Weakly supervised localization model for plant disease based on Siamese networks.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Plant diseases pose a significant threat to crop yield and agricultural productivity.
• Current image-based disease detection methods struggle with symptom variability, leading to high false alarm rates.

Purpose of the Study:

• To develop an efficient, weakly supervised model for precise agricultural disease localization.
• To improve the accuracy and reduce false alarms in plant disease detection systems.

Main Methods:

• Utilized a Siamese neural network with a weight-sharing mechanism to capture visual differences in diseased plants.
• Developed and integrated the Agricultural Disease Precise Localization Class Activation Mapping (ADPL-CAM) algorithm for accurate localization.
• Evaluated the model's performance on various network architectures, including ResNet50 and SPDNet.

Main Results:

• ADPL-CAM demonstrated superior performance across all tested network architectures compared to GradCAM and SmoothCAM.
• Achieved a 3.96% higher top-1 accuracy and 27.09% higher average Intersection over Union (IoU) than GradCAM on ResNet50.
• On the SPDNet architecture, ADPL-CAM reached 54.29% top-1 accuracy and 67.5% average IoU, outperforming other methods.

Conclusions:

• The developed Siamese network model with ADPL-CAM effectively localizes plant diseases.
• The proposed method offers accurate and prompt identification and localization of diseased plant leaves.
• This approach has the potential to significantly enhance disease management strategies in agriculture.