Boll weevil (Anthonomus grandis), a major pest in cotton, has been eradicated from almost all states in the US. Volunteer cotton (Gossipium hirsutum), growing between seasons provides the perfect habitat for boll weevil to survive during winter allowing the spread of the pest. The boll weevil eradication program in South Texas works extensively trying to extirpate this pest, and the early detection of volunteer cotton in grain fields is detrimental for this process. In this study, we investigated images collected from a five-band multispectral camera mounted on an Unmanned Aerial Vehicle (UAV) to detect volunteer cotton in a cornfield. The core objective of this study was to compare accuracies of difference image classification techniques in detection of volunteer cotton in cornfield. In this study, we used second-order co-occurrence filter with 3x3 moving matrix in eight directions (0 degrees, 45 degrees, 90 degrees, 135 degrees, 180 degrees, 235 degrees, 270 degrees, 315 degrees) to extract eight textural features (mean, variance, homogeneity, contrast, dissimilarity, entropy, second moment and correlation) for each direction. Parallelepiped, Maximum Likelihood and Mahalanobis distance supervised classifications were used in the direction of maximum obtained accuracy (270 degrees) for the five-band stacked and NDVI images. Overall accuracy and Kappa coefficients were determined and compared for all three classified results. Five-band stacked image resulted in the highest overall accuracy of 91.75% and Kappa coefficient of 0.87 for Mahalanobis distance classification whereas NDVI resulted in the highest overall accuracy of 85.73% and Kappa coefficient of 0.76 for the same classification. Maximum likelihood was found to be the best for classifying cotton with a class accuracy of 70.07% using five-band stacked image while Mahalanobis distance was found to be the best for classifying cotton with a class accuracy of 54.93% for NDVI image.
