STATISTICAL AND NEURAL NETWORK CLASSIFIERS FOR CITRUS DISEASE DETECTION USING MACHINE VISION

Abstract

Listen

Translate article

The citrus industry is an important constituent of Floridas overall agricultural economy. Proper disease controlmeasures must be undertaken in citrus groves to minimize losses. Technological strategies using machine vision and artificialintelligence are being investigated to achieve intelligent farming, including early detection of diseases in groves, selectivefungicide application, etc. This research used a texture analysis method termed the color co-occurrence method (CCM) todetermine whether classification algorithms could be used to identify diseased and normal citrus leaves. Normal and diseasedcitrus leaf samples with greasy spot, melanose, and scab were collected in the field and brought to the laboratory for thedevelopment of suitable segmentation and classification algorithms. Four feature models were created for classificationanalysis using varying subsets of a 39-variable texture feature set. The classification strategies used were based on aMahalanobis minimum distance classifier, using the nearest neighbor principle, as well as neural network classifiers basedon the back-propagation algorithm and radial basis functions. The leaf sample discriminant analysis using the Mahalanobisstatistical classifier and the CCM textural analysis achieved classification accuracies of over 95% for all classes (99% meanaccuracy) when using hue and saturation texture features. Likewise, a back-propagation neural network algorithm achievedaccuracies of over 90% for all classes (95% mean accuracy) when using hue and saturation features. It was concluded thatthe Mahalanobis statistical classifier and the back-propagation neural network classifier performed equally well when usingten hue and saturation texture features selected through a stepwise variable reduction method. Future studies will seek toapply the developed algorithms in a natural citrus grove environment.