Image Analysis based on Spectral and Spatial Grouping B. Naga Jyothi 1, K.S.R. Radhika 2 and Dr. I. V.Murali Krishna 3 1 Assoc. Prof., Dept. of ECE, DMS SVHCE, Machilipatnam, A.P., India 2 Assoc. Prof., Dept. of CSE, DMS SVHCE, Machilipatnam, A.P., India 3 Director (Retd.,), IST/R&D Center, JNTU, Hyderabad, India. ABSTRACT The paper aims to present image classification using region wise information. Effort is made to use spectral labelled & classified image as well as spatial information as the basis for classification.using a selected sample set, the image is first divided into several groups in the spectral space using mahalanobis distance as a measure of similarity.the output is then segmented by growing pixels with equal IDs basing on some specific connectivity. The entire image information is then available as region information. Each region is defined by a no. of attributes such as unique ID, size,the list of all its member pixels,the mean intensity and covariance matrix of the spectral values in that region. A structure is used to hold the segmentation information and the entire data is saved as a file. Secondly, a region labelled image consisting of region IDs is created. Finally the regions are classified using Euclidean distance measure. Keywords: Spectral Classification, Region Labelling, Euclidean Distance, Training Samples 1. INTRODUCTION Depending on the image primitive used viz, pixel or object based, image classification methods are of two main categories. Pixel based classification methods use only the spectral patterns to classify the individual pixels. Object based methods try to group pixels into objects using an image segmentation process based on some similarity chosen and then use both the spectral, spatial and contextual information of these objects to classify the whole image. Object based methods eliminates mixed pixel problem suffered by most pixel based methods and hence are superior ways of image classification. 1.1 SUPERVISED CLASSIFICATION In supervised classification, the identity of land cover types are known prior through some means such as aerial photography, map analysis and personal experience etc. The analyst locates specific sites that represent homogeneous examples in the remotely sensed data. These areas are referred as training samples and the spectral characteristics of these areas are used to train the classification algorithm of the remainder of the image. Multivariate statistical parameters such as mean, standard deviation, covariance matrices, correlation matrices are calculated for each training site. Every pixel both within and outside the training sites is then evaluated and assigned to the class of which it has the highest likelihood of being a member. Thematic classification of an image involves the following steps: [Ref: 3] Feature extraction: Transformation of the multispectral image by a spatial or spectral transform to a feature image. Examples are selection of subset of bands, a PCT to reduce the data dimensionality, or a spatial smoothing filter. This step is optional i.e., the multispectral image can be used directly, if desired Training: Selection of pixels to train the classifier to recognize the different themes, or classes, and determination of decision boundaries which partition the feature space according to the training pixel properties. This step is either supervised by the analyst or unsupervised with the aid of a computer algorithm. For supervised training, the analyst must select representative pixels for each of the categories. It is important that the training area be a homogenous sample of the respective class, but at the same time include the range of variability for the class. Labelling: Application of the feature space decision boundaries to the entire image to label all the pixels. If the training was supervised, the labels are already associated with the feature space regions; if it was unsupervised, the analyst must now assign labels to the regions. The output map consists of one label for each pixel. 1.2 INPUT IMAGE DETAILS The input image [Ref.1: www.nasa.gov] is Landsat 7 pan image of Pyongyang, North Korea acquired in September 2002. [Fig:1] This is a natural-color image using ETM+ bands 3, 2, 1. In this image, Pyongyang appears grey in color Volume 2, Issue 3, March 2013 Page 486
and is surrounded by vegetation (in green). The Taedong River, shown in dark blue, almost black travels through the city. 2. METHODOLOGY Fig: 1 Input Image 2.1 SPECTRAL CLASSIFICATION [ Ref.2 Lonesome M. Malambo 2009]. Spectral grouping is done by determining the closeness of each image pixel to each of the samples selected from the input image. Samples are selected based on the expected number of classes in the image and are assigned different IDs. Each pixel is assigned to its closest sample ID based on Mahalanobis distance measure of closeness. The Mahalanobis distance D as defined below, is used a measure of closeness or similarity.. (1) In (1), x is the pixel spectral vector, μ is the mean spectral vector of a sample in a multi band image, Σ is the covariance matrix of the sample, T denotes the transpose of the matrix. The output is an image composed of sample IDs. 2.2 IMAGE SEGMENTATION Spatial grouping uses the image created from the spectral grouping as the input. Equal IDs are grown in the region growing process with specific neighbor connectivity. A structure is used to hold the segmented image information. Each region is defined by an unique ID, list of its member pixels, mean and covariance of intensities in the region. Finally a region label image comprising of region IDs is obtained. The region wise information can be saved to a file for further processing. 2.3 IMAGE CLASSIFICATION Specific regions are selected to serve as training samples for region classification. Regions are classified using Euclidean distance. Each region is compared to the training samples and is assigned to the closest class. 3. IMPLEMENTATION The programming uses MATLAB 7.0.1 [Ref.2 Lonesome M. Malambo 2009]. The sequence of steps involved in the developed code are as follows: The loaded input image is smoothed using a Gaussian filter to reduce noise. Samples are selected from different classes Viz., vegetation, water, settlement and barren regions [Fig:2] and their statistics are computed [ Table:1 ] The samples are Labelled and Mahalanobis distance is calculated between each pixel and each instance of sample set. Pixels close to a sample are assigned the same sample ID. The result is an image composed of sample IDs [Fig:3 spectral classified image] Region growing is done with only equal ID pixels using 8-neighbor connectivity. Segmented information regarding individual regions is loaded into a structure. Each region has its own ID, member pixels, mean and covariance of intensities in it. [ Table: 2 ] A region label image with region IDs is formed. Then segmented image is observed [Fig:4]. For region classification,training Samples are selected from different Regions Viz., vegetation, water, settlement and barren regions using MATLAB s getpts function An image region is classified based on its Euclidean distance to a training sample. Regions close to a sample are assigned the same sample ID. Volume 2, Issue 3, March 2013 Page 487
Finally the result is a region (object) classified image. [Fig:5]. 4. RESULTS Fig: 2 Selected samples from different classes in the input image Sample 1: Vegetation Sample 2: water Sample 3: Settlement Sample 4: Barren Fig: 3 Output of Spectral Classification Vegetation: dark blue Water: light blue Settlement: yellow Barren: red Table: 1 Sample statistics for spectral Classification Volume 2, Issue 3, March 2013 Page 488
Fig: 4 Output of region growing process & segmented image Vegetation: yellow Water: light blue Settlement: dark blue Barren: orange Table: 2 Sample set of Region wise information Fig: 5 Output of region classification process & Final region (object) classified image Vegetation: light blue Water: dark blue Settlement: yellow Barren: red Volume 2, Issue 3, March 2013 Page 489
5. CONCLUSION Good results have been obtained as can be seen by comparing the input and output classified images. This is because the image is classified on the region (object) level and usually more information is used for classification.the final classified result is influenced by the samples considered in the spectral grouping, final classification process and also the number of samples taken for each class. REFERENCES [1] www.nasa.gov + NASA Home Multimedia. http://landsat.gsfc.nasa.gov/images/archive/c0018.html [2] Lonesome M.Malambo A region based Approach to Image Classification, Applied Geoinformatics for Society and Environment 2009-Stuttgart University of Applied Sciences. [3] Robert A. Schowengerdt. Remote Sensing: Models and Methods for Image Processing (3rd Ed.). Academic Press USA, pg: 388,396. [4] John R. Jensen Introductory Digital Image Processing, A Remote Sensing Perspective, Third Edition, PPH [5] B. Naga Jyothi, Dr. G. R. Babu, Dr. I. V. Murali Krishna, Thematic classification of multispectral imagery, International Journal of Electronics and computer Science Engineering, V1N2-181-190, 2012 [6] KSR Radhika, B.Naga Jyothi, et.al. Accuracy Assessment of per pixel Based Classification Conference Proceedings, NC-Velasiem-2k12: Pg.73-77 [7] MathWorks.Image Processing Toolbox 7.0.1. Mathworks.com [8] D. LU and Q. WENG A survey of image classification methods and techniques for improving classification performance International Journal of Remote Sensing Vol. 28, No. 5, 823 870, 10 March 2007 ACKNOWLEDGEMENTS The authors acknowledge "Landsat imagery courtesy of NASA Goddard Space Flight Center and U.S. Geological Survey" or "USGS/NASA Landsat" for the images. Volume 2, Issue 3, March 2013 Page 490