A Deep Learning-Based Mobile Application for Classifying Rice Crop Diseases in Labo, Camarines Norte

Abstract

The primary concern of the rice farming community is the early detection of rice crop disease. Rice crop disease can be detected with high accuracy with the availability of advanced digital cameras and smartphones to improved image acquisition modes and deep learning methods such as convolutional neural networks (CNN). This study used a qualitative approach employing focus group discussions with selected farmers and an online meeting with the Department of Agriculture (DOA). Also compared and evaluated different optimizers using several optimization techniques namely Stochastic Gradient Descent with Momentum (SGDM), Root Mean Squared Propagation (RMSProp), Nesterov-accelerated Adaptive Moment Estimation (Nadam), and Adaptive Moment Estimation (Adam) in different dataset partitioned by 80/20%, 60/40%, 50/50%, 40/60%, and 20/80% using cv2 module from OpenCV library. Furthermore, presents the hardware and software to developed a free, easy-to-use and widely accessible mobile application that can efficiently and accurately diagnose 22 types of diseases and a healthy leaf sample. The experiment results show that Nadam optimizer achieve a maximum accuracy of 97.67-100.00% in the 80/20 partition, 88.17-100% in the 60/40 partition, 84.93-100% in the 50/50 partition, 64.67-100% in the 40/60 partition, and 37.03-99.90% in the 20/80 dataset partition. Therefore, the android application “Rice Crop Diseases Classification” can accurately classify rice diseases using Nadam optimizers including healthy rice. Addiditonally, despite employing various dataset partitioning methods, it achieves the highest accuracy from both low and high records using 80 by 20% dataset partitioned.