Paddy Pest Recognition Based on Deep Learning in Sri Lanka

Abstract

Regarding the majority of crops, one of the significant factors affecting crop yield is pests’ disasters. Since most pest species are highly related, pest detection on field crops, such as soybean, rice, and other crops, is likewise challenging than generic object detection. Presently, distinguishing pests in crop fields relies on manual classification, which is highly time-consuming and costly. This work proposes a convolutional neural network model to resolve the problem of the multi-classification of crop pests. The model can fully use the advantages of the neural network to extract multifaceted pest features comprehensively. This system identifies ten paddy pest’s species Common Thrips, wheat phloeothrips, wireworm, grain spreader thrips, rice leafhopper, rice water weevil, small brown planthopper, rice gall midge, yellow rice borer, and paddy stem maggot, which are multi-scale which improves the mean average precision (mAP) to 82.38%; the result increased by 1% with the original model. We embrace this model to a versatile stage to allow each rancher to utilize this program to analyze bothers progressively and give ideas on bug control. We planned a paddy bug imaging data set with ten ordinary paddy bothers and contrasted the highlights of many models with a pick as the ideal model. Besides, for the high mAP, we have utilized information expansion (DA) and added a dropout layer. The analyses are executed on the Android application we created. The outcome shows that our methodology outperforms the first model obviously and is helpful for a coordinated bug the executives. This application has made ecological flexibility, reaction speed, and precision by standing out from the past works. It has the assistance of minimal expense and primary activity, reasonable for the irritation checking mission.