Using a Two-Stage HOG-SVM / CNN Model to Identify and Classify Forms of Brown Planthoppers

Abstract

Approximately ten percent of rice crop yields throughout the Asia-Pacific region are reduced due to pests called brown planthoppers (BPH). We use a two-stage model to identify BPH from rice crop images and use these to determine the form of each BPH in the image, which has implications for predicting potential BPH outbreaks. Using a unique form of concentric Histograms of Oriented Gradient (HOG) descriptors and SVM classifiers, we can obtain to identify BPH with a recall of 96.56% and an FDR (false detection rate) of 2.91%, surpassing other efforts on similar datasets. Applying a VGG-19 CNN architecture, we achieved a classification accuracy of 92.76%for the three BPH forms. These outcomes provide a foundation for other efforts in pest identification and insect lifecycle detection.