FEATURE EXTRACTION BASED ON OBJECT ORIENTED ANALYSIS

Transcription

1 FEATURE EXTRACTION BASED ON OBJECT ORIENTED ANALYSIS Wong T. H. 1, S. B. Mansor 1, M. R. Mispan 2, N. Ahmad 1, W. N. A. Sulaiman 1 1 Spatial and Numerical Modeling Laboratory Institute of Advanced Technology, University Putra Malaysia, Serdang, Selangor, Malaysia Tel.: + (603) Fax: + (603) krwong@tm.net.my 2 Strategic, Environment and Natural Resources Research Centre (SENR) MARDI, G.P.O.BOX Kuala Lumpur, Malaysia Tel: Fax : radzali@mardi.my ABSTRACT: Most traditional pixel-based classification approaches are based exclusively on the digital number of the pixel itself. Thereby only the spectral information is used for the classification. The situation comes into worse when extracting certain interested features only. As a result, it produces the salt-and-pepper-effect pseudocolor layer. An object-oriented classification is suggested in order to overcome the limitations stated above. The existing software, ecognition allows the polygon based classification process. It is based on fuzzy logic, allows the integration of a broad spectrum of different object features, such as spectral values, shape and texture. This study demonstrated the extraction of Penang Bridge from the Landsat TM image even with its moderate resolution. The land use classification result as well as the spatial information can be exported to GIS environment for further analysis. 1. INTRODUCTION The traditional pixel-based classification normally can be performed by unsupervised and supervised classification. Unsupervised classification has few disadvantages where analyst has little control over image classes - making inter-comparison of data difficult. Besides, spectral properties keep on changing over time, therefore the relationship between spectral response and information class are not constant and detailed spectral knowledge of surfaces may be necessary (Campbell, 1996). An experiment has been carried out to compare the unsupervised and supervised classification. The results showed that the supervised classification achieve higher accuracy than unsupervised classification if the analysts have long known well about the study area. In addition also, supervised classification allows the users to pick out the different regions and different between features that look similar, like clouds and snow. In the unsupervised classification, similar regions were often lumped together (McCready and Hautaniemi, 2000). In a supervised classification, the identity and location of some of the land cover types, such as urban, agriculture or wetland are known a priori (before the fact) through a combination of fieldwork, analysis or aerial photography, maps and personal experience (Mausel et al., 1990). The analyst attempts to locate specific sites in the remotely sensed data that represent homogeneous examples of these known land cover types. Various supervised classification algorithms may be used to assign an unknown pixel to one of a number of classes. Among the most frequently used classification algorithms are the parallelepiped, minimum distance and maximum likelihood decision rules. Among these algorithms, the maximum likelihood method is generally preferred (Campbell, 1998; Avery and Berlin, 1992). It becomes most commonly used classifier due to its higher accuracy levels. Initially it was not used as much purely due to the large amount of time the images took to process. Now with faster computers, it is used much more frequently. It is generally accepted that this is the most accurate form of classification if compare to parallelepiped and minimum distance algorithms (Curran, 1985).

2 Anyway, traditional supervised classification has its disadvantages too. The most common error occurred when assigning the classes on training data, in the shape of polygon, pixels are assigned to a specific class if they fall within the polygon regions are allocated to the appropriate categories. Problems occur when pixels fall outside the specific regions or within overlapping regions. This will result in misclassification of data. In the other words, it is very reliant on the accuracy of the training data. Therefore, with box classification the possibilities of having misclassification or unclassified data were high (Kardono, 1992). In contrast to classic image processing methods, the basic processing units of object-oriented image analysis are image objects or segments and not single pixels, moreover, classification acts on image objects. One motivation for the object-oriented approach is the fact that, in many cases, the expected result of most image analysis tasks is the extraction of real world objects, proper in shape and proper in classification. This expectation cannot be fulfilled by traditional, pixel-based approaches (Baatz and Shape, 1999). Typically, they have considerable difficulties dealing with the rich information content of Very High Resolution (VHR) or moderate resolution such as Landsat TM or Spot data; they produce a characteristic, inconsistent salt-and-pepper classification, and they are far from being capable of extracting objects of interest. The concept is that important semantic information necessary to interpret an image is not represented in single pixels, but in meaningful image objects and their mutual relations. Image analysis is based on contiguous, homogeneous image regions that are generated by initial image segmentation. Connecting all the regions, the image content is represented as a network of image objects. These image objects act as the building blocks for the subsequent image analysis. In comparison to pixels, image objects carry much more useful information. Thus, they can be characterized by far more properties such as form, texture, neighborhood or context, than pure spectral or spectral derivative information (Baatz and Shape, 2000). These additional information must be describable in an appropriate technique to derive improved classification results. 2. OBJECT ORIENTED CLASSIFICATION Analysis of an image in the object-oriented approach involved classifying the image objects according to class descriptions organized in an appropriate knowledge base. This technique was created by means of inheritance mechanisms, concepts and methods of fuzzy logic and semantic modeling. The process of the object oriented classification mainly involved two sections, they are segmentation and classification. 2.1 Multiresolution Segmentation Segmentation is not an aim in itself. As regards the object-oriented approach to image analysis, the image objects resulting from a segmentation procedure are intended to be rather image object primitives, serving as information carriers and building blocks for further classification or other segmentation processes. In this sense, the best segmentation result is the one that provides optimal information for further processing (Hofmann, Puzicha and Buhmann, 1998). The ecognition software performs a first automatical processing - segmentation - of the imagery. This results to a condensing of information and a knowledge-free extraction of image objects. The formation of the objects is carried out in a way that an overall homogeneous resolution is kept. The segmentation algorithm does not only rely on the single pixel value, but also on pixel spatial continuity (texture, topology). The organized images objects carry not only the value and statistical information of the pixels of which they consists, but also information on texture and shape as well as their position within the hierarchical network (Humano, 2000; Manakos, 2001). Figure 1 below shows the concept of segmentation, in which where mainly three different levels of image objects have been created representing different scales. All of the image objects were automatically linked to a network after the segmentation process. Each image object knows its neighbors, thus affording important context information for later analysis. Subsequently, repetition of segmentation with different scale parameter creates a hierarchical network of image objects. Each image object knows its super-object and its sub-objects. The basic difference, especially when compared to pixel-based procedures, is that

3 object oriented analysis does not classify single pixels, but rather image objects which are extracted in a previous image segmentation step. Fig. 1: Hierarchical net of image objects derived from image segmentation level 1 (10 pixels scale parameter), level 2 (20 pixels scale parameter) and level 3 (30 pixels scale parameter) When dealing with certain features extraction, the researchers must have well known about the sensitivity of the bands to be set as priority input for segmentation. Wrong bands or unnecessary bands input will result deficient image objects output. Bear in mind that the classification will be based on the segmented image objects, if the interested features are not in the proper segmented objects, it won t be able to give satisfying result finally. In this study, the Penang Bridge is the interested feature that going to be extracted (fig.2). (a) (b) Fig. 2: Different bands input for segmentation produces different image objects: (a) Band 1, 2 and 3 as bands input; (b) Band 2, 3, 4, 5 and 7 as bands input

4 2.2 Classification ecognition supports different supervised classification techniques and different methods to train and build up a knowledge base for the classification of image objects. The frame of knowledge base for the analysis and classification of image objects is the so-called class hierarchy. It contains all classes of a classification scheme. The classes can be grouped in a hierarchical manner allowing the passing down of class descriptions to child classes on the one hand, and meaningful semantic grouping of classes on the other. This simple hierarchical grouping offers an astonishing range for the formulation of image semantics and for different analysis strategies. The user interacts with the procedure and based on statistics, texture, form and mutual relations among objects defines training areas. The classification of an object can then follow the "hard" nearest neighbourhood method or the "soft" method using fuzzy membership functions (Manakos, 2001). Under the soft method, each class of a classification scheme contains a class description. Each class description consists of a set of fuzzy expressions allowing the evaluation of specific features and their logical operation. A fuzzy rule can have one single condition or can consist of a combination of several conditions that have to be fulfilled for an object to be assigned to a class. The fuzzy sets were defined by membership functions that identify those values of a feature that are regarded as typical, less typical, or not typical of a class, i.e., they have a high, low, or zero membership respectively, of the fuzzy set (Mitri and Gitas, 2002). It can accept any external input as a parameter to classify the desire classes. In this paper, the vegetation and urbanization area is classified by the external input of Normalized Difference Vegetation Indexes (NDVI) Normalized Difference Vegetation Indexes (NDVI) The NDVI for Landsat TM multispectral data is generated as follows (Marsh et al., 1992; Larsson, 1993): TM 4 TM 3 NDVI TM = TM 4 + TM 3 Generally, NDVI shows the results where the brighter the pixel is, the greater the amount of photosynthesizing vegetation present. The NDVI provides a measure of the amount and vigour of vegetation at the land surface. It is a non-linear function that varies between 1 to +1 (undefined when VIR and VIS are zero). The magnitude of NDVI is related to the level of photosynthetic activity in the observed vegetation. Values of NDVI for vegetated land generally range from about 0.1 to 0.7 with values greater than 0.5 indicating dense vegetation. In general higher values of NDVI indicate greater vigour and amounts of vegetation. The reason NDVI is related to vegetation is that healthy vegetation reflects very well in the near infrared part of the spectrum (Khali et al, 2002). 3. STUDY AREA The Penang Island is selected for this study. It is located at the North-West of Peninsular Malaysia. The image area is situated within latitudes E to E and longitudes N to 5 29 N. Satellite data required for this study was Landsat TM 7 (Path 128, Row 56), 2000 which obtained from Malaysian Centre of Remote Sensing (MACRES). The satellite image has been performed geometric correction and registered to RSO projection before carried out the classification. The registered subset image is shown in Figure 3.

5 Fig. 3: Penang Island Landsat TM image year 2000 with composite color (Band TM 4, TM 5, TM 3) 4. METHODOLOGY Import Multiresolution Segmentation Create General Classes 1. Fuzzy Logic (Membership Function) 2. Standard Nearest Neighbor (Training samples) Classify Without Related Classes Generate Child Classes (create desire classes) Classify with related classes Segmentation Based Classification Export Fig. 4: Schematic diagram illustrates the classification flow in ecognition.

6 In this paper, the process can be divided into few simple steps. After bringing the image into ecognition, the image will be applied the multiresolution segmentation. The consideration of the bands should be based on the types of features that are going to extract. Figure 5 shows the grey scale of each band except thermal band (band 6) where we are not interested on the temperature in this case. Note that band 4 shows a clear coastal area boundary at the west coast of Penang Island; band 5 and band 7 shows the clear structure of Penang Bridge. (a) (b) (c) (d) (e) (f) Fig. 5: Grey scale display: (a) band 1; (b) band 2; (c) band 3; (d) band 4; (e) band 5; (f) band 7 After satisfied with the segmentation result, the water body and land was carried as first priority, by assigning ratio band 1 lower than 0.17 for land, and otherwise. Secondly, NDVI differentiation of vegetation area and urbanization area was carried by assigning the membership function above 0.05 for vegetation area, and otherwise for urbanization area. Further classification process can be carried out to generate the desire classes, for example, urban area and open or under construction area, from urbanization area subclass. By looking at the primary result shown in fig. 6, Penang Bridge and coastal area fall under water body main class. To extract the Penang Bridge, feature view function in ecognition gives absolutely control over the image. These child classes can be generated by its full range of fuzzy logic functions availability. For Penang Bridge, 3 combinations have been used for extraction. They are ratio band 7, mean band 7 and standard derivation band 7 (Fig. 7). Throughout the feature view as well, carry out the coastal extraction by using mean band 4 as a function parameter. The costal area falls within 42 and 52 in mean band 4. Mean, ratio and standard derivation features are features that concerning the pixel channel values of an image object (spectral features).

7 Fig. 6: Primary classification results of Penang Island (a) (b) (c) Fig. 7: Feature view on (a) mean band 7; (b) standard derivation band 7; (c) ratio band 7 The segmentation-based classification is then applied to the image to call out the desire classes to merge together to new level. The cleaning process can be carried out in this new level as well as performing the further details classification. The standard nearest neighbors classification can be carried out as well as the fuzzy logic function to assign those desire classes for exporting. The pond and ocean were differentiated by considering the area function as parameter. The complete class hierarchy is shown in fig. 8. The final step is exporting the classification results as well as its attributes table to GIS environment for further analysis.

8 Fig. 8: Complete Class Hierarchy of Penang Island 5. RESULTS AND DISCUSSION The pixel-based supervised classification is also carried out to make the comparison. Since the class for Penang Bridge involved small clusters, the line sampling at the training site will be considered. The results from both traditional pixel-based maximum likelihood classification and object-oriented classification as shown below: (a) (b) Legend Vegetation Area Urban Area Ocean Open or Under Construction Area Penang Bridge Coastal Area Pond Fig. 9: Comparison between (a) Pixel-based classification; (b) Object oriented classification.

9 The part (a) from Fig. 9 is the output of pixel-based maximum likelihood supervised classification. As a result, the classified image produced salt and pepper image or lot of small clumps (< 10 pixels) appeared in the classified image. Vice versa, the classified image derived from polygon-based classification is closer to human visual interpretation. The pixel-based classification results, which only based on the spectral mean of the digital number itself, is no way to differentiate Penang bridge with similar spectral values of Ocean. There are still existing Penang Bridge class clusters located at other location (Black dots surround the Island). But if look carefully on Penang Bridge at Fig. 9, part (b), it can be clearly seen that Penang Bridge is completely extracted. In this case, the Penang Bridge is falling under water body mother classes. Due to this, it is not difficult to extract certain class, which has been narrow down step by step from coarse classes into detail classes. It is totally depending on user needs. Obviously, only polylines able be extracted from raster image so far. In ecognition, the classified result can be exported in smooth perfect polygons vector format, and these polygons are attached 100% to each other. The spatial data that exported together with the classified polygons helps the polygons to recognize their own classes. This integration of remote sensing and GIS ability saved time to connect the polylines, convert to polygons and assigning the classes on these polygons. 7. CONCLUSION In this paper, the result of object oriented analysis is satisfied for land use classification as well as feature extraction and towards produce the land use map. The possibility of performing classification based segmentation obviously improved the classification result obtained at the second level. Furthermore, the use of membership functions resulted in a reduction in the number of misclassified pixels. Discriminating among the different confusion classes was possible using the contextual and spectral information supplied by the images. All steps involved in the image analysis could be recorded as a complete procedure. Thus, the whole strategy for solving a particular problem can be applied to other data of the same type especially applying onto time series data. Besides, it constitutes an important step towards the integration of remote sensing and GIS, by providing operational means of interpretation highresolution data. This technique is recommended to test on VHR data such as Ikonos image or Aerial photos especially in town area where more details classes can be generated. REFERENCES AVERY, T.E. and BERLIN, G.L. (1992). Fundamentals of Remote Sensing and Airphoto Interpretation.. Prentice Hall, Upper Saddle River, New Jersey. Chapter 6. BAATZ, M.A.S. (1999). Object-Oriented and Multi-Scale Image Analysis in Semantic Networks. In: Proc. Of the 2nd International Symposium on Operationalization of Remote Sensing(Enschede. ITC). BAATZ, M.A.S. (2000). Multiresolution segmentation - an optimization approach for high quality multiscale image segmentation. Angewandte Geographische Informationsverarbeitung XI. CAMPBELL, J. (1998). Map Use and Analysis. WCB/McGraw Hill. 3rd Ed., Chapter 18, pp CAMPBELL, J.B. (1996). Introduction to Remote Sensing. New York: Guilford Publications. 622 p. CURRAN, P.J. (1985). Principles of Remote Sensing. Longman Scientific and Technical Group, Essex, England. 282pp. HOFMANN, T. PUZICHA, J, and BUHMANN, J. (1998). Unsupervised texture segmentation in a deterministic annealing framework. In: IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 13: HUMANO, A. (2000). ecogntition and Change Detection-Integrating Aerial Photos and Satellite Images. ecognition Application Notes, Vol. 1, No. 2, September JENSEN, J. R. (1996). Introductory Digital Image Processing A Remote Sensing Perspective Prentice Hall, Upper Saddle River, New Jersey. 2 nd Ed.

10 KARDONO, P. (1992) The Use of Remote Sensing Techniques in the Detection of Upland Vegetation Communities in the North York Moors National Park. Mlandresreviewpdf/8.4.pdf. Unpublished PhD Thesis, University of St. Andrews. KHALI, A. H., YUSMAH, M. Y., SHAJOERIL, T. and ADB AZHAN, S. I. (2002). The Application of TiungSAT-1 Data for Forest Vegetation Mapping A Case Study. TiungSAT-1 First User Group Seminar, 24 th September (9). LARSSON, H. (1993). Regression for Canopy Cover Estimation in Acacia Woodlands Using Landsat TM, MSS and SPOT HRV XS Data. International Journal of Remote Sensing, 14 (11): MANAKOS, I. (2001). ecognition and Precision Farming. ecognition Application Notes, Vol. 2, No 2, April MARSH, S. E., WALSH, J.L., LEE, C. T., BECK, L. R. and HUTCHINSON, C. F. (1992), Comparison of Multi-Temporal NOAA-AVHRR and SPOT-XS Satellite Data for Mapping Land-cover Dynamics in the West African Sahel. International Journal of Remote Sensing, 13(16): MAUSEL, P. W., KAMBER, W. J. and LEE, J. K. (1990). Optimum Band Selection for Supervised Classification of Multispectral Data. Photogrammetric Engineering & Remote Sensing, 56(1): MCCREADY, M. and HAUTANIEMI, T. (2000), Exercise 3: Supervised Classification Department of Geoscience, University of Massachusetts, USA. MITRI, G.H. and GITAS, I.Z. (2002). The development of an object-oriented classification model for operational burned area mapping on the Mediterranean island of Thasos using LANDSAT TM images. Forest Fire Research & Wildland Fire Safety, Viegas (ed.) Millpress, Rotterdam, ISBN