temporal Scale: Across the wheat growing season, covering various growth stages.

Methodology and Data

• Models used: Lightweight multimodal data fusion network with MobileNetV3-Small as the backbone. Compared against MobileNetV3-Large, ResNet-18, MNASNet, EfficientNet-B0, and ConvNeXt-Tiny.
• Data sources: Unmanned Aerial Vehicle (UAV)-acquired RGB images, multispectral (MS) data, digital surface models (DSM), and derived spectral vegetation indices (VI).

Main Results

• The proposed multimodal fusion model achieved a growth stage recognition accuracy of 99.57%, precision of 99.58%, recall of 99.57%, and F1 score of 99.57%.
• It significantly improved stage differentiation in critical transitions, such as Booting to Heading, reducing misclassification risks.
• MobileNetV3-Small demonstrated the best trade-off between accuracy and resource efficiency, featuring 1.53 million parameters and an inference time of 6.03 milliseconds on an RTX 4090 and 25.11 milliseconds on a Jetson Orin NX.
• The high efficiency enables real-time deployment on resource-constrained edge devices.

Contributions

• Introduces a novel lightweight multimodal data fusion network for wheat growth stage recognition, addressing the challenges of distinguishing phenologically adjacent stages and computational overhead.
• Demonstrates that fusing RGB, multispectral, and DSM data significantly enhances recognition accuracy compared to single-modal approaches.
• Proposes an efficient deep learning solution (MobileNetV3-Small backbone) suitable for real-time deployment on edge devices, overcoming the computational constraints of complex networks.
• Provides a robust and accurate solution that supports timely agronomic decision-making in precision agriculture.

Funding

Not specified in the provided text.

Citation

@article{Zhang2025Automatic,
  author = {Zhang, Lulu and Zhang, Bo and Zhang, Huanhuan and Chen, Libang and Zhang, Zhenpeng and Cai, Jianrong and Wu, Chundu and Wang, Xiaowen},
  title = {Automatic recognition of wheat growth stages with a lightweight multimodal data fusion network},
  journal = {Computers and Electronics in Agriculture},
  year = {2025},
  doi = {10.1016/j.compag.2025.111373},
  url = {https://doi.org/10.1016/j.compag.2025.111373}
}