Diagnosis and Optimization of Citrus Pests Using a Lightweight Deep Learning Model

Pest outbreaks in citrus reduce the quality of the crop and reduce production, so it is necessary to detect outbreaks in a timely manner. Image classification models commonly used to diagnose crop dis- eases are highly accurate but costly to train and require high-performance equipment. Since devices utilized in the agricultural field are limited in terms of memory and power, efficient deep learning mod- els are needed. Therefore, in this paper, we studied a lightweight model for disease diagnosis in crops. We implemented a disease diagnosis model for crops using tangerine fruit and leaf images. The per- formance of the implemented models was compared through accuracy versus FLOPs, and MobileNet V3 Small was the best among the lightweight models. We further applied data balancing and performance enhancement techniques. When data balancing was performed, the average accuracy of the model de- creased, but the classes were more evenly classified. As for the performance improvement, the Label smoothing technique improved the accuracy by 1.2%. Therefore, it is recommended that the citrus pest diagnosis model for mobile devices should be preprocessed with data balancing and smoothing techni- ques and implemented using the MobileNet V3 Small model for optimal user experience and accuracy.