Development and Evaluation of a Noise Reduction Algorithm for Improvement of Hypserspectral Image Classification

Hyperspectral Imagery as a modern technology of remote sensing is a valuable source for various applications in geo-sciences such as land cover mapping, mine exploration and environmental monitoring. However, because of hardware and technological issues, these data have inherent problems. Problems such as various noises, data redundancy, absorption and damaged bands, etc., which are caused by different factors. Since the improvement of hardware system in hyperspectral sensors is very expensive, image processing methods like noise reduction and feature extraction are less expensive and more important. The most recent and competent method is multi-hypothesis prediction. The multi-hypothesis prediction method acts based on forming a linear combination from neighbors of a pixel, relying on the belief that the target pixel is most likely to its neighbors. As its drawback, this method does not use the convenient way to select the bands with maximum likelihood. The aim of this research is to apply the multi-hypothesis prediction method and to select the spectral bands, based on linear regression. The experiments revealed applying bands grouping using linear regression method improved greatly the signal to noise ratio which are noticeable in classification results. This approach was used because of its good flexibility in determining the coefficient similarity of spectral bands. The approach consists of several steps: first, the proposed method was applied on the data to reduce the noise. Second, a reasonable number of features were extracted based on existing classes and using Feature Space Discriminant Analysis method. And finally, the extracted features were classified. The data sets used in this research are the tip data available for hyperspectral image processing experiments which were collected by Bask University (Spain). The data sets include the images of training sites of Indian Pines and Pavia University farms which were acquired by AVIRIS and ROSIS sensors respectively. These data were selected to have variety option in bands and spatial resolutions, so that they can show the effectiveness of the proposed method more effectively. The results of implemented method using the Support Vector Machine and K Nearest Neighbor classification on two sites showed the overall accuracy of 95.82, 99.43, 92.89 and 98.88 respectively. This means an improvement of 0.3 and 0.4 by SVM classifier and 8.22 and 2 by KNN classifier for two sites respectively. The results revealed the good performance of applied method. The advantage of the proposed method appears when the accuracy of the KNN and SVM Classification approaches to each other. The results showed that the method was able to improve, significantly, the KNN classification more than SVM classification. From this we can conclude that the proposed method, regardless of the type of data and classification method, can greatly increases the classification accuracy. This is because of the proper manner of the method in data processing, regardless of the classification approach. Since each classification method responds to a particular type of data, each classification method has its own specific results for a given type of data. To overcome this problem, the data were matched with each specific classification method.