Insects Identification with Convolutional Neural Network Technique in the Sweet Corn Field

ALDEANSYAH PRIMA N, Dr. Rudiati Evi Masithoh, S.T.P., M.Dev.Tech; Asst. Prof. Choatpong Kanjanaphachoat, Ph.D

2021 | Skripsi | S1 TEKNIK PERTANIAN

Image recognition for automatic identification is one of the popular methods used these days especially with the advanced development of computers and artificial intelligence. Some insects are harmful to plant growth which can cause losses for farmers or plantations, but some others are beneficial for plants. Therefore, a method to identify the type of insects to obtain accurate and precise results is of importance. Nowadays, with increasing computer technology, an automatic object identification system with increased accuracy, improved speed, and less cost has been developed. Deep learning is a popular method recently in image recognition since it can increase the effectiveness of object identification. However, it requires large image data which can be solved by using a Convolutional Neural Network (CNN) technique. The implementation of CNN for image identification or classification can be done by collecting a large-scale dataset that contains hundreds to millions of images for network training because of the need for learning many parameters involved in the network. This research is conducted to develop and apply the CNN model to identify eight species of insects in the sweet corn field. Those insects were Calomycterus sp., Rhopalosiphum maidis, Frankliniella williamsi, Spodoptera frugiperda, Spodoptera litura, Ostrinia furnacalis, Mythimna separata, and Helicoverpa armigera. The CNN model in this research was built with four convolutional layers, which consist of Conv2D, batch normalization, max pooling, dropout sublayer, and a fully-connected layer. A total of 5568 images were trained with 5 trials and 100 epochs, then tested with 40 images. The result of this research shows that the CNN model has succeeded in identifying images of sweet corn insects with 19,00 +/- 5,18 percent of prediction accuracy for images with background, and 82,00 +/- 8,37 percent of prediction accuracy for images with no background, with training accuracy value 93,30 +/- 1,29 percent.

Image recognition for automatic identification is one of the popular methods used these days especially with the advanced development of computers and artificial intelligence. Some insects are harmful to plant growth which can cause losses for farmers or plantations, but some others are beneficial for plants. Therefore, a method to identify the type of insects to obtain accurate and precise results is of importance. Nowadays, with increasing computer technology, an automatic object identification system with increased accuracy, improved speed, and less cost has been developed. Deep learning is a popular method recently in image recognition since it can increase the effectiveness of object identification. However, it requires large image data which can be solved by using a Convolutional Neural Network (CNN) technique. The implementation of CNN for image identification or classification can be done by collecting a large-scale dataset that contains hundreds to millions of images for network training because of the need for learning many parameters involved in the network. This research is conducted to develop and apply the CNN model to identify eight species of insects in the sweet corn field. Those insects were Calomycterus sp., Rhopalosiphum maidis, Frankliniella williamsi, Spodoptera frugiperda, Spodoptera litura, Ostrinia furnacalis, Mythimna separata, and Helicoverpa armigera. The CNN model in this research was built with four convolutional layers, which consist of Conv2D, batch normalization, max pooling, dropout sublayer, and a fully-connected layer. A total of 5568 images were trained with 5 trials and 100 epochs, then tested with 40 images. The result of this research shows that the CNN model has succeeded in identifying images of sweet corn insects with 19,00 +/- 5,18 percent of prediction accuracy for images with background, and 82,00 +/- 8,37 percent of prediction accuracy for images with no background, with training accuracy value 93,30 +/- 1,29 percent.

Kata Kunci : deep learning, CNN, insects, image classification