Evaluation of support vector machine and artificial neural networks in weed detection using shape features

Weed detection is still a challenging problem for robotic weed removal. Small tolerance between the cutting tine and main crop position requires highly precise discrimination of the weed against the main crop. Close similarities between the shape features of sugar beet and common weeds make it impossible to define an exclusive feature to be able to efficiently detect all the weeds with acceptable accuracy. Therefore in this study, it was tried to integrate several shape features to establish a pattern for each variety of the plants. To enable the vision system in the detection of the weeds based on their pattern, support vector machine and artificial neural networks were employed. Four species of common weeds in sugar beet fields were studied. Shape feature sets included Fourier descriptors and moment invariant features. Results showed that the overall classification accuracy of ANN was 92.92%, where 92.50% of weeds were correctly classified. Higher accuracies were obtained when the SVM was used as the classifier with an overall accuracy of 95.00% whereas 93.33% of weeds were correctly classified. Also, 93.33% and 96.67% of sugar beet plants were correctly classified by ANN and SVM respectively.