Techniques of Texture Analysis for Identification of Crocidolomia pavonana, Spodoptera exigua, Spodoptera litura

Abstract

Cabbage and green onion are two common vegetables cultivated by farmers in highland. Among major pests causing significant damage are Crocidolomia pavonana on cabbage and Spodoptera exigua on green onion. In addition, the generalist Spodoptera litura occasionally found attacking these vegetables. For unexperienced farmers as well as extension agents, the three caterpillars are not easily differentiated; although an accurate identification is required for a better control of the pests. Therefore, research was conducted with the objectives to develop identification system of the caterpillars based on computer vision technology. The three species of caterpillars have a unique texture and distinct feature from each other; and therefore, differentiating the species can be based on techniques of texture analysis. For this purpose, we used combination of several texture analysis: entropy feature, GLCM, Haralick features and Tamura feature. Each species consisted of 45 images, and consequently the total images used in this research were 135 images. Data work splitted using k-fold cross validation technique into training data (80%) and testing data (20%). Probabilistic Neural Network (PNN) was applied for classifying the three species of caterpillars. Research was conducted using GLCM with 1 pixel distance and 8 angles; Haralick features consisting of entropy, homogeneity, information of correlation 1, information of correlation 2; and Tamura feature consisting of coarseness; and entropy feature techniques. Our research revealed that Tamura features and information of correlation 1 were able to differentiate S. litura from other species. Entropy feature was able to differentiate C. pavonana with S. litura. Haralick entropy was able to differentiate C. pavonana with S. Exigua. Homogeneity was able to differentiate S. Exigua with S. litura whereas information of correlation 2 was not able to differentiate among the three species. However, combination of four techniques were able to improve the accuracy of caterpillar identification system. System performance obtained from the average accuracy of the entire fold reached 77.03%. The best accuracy (88.89%) was obtained when k = 5. Furthermore, we tested the system using classification model based on the fifth fold. Error occurred when C. pavonana incorrectly identified as S. exigua, and S. exigua incorrectly identified as S. litura. The error in identification were caused by inappropriate cropping of caterpillar images, the blur of image, and the caterpillar was not in advance stage. However, in general the pest identification system developed through this research can be used by farmers and extension agents who have difficulties in differentiating the three species of caterpillars.