A novel framework GRCornShot for corn disease detection using few shot learning with prototypical network.

Abstract

Precision and timeliness in the detection of plant diseases are important to limit crop losses and maintain global food security. Much work has been performed to detect plant diseases using deep learning methods. However, deep learning techniques demand a large quantity of data to train the models for diagnosis and further classification. Few-shot learning has surfaced to remove the drawbacks of deep learning methods. Therefore, the proposed work presents a novel GRCornShot model for corn disease diagnosis using few-shot learning with Prototypical Networks based on metric learning. Metric Learning calculates the distance to measure the similarity between the data points. Hence, addressing the challenge of limited labeled data, GRCornShot effectively classifies healthy and corn diseases. Furthermore, the Gabor filter is incorporated into the backbone network ResNet-50 to extract the texture features and to enhance the classification performance. The experiments show the promising application of few-shot learning in agronomic applications, providing a robust solution for detecting corn diseases precisely with minimal data requirements. Using a 4-way 2-shot, 3-shot, 4-shot, and 5-shot learning strategy, GRCornShot achieves impressive accuracy of 96.19%, 96.54%, 96.90%, and 97.89%, respectively.

Plant phenotyping relevance

トウモロコシ葉の病害状態を画像から分類する少数ショット深層学習手法を開発し、精度を評価しており、植物フェノタイピング手法が中心である。

Code and data availability

The paper's corn disease detection experiments use two publicly available image datasets (Roboflow corn-disease and the Kaggle corn/maize leaf disease dataset), explicitly linked in the Data Availability statement. No author analysis code or trained model checkpoint is deposited.