Sustainable AI for plant disease classification using ResNet18 in few-shot learning

Abstract

Addressing the critical challenge of reduced crop production caused by plant diseases is essential to safeguard agricultural yield and quality. Conventional methods like visual inspection and laboratory testing are both time-consuming and costly. Although Modern AI-based deep learning techniques are promising, their potential in fields such as plant disease identification often remains unexplored due to the requirement of large and expert-labeled data. To mitigate these challenges, it is imperative to explore sustainable approaches that require minimal data while maintaining high accuracy in classification tasks. This research proposes a novel few-shot learning (FSL) framework employing a minimum sample size of 1 image and a maximum of 10 images per class for the accurate classification of plant diseases. The architecture incorporates a pre-training phase based on transfer learning as a feature extractor, followed by meta-learning using Prototypical Networks (ProtoNets) for class prototype computation and distance-based classification. The study evaluates the effectiveness of the proposed approach on the PlantVillage as well as rice disease datasets, performing comparative analyses among different transfer learning models such as ResNet18, ResNet50, and Vision Transformers in combination with Prototypical Networks under various N-way classification tasks (3-way, 5-way, and 10-way) and support sample (K-shot) settings (K = 1 , K = 5 , K = 10 ). The experimental findings indicate that the proposed combination of pretraining through ResNet18 with Prototypical Networks achieved an impressive accuracy of 93% and 75% on PlantVillage. The proposed model’s performance was further evaluated on rice disease data where it achieves the average accuracy of 75%. Specifically, the proposed model demonstrated the ability to classify 10 distinct plant diseases with high accuracy when provided with a suitable sample size per class. The proposed framework offers a substantial advancement in sustainable AI for plant disease recognition by enhancing the model generalization, enabling accurate classification across numerous classes with minimal sample size, and addressing data scarcity in AI-driven agricultural solutions. • The study demonstrates the effectiveness of Few-Shot Learning in plant disease classification, particularly in scenarios with scarce labeled data, providing a scalable solution for data-constrained environments. • The proposed transfer learning approach (ResNet18) was applied to novel tasks involving up to 10 classes on the PlantVillage dataset and achieved impressive accuracy rates. • The study contributes to the understanding of how increasing the number of support images can significantly improve classification accuracy, even when dealing with complex rice disease data characterized by noisy backgrounds. This finding reinforces the critical role of support data in enhancing model generalization, highlighting its importance for real-world applications in agricultural disease recognition where annotated data is often limited or challenging to obtain.

Topics & Concepts