Detection of Huanglongbing disease based on intensity-invariant texture analysis of images in the visible spectrum

• Published in: Computers and electronics in agriculture Vol. 162; pp. 825 - 835
• Main Authors: Gómez-Flores, Wilfrido, Garza-Saldaña, Juan José, Varela Fuentes, Sóstenes E
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier B.V 01.07.2019; Elsevier BV
• Subjects: Algorithms; Applications programs; Citrus; Citrus trees; color; Color imagery; crop production; Feature extraction; Gray-level co-occurrence matrix; greening disease; Huanglongbing; Illumination; Image analysis; Image classification; Image detection; Invariants; leaves; lighting; Medical imaging; Mobile computing; nutrient deficiencies; Random forest; Ranklet transform; Texture; Texture analysis; trees; Visible spectrum; Ranklet transform; Texture analysis; Random forest; Huanglongbing; Gray-level co-occurrence matrix
• ISSN: 0168-1699; 1872-7107
• DOI: 10.1016/j.compag.2019.05.032

Huanglongbing (HLB) is considered one of the most threatening diseases for citrus production and has caused economic damage in many countries around the world. Hence, citrus producers need in-field tools for the early detection of HLB to control its spread. Technological approaches are widely used for in-field HLB detection: algorithms of image analysis have demonstrated their potential to detect HLB in color images automatically. Specifically, texture features have been commonly used for characterizing HLB and other nutrient deficiencies. However, because the systems for HLB detection are intended to operate in-field, illumination of the environment is unlikely to be always the same; thus, texture analysis can be sensitive to the illumination. To overcome this limitation, a method for HLB detection based on intensity-invariant texture analysis is presented in this study. The ranklet transform is used to convert the input image to an intensity-invariant representation, from which common texture features are extracted. A random forest classifier is used to distinguish between distinct classes of citrus leaves, including healthy, nutritionally deficient, and HLB. The experimental results show the robustness of the proposed approach to different types of illumination: the classification performance remains stable independent of the brightness of the input image. An accuracy of about 95% in distinguishing between HLB-infected and HLB-negative classes was achieved, and an accuracy of about 81% in identifying between six classes of citrus leaves. These results reveal the potential of the proposed approach to be implemented within a mobile application that can be used in-field for HLB detection in symptomatic citrus trees.