A Comparative Analysis of Lightweight and Non-Lightweight Deep Learning Models for Agricultural Pest Identification

Abstract

A key component of effective pest control is pest recognition, which involves the exact identification of pests and the prediction of outbreaks through the analysis of data from statistics, mathematics, and biology. More effective identification techniques are required as traditional field surveys, despite their depth, are labor-intensive and have a limited reach. This skill is crucial for the early detection of pests and diseases, enabling prompt actions to prevent their spread. With the advancement of artificial intelligence, approaches like Convolutional Neural Networks (CNNs) can handle vast amounts of visual data gathered by drones, cellphones, and other devices for large-scale agricultural applications, allowing for automatic, real-time monitoring. However, the bulkiness of standard CNN models makes them challenging to deploy on small mobile devices. To achieve effective and precise pest detection in a mobile environment, this study analyzes various lightweight and non-lightweight models for pest recognition. Results indicate that lightweight models show potential for mobile deployment with reasonable accuracy. Furthermore, with additional training and optimization, lightweight models can be effectively utilized for mobile pest identification, offering a promising direction for integrating advanced pest prediction technologies into practical agricultural applications.