Development and evaluation of an automated classification and counting system for rice planthoppers captured on survey boards.

Rice planthoppers are the most economically important insect pests of rice in Asia. Traditional surveys to examine their abundance and composition in paddy fields involve human visual inspection, which requires considerable time and effort by expert entomologists. We previously developed a deep learning-based object detection system which can detect rice planthopper individuals from scanned images of survey boards, categorize, and count planthopper individuals by 18 categories, based on species, developmental stages, adult sexes, and adult wing-forms, with a mean average precision (mAP) of 79%. In this study, we modified the system by reconsidering the categories of planthopper individuals to be counted and by additional supervised training. The modified system can count rice planthopper individuals captured on survey boards across 17 categories with a mAP of 91%. We also showed that by using the system developed here, classification and counting of rice planthopper individuals can be completed in approximately six minutes per survey board, which can take more than an hour for human experts. Thus, this high-performance system can greatly save time and reduce labor costs for monitoring the occurrence, reproduction, and population growth of the rice planthoppers in paddy fields, which could increase the efficiency of their management.