A maturity classification model for winter jujubes based on DSAF-ResNet.

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Accurate, non-destructive winter jujube maturity classification is essential for optimizing harvest timing and ensuring fruit quality.
• Current methods often lack the precision needed to differentiate subtle maturity stages, impacting post-harvest processes.

Purpose of the Study:

• To develop and evaluate a novel deep learning model for accurate, non-destructive classification of winter jujube maturity.
• To investigate the efficacy of fusing hyperspectral and texture features for enhanced maturity assessment.

Main Methods:

• A dual-stream attention-fused residual network (DSAF-ResNet) was proposed, integrating hyperspectral imaging and Gray-Level Co-occurrence Matrix (GLCM) texture features.
• The network incorporated RepVGGBlock and SimAM attention mechanisms within a dual-stream architecture.
• Model performance was validated using test accuracy, precision, and recall metrics.

Main Results:

• The fused multimodal approach significantly improved classification performance compared to single-modality inputs.
• The DSAF-ResNet achieved high test accuracy (97.24%), precision (97.31%), and recall (97.24%).
• Ablation studies confirmed the effectiveness of individual network components and the fusion strategy.

Conclusions:

• The DSAF-ResNet offers an effective and scalable framework for non-destructive fruit maturity classification.
• This approach enhances intelligent agricultural practices and supports precision agriculture by enabling robust maturity assessment.
• The model demonstrates excellent generalization and stability, even with imbalanced datasets.