The Application Of Genetic Programming Algorithms On Hyperspectral Data Of Uranium Deposits

Remote sensing is one of the basic techniques to obtain the earth’s surface information via satellite imagery. Hyperspectral remote sensing is a newer technique with each hyperspectral image containing more bands of informative numbers hosting clearer resolution and better color contrast than a multispectral image, causing the hyperspectral technique to play a more important role in the field of geological survey than their multispectral counterpart. Many researchers and scientists used the hyperspectral technique to detect and identify minerals, terrestrial vegetation as well as manmade and other kinds of objects. Evolving from a multispectral image, a hyperspectral image is about10nm of its spectral resolution and contains more than200bands of numbers. Although the hyperspectral technique is the forefront technology, hyperspectral data contain redundant information and it is important to choose appropriate approaches to process the hyperspectral data. There are two ways currently in use:dimension lowering and best bands selection. In this passage, the best bands selection method was used and the application of image classification using bands selected by experience and bands selected by algorithms were compared.Genetic programming, as one of classic machine learning algorithms, can effectively classify or predict according to an input data. Now this algorithm is widely used in data mining fields. Genetic programming mimics the evolutionary strategies of organisms in nature and searches for the best solution by selection, crossover and mutation operation. Genetic programming algorithms use hierarchical structure to restore their expression, fitness function to produce an expression and then classify data with the outcome. The strategies for genetic programming algorithms to classify or predict an input data are different. For those two-class classification and regression problems, one classifier is needed and is the only expression used. However, more classifiers are in need if it is a multiclass classification problem. By using the associate strength matrix, each genetic programming classifier expression needs to calculate a separated weight. Then heuristic rules are applied as a complement to improve the accuracy of the final result by effectively avoiding the disturbance of noise data. In this paper, genetic programming algorithms are first used to distinguish different minerals on earth surface. Then they are used to predict hematite content in uranium minerals. In the end, genetic programming algorithms are used to find similar places of a known mineral deposit in a hyperspectral image. Furthermore, the hematite ratio method is also used to find the potential location of the mineral deposit in the hyperspectral image. The results of these two methods are combined to pinpoint potential mineral deposits on the hyperspectral image.In first three chapters, the preprocessing of hyperspectral data and basic knowledge of genetic programming in classification are introduced. In chapter four, genetic programming algorithms are used to classify some minerals according to their spectral curves. A genetic programming classification algorithm and the hematite ratio method are combined together in chapter five, which is used to identify places of interest on a hyperspectral image. Since the location of a mineral deposit is known, data mining algorithms can also be used to find out similar places in a hyperspectral image. The result of the hematite ratio method is combined with data mining algorithms in order to find distribution of places of interest in a hyperspectral image.The innovations of this paper are as follows:(1) genetic programming method is proposed to classify different types of minerals according to their spectral curves;(2) a new way is found so as to identify mineral deposits in a hyperspectral image:that is combined a data mining algorithm with hematite ratio methods. The visibility of this new way is verified by comparing the search result with the location of another mineral deposit which was already explored.