Evaluating Deep Learning Architectures for Potato Pest Identification: A Comparative Study of NasNetMobile, DenseNet, and Inception Models

Abstract

Manual potato pest identification that is still applied today is often time-consuming and highly dependent on farmer skills in the field. This causes delays in taking action and inaccurate reporting, especially in pest emergencies. In addition, these limitations slow down the response to pest control which ultimately risks reducing crop yields and farmer income. This study aims to develop a more accurate, fast, and consistent deep learning-based approach to identify potato pests, in order to support practical solutions that farmers can implement independently. This study contributes by comparing three deep learning architecture models, namely NasNetMobile, DenseNet, and Inception which are designed to identify pest images. The potato pest image dataset used was collected from various sources equipped with an augmentation process to increase data diversity. The model was drilled using transfer learning techniques to utilize previously learned features on a large dataset. The evaluation model was carried out comprehensively based on accuracy, precision, and inference time efficiency. The results showed that the DenseNet model achieved the highest accuracy of 97% with an inference time of 11 seconds, and this model maintained a relatively stable performance and was superior several times compared to other models. Based on these results, DenseNet was chosen as the most effective and reliable model to be developed for practical applications in the field. This study provides practical implications in the form of providing a model that can be integrated into a mobile-based application that is easy to use by farmers, including in remote areas. This allows farmers to identify pests independently without requiring in-depth technical expertise. In addition, this study is a new benchmark for the development of artificial intelligence-based pest identification systems in other crops and opens up opportunities for integration with IoT-based technologies to support sustainable agricultural practices.