Convolutional Neural Network for Automatic Identification of Plant Diseases with Limited Data

Abstract

Automated identification of plant diseases is very important for crop protection. Most automated approaches aim to build classification models based on leaf or fruit images. These approaches usually require the collection and annotation of many images, which is difficult and costly process especially in the case of new or rare diseases. Therefore, in this study, we developed and evaluated several methods for identifying plant diseases with little data. Convolutional Neural Networks (CNNs) are used due to their superior ability to transfer learning. Three CNN architectures (ResNet18, ResNet34, and ResNet50) were used to build two baseline models, a Triplet network and a deep adversarial Metric Learning (DAML) approach. These approaches were trained from a large source domain dataset and then tuned to identify new diseases from few images, ranging from 5 to 50 images per disease. The proposed approaches were also evaluated in the case of identifying the disease and plant species together or only if the disease was identified, regardless of the affected plant. The evaluation results demonstrated that a baseline model trained with a large set of source field images can be adapted to classify new diseases from a small number of images. It can also take advantage of the availability of a larger number of images. In addition, by comparing it with metric learning methods, we found that baseline model has better transferability when the source domain images differ from the target domain images significantly or are captured in different conditions. It achieved an accuracy of 99% when the shift from source domain to target domain was small and 81% when that shift was large and outperformed all other competitive approaches.

Plant phenotyping relevance

植物画像から病害状態を推定するCNN手法の開発と比較評価が研究の中心であり、植物病害フェノタイピング手法に該当する。

Code and data availability

The paper's few-shot plant disease classification uses two public image datasets, both explicitly linked in the Data Availability Statement: PlantVillage (source domain) and the coffee leaf dataset (target domain). No author code or models are shared.