Real-Time Detection and Classification of Scirtothrips dorsalis on Fruit Crops with Smartphone-Based Deep Learning System: Preliminary Results

Abstract

Simple Summary This study developed a real-time thrips detection application to classify the Scirtothrips dorsalis from other thrips species, and was optimized for working on embedded devices, such as smartphones. The performances of several deep learning models, including YOLOv5, Faster R-CNN, SSD MobileNetV2, and EfficientDet-D0, were evaluated based on their accuracy and speed in detecting thrips on yellow sticky traps. The models were trained and tested on two datasets containing thrips and non-thrips insects captured under different lighting conditions (reflectance, transmittance, and reflectance + transmittance). The developed application used the EfficientDet-D0 model trained on the transmittance dataset to detect and visualize the presence of thrips in real-time, making it a valuable tool for the monitoring and management of thrips in agriculture, with practical implications for pest control and crop yield improvement. This study proposes a deep-learning-based system for detecting and classifying Scirtothrips dorsalis Hood, a highly invasive insect pest that causes significant economic losses to fruit crops worldwide. The system uses yellow sticky traps and a deep learning model to detect the presence of thrips in real time, allowing farmers to take prompt action to prevent the spread of the pest. To achieve this, several deep learning models are evaluated, including YOLOv5, Faster R-CNN, SSD MobileNetV2, and EfficientDet-D0. EfficientDet-D0 was integrated into the proposed smartphone application for mobility and usage in the absence of Internet coverage because of its smaller model size, fast inference time, and reasonable performance on the relevant dataset. This model was tested on two datasets, in which thrips and non-thrips insects were captured under different lighting conditions. The system installation took up 13.5 MB of the device's internal memory and achieved an inference time of 76 ms with an accuracy of 93.3%. Additionally, this study investigated the impact of lighting conditions on the performance of the model, which led to the development of a transmittance lighting setup to improve the accuracy of the detection system. The proposed system is a cost-effective and efficient alternative to traditional detection methods and provides significant benefits to fruit farmers and the related ecosystem.