Selective application of herbicide in agricultural cropping systems provides both economic and environmental benefits. Implementation of this technology requires knowledge of the location and density of weed species within a crop. In this study, two image classification techniques (Artificial Neural Networks (ANNs) and Maximum Likelihood Classification (MLC)) are compared for accuracy in weed/crop species discrimination. In the summer of 2005, high spatial resolution (1.25mm) ground-based hyperspectral image data were acquired over field plots of three crop species seeded with two weed species. Image data were segmented using a threshold technique to identify vegetation for classification. The ANNs consistently outperformed MLC in single-date and multitemporal classification accuracy. With advancements in imaging technology and computer processing speed, these network models would constitute an option for real-time detection and mapping of weeds for the implementation of site-specific herbicide management. / xii, 106 leaves : ill. (col. ill.) ; 29 cm

An understanding of the distribution of the Bracken fern (Pteridium aquilinum (L.) Kuhn) is critical for providing an appropriate management strategy. In this regard, remote sensing can play a critical role in mapping and modelling such distribution. In this study, an integrated approach using the random forest, maximum likelihood and vegetation indices was developed and tested to determine the capability of WorldView-2 multispectral eight band image in characterising the Bracken fern. Results based on the WorldView-2 were further compared to SPOT-5 multispectral (MS) image findings. The WorldView-2 (WV-2) image was spectrally resized to four traditional bands (blue, 450-510nm; green, 510-580 nm; red, 630-690 nm and NIR1, 770-895 nm) and four additional bands (coastal blue, 400-450 nm; yellow, 585-625 nm; red-edge, 705-745 nm and NIR2, 860-1040 nm) to evaluate the practicality of the spectral resolution in mapping the Bracken fern. The results from this analysis showed that the spectrally resized additional bands were more successful in general land cover mapping and characterising the Bracken fern. The result’s overall accuracy was 79.14% while the user’s and producer’s accuracies were 97.62% and 91.11% respectively. The second part of the study sought to improve the classification accuracy by applying a robust machine learning algorithm, the random forest. Since the random forest does not automatically choose the optimal bands, the backward variable elimination technique was employed to identify the optimum wavelengths in WV-2 for the identification of the Bracken fern. Respective out-of-bag (OOB) errors of 13.1% and 9.17% were achieved when the WV-2’s eight bands and optimally selected bands (n= 5) were used. These bands lie in the green (510-580nm), near-infrared1 (770-895nm), red-edge (705-745nm), near-infrared2 (860-1040nm) and the coastal blue (400-450nm) regions of the electromagnetic spectrum. These findings confirm the importance of the additional bands in vegetation analyses. The vegetation indices computed from these regions of the spectrum were superior to those in the visible region. The classification accuracy using WV-2 bands was superior to that from the commonly used SPOT 5 image. / Thesis (M.Sc.)-University of KwaZulu-Natal, Pietermaritzburg, 2014.

Thesis (M.S.)--Mississippi State University. Department of Plant and Soil Sciences. / Title from title screen. Includes bibliographical references.