Zero-Shot Object Detection in Smart Farming for Pest/Non-Pest Classification

Abstract

Pest infestations cause substantial crop loss and financial harm, making them a key barrier to farm sustainability and production. Chemical pesticide treatment and labor-intensive inspection have been the mainstays of traditional pest identification, both of which are resource-intensive and environmentally taxing. The advancement of artificial intelligence (AI) in image processing provides a revolutionary approach for the automatic and real-time identification of pests. This research explored a deep learning-based pest/non-pest classification through the utilization of zero-shot object detection models, specifically Grounding DINO (GD) and Apple Image Model-V2 (AIM-V2). The models were evaluated on a dataset containing 523 high-quality images of pests that were randomly drafted out from a large dataset, including both pest and non-pest insects. The research assessed their effectiveness in detecting pests with key performance metrics, such as accuracy, precision, recall, and F1-score. Results indicated that AIM-V2 had an accuracy of 72.88%, surpassing the 62.69% accuracy of GD. While GD demonstrated greater precision in pest detection, this indicates that it was more robust in complex agricultural environments. A study was also conducted to compare predictions from the proposed models with farmer judgment, revealing that AIM-V2 was more consistent with local conditions. The study points toward the potential of zero-shot object detection models for enabling smarter farm practices.