Multisource information fusion method for vegetable disease detection | JoVE Visualize

Area of Science:

Agricultural Science
Computer Vision
Artificial Intelligence

Background:

Automated detection of vegetable diseases is crucial for improving crop quality and farm profitability.
Existing methods struggle with complex greenhouse backgrounds, diverse diseases, and subtle visual differences, leading to low recognition rates.
Challenges include insufficient validation datasets and poor accuracy in pinpointing lesions and quantifying infection severity.

Purpose of the Study:

To develop an advanced automated system for precise detection and identification of vegetable diseases in greenhouse environments.
To address limitations of current methods by creating a robust model capable of handling complex visual data and improving recognition accuracy.
To provide technical support for comprehensive vegetable disease management and the advancement of smart agriculture practices.

Main Methods:

Established a self-built vegetable base in Shouguang ('Vegetable Town') for large-scale data collection of greenhouse disease images.
Introduced the Space-Time Fusion Attention Network (STFAN) to integrate multi-source information and enhance model resilience.
Proposed the Multilayer Encoder-Decoder Feature Fusion Network (MEDFFN) with Boundary Structure Loss to improve feature representation and boundary detail.

Main Results:

The proposed Multisource Information Fusion Method for Vegetable Disease Detection (MIFV) achieved improved mean Average Precision (mAP) by 3.43% over YOLOv7-tiny, 3.02% over YOLOv8n, and 2.15% over YOLOv9 on the VDGE dataset.
MIFV demonstrates excellent real-time performance and detection accuracy with 39.07 million parameters and 108.92 GFLOPS computational complexity.
The model achieved precise disease detection and identification by extracting high-resolution, multi-source feature representations.

Conclusions:

The MIFV model offers a significant advancement in automated vegetable disease detection, outperforming mainstream algorithms in accuracy and efficiency.
This technology provides a cost-effective solution for swift and accurate identification of greenhouse vegetable diseases.
The research contributes to the advancement of smart agriculture through enhanced disease prevention and control strategies.