A Mobile-Based Deep Learning Model for Cassava Disease Diagnosis.

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Pathology

Background:

• Traditional plant disease phenotyping relies on visual diagnostics, requiring specialized expertise.
• Convolutional neural network (CNN) models offer potential for automated plant disease identification.
• Deploying CNNs on mobile devices introduces challenges like variable lighting and orientation.

Purpose of the Study:

• To train and evaluate a CNN object detection model for identifying cassava foliar diseases under real-world conditions.
• To assess the model's performance in detecting varying disease severities (mild and pronounced) for early symptom detection.
• To investigate the impact of mobile deployment, including image and video input, on model accuracy.

Main Methods:

• A CNN object detection model was trained to identify cassava foliar disease symptoms.
• The model was deployed in a mobile application and tested in a field setting in Tanzania.
• Performance was evaluated using 720 images and videos of diseased cassava leaflets, considering two severity levels.

Main Results:

• A decrease in model performance (F-1 score) was observed for real-world images and videos compared to training data.
• Performance dropped by 32% for pronounced symptoms in real-world images, primarily due to reduced model recall.
• Model performance varied between image and video input, indicating data type as a critical factor.

Conclusions:

• Real-world conditions significantly impact CNN model performance for plant disease phenotyping on mobile devices.
• Optimizing model recall is crucial for achieving desired performance in practical applications, especially for early disease detection.
• Consideration of input data type (image vs. video) is essential for designing effective mobile CNN applications in agriculture.