|
Comparison of Pixel and Object Oriented Based Classification of Fused Images
 Dr. M. Seetha Professor GNITS India smaddala2000@yahoo.com Dr.I.V.Muralikrishna Ex-Director IST JNTU India ivm@ieee.org Dr. B.L.Malleswari Professor GNITS India blmalleswari@gmail.com Nagaratna .P. Hegde Associate Professor VCE India nagaratnap@yahoo.com Dr. B. L. Deekshatulu Visiting Professor HCU India Abstract Various studies have been carried out to find an appropriate method to classify the remote sensing data. Most traditional pixel-based classification approaches are based exclusively on the digital number of the pixel itself. The object oriented techniques offer the suitable analyzed to classify the satellite data. In the object-oriented approach, images were segmented to homogenous area by suitable parameters in some level. Classification process is applied on the LISS-III and Fused images. The classification is predominantly based on shape and neighborhood related features which will be exemplified by the extraction of land cover classes with a region-growing rule base. Classification results of object oriented image analysis and pixel based are compared. Object oriented image analysis approach obtained higher overall accuracy and kappa statistic than those by pixel based image analysis approach. 1. Introduction Remotely sensed image analysis is a challenging task and is accomplished by digital image classification. The classification of the complex structures from high resolution remote sensing imagery causes obstacles due to their spectral and spatial heterogeneity. The classification methodology is predominantly based on shape and neighborhood related features, will be exemplified by the extraction of the spectral classes with a region-growing rule-base. Traditionally used pixel based classification methods are based on conventional statistical techniques. Though this approach performs well but the ability for resolving inter-class confusion is limited. Pixel-by-pixel methods are slower and less accurate while classifying images of urban environments, which consist of a mosaic of small-scale features made up of different materials. As a result, in recent years, alternative strategies have been proposed, particularly the use of artificial neural networks, decision trees, methods derived from fuzzy set theory, and incorporation of secondary information such as texture, context and terrain features. The fused images are frequently classified to obtain spectral classes. The classification is similar to the frequency of each spectrum but only the most dominant spectra are considered in classification. The classification results of fused images can be analyzed to understand the effect of spectral error(s). The low errors in classification correspondingly leads to the better quality of the fused images are summarized. A study in a small area using QuickBird data has been accomplished to compare the object-oriented with pixel-based classification approach. Object oriented classification do not classify single pixels but objects created in multi-resolution segmentation process, which allows use of, not only spectral responses but also texture, context and information from other object layers. The fuzzy logic rules are applied to the construction of class hierarchy. The two technically and theoretically different image processing techniques based on methods that derive spatially explicit multiscale contextual information from a single resolution of remote sensing imagery are compared. The classification process is proposed using object-oriented image analysis to extract information from satellite image to detect tropical deforestation . With fuzzy sets concept each pixel may have fuzzy membership with more than one class expressed as the degree of its membership to each class (values range between 0 and 1). The untrained classes will only display membership to trained classes, which can introduce a significant bias to classification accuracy. Unsupervised classification based on ISODATA algorithm to provide priori knowledge on the possible candidate spectral classes exists in the experimental area. The results show that the object-oriented approach gave more accurate results than those achieved by pixel-based classification algorithms. A predicate is provided for measuring the evidence for a boundary between two regions using a graph-based representation of the image. New modules were developed to support improved and semi-automated geocoding of vertical imagery. The high resolution image classification is presented based on fuzzy rules by the means of descriptors such as: form, texture and relations between objects and sub-objects. The flexibility of object-compressed BSQ (band sequential) format and the performance of the format are investigated using selected methods in ENVI. Image classification based on unsupervised image segmentation algorithm and its combination with MPEG-7 low-level descriptors and a Bayes classifier . Experimental results using different pairs of classes and corresponding data sets demonstrate the efficiency of the proposed approach. Wavelet based image fusion of SAR and Landsat data for tropical land cover mapping is demonstrated it is opined that a statistical correlation of the original data with the fused data would have supported the study. The study on classification reveals that considerable research has been done on the classification of multispectral images obtained from various sensors. Also, the accuracy assessment using the standard parameters like overall accuracy, producer’s accuracy and Kappa coefficient is performed. There is a dearth on the classification of fused images and accuracy assessment using IRS 1D LISS III image using all the classification techniques. Therefore, it is imperative for the analysis of classification techniques of fused images. The fused images obtained by different fusion techniques alter the spectral content of the original images. Therefore, the spectral separabiltiy of the classes was analyzed by the classification of fused images. In this paper, the traditional pixel based image classification and the newly developed object oriented image classification techniques were employed on LISS III image and fused images (data set1- part of Hyderabad image of size 256X256) obtained from Brovey transform, Multiplicative, PCA, Wavelet Transform and Lifting Wavelet transform techniques of image fusion. The classification accuracy of the original multispectral and fused images was assessed with parameters of overall accuracy, and kappa statistic. 2. Image Classification Techniques 2.1 Pixel Based Image Classification The pixel based image classification approach classifies remote sensing images according to the spectral information in the image and the classification manner is “pixel by pixel” and one pixel can only belong to one class. The unsupervised classification and supervised classification techniques of pixel-based classification are elucidated as given below 2.1.1 Unsupervised Classification using K-Means Unsupervised classifiers do not consider training data as the basis for classification. These classifiers examine the unknown pixels in an image and aggregate them into a number of classes based on the natural groupings or clusters present in the image values. The “K-means” approach accepts the number of clusters to be located in the data, from the analyst. The algorithm then arbitrarily locates number of cluster centers in the multidimensional measurement space. Each pixel in the image is next assigned to the cluster whose arbitrary mean vector is closest. After all pixels have been classified in this manner, revised mean vectors for each of the clusters are computed. The revised means are used as the basis to reclassify the image data. The procedure continues until there is no significant change in the location of class mean vectors between successive iterations of the algorithm. When this point is reached, the analyst determines spectral signatures identity of each spectral class. 2.1.2 Supervised Classification Using Maximum Likelihood Classifier Supervised classification is more closely controlled by the user than unsupervised classification. In this process, pixels which represent patterns are selected that user recognizes and patterns identified from other sources. Prior to selection of training samples, knowledge of the data, the classes desired and algorithm are assimilated. There are three basic stages involved in the supervised classification method: training stage, classification stage and accuracy assessment stage. The maximum likelihood decision rule is based on a normalized (Gaussian) estimate of the probability density function of each class. The maximum likelihood classifier quantitatively evaluates both variance and covariance of the category spectral response patterns while classifying an unknown pixel. 2.2. Object Oriented Image Classification The most evident difference between pixels based image classification and object oriented image classification is that, firstly the basic processing units are image objects or segments, not single pixels in object oriented image classification. Secondly, the classifiers in object oriented image classification are soft classifiers that are based on fuzzy logic. Soft classifier use membership to express an object’s assignment to a class. The membership value usually lies between 1.0 and 0.0, where 1.0 expresses a complete assignment to a class and 0.0 expresses absolutely improbability. The degree of membership depends on the degree to which the objects fulfill the class-describing conditions. The advantage of these soft classifiers lies in their possibility to express uncertainties about the classes’ descriptions. The basic processing units in object oriented image classification are objects or pixel clusters, with object oriented approach to analyze images; the initial step is always to form the processing units by image segmentation. The object oriented classification is performed by segmenting the image using region growing algorithm and applying fuzzy classification. 2.2.1 Fuzzy based Object Oriented Classification In object oriented image analysis the classifier is soft classifier (for example fuzzy system), which uses a degree of membership to express an object’s assignment to a class. The membership value usually lies between 1.0 and 0.0, where 1.0 expresses full membership (a complete assignment) to a class and 0.0 expresses absolutely non-membership. The degree of membership depends on the degree to which the objects fulfill the class-describing conditions. The main advantage of this soft classifier lies in their possibility to express uncertainties about the classes’ descriptions. It makes it also possible to express each object’s membership in more than just one class or the probability of belonging to other classes, but with different degrees of membership. With respect to image understanding these soft classification results are more capable of expressing uncertain human knowledge about the world and thus lead to classification results which are closer to human language, thinking and mind. 3. Accuracy Assessment The classification accuracy has been assessed using overall accuracy, and kappa statistic. In this context, the “accuracy” means the level of agreement between labels assigned by the classifier and class allocations on the ground collected by the user as test data. With error matrix, error of omission and commission can be shown clearly and also several accuracy indexes such as overall accuracy can be assessed. The following is the detailed description about the three accuracy indexes and their calculation methods. 3.1 Overall accuracy Overall accuracy is computed by dividing the total number of correctly classified pixels (the sum of the elements along the main diagonal) by the total number of reference pixels. From the error matrix, the overall accuracy can be calculated as the following:  3.2 Kappa coefficient Kappa coefficient provides a difference measurement between the observed agreement of two maps and agreement that is contributed by chance alone. A Kappa coefficient of 90% may be interpreted as 90% better classification than would be expected by random assignment of classes.  Where SUM = sum across all rows in matrix Xi+ = marginal row total (row i) X+i = marginal column total (column i) n = number of observations takes into account the off-diagonal elements of the contingency matrix (errors of omission and commission). 4. Results and Discussions The classification results of the pixel based unsupervised and supervised and object image classification approaches for data set 1 were shown below (Table-1). The K means clustering method uses the minimum spectral distance to form clusters. During this process, the maximum number of iterations is 24 and the convergence threshold is 0.950. The supervised classification process was performed using parametric decision rule with the maximum likelihood classifier. For classification accuracy assessment 50 random points are used.  Table: 1 Overall accuracy and Kappa statistic of unsupervised, supervised and object oriented classification of LISS III and fused images for data set 1  Figure 1: Overall accuracy of unsupervised, supervised and object oriented classification of LISS III and fused images for data set 1 of table 1  Figure 2: Overall accuracy of unsupervised, supervised and object oriented classification of LISS III and fused images for data set 1 of table 1 Hence it is depicted that the unsupervised classification of Lifting Wavelet fused image has high overall accuracy and kappa coefficient than the other unsupervised classified fused images.  Table 2: Spectral information of Unsupervised classified LISS-III and Fused images for data set 1  Table 3: Spectral information of supervised classified LISS-III and fused images for data set 1  Table 4: Spectral information of object oriented classified LISS-III and fused images for data set 1 The tables 2, 3, 4 shows the number of pixels assigned for each spectral class in LISS III and fused images with unsupervised, supervised and object oriented approaches respectively for data set 1.Figure 1shows overall accuracy of unsupervised, supervised and object oriented classification of LISS III and fused images for data set 1 oftable 1.Figure 2 shows overall accuracy of unsupervised, supervised and object oriented classification of LISS III and fused images for data set 1 of table 1 It is observed that Lifting Wavelet transform had large amount of pixels for all the spectral classes. The results of object-oriented image classification were compared with the results of pixel based image classification techniques. The results of the classified LISS III and fused images were analyzed with respect to Overall accuracy, Kappa statistic and spectral classes. It is observed that the object oriented image classification yielded more accurate results than pixel based image classification techniques. The Lifting Wavelet fused classified image produced more accurate results than other classified fused images. The comparative analysis indicates that the object oriented classification on the Lifting Wavelet fused image is superior to all other classification techniques on original and fused images. Amongst all the classified images, the Lifting Wavelet fused images based on object oriented classification achieved highest Overall accuracy and Kappa statistic. The numbers of pixels that are assigned to each class in unsupervised, supervised and object oriented classification of LISS III and fused images indicate that object-oriented image classification in Lifting Wavelet Transform furnished superior results than pixel based image classification. 5. Conclusions Image fusion techniques alter the spectral content of the original images. Hence, the spectral separabiltiy of the classes was analyzed by the classification of fused images. The pixel and object oriented based classification process has been conducted on LISS-III and five types (Brovey, PCA, Multiplicative, Wavelet and Lifting Wavelet) of fused images. Fuzzy theory used in object oriented image classification can classify an image more logically to express the real world situation than the classification result by the hard classifier using pixel based image classification. The classification accuracy of the original multispectral and fused images was assessed with parameters of overall accuracy, and kappa statistic. Amongst all the classified images, the Lifting Wavelet fused images based on object oriented classification achieved highest Overall accuracy and Kappa statistic. Also, the numbers of pixels that are assigned to each class in object oriented in Lifting Wavelet Transform are more than pixel based image classification. The comparative study concludes that object oriented image classification is superior to pixel based image classification. The study can be extended further by improving the classification method based on segmentation. References |