In: Stilla U, Gamba P, Juergens C, Maktav D (Eds) JURSE 2011 - Joint Urban Remote Sensing Event --- Munich, Germany, April 11-13, 2011
Using Hyperspectral Remote Sensing Data in Urban Mapping Over Kuala Lumpur Ebrahim Taherzadeh and Helmi Z.M. Shafri* Institute of Advanced Technology (ITMA) and Faculty of Engineering Universiti Putra Malaysia (UPM) 43400 Serdang, Selangor Malaysia
[email protected] *Corresponding author:
[email protected] technology before the relevant authorities can be convinced to adopt hyperspectral remote sensing as a source of information in urban-related applications. For that purpose, this study will assess the capability of hyperspectral remote sensing to identify and detect surface materials in the urban area with specific focus on roof type materials detection. Roofs play a key role in protecting occupants of buildings and interiors from external weather conditions, primarily heat, rain and urban water quality.
Abstract— Hyperspectral remote sensing has great application potential for analyzing complex urban scenes. In this study, airborne hyperspectral data over part of Kuala Lumpur, Malaysia were used to classify 14 urban classes. In order to do the classification, Support Vector Machine (SVM) was used. Some filters (Lee and Enhanced Lee) were used before performing the classification. Consequently, the results showed that the overall accuracy is improved (3%-4%) when the filters were applied to the image. The overall accuracy for classification of the study area using SVM is 89% with Kappa coefficient 0.88 without filtering. The use of Lee and Enhanced Lee filters improved the accuracy to 92 and 93.6% respectively. This study serves as a pioneering effort in the application of hyperspectral sensing for urban area in Malaysia.
I.
Thus for classification, Support Vector Machines (SVM) methods was used. SVM classification has given good accuracies when applied to hyperspectral images [8, 9] based on the literature. In order to further improve the classification results, some filters such as Lee and Enhanced Lee, were applied which are available in ENVI 4.7 software.
INTRODUCTION
Urban areas are characterized by a large variety of artificial and natural surface materials influencing ecological [1], climatic and energetic [2] conditions. Due to rapid expansion and development of urban centers and cities we require and need new methods for frequent updating of existing databases instead of standard methods which are mostly based on field investigations and visual interpretation of aerial photographs.
II.
MATERIALS AND METHODS
A. Study area Test site was chosen in a part of Kuala Lumpur which contains a mixture of historical and modern buildings. The test site comprises a variety of typical urban structures, such as residential quarters, railway stations and recreational areas. AISA Eagle airborne hyperspectral sensor was deployed in 2009 and collected the data in 128 bands from 400 to 1000 nm with approximately 5 nm spectral resolution with 1 m spatial resolution (Figure 1).
Hyperspectral data sets are generally composed of about 100 to 200 or much more spectral bands with relatively narrow bandwidth (5-10 nm), whereas, multispectral data sets are usually composed of about 5 to 10 relative wider bands (70400 nm). With the high spectral resolution, hyperspectral remote sensing has great application potential for analyzing complex urban scenes. With the recent advent of hyperspectral sensors these data provide us high spectral and spatial resolution data which offer improved spectral and urban mapping capabilities [3, 4, and 5]. Still, there are some limitation for mapping urban area using hyperspectral data such as lack of suitable spectral library for urban material, shadow, non roof coverage elements and spectral similarity. Potential applications related to urban planning and management include mapping impervious surfaces for flood management and urban water quality [6], roof types for energy use and fire danger [2, 7]. In Malaysia, airborne hyperspectral remote sensing has just been recently introduced and is yet to be exploited for urban mapping in a routine basis. Thus, it is important to test the feasibility of the
,(((
405
In: Stilla U, Gamba P, Juergens C, Maktav D (Eds) JURSE 2011 - Joint Urban Remote Sensing Event --- Munich, Germany, April 11-13, 2011
Support vector machine: The SVM represents a group of theoretically superior machine learning algorithms. The SVM employs optimization algorithms to locate the optimal boundaries between classes. Statistically, the optimal boundaries should be generalized to unseen samples with least errors among all possible boundaries separating the classes, therefore minimizing the confusion between classes. In this paper, SVM classification method was used in terms of using 14 classes and the parameters of SVM set as below: i. Gamma in Kernel function set as 0.008, this field to set the gamma parameter used in the kernel function. This value is a floating point value greater than 0. ii. Penalty parameter was chosen 100, this parameter control the trade-off between allowing training errors and forcing rigid margins. Increasing the value of the penalty parameter increases the cost of misclassifying points and causes ENVI to create a more accurate model that may not generalize well.
Figure 1. RGB image of the original acquisition (False color)
B. Categorization of urban material Based on the hyperspectral image, the test site was categorized for classification of different materials in the study area. To do this, field survey has been conducted and finally 14 classes were defined in order to generate ground truth map. Training and testing pixels were determined based on random sampling and were separated to avoid bias the accuracy assessment process.
iii. Classification threshold was chosen 0, this field use to set the probability that is required for the SVM classifier to classify a pixel. Pixels where all rule probabilities are less than this threshold are unclassified. The range is from 0 to 1.
C. Preprocessing Before processing the hyperspectral image, some corrections should be applied which include atmospheric correction and geometric correction. In this study, due to the lack of some meteorological parameters during the flight, the Quick Atmospheric Correction extension (QUAC) was used which is available in ENVI 4.7 software.
Applying Lee and Enhanced Filters: The Lee filters were used to smooth noisy (speckled) data that have an intensity related to the image scene and that also have an additive and/or multiplicative component. Lee filtering is a standard deviation based (sigma) filter that filters data based on statistics calculated within individual filter windows. Unlike a typical low-pass smoothing filter, the Lee filter and other similar sigma filters preserve image sharpness and detail while suppressing noise. The pixel being filtered is replaced by a value calculated using the surrounding pixels.
D. Processing In this section, in order to improve classification accuracy some filters such as Lee and Enhanced Lee were applied and the results of filtering were used as input data for SVM. Based on the field surveying and endmember extraction, the training and testing data were provided as per table 1.
Accuracy assessment: Generally, classification accuracy refers to the extent of correspondence between the remotely sensed data and reference information [10]. In order to assess the accuracy of land cover maps extracted from the hyperspectral image, the testing data which is provided based on the field survey (table 1) were used and the results were recorded in a confusion matrix. A non-parametric Kappa test was also used to measure the classification accuracy.
Table 1: Training and testing pixels Classes Training Testing water 560 509 grass 290 267 Tree species 1 313 267 Tree species 2 305 312 Tree species 3 189 151 Polycarbonate roof 44 41 tarmac 394 392 Zink roof 577 563 Clay roof 405 310 pavement 72 53 Shadow 93 84 unclassified 180 95 Asbestos Roof 112 139 Cement 245 219
III.
RESULTS AND DISCCUSION
The result of SVM classification without applying filters is presented in fig 2. The overall accuracy is 89% with 0.88 Kappa coefficients (Table 2). But the problem is that it is so noisy and not clear for visual interpretation. In addition some classes such as some roof materials were mixed with tarmac and pavement.
406
In: Stilla U, Gamba P, Juergens C, Maktav D (Eds) JURSE 2011 - Joint Urban Remote Sensing Event --- Munich, Germany, April 11-13, 2011
Figure 2: Result of SVM classification without filtering Table 2: The accuracy of 14 classes in percent (without filtering) Class Water Grass Tree species 1 Tree species 2 Tree species 3 Polycarbonate roof Tarmac Zinc roof Clay roof Pavement Shadow Unclassified Asbestos Roof Cement Overall Accuracy
Figure 3. Classification result after applying Lee filtering
Prod. Accuracy (%) User Accuracy (%) 99.8 100 94.76 97.68 92.88 81.31 85.90 89.04 84.11 96.21 92.68 100 94.13 74.40 74.42 87.66 98.06 97.44 43.4 71.88 84.52 100 100 100 100 97.20 80.37 75.86 89
Table 3: Classification accuracy after applying Lee filter Class Water Grass Tree species 1 Tree species 2 Tree species 3 Polycarbonate roof Tarmac Zinc roof Clay roof Pavement Shadow Unclassified Asbestos Roof Cement Overall Accuracy
After applying Lee filtering (Figure 3), the overall accuracy is improved to 92% with 0.91 kappa coefficient. The result of each class is presented in Table 3.
Prod. Accuracy (%) User Accuracy (%) 100 100 97.38 97.01 94.76 85.47 89.42 92.38 86.09 99.24 92.68 100 95.92 81.92 81.88 91.29 99.35 97.78 60.38 72.73 89.29 100 100 100 100 98.79 80.82 79.37 92
Then the Enhance Lee filtering was applied to the data and then the result of this classification is presented as in Figure 4 and for each class, the accuracy is presented as in Table 4.
407
In: Stilla U, Gamba P, Juergens C, Maktav D (Eds) JURSE 2011 - Joint Urban Remote Sensing Event --- Munich, Germany, April 11-13, 2011
IV.
CONCLUSION
Classification of hyperspectral data from urban areas in Kuala Lumpur, Malaysia has been discussed and the use of hyperspectral data shows good potential for mapping the urban area. The results show that applying Lee and Enhanced Lee filters before classification with SVM provide us the suitable accuracy (overall accuracy is improved 3%-4%) to detect roof materials in urban area. This feasibility study is encouraging in terms of the use of hyperspectral data for Malaysian urban environment. However, lots of other investigations will need to be conducted to properly assess the practical and cost-effectiveness aspect of the technology before it can be adopted for routine implementation of urban monitoring for a developing country like Malaysia. Future work should be done to further improve the effectiveness of the technology such as combining spectral and spatial information to reduce misclassifications and better edge detection. ACKNOWLEDGMENT The authors would like to acknowledge the funding from the Ministry of Higher Education, Malaysia for this study and the computing facilities provided by UPM.
Figure 4. Classification result after applying Enhanced Lee filtering Table 4: Classification accuracy after applying Enhanced Lee filter Class Water Grass Tree species 1 Tree species 2 Tree species 3 Polycarbonate roof Tarmac Zinc roof Clay roof Pavement Shadow Unclassified Asbestos Roof Cement Overall Accuracy
Prod. Accuracy User Accuracy (%) (%) 100 99.80 98.13 96.32 95.88 86.78 90.38 93.69 85.43 100 92.68 100 98.98 87.78 87.39 95.16 99.35 97.78 66.04 68.03 89.29 100 100 100 100 97.89 80.37 80 93.59
REFERENCES [1]
C. L. J. Arnold, and C. J. Gibbons, “Impervious Surface Coverage: The emergence of a key environmental indicator”, Journal of the American Planning Association, Vol. 62,No. 2,1996, pp. 243í258. [2] T. Oke, Boundary layer climates. New York: Routledge,1987. [3] E. Ben-dor, N. Levin, and H. Saaroni, “A spectral Based Recognition of the urban environment Using The Visible and Near-Infrared specral Regin (0.4-1.1 um)”,International Journal Of Remote Sensing,Vol. 22,No. 11, 2001, pp. 2193-2218. [4] G. F. Hepner, B. Houshmand, I. Kulikov, and N. Bryant, “ Investigation of the integration of AVIRIS and IFSAR for urban analysis”, Photogrammetric Engineering and Remote Sensing, Vol. 64,No. 8,1998,pp. 813– 820. [5] S. Roessner, K. Segl, U. Heiden, and H. Kaufmann, “Automated differentiation of urban surfaces based on airborne hyperspectral imagery”, IEEE Transactions on Geoscience and Remote Sensing, Vol. 39,No. 7, 2001,pp. 1525í1532. [6] M. K. Ridd, “Exploring a V– I –S (vegetation – impervious surface–soil) model for urban ecosystem analysis through remote sensing: comparative anatomy for cities”, International Journal of Remote Sensing, 16 ,1995, pp. 2165– 2185. [7] J. P. Woycheese, P. J. Pagni, and D. Liepmann,”Brand lofting above large-scale fires”. Proc. 2nd international conference on fire research and engineering (ICFRE2), 1997, Gaithersburg, MD pp. 137–150. [8] G. Camps-Valls, and L. Bruzzone, “Kernel-based methods for hyperspectral image classification”, IEEE Trans. Geos. And Remote Sensing, vol. 43,No. 6,2005, pp. 1351–1362. [9] M. Fauvel, “Spectral and spatial methods for the classification of urban remote sensing data”. Ph.D. dissertation, Grenoble Institute of Technology. [10] R. G. Congalton, “A review of assessing the accuracy of classifications of remotely sensed data”, Remote Sensing of Environment,Vol. 37,No. 1,1991,pp. 35–46.
The results show that by applying Lee and Enhanced Lee filtering we can improve our classification accuracy. The maximum overall accuracy that can be achieved is 93.59% which is related to the use of Enhanced Lee filtering. In this paper as mentioned the special focus is to detect the roof materials. For the detection of the polycarbonate roof, there is not much of the difference in applying the filters. But for the discrimination between Zinc roof and other roof types, the results show improvement in detection accuracy after applying the filters. Still there are some limitations in this paper for example, in some parts of the image, road and some roof materials cannot be distinguished due to spectral similarly. Furthermore, due to overhead coverings by trees, some classes (such as roof and tarmac) cannot be detected accurately.
408
