APPLY K-MEANS IN OUTLIER REMOVAL TO IMPROVE RICE LEAF DISEASE CLASSIFICATION PERFORMANCE USING MOBILENETV3 MODEL

Abstract

The population is continuously increasing, urbanization is accelerating, and agricultural land is shrinking. Diseases on rice leaves cause significant yield losses, necessitating early diagnosis to mitigate their impact on productivity and ensure food security. The application of technology in detecting and diagnosing rice leaf diseases is essential. This study proposes a method to enhance the accuracy of the MobileNetV3 deep learning model. The technique involves removing duplicate and outlier images to improve the efficiency of the MobileNetV3 model using K-Means. The dataset used in the experiment is acquired from a secondary source. The data consists of 5932 images of four common diseases on rice leaves. Three sets of data are created (Set-1, Set-2 and Set-3), corresponding to the outlier thresholds set at 0.00, 0.05, and 0.06, respectively. The results show a considerable increase in accuracy when applying this method. The Top-1 accuracy rose from 80.11% to 86.85% and 87.69%, respectively.