STUDY OF SOME IMFORMATION EXTRACTION TECHNIQUES FOR HYPERSPECTRAL IMAGING

Abstract

Remote sensing is an art and science of acquiring information about earth surface without actual contact with the earth surface. It gives data in the form of images which comprises of finite number of elements known as pixels. Such images which are obtained in from remote sensing needs to be processed and the easy and effective way is to do it digitally. Hence the field of digital image processing refers to processing of digital images by means of a computer. Digital image processing plays a key role in the field of image processing, image analysis, image enhancement and coding with a wide range of applications such as, classification, target/anomaly detection, mineral identification, super resolution mapping, etc. In remote sensing, the digital image processing methods are used for deriving the information from airborne and satellite borne images. Remote sensing community has utilized the digital images obtained from airborne and satellite borne sensors and has made enormous progress in recent years. Based on the sensor specifications the digital image produced by each sensor varies in spatial, spectral and radiometric resolutions. Other components like dimension of the image, temporal resolution of the image, swath width of the image, etc., plays secondary role in some types of applications. Due to variation in spectral resolution we have panchromatic (single band), multi-spectral (few bands less than 10) and hyperspectral (100s of bands) data are available. In this research, the study has been done for hyperspectral data. The land use land cover classes on the earth's surface have different physical characteristics. Particularly, for a hyperspectral data, due to coarse spatial resolution pixels may contain two or more classes. Hence classification of hyperspectral data is a way to extract useful information from it. But due to the large spectral dimension of hyperspectral data, it suffers while doing processing. Redundant information from hyperspectral data are carried every time. Also one of the important properties of hyperspectral data is that the neighbouring bands of hyperspectral data are highly correlated. Hence due to the collective reasons the hyperspectral data needs to be reduced. To reduce the hyperspectral data, in this research, the feature extraction techniques have been employed. The two feature extraction techniques employed here are wavelet based feature extraction and PCA based feature extraction. In wavelet based feature extraction three wavelet transforms have been used, namely, Haar wavelets, Daubechies wavelets and Coiflets wavelets. These wavelets along with its sub classes have also been utilized. For each of the sub classes, decomposition [ii] upto 4 levels have been performed. In case of PCA based feature extraction two prominent techniques have been used, segmented PCA and spectrally segmented PCA. The evaluation of the feature extraction techniques has been done by calculating the classification accuracy. Also a study has been made to observe the duration of feature extraction, duration of classification, number of extracted features, relation between classification accuracy and number of reduced features, etc. Once the features have been extracted from hyperspectral data, now it is ready for performing classification. Even though the number of bands reduced to significant level the algorithms used for multi-spectral classification are incapable of producing useful information. Hence for classification of hyperspectral data few algorithms exist in literature. For hyperspectral data also, the classification may be per pixel and sub pixel. By incorporating training pixels if the classification is done then it is called as supervised classification while with no help from training data if classification is done then it is unsupervised classification. Also if the statistical parameter extracted from the data are used then it is parametric classification and if not it is called as non-parametric classification. In this research all types of these classifications are touched and they are support vector machines (SVM) which is supervised, per pixel and non-parametric classifier while linear mixture model (LMM) is also supervised and non-parametric but sub pixel classification. Finally, the independent component analysis (ICAMM) which is parametric classification unsupervised but produces sub pixel outputs. The information extractions by each of these classification techniques are useful in one or the other way. To have a study about the performance of these techniques, three hyperspectral datasets have been taken. Two AVIRIS datasets with different spatial resolutions 4m and 20 m while one HYPERION dataset with spatial resolution 30m. The 20m spatial resolution dataset is named as dataset II covers by most of the region by vegetation and the classes are almost crisp. The 4m resolution data from HYPERON is named as dataset III which covers Roorkee and its surroundings covers by urban area and vegetation. Also some small classes like barren land, sand, etc., are also present in the dataset. Moreover the classes are not crisp enough to retrieve them back easily because of the nature of the land cover and the coarse spatial resolution. The next dataset is again from AVIRIS sensor but its airborne and hence a fine spatial resolution of 4m data is available. This data is over San Diego Naval Station and has many classes whose spectral signatures are almost [iii] similar. The study of classification algorithms has been made on these three hyperspectral datasets. When per pixel classification produces thematic map in which each pixel has been allotted to one and only one class while in sub pixel classification the fraction abundance of each class present in each pixel has been estimated. But that is not suffice until the spatial location has been estimated at sub pixel level. A solution to this problem may be achieved by super resolution mapping. In super resolution mapping, every pixel is divided into a specified zoom factor and each divided portion (sub pixel location) is filled by unique class thus making the original data into much finer spatial resolution. In this research, a novel super resolution mapping algorithm has been proposed, named as Pixel Filling Algorithm, to split each of the pixels and by gathering fractional abundance information from neighbouring pixels each sub pixel location may be filled by class values. This algorithm has been compared with existing super resolution mapping algorithm named as Pixel Swapping Algorithm. Due to the novelty of the pixel filling algorithm, a synthetic dataset has been generated of dimension 45x60. This dataset has been reduced by two low-pass filters, 3x3 and 5x5, to convert the data to dimensions 15x20 and 9x12 respectively. Now if super resolution mapping algorithm has been applied on the reduced dataset by appropriate zoom factors, 3 and 5, respectively then it is expected to get back the original data of dimension 45x60. At every stage of information extraction, the accuracy assessment has been performed. The feature extracted hyperspectral data are classified by SVM and then the conventional error matrix based accuracy assessment has been used. For sub pixel classification techniques the fuzzy error matrix based accuracy assessment has been performed. Finally, for super resolution mapping algorithms, the conventional error matrix based accuracy assessment has been performed but by taking three types of testing samples. Among the feature extraction techniques studied here, Daubechies wavelets perform better in extracting useful information from hyperspectral data. The second level decomposition is better in both accuracy wise and feature reduction wise. Since for all the datasets only the first two decomposition levels give better overall classification accuracies the sub pixel classifications have been performed only to the first two decomposition levels [iv] and that too only for Daubechies and its sub classes. In LMM has two solutions, constrained and unconstrained and there is no significant classification accuracies between them. The accuracy comes to be around 67%, 62% and 61% respectively for dataset II, dataset III and dataset IV. The ICAMM too extracts sub pixel information but due to unsupervised nature it retrieves some classes which are small enough to collect pure pixels from the three hyperspectral datasets. From dataset II, ICAMM retrieves railway track and non-metallic road and from dataset III, it retrieves noise free water body while from dataset IV it retrieves cylindrical drum, aircrafts, etc. In this way, each algorithm proves that it is superior over the other. The results for super resolution mapping have been analysed for the two algorithms, the proposed pixel filling algorithm and pixel swapping algorithm. The pixel filling algorithm performs well in super resolving mixed pixels which are having complicated boundaries and simple boundaries while the pixel swapping works well for only classes having linear boundaries. The accuracy for pixel filling and pixel swapping algorithms has been given in ordered pairs (for easy comparison) (92%, 98%), (96%, 90%), (97%, 91%) for datasets II, III, IV respectively for zoom factor 3 while for zoom factor 5 it is (91%, 94%), (90%, 70%), (90%, 70%) respectively. For dataset I, the overall accuracy of super resolved image by pixel filling algorithm gives 90.7% and 72.7% for zoom factors 3 and 5 respectively. Also the time required to perform super resolution mapping via pixel resolution mapping takes less than 10 seconds while by pixel swapping algorithm it takes more than a minute. Both in accuracy wise and duration for super resolution wise pixel filling algorithm is better than pixel swapping algorithm.