Paddy Leaf Diseases Image Classification using Convolution Neural Network (CNN) Technique

• Published in: 19th IEEE Student Conference on Research and Development: Sustainable Engineering and Technology towards Industry Revolution, SCOReD 2021
• Main Author: Zainorzuli S.M.; Afzal Che Abdullah S.; Abidin H.Z.; Ahmat Ruslan F.
• Format: Conference paper
• Language: English
• Published: Institute of Electrical and Electronics Engineers Inc. 2021
• Online Access: https://www.scopus.com/inward/record.uri?eid=2-s2.0-85124396924&doi=10.1109%2fSCOReD53546.2021.9652688&partnerID=40&md5=828aa9757a4bfeaaa9e9a1daa68fbb19

Rice is Malaysian daily food consumption thus it is very important to ensure its production. There is a large manufacturer of rice in Malaysia every year to contain the need of millions of Malaysians but still not sufficient. Pests and diseases play an important role that contributes to the reduction of rice production. The damage by the diseases will become severe as the rice grain grows. In the early days, a paddy disease expert is required to identify and diagnosis the paddy leaf disease. The paddy disease expert will obtain several samples of paddy leaf images from the farmer. Afterwards, the required sample was sent to the biotech laboratory so that the affected leaf can be analyzed. The process for this method was time-consuming, inconvenient for the farmer and on top of that it is very costly. However, with the application of Deep Learning method the diseases can be detected at the early stage. Thus, precaution measures can be taken to lessen the damage as soon as possible. The objective of this work is to identify the types of paddy leaf diseases by using three types of Convolution Neural Network (CNN)models which are AlexNet, GoogleNet and ResNet-50. A preliminary study was conducted to select the appropriate model and to obtain the optimum parameter by using four different types of paddy leaf diseases datasets obtained from the Kaggle database. Then, the accuracy of the three CNN models were compared to determine the best method. Hence, the result shows the highest accuracy at 89.82% by setting the optimal configuration namely learning rate at 0.001 and number of epochs at 30. © 2021 IEEE.