Pest classification: Explainable few-shot learning vs. convolutional neural networks vs. transfer learning

Abstract

Pests significantly threaten plant yield and overall agricultural productivity, leading to reduced output in the farming industry. Accurate and automated detection of crop insect pests is crucial for effective pest control and optimal utilization of agricultural resources. This study addresses the problem of limited datasets in pest detection by exploring the potential of Explainable Few-Shot Learning (FSL), a machine learning approach that not only enables learning from a small amount of data but also provides interpretable insights into the decision-making process. Unlike traditional pest detection studies that rely on large labeled datasets or black-box models, this research introduces an advanced methodology by integrating explainability techniques such as Grad-CAM into FSL models, specifically Prototypical Network and Siamese Network. This dual approach ensures high accuracy with minimal training data while identifying key image features influencing predictions, thereby enhancing transparency and trust. A comparative analysis was conducted against Convolutional Neural Network (CNN) and transfer learning models using full pest images, half pest images, and Malaysian pest images. This study found that Explainable FSL achieved the highest accuracy of 99.81 % in various scenarios, including 9-way 1-shot, 3-shot, 5-shot, and 10-shot configurations, outperforming both CNN and transfer learning models. These findings demonstrate that Explainable FSL models can significantly improve the accuracy, transparency, and efficiency of pest detection systems, even with limited data. By advancing both the detection capabilities and interpretability of Artificial Intelligence (AI) systems, this research provides a novel contribution to smart agriculture, enabling robust pest detection systems tailored to real-world, data-scarce scenarios.