Evaluation of Neural-network Classifiers for Weed Species Discrimination
摘要
The potential environmental impact of herbicide utilisation has stimulated research into new methods of weed control. Selective herbicide application offers the possibility to reduce herbicide usage while maintaining weed control. The research reported utilised colour co-occurrence method (CCM) texture analysis techniques to evaluate three different neural-network classifiers for potential use in real-time weed control systems. An image data set consisting of 33 unique texture features for each image in a six class data set (40 images per class) was generated for the following classes; foxtail, crabgrass, common lambsquarter, velvetleaf, morning-glory, and clear soil surface. The data sets were evaluated using a stepwise variable selection procedure to provide six unique data models separated according to hue-saturation-intensity colour features. A comparison study of the classification capabilities of three neural-network models (backpropagation, counterpropagation, and radial basis function) was conducted. It was found that the backpropagation neural-network classifier provided the best classification performance and was capable of classification accuracies of 97%, which exceeded traditional statistical classification procedure accuracy of 93%. When comparing the three neural-network methodologies, the backpropagation method not only achieved a higher classification accuracy, but also had less computational requirements.