Study On Land Coverage Classification Of Hyperspectral Image Based On Improved DNA Coding

In recent years, land cover classification of remote sensing hyperspectral data has become an interesting area. We formed a series of hyperspectral image classification algorithms especially of its own characteristics, which based on the traditional classification algorithm. The spectral curves and the surface features statistical characteristics are two ways to identify. The former that based on the spectral reflectance or transmission curve to classify is the most characteristic. These methods combine the known spectral data and matching algorithm to identify the type of ground cover classes, in which coding matching classification is an important form. Classical encoding methods include spectral classification match binary coding, single threshold and multi-thresholds coding, etc. My research is based on the hyperspectral images by matching features to classify. The results are listed as following:Firstly, DNA computing is introduced into the hyperspectral classification. DNA computing is a new way of calculation, which is based on a large number of DNA molecules in parallel operation and intensify process technology to produce a combined result, and extract/detect the results, furthermore, DNA computing has a very huge storage capacity. That’s the reason why DNA computing thought is introduced into the spectral matching algotithm. It is a good way to reduce the time spent on training or classification and the complexity even to improve the classifacation accuracy that converting the known feature spectral data to the corresponding DNA chain parameters, which based on molecular level of genetic information coding model mechanism and the control mechanism by setting the optimal threshold.Secondly, based on the original DNA encoding matching algorithm(Jiao etc,2010), my research improve the way of DNA encoding. In this paper DNA enconding methods are based on the average reflectance values and the values of reflectance slope. The training samples are selected by Pixel Purity Index, meanwhile, reasonable coding deciation threshold and interative threshold are settled. We conducts experiments on the California Salinas Valley dataset, trying to classify the sixteen categories.Lastly, I compare the proposed method with different classifiers in the fields of overall precision and Kappa coefficient. This experiment shows the higher accuracy and lesser time of the DNA computing method than others in the large number of hyperspectral data.