Towards practical AI for agriculture: A self-supervised attention framework for Spinach leaf disease detection

Area of Science:

• Agricultural Science
• Computer Science
• Plant Pathology

Background:

• Malabar spinach is vital in Bangladesh but vulnerable to Alternaria leaf spot and straw mite.
• Accurate disease identification is crucial for maintaining crop yield and quality.

Purpose of the Study:

• To develop an efficient and interpretable deep learning framework for automatic Malabar spinach leaf disease classification.
• To evaluate the performance of various deep learning architectures, including CNNs and Transformers, under limited data conditions.
• To address annotation scarcity using self-supervised pretraining.

Main Methods:

• A curated dataset of Malabar spinach images was created, categorized into healthy, Alternaria, and straw mite classes.
• Models like SpinachCNN, Spinach-ResSENet, SpinachViT, and SwinV2-Base were trained and evaluated.
• Self-supervised pretraining with SimSiam was employed on unlabeled data, followed by supervised fine-tuning.
• Domain-optimized models incorporated attention mechanisms (Squeeze-and-Excitation, Convolutional Block Attention Modules).

Main Results:

• The SimSiam-CBAM-ResNet-50 model achieved 97.31% test accuracy and high ROC-AUC, demonstrating robustness to noise.
• Transformer-based models like SwinV2-Base showed slightly higher accuracy but required extensive pretraining and more parameters.
• Interpretability methods (Grad-CAM, LayerCAM) confirmed model focus on relevant pathological regions.

Conclusions:

• The proposed SimSiam-CBAM-ResNet-50 offers a parameter-efficient and deployable solution for Malabar spinach disease detection.
• Self-supervised learning effectively mitigates annotation scarcity in agricultural datasets.
• Deep learning provides a powerful tool for precision agriculture, aiding in disease management and crop protection.