Pixel quantification and color feature extraction on leaf images for oil palm disease identification
Anindita Septiarini, Hamdani Hamdani, Tiya Hardianti, Edy Winarno, Suyanto Suyanto, Edy Irwansyah
Abstract
Oil palm disease can be identified by the appearance of yellowish spots on the leaf area. It causes a decrease in the quality and quantity of palm oil. Therefore, this work has developed a method for leaf disease identification using image processing to quantify the infected area’s pixels and extract the color features. The region of interest (ROI) detection was initially performed by applying Otsu thresholding based on $\mathrm{L}^{\star}\mathrm{a}^{\star}\mathrm{b}$ color space to obtain the sub-image called ROI image. Afterward, pre-processing was performed by converting RGB to several color spaces and using contrast stretching. The features extracted the mean value intensity of each channel on five color spaces and counting the pixel number of the infected area. Those features were reduced using correlation-based feature selection (CFS) followed by k-nearest neighbors (KNN) classification. The dataset used consists of 100 leaf images (50 healthy and 50 unhealthy). The method performance achieved accuracy, precision, and recall of 99%, 98%, and 100%.