Segment Anything Model and Fully Convolutional Data Description for Plant Multi-Disease Detection on Field Images

Abstract

Researchers have designed various models trained on public or private datasets for plant disease detection to help farmers remedy crop yield losses on their farms due to plant diseases. Plantvillage is the most widely used plant disease dataset with laboratory images captured under controlled conditions with a single leaf on each image and a uniform background. Models trained on such datasets have extremely low classification accuracies when running on field images captured directly from plantations with various interwoven leaves, complex backgrounds, and different lighting conditions. In this study, we propose a model ensemble solution for the accurate identification and classification of plant diseases using field images. The model uses Segment Anything Model to efficiently circumscribe all identifiable objects in the image. Image Processing techniques are then used to isolate the identified objects from the original image. Background objects are separated from actual leaf objects using Fully Convolutional Data Description, which is an explainable deep one-class classification model for anomaly detection. Finally, the selected leaves are submitted to a Plantvillage-trained classification model for inference. Our model can detect diseases appearing on individual leaves of the same image and improves classification accuracy by more than 10% on public field plant disease datasets such as PlantDoc, thus providing a reliable solution for farmers and practitioners.

Keywords

• Field plant images
• fully convolutional data description
• laboratory images
• plant disease dataset
• plant disease detection and classification
• segment anything model