Early Detection of Jujube Shrinkage Disease by Multi-Source Data on Multi-Task Deep Network

Area of Science:

• Agricultural Science
• Plant Pathology
• Remote Sensing

Background:

• Shrinkage disease significantly impacts jujube quality and yield in arid regions.
• Current detection methods using RGB and hyperspectral imaging (HSI) struggle with early, subtle symptoms.

Purpose of the Study:

• To develop a non-destructive, high-precision method for early detection of jujube shrinkage disease.
• To enhance detection sensitivity by fusing multi-source imaging data.

Main Methods:

• Collected 317 'Junzao' jujube fruits across various infection and growth stages.
• Extracted morphological features from RGB images and identified key wavelengths from HSI using PCA, SPA, and GA.
• Developed a hybrid CNN-MLP architecture with dynamic feature weighting for multimodal data fusion.

Main Results:

• The proposed fusion method significantly improved detection accuracy and F1 scores compared to standalone MLP and CNN models.
• Achieved approximately 8.0% and 5.4% higher accuracy, and 38.6% and 32.4% higher F1 scores, respectively.
• Demonstrated robust performance across different growth stages and fruit colors.

Conclusions:

• Multi-source data fusion offers a superior approach for early jujube shrinkage disease detection.
• The developed CNN-MLP model provides a scalable solution for postharvest quality assessment.
• This technology can aid in improving jujube production quality and market value.