Color image texture analysis and neural network classification of weed species

The environmental impact of herbicide utilization has stimulated research into new methods of weed control, such as selective herbicide application on infested crop areas. This research utilized color co-occurrence texture analysis techniques (CCM) to determine the effects of plant maturity on weed species classification accuracy, and evaluated neural network classifiers for potential use in real time control systems.; The weed species evaluated were (1) foxtail, (2) crabgrass, (3) common lambsquarter, (4) velvetleaf, and (5) morningglory, with soil image subsets added to each data set. A program termed GCVIS was developed to transform the red-green-blue (RGB) color format images into hue-saturation-intensity (HSI) format and generate CCM texture feature data from the digitized images. Statistical procedures were used to determine the discriminant powers of the texture features and then conduct discriminant analysis to determine the classification accuracy for each of three data sets. The between species discriminant analysis showed that the CCM texture statistics classification procedure was able to classify between five species and soil with an accuracy of 93% using hue and saturation statistics, only. The species/maturity discriminant analysis was able to classify between maturity and species with a level of accuracy above 97% for two maturity data sets (four species by two maturities, and two species by three maturities), when using all three HSI texture statistics, and above 88% while using hue and saturation statistics, only.; Finally, a comparison study of the classification capabilities of three neural network models (back propagation, counter propagation, and radial basis function) was conducted. It was found that the back propagation neural network classifier is capable of classification accuracies of 97% (data set using the second maturity level images with hue and saturation statistics), which exceeded traditional statistical classification procedure accuracy of 93%. In comparison with the other three neural network methodologies, the back propagation method achieved a higher classification accuracy with less computational requirements. A significant accomplishment has been the elimination of the intensity HSI feature requirement from the model. Elimination of intensity should promote color stability during ambient light variations due to cloud cover and solar position.