Full Text - Research Publishing Group
Transcription
Full Text - Research Publishing Group
Research Article ISSN 2277–9051 International Journal of Remote Sensing and GIS, Volume 2, Issue 1, 2013, 52-60 © Copyright 2013, All rights reserved Research Publishing Group www.rpublishing.org Assessment of Various Block Truncation Coding Based Remote Sensing Image Classification Techniques Pravada S. Bharatkar1 and Rahila Patel2 1 Department of Computer Science, RCERT (Nagpur University), Chandrapur, India Email : [email protected] 2 Department of Computer Science, RCERT (Nagpur University), Chandrapur, India Email : [email protected] Received February 10, 2013; received in revised form March 22, 2013; accepted March 23, 2013 Abstract This work presents a new interactive approach based on block truncation coding (BTC) to classify various regions in remote sensing (RS) imagery. In the classification, classified maps are the main product. As such, there exist a several techniques in literature for RS image classification, among which, very few have proved good precision in classifying RS image. Hence, the thirst of better and faster image classification is increasing and earned enormous attention. Developing computationally efficient algorithm for image classification without compromising the classification accuracy is of primary importance. Therefore, we are experimenting RS image classification by incorporating BTC approach in existing supervised and unsupervised classification technique. Five themes of land use/land cover classes (LULC) namely, Agricultural land, barren land, shrubs, waste land with scrub, and water body are considered and assess the accuracy of classification in terms of overall accuracy and kappa statistics. The results depicted that the supervised classification is superior to unsupervised classification technique. It is revealed from the present work, the reconstructed RS image by amalgamating BTC approach in the existing classification techniques can be a novel classification approach to map LULC in an efficient and accurate way. Key words: Block Truncation Coding (BTC) approach, remote sensing (RS), classification techniques, land use/land cover (LULC), supervised classification technique, unsupervised classification technique 1 Introduction The science and technology of Remote Sensing (RS) has emerged as one of the most fascinating and challenging subject over the past three decades. RS offers unique capabilities for understanding, monitoring, forecasting, managing and decision making about our planet’s resources via classification. The goal of classification is to group image cells into various classes/clusters to prepare classified maps. These maps are basic source of information for many applications in the field of agriculture, water resources, and environment. However, classifying RS image into thematic maps remains a challenge due to complexity of the landscape, the availability of reference data, the selected remotely sensed data, image-processing and image classification techniques, and the analyst’s experiences. Generally, the content of image such as color, texture and shape and size plays an important role in semantic image classification. The proper selection of image features and classification techniques are challenging task of classification, so many researchers delivered different classification techniques for image classification; such as supervised classification techniques i.e. box or parallelepiped, minimum distance to mean, mahalanobis and maximum likelihood, etc. and unsupervised clustering. Each of these techniques has some limitation necessitating for development of computationally efficient algorithms for image classification without compromising the classification accuracy. In past, numerous scientists had made efforts in developing various classification approaches for improving the classification accuracy (Gong and Howarth, 1992; Kontoes et al., 1993; Foody, 1996; Atkinson et al., 1999; Pal and Mather, 2003; Lu and Weng, 2007; Silakari et al., 2009; Blaschke, 2010). Recently, Kekre et al. (2012) have made an effort on Block Truncation Coding (BTC) approach for classification of photographic image of human being, animals, and natural scenery and found a better classification performance. Several studies also experimented on different aspects with BTC approach by various researchers (Maheswary and Srivastava, 2009; Silva et al., 2011; Rawat and Patil, 2012 and Samathal and Mohanraj, 2012). Based on the critical reviews on RS image classification technique by author (Bharatkar and Patel, 2012), the possibility of 52 Bharatkar and Patel/International Journal of Remote Sensing and GIS, Volume 2, Issue 1, 2013, 52-60 BTC approach in RS image classification is pointed out and they recommended that instead of using all pixel data of image as feature vector for extracting the information from image, the threshold values of pixels based on BTC, can be used for better classification of RS image. Therefore, the present work has undertaken to amalgamate the BTC approach in the existing RS image classification technique to develop a novel classification approach for mapping LULC in an efficient and accurate way. 2 Proposed Methodology The present study proposes an evaluation of various RSI classification methods to extract the features of the BTC based reconstructed RS image for effective land use mapping. The RS image and scanned survey of India (SOI) toposheet were first rectified using the GIS software namely, Integrated Land and Water Information System (ILWIS) to prepare the base map of the watershed boundary, drainage and road maps of study area for ground verification. The procured RS image was reconstructed by using BTC approach described as follows. 2.1 BTC Approach In BTC approach, first divided the procured RS image into n*n non overlapping blocks of R, G, B components and computed the interband average image (IBAI); which is the average of all components (R, G, and B). After this, mean of interband average is taken as threshold (MR, MG, and MB). By using three independent R, G, and B components of image to calculate three different thresholds and then apply BTC to each individual R, G and B components. The thresholds be MR, MG, MB can be computed as given below; MR 1 m n R(i, j ) m n i1 j 1 MG 1 m n G (i, j ) m n i 1 j 1 MB 1 m n B(i , j ) m n i 1 j 1 Now three binary bitmaps are compute as BMr, BMg, BMb. If a pixel in each component (R, G and B) is greater than or equal to the respective threshold the corresponding pixel position of the bitmap have a value of 1 otherwise it is having the value of 0. 1, if……..R(i, j)>MR BMr(i, j)={ 0,…….if…. R(i, j)<=MR 1, if……..G(i, j)>MG BMg(i, j)={ 0,…….if…. G(i, j)<=MG 1, if……..B(i, j)>MB BMb(i, j)={ 0,…….if…. B(i, j)<=MB Two mean colors one for the pixel greater than or equal to the threshold and other for the pixels smaller than the threshold are also calculated .The UM (Upper mean) and LM (Lower mean) can be calculated as; UR m n 1 * BMr (i, j ) * R (i, j ) BMr (i, j ) i 1 j 1 UG m n 1 * BMg (i, j ) * G (i, j ) BMg (i, j ) i 1 j 1 UB m n 1 * BMb(i, j ) * B (i, j ) BMb(i, j ) i 1 j 1 53 Bharatkar and Patel/International Journal of Remote Sensing and GIS, Volume 2, Issue 1, 2013, 52-60 LR LG LB m 1 m * n BMr (i, j ) n * {1 BMr (i, j )} * R(i, j ) i 1 j 1 m n m *n 1 * {1 BMg (i, j )} * G (i, j ) m * n BMg (i, j ) i 1 j 1 m 1 m * n BMb(i, j ) n * {1 BMb(i, j )} * B (i, j ) i 1 j 1 These upper mean and lower mean together form a feature vector or the signature of the newly reconstructed BTC based RS image as shown in Figure 1. Figure 1: Spectral representation of original and reconstructed (BTC) RS image in ILWIS The procured and reconstructed RS images along with scanned SOI toposheet no. 47 J/5 were first rectified by projecting them into a plane by using map projection. For this purpose, a co-ordinate system was created with Universal Traverse Mercator (UTM) projection for zone 43 (in which the study area lies), using Everest (India, 1956) Ellipsoid and Everest (India, Nepal) datum under the GIS image processing utilities of ILWIS (Integrated Land and Water Information System), GIS (Geographical Information System) software. The procured and rectified RS images were further interpreted for LULC classification using the various classification techniques described in the subsequent section. The nationwide classification system prepared by NRSC, Hyderabad was adopted for LULC mapping (NRSC, 2011). The flow chart shows the complete procedure after incorporating the BTC approach in the existing algorithm is shown in Figure 2. The field verification for accuracy assessment, the rectified toposheet was used to delineate the permanent features in the watershed like roads, rural settlement area, and drainage networks, etc. The base map of the watershed boundary at 1:50,000 scales were prepared using the location of various contour and drainage lines. For accuracy assessment, reference data or ground truth data was taken using the same schemes used in the classification efforts. Since, ground based data is assumed to be 100% correct in accuracy assessments, due care was taken during the data collection. The classified LULC maps of original and reconstructed images using various classification methods have been evaluated in terms of some statistical accuracy measures. One basic accuracy measure is the overall accuracy (OA), which is calculated by dividing the correctly classified pixels (sum of the values in the main diagonal) by the total number of pixels checked. Besides OA, classification accuracy of individual classes can be calculated in a similar manner. Since, OA does not indicate how the accuracy is distributed across the individual categories, more specific accuracies are needed. One may be the user’ s accuracy (UA) which is the ratio between the number of correctly classified pixels and the classified totals pixels of particular LULC class, while another may be the producer's accuracy (PA). PA is the ratio between the number of correctly classified pixels and the reference total pixels. A more appropriate way of presenting these individual classification accuracies may be in terms of commission error (CE) and omission 54 Bharatkar and Patel/International Journal of Remote Sensing and GIS, Volume 2, Issue 1, 2013, 52-60 error (OE). Kappa coefficient provides a difference measurement between the observed agreement of two maps and agreement that is contributed by chance alone. A kappa coefficient above 75% may be interpreted as better classification than would be expected by random assignment of classes (Susana et al., 2010). Figure 2: Flow chart of BTC based classification algorithm 2.2 Image Classification Techniques Image classification technique, which is computer based interpretation of remotely sensed images, based solely on the detection of the spectral signatures (i.e., spectral response patterns) of various LULC classes. The computer classification technique of RS images are divided into unsupervised classification and supervised classification (Pooja et al., 2011; Zhang and Wang, 2012). Unsupervised classification is the process that for remote sensing images without prior knowledge, only depends to the statistical difference of combination of different spectroscopic data, and then validates ground objects according to properties of various classified objects. In unsupervised image classification, no training stage is required, but different algorithms are used for clustering. Numerous factors affect the classification results, among which important ones being the objective of classification, the spectral and spatial characteristics of the data, the natural variability of terrain conditions in geographic region, and the digital classification technique employed (Mishra et al., 2011). In supervised image classification training stage is required, which means first we need to select some pixels form each class called training pixels. Find the characteristics of training pixels and also find other pixels which have same characteristics. The software system is then used to develop a statistical characterization of the reflectance for each information class (Congalton and Green,1999). This stage is often called signature analysis and may involve developing a characterization as simple as the mean or the range of reflectance on each band, or as complex as detailed analyses of the mean, variances and covariances over all bands. Once a statistical characterization has been achieved for each information class, the image is then classified by examining the reflectance for each pixel and making a decision about which of the signatures it resembles most. There are several techniques for making these decisions, called classification techniques as described below. 2.2.1 Clustering It is an unsupervised classification technique, in which no training stage is required and can used to determine the natural spectral groupings present in the image. It is a rather quick process in which image data is grouped into spectral clusters based on the statistical properties of all pixel values. It is an automated classification approach. 55 Bharatkar and Patel/International Journal of Remote Sensing and GIS, Volume 2, Issue 1, 2013, 52-60 2.2.2 Box or parallelepiped classifier It is a very simple supervised classification algorithm. In this, two image bands are used to determine the training area of the pixels in each band based on maximum and minimum pixel values. Although parallelepiped is the most accurate of the classification techniques, it is not most widely used. It leaves many unclassified pixels and also can have overlap between training pixels (Perumal and Bhaskaran, 2010; Xu and Wei, 2012). This classification is based on simple Boolean “and/or” logic that decides the particular brightness value in the image belongs to a specified class. 2.2.3 Minimum distance to mean classifier This is simple in principle based on Euclidean distances towards class means only. For the spectral values of a pixel to be classified, the distances towards the class means are calculated. If the shortest (Euclidean) distance to a class mean is smaller than the user-defined threshold, then this class name is assigned to the output pixel, else the undefined value is assigned. 2.2.4 Mahalanobis distance classifier This classifier is based on the distances towards class means and the variance-covariance matrix of each class. For the spectral values of a pixel to be classified, the distances towards the class means are calculated as Mahalanobis distance (Perumal and Bhaskaran, 2010). The Mahalanobis distance depends on the distances towards class means and the variance-covariance matrix of each class. The class name with the shortest Mahalanobis distance is assigned, if this distance is smaller than the user-defined threshold value else, the undefined value is assigned. 2.2.5 Maximum likelihood classifier (MLC) It is perhaps the most widely used classification method of classification in remote sensing in which a pixel with the maximum likelihood is classified into the corresponding class (Xu and Wei, 2012). MLC algorithm uses Bayes’ rule and a classification method that minimum incorrect probability in terms of statistical rules. This Classification uses the training data by means of estimating means and variances of the classes, which are used to estimate probabilities and also consider the variability of brightness values in each class (Perumal and Bhaskaran, 2010). 3 Study Area and Data Source The Indian Remote Sensing (IRS) 1D LISS (Linear Imaging Self Scanning) III satellite image of Ralegaon Siddhi watershed having values in bands G, R, NIR and SWIR with a swath of 141kms in the format of LGSOWG ((Landsat Ground Station Operators Working Group) of the electromagnetic spectrum with a spatial resolution have been procured from NRSC (National Remote Sensing Centre), Hyderabad. The dataset consists of 2282 x 2507 pixels. The advantage of using this dataset is the availability of the referenced date set (Table 1) produced from field survey, which is used for the classification accuracy purpose. Table 1: Reference data set Sr. No. 1 LULC classes Total nos. of ground truth pixels 214 2 shrubs 82 3 Waste land with scrub 111 4 Barren land 59 5 Water body 74 Agriculture Land Total 540 The study area is a part of Dudh nadi sub-basin in the drought prone region of Parner taluka of Ahmednagar district of Maharashtra state (India). The watershed is lying between 18054' N to 18057' N and longitudes of 74023' E to 74027' E as shown in Figure 3. 56 Bharatkar and Patel/International Journal of Remote Sensing and GIS, Volume 2, Issue 1, 2013, 52-60 Figure 3: Location of the study area 4 Result Analysis In the present work, RS (original and reconstructed) images were processes through image processing capability of ILWIS GIS to derive an effective LULC map using the various classification techniques. The five major LULC classes are considered namely, agricultural land, barren land, shrubs, wasteland with scrub and water body. The confusion matrices for all the supervised and unsupervised classification techniques have been computed. As the results reflected that supervised maximum likelihood classification (MLC) technique performed better than other classification methods; the confusion matrix for this classification method is only as given in Table 2 and classified map is shown in Figure 4. Table 2: Confusion matrix for supervised maximum likelihood classification technique LULC class Agriculture land Barren land Shrubs Waste land with scrub water body Column Total Agriculture land Barren land Shrubs Waste land with scrub water body Column Total Agriculture land Barren Shrubs Waste land land with scrub Original RS image 214 2 0 47 0 55 0 6 0 0 82 0 0 2 0 58 0 0 0 0 214 59 82 111 Reconstructed RS image (BTC) 214 0 0 50 0 56 0 0 0 0 82 0 0 3 0 61 0 0 0 0 214 59 82 111 57 Water body Row Total 5 0 0 0 69 74 268 61 82 60 69 540 2 0 0 0 72 74 266 56 82 64 72 540 Bharatkar and Patel/International Journal of Remote Sensing and GIS, Volume 2, Issue 1, 2013, 52-60 74024’39’’ 74025’48’’ 74026’56’’ 18054’43’’ 18055’14’’ 18055’50’’ 18056’20’’ 74023’31’’ Figure 4: Classified LULC map derived by maximum likelihood classification techniques using reconstructed (BTC) RS image Based on the confusion matrix, various statistics such as classification errors (OE and CE) and accuracies (PA and UA) have been computed for all classification methods. The graphical representation of errors and accuracies using supervised MLC technique is shown in Figure 5. Figure 5: Graphical representation of classification errors (E) and accuracies (A) for supervised maximum likelihood classification technique by using original (O) and reconstructed (R) RS images The commission error (CE) is also called a misclassification error reflecting overestimation, while the offdiagonal elements omitted during classification is called omission error (OE measure of under estimation). As stated earlier, the user’s accuracy is ratio between the number of correctly classified and the row total. Since, the users are concerned about what percentage of the classes has been correctly classified; user’s accuracy reflects the accurate classification of individual LULC class. The producer's accuracy, which reflects the exact classification of particular land use/cover class, is the ratio between the number of correctly classified and the column total. As can be seen from the Figure 5, there is an inverse relationship existed between accuracy and 58 Bharatkar and Patel/International Journal of Remote Sensing and GIS, Volume 2, Issue 1, 2013, 52-60 error. The minimum errors (OE and CE) and maximum accuracies (PA and UA) are found for classification of reconstructed RS image using supervised classification. The accuracies are maximizes with the minimization of errors from Box to maximum likelihood classifier in proceeding with Mahalanobis and minimum distance to mean classifiers for mapping most of the LULC classes. The CE for agriculture land and OE for waste land are comparatively more for all classification methods due the reasons that the misclassification of some of training pixels of waste land as agricultural land. The shrub is better classified in all classification techniques due to its distinct features observed in the image. The overall accuracy (OA) representing the sum of all correctly classified pixels divided by the total number of test pixels and kappa coefficient (difference measurement between the observed agreement of two maps and agreement that is contributed by chance alone) were also computed as presented in Table 3 for both original and reconstructed RS image using all classifiers considered for this study. Table 3: Overall accuracy and Kappa coefficient for various classification techniques Name of the Classification Technique Clustering Box Minimum distance Mahalanobis Maximum likelihood Original RS Image Reconstructed RS Image (BTC) Overall Kappa Overall Kappa Accuracy (%) Coefficient Accuracy (%) Coefficient Unsupervised classification technique 60.78 0.477 61.48 0.496 Supervised clustering technique 58.15 0.443 58.70 0.444 81.85 0.762 82.04 0.764 84.26 0.781 85.19 0.794 88.52 0.842 89.81 0.860 It is seen from the Table that the maximum likelihood classifier using reconstructed (BTC) RS image gives higher overall accuracy of 89.81% with excellent kappa value of 0.860 as compared to other classifiers. 5 Conclusions The present study focuses on the classification accuracy of the various RS classification techniques. The procured IRS LISS III image covering Ralegaon Siddhi watershed is reconstructed with novel approach of BTC and tested for its performance based on overall accuracy and kappa coefficient. The supervised and unsupervised classification techniques were employed on original and BTC based reconstructed RS image for their classification into various LULC classes. It is found that, the classification of reconstructed (BTC) RS image gives minimum errors with maximum accuracies for maximum likelihood supervised classification method. The supervised classification method is more reliable as compare to unsupervised clustering technique. The value of overall accuracy of 89.81% and kappa value of 0.860 for classification of reconstructed (BTC) RS image shows a better performance of maximum likelihood classification technique. Hence, it can be concluded from the study that the incorporation of BTC approach in the existing algorithm improves the overall classification accuracy by minimizing the errors. References Aplin, P., Atkinson, P. M. and Curran, P. J.(1999). Per-field classification of land use using the forthcoming very fine spatial resolution satellite sensors: problems and potential solutions. In P.M. Atkinson and N.J. Tate (Eds.), Advan. in Remote Sensing and GIS Analysis, New York: John Wiley and Sons., 219– 239. Bharatkar, P. S., Patel, R. (2012). A survey on RSI classification techniques. International Journal of Advanced Research in Computer Science (ISSN No. 0976-5697), 3(7), 218-223. Blaschke, T. (2010). Object based image analysis for remote sensing. Photogrammetry and Remote Sensing, 65(1), 2-16. ISPRS International Journal of Congalton, R. G. and Green, K.(1999). Assessing the Accuracy of Remotely Sensed Data, Principles and practices (Boca Raton, London, New York: Lewis Publishers). 59 Bharatkar and Patel/International Journal of Remote Sensing and GIS, Volume 2, Issue 1, 2013, 52-60 Foody, G. M. (1996). Approaches for the production and evaluation of fuzzy land cover classification from remotely-sensed data. International Journal of Remote Sensing, 17, 1317–1340. Gong, P., Howarth, P. J. (1992). Frequency-based contextual classification and gray-level vector reduction for land-use identification. Photogrammetric Engineering and Remote Sensing, 58, 423–437. Kekre, H. B., Thepade, Sudeep , Das, R. K. and Ghosh, S. (2012). Image classification using block truncation coding with assorted colour spaces. International Journal of Computer Application, 44(6), 9-14. Kontoes, C., Wilkinson, G. G., Burrill, A. and Goffredo, S., and Megier, J. (1993). An experimental system for the integration of GIS data in knowledge-based image analysis for remote sensing of agriculture. International Journal of Geographical Information System, 7, 247–262. Lu, D. and Weng, Q.(2007). A survey of image classification methods and techniques for improving classification performance. International Journal of Remote Sensing, 28(5), 823-870. Maheswary, Priti, Srivastava, Namita (2009). Retrieval of remote sensing images using colour & texture attribute. International Journal Computer Science and Information Security, 4(1 & 2). Mishra, Jayant, Sharma, Anubhav and Chaturvedi, Kapil (2011). An unsupervised cluster-based image retrieval algorithm using relevance feedback. International Journal of Management and Information Technology, 3(2),9-16. NRSC (2011). Natural Resource Census - Land Use Land Cover Database. Technical Report – Ver.1, National Remote Sensing Centre, NRSC-RS&GISAA-LRUMG-LUCMD-N0V.,2011-TR-316 Pal, M. and Mather, P. M. (2003). An assessment of the effectiveness of decision tree methods for land cover classification. Remote Sensing Environment, 86, 554–565. Perumal, K. and Bhaskaran, R. (2010). Supervised classification performance of Multispectral images. Journal of Computing, 2 (2), 124-128. Pooja, A. P., Jayanth, J. and Koliwad, Shivaprakash (2011). Classification of RS data using decision tree approach.International Journal of Computer Applications (0975 – 8887), 23(3),7-11. Rawat, Sunita and Patil, Dharmaraj (2012). Content based image retrieval using block truncation coding. World Journal of Science and Technology, 2(3), 34-37. Samathal, S., Mohanraj, N. (2012). BTC with K means classifier using color image clustering. Journal of Computer Application, 5(EICA). Silakari, Sanjay, Motwani, Mahesh and Maheshwari, Manish (2009). Color image clustering using block truncation algorithm, International Journal of Computer Science, 4(2), 31-35. Silvia, K. Sneha, Vamsidhar, Y. and Sudhakar, G. (2011). Colour image clustering using K-Means. International Journal of Computer Science and Technology, 2(1), 11-13. Susana, M. Vieira, Uzay, Kaymak Joao, and Sousa M. C. (2010). Cohen Kappa Coefficient as a performance Measure for Feature Selection”, Proceedings IEEE International Conference, 2010, doi: 978-1-42448126-2/10. Xu, Mengxi and Wei, Chenglin. (2012). Remotely sensed image classification by complex network eigenvalue and connected degree. Hindawi Publishing Corporation Computational and Mathematical Methods in Medicine. doi:10.1155/2012/632703. Zhang, Hebing, and Wang, Shidong (2012). Research progress of computer automatic classification technology and methods based on remote sensing images, The 2nd Intern. Conf. Comp. Appl. Sys. Model. Published by Atlantis Press,Paris, France. 60