The Detection of Vegetation Species in Remote Sensing Imaging

Remote sensing classification is a complicated process and requires the perception of many factors. The main image classification factors that can be taken into consideration to improve classification accuracy may include placement of a sufficient classification system, chosen of training samples, feature extraction, and selection of appropriate classification techniques. The focus of the research in this thesis is on the monitoring of cereal crops, in particular infestation of wheat crops by black grass, a pernicious weed. This is regarded as the major pest for cereal crops, at least in East Anglia, one that causes significant loss of income for farmers. Although annual rather than perennial, a single back grass plant is able to produce many thousands of seeds, which are small enough to be distributed by the wind, though it is believed that the principal means by which it spreads from field to field is by sticking inside farm machinery such as combine harvesters. Anecdotal evidence from farmers is that black grass plants tend to grow directly adjacent to wheat plants, with their roots enveloping those of the wheat and thus effectively strangling the wheat. Black grass is resistant to most weed-killers, and spraying with those weed-killers that are allowed requires a licence — this makes eradicating it difficult. When young black grass plants are potentially detectable because they do not grow in the regular pattern of the wheat plants; but after germination, they tend to be smaller than adjacent wheat plants until they reach maturity, at which time they are slightly taller and with their characteristic black seed-heads. Thus the purpose of this thesis is to explore the detection of black grass infestations of wheat crops. Imagery of fields infested with black grass plants is simply not available at the kinds of spatial resolution needed for detection, so an important part of the research was the collection of imagery from a drone. This was done over more than one year and at critical points in the growth of the crop and its ‘companion’ weed, so that detection at different stages of the growth cycle is potentially possible. However, it should be pointed out that black grass is difficult to detect when immature, except by walking amongst the plants and examining them individually. The thesis starts by reviewing remote sensing image principles in general terms. It goes on to discuss the use of Image classification in remote sensing and the major steps that may be involved in the image classification process. It sheds the light on the two main approaches that used for classification in remote scening; supervised and unsupervised learning with some detail of the particular techniques used in this thesis, ranging from conventional statistical ones to Genetic Programming and Convolutional Neural Networks, are presented along with a review of their application in remote sensing. The thesis presents two experimental work that used different conventional remote sensing classification techniques. The former sheds light on the regional agricultural land texture classification using grey level cooccurrence matrices (GLCMs) for features extraction and using SVM and Decision Tree Induction techniques for classification. The later focuses on outcomes gained from three conventional supervised techniques that used for remote sensing images classification; random forest, SVMs, and decision tree classifiers. The thesis focuses on analysis by Vegetation Indices (VIs) and sheds light on the fundamental aspects of the most common indices that analysts use in remote sensing; NDVI, CRI2, SG index, and RG Ratio index. The thesis presents two genetic programming toolkits developed at Essex University: the Jasmine vision system builder and ELVS (evolutionary learning vision system) to detect the tree crown regions and consequently measure its efficiency to solve the black grass detection problem. In addition, it presents the relevant fundamentals of convolutional neural networks (CNNs) that used for remote sensing images classification. Five deep convolutional neural networks (Google-NET, VGG16, VGG19, ResNET50, ResNET101) and shallow convolutional neural networks have been trained to five different classes of ROI images including unknown range: wheat, black grass, road, and bushes. Finally, the conclusions drawn from this research and makes suggestions for further work have been presented.