Genetic Algorithm Tuned SVM Classifier for Weed Species Recognition

Weed detection using image processing is a progressive research area which can revolutionize crop production and increase efficiency in herbicide usage. A SVM classifier was applied in our research to classify and to detect the weed type for weed scouting and spot weeding purposes. However, parameter fine tuning and feature selection for SVM is a complex procedure that is still an active area of research. A wrapper type feature selection and parameter fine tuning were proposed for the SVM training using Genetic algorithm to reduce the features and consequently over fitting of classification. The resulting classifier requires 2 features out of the 68 features. The performance of the classifier were optimal with additional 200 test samples with 96% correctly classified for positive label and 100% for negative label.