Lightweight Few Shot Model for Plant Leaf Disease Classification Using an Aggregated Loss Function

Background: It is essential to monitor agriculture's growth and health status of plants to help farmers take prompt actions regarding any plant disease. An automatic plant disease classification algorithm is necessary to identify and prevent diseases in plants rapidly. Deep learning and convolutional neural networks consisting of millions of tunable parameters give good performance in the plant disease classification tasks compared to the manual feature extraction methods. This breakthrough of deep learning is possible due to the availability of large annotated datasets, and the lack of annotated samples may result in performance degradation. To tackle this problem, the researchers are proposing few-shot learning concepts. Few-shot classification aims to identify unseen classes while training with small labeled data.Aim: This paper aims to classify plant diseases by considering a few samples and lightweight transfer learning-based deep learning models that could be effectively installed in the resource constraint devices.Method: This paper proposes an aggregated loss function for a few-shot classification to classify plant diseases using the plant village dataset effectively. The dataset contains 38 classes of leaf diseases, having 54,303 leaf samples. Different domain splits were considered to divide the dataset into the source and target domains. A large quantum of experimentation is conducted on the target dataset considering various sample sizes. MobileNetV2 is regarded as the base model for this work, and the proposed enhanced loss function is used to classify the samples effectively.Results: This work considers four domain splits for different comparison purposes, and it is found that an average improvement in accuracy is 1.49% for split-1, 16.25% for split-2, 2.9% for split-3, and2.1% for split-4 using the proposed TC-loss and the transfer learning-based lightweight model for the target domain data (K-ways, N-shot). Also, the source domain images (containing more than 1000 images per class) are evaluated. It is seen that the proposed method outperforms the existing literature in the source domain analysis.Conclusion: Proposed work is compared with several state-of-the-art works of literature using different evaluation metrics, sample sizes, and loss functions, and satisfactory result is obtained in each case