Road Extraction from Satellite Images

Transcription

1 Chapter 5 Road Extraction from Satellite Images 5.1Introduction With the emergence of google map and GPS navigation systems, road network extraction from satellite images has become an important research area. Especially, road network is one of the most important features in GIS databases. Techniques to extract road feature from an urban scene have numerous applications in urban mapping, urban planning and Geo-Information Engineering. The recent availability of commercial high-resolution satellite imaging sensors provides a new data source for road extraction. The focus of our research work is to develop a segmentation algorithm which accurately segments road features from complex urban scenes irrespective of its high curvature. Also we have done an attempt for the automatic extraction of roads even from low resolution, noisy and blurred satellite images. 5.2 Existing Techniques The classification and extraction of road features are greatly facilitated by the very fine details in urban areas and its linear feature. The methods presented by Haverkamp, D. (2002) and Mohammadzadeh, A. et al. (2004) are using these qualities of road networks. Using linearity as one of the key feature for road extraction, later a number of methods have been developed for road extraction considering other features also from high resolution satellite imagery but still the 73

2 Study & Analysis of Satellite Images for the Extraction of Structural Features percentage accuracy is not so good in case of dense urban areas. The quality percentage of the methods developed by Movaghati, S. et al. (2010), Xu, L. et al. (2012) and Chaudhuri, D. et al. (2012) is superior but we can t say that the methods are fully automatic. In order to cope with the high complexity of real scenes and ensuring fully automatic algorithms, soft computing methods can be used. The use of fuzzy based systems (Amini, Jalal., 2006), training using neural networks (Mokhtarzade, M. & Zoej,M. J. V., 2007) and genetic algorithms (Mokhtarzade, M. et al., 2008) for road network extraction are discussed in many works to reduce human intervention. Researchers from remote sensing community have contributed a lot in this area, but none of them provide a solution for the fully automatic extraction of roads from low resolution, noisy and blurred images. Most of the works are concentrated on extraction of features from high resolution satellite images. For the extraction of roads from distorted images as accurately as obtained from a high resolution image itself, some extra pre-processing and intelligent techniques are necessary. This chapter provides a new approach suitable for low resolution images which combines a wavelet based watershed segmentation method with fuzzy inference system. The method can be applied for high resolution satellite images as well. 5.3 Study Region The proposed road extraction method has been applied for a number of images which includes both real time satellite images and images obtained from Google Earth software. Of these, six of them with different features are chosen to test the validity of the method. Fig.5.1 is a random satellite image degraded due to various unknown factors. This image is chosen as the worst case to test the extend of our proposed method. Fig.5.2 and 5.3 are RGB images obtained from Google Earth software. We have purposefully degraded both test images by adding speckle noise of 0.5 (fig.5.4) and gaussian blur of radius 3 pixels (fig.5.5) respectively to assess the susceptibility of our algorithm, because these two are the most common distortion phenomenon occurring in satellite images. Both images are taken from the city of Washington, D.C having spatial resolution of 1m and 60cm respectively. Fig. 5.6 is a low resolution (5.8m) level 3A panchromatic IRS-ID image of Hyderabad city situated between the latitude 17 22' 31.15" N to 17 22' 18.56" N and the longitude 74

3 5. Road Extraction from Satellite Images 78 28' 27.21" E to 78 28' 12.67" E taken on 7 th October, The image is obtained from National Remote Sensing Centre (NRSC), India ( Figure 5.1: Degraded Satellite Image Figure 5.2: RGB Satellite Image (1m) Figure 5.3: RGB Satellite Image (60cm) Figure 5.4: Image Affected Figure 5.5: Image Affected by by Speckle Noise of 0.5 Guassian Blur of radius 3 75

4 Study & Analysis of Satellite Images for the Extraction of Structural Features Figure 5.6: IRS-1D Image of Hyderabad city (5.8m) 5.4 Road Extraction Method The method proposed in this work is based on image filtering and denoising using wavelet transforms in the pre-processing stage. The second stage is the segmentation of the image using morphological watersheds. Image pixels are classified as road or non-road using a fuzzy based classifier in the final stage. The flow chart of the designed system is given in fig Pre-processing using Wavelet filters Time frequency analysis plays an important role in image processing. Because of its good time-frequency localization characteristics the Wavelet Transform (WT) has gained widespread acceptance in signal and image processing. Wavelet filters are especially suitable for applications where scalability and tolerable degradation are important. Wavelet transform decomposes a signal into a set of basis functions. These basis functions are called wavelets. Wavelets are obtained from a single prototype wavelet called mother wavelet by dilations and shifting. The wavelet transform is computed separately for different segments of the image at different resolutions. It is designed to give good time resolution and poor frequency resolution at high frequencies and good frequency resolution and poor time resolution at low frequencies. Images are signals having high frequency components for short durations and low frequency components for long duration and therefore wavelet analysis plays a leading role in the pre-processing stage of images. Since satellite images are finite energy functions a wavelet transform exists. 76

5 5. Road Extraction from Satellite Images Read input PAN/ RGB degraded satellite image as Aj Initialize j=0 Wavelet filtering of Aj using db1 j= p? No Yes Image gradient Merging of regions based on extended minima transform Binary image imposed on original Watershed transform Image reconstruction using db1 j=0? Segmented Image Mean Standard Deviation Hough Transform Fuzzification Mamdani Fuzzy Inference system Rules (11) Defuzzification Extracted Roads Figure 5.7: Flow Chart of the Proposed Method 77

6 Study & Analysis of Satellite Images for the Extraction of Structural Features The intensity of the features to be segmented in an urban area is not a constant value, because roads, buildings, vehicles and shadows can cause rapid changes to the image intensity. Because of the above mentioned features, wavelet filters are used to smooth images in our research (Mallat, S., 1989) Discrete wavelet transform Discrete wavelet transform (DWT), transforms a discrete time signal to a discrete wavelet representation. The wavelet function is given by two functions that is, a scaling function Φ(t) and wavelet function Ψ(t), which represents a low pass filter h(n) and high pass filter g(n) respectively. If L and H represent low pass and high pass responses, row wise and column wise convolution of the original image A is done by the following equations (5.1 and 5.2). K 1 A, L n A 1, L 2n k H k k 0 [] = [ ] [] = (5.1) K 1 A, H n A 1, H 2n k G k k 0 [] = [ ] [] = (5.2) The daubechies wavelet (Daubechies, 1992) functions used for implementing wavelet filtering in this work is db1 (Haar wavelet function). The FIR filter coefficients of haar wavelet function are given in fig.5.8 and fig.5.9. Haar system is the unique one that satisfies bi-orthogonality, symmetry and compact support properties at the same time. It is also the earliest coined wavelet function. Haar system has a symmetric scaling function, an anti-symmetric wavelet function, a single vanishing moment and has finite support length in the interval [0,1]. Φ(t) H(n) Figure 5.8 Haar Scaling Function and its FIR Response 78

7 5. Road Extraction from Satellite Images Ψ(t) G(n) Figure 5.9 Haar Wavelet Function and its FIR Response Wavelet Filtering Many methods are developed for wavelet filtering of image including the well known mallat s algorithm and a` trous algorithm Mallat s algorithm In order to understand the multi resolution analysis concept based on Mallat s algorithm, it is very useful to represent the wavelet transform as a pyramid, as shown in fig The basis of the pyramid is the original image, with C columns and R rows. Figure 5.10: Pyramidal Representation of Mallat s Wavelet Decomposition Algorithm Each level of the pyramid, which is only accessible from the immediately lower level, is an approximation to the original image. When climbing up in the pyramid, the successive approximation images have a coarser spatial resolution. At the N th level, the approximation image has C/2 N columns and R/2 N rows because a dyadic wavelet transform with subsampling or decimation is applied (Mallat, 1989). These approximation images are computed using scaling functions related to the 79

8 Study & Analysis of Satellite Images for the Extraction of Structural Features Mother Wavelet function ψ(t) (Daubechies, 1992; Mallat, 1989). The difference between the information from two successive levels of the pyramid, e.g. between the original image A j 2 at a resolution 2 j and the approximation image A j-1 2 at a resolution 2 j-1, is given by the wavelet transform and computed using the wavelet functions. Three wavelet coefficient images, DH j-1 2, DV j-1 2 and DD j-1 2 pick up, respectively, the horizontal, vertical and diagonal detail that is lost between the images A j j-1 2 and A 2 and contain the features with sizes comprised between 2 j and 2 j-1 resolution (nonredundant DWT algorithm). If the original image has C columns and R rows in the first level of decomposition, the approximation and the wavelet coefficient images obtained in the first level of decomposition by applying this multi resolution algorithm have C/2 columns and R/2 rows. When the inverse wavelet transform is applied, the original image A j 2 can be reconstructed exactly from the approximation image (A j-1 2 ) and the wavelet coefficients (DH j-1 2, DV j-1 2 & DD j-1 2 ) respectively (fig.5.11). A 2 j-1 A 2 j DH 2 j-1 DV 2 j-1 H-scaling function coefficients G-wavelet function coefficiens DD 2 j-1 Figure 5.11 Ordinary Wavelet Decomposition of an Image The a` Trous Algorithm Another discrete approach of the wavelet transform is the a` trous algorithm (Holschneider,M. et al., 1989) which is used in our work. In this case, the image decomposition scheme cannot be represented with a pyramid as in Mallat s algorithm but with a parallelepiped. The basis of the parallelepiped is the original j image, A 2 at a resolution 2 j, with C columns and R rows. Each level of the parallelepiped is an approximation to the original image, as in Mallat s algorithm. When climbing up through the resolution levels, the successive approximation images have a coarser spatial resolution but the same number of pixels as the original image, as shown in fig

9 5. Road Extraction from Satellite Images Figure Parallepiped Representation of the a` Trous Wavelet Decomposition Algorithm If a dyadic decomposition approach is applied, the resolution of the approximation image at the N th level is 2 j-n. These approximation images are computed using scaling functions. The spatial detail that is lost between the images A j-1 j 2 and A 2 is collected in just one wavelet coefficient image, w j-1 2, frequently called wavelet plane. This wavelet plane, which globally represents the horizontal, vertical and diagonal spatial detail between 2 j and 2 j-1 resolution, is computed as the difference between A j-1 2 and A j 2, i.e. two consecutive levels of the parallelepiped. When the inverse transform is applied, the original image A j 2 can be reconstructed exactly by adding to the approximation image A j-1 2 the wavelet plane w j-1 2. In contrast to Mallat s algorithm, the a` trous algorithm allows shiftinvariant discrete wavelet decomposition. All the approximation images obtained by applying this decomposition have the same number of columns and rows as the original image. For the practical implementation of the a` trous algorithm, a twodimensional filter associated to the scaling function is used. As we filter to obtain coarser approximations of the original image, the above filter must be filled with zeros, in order to match the resolution of desired level. Inserting zeros in the filters creates holes (trous in French). Now filtering is performed in a similar way over these filters with new coefficients (fig.5.13). Why Trous Algorithm? 1. No need for down sampling. 2. Obeys linear additive reconstruction. 81

10 Study & Analysis of Satellite Images for the Extraction of Structural Features This means that original image is reconstructed by adding the detail coefficients of each level to the smoothened image. 3. Ease of interpretation. H -scaling function with zero inserted between the coefficients G -wavelet function with zero inserted between the coefficients Figure 5.13 A Trous Wavelet Decomposition of an Image If the smoothing operation is stopped at resolution level p, reconstruction of the original image A is achieved by adding the detail coefficients DD j of each level 1 < j < p to the wavelet filtered image A p (eqn.5.4) where k 1 & k 2 are dummy integers in which the summations are made. The wavelet filtered image c p is given in eqn.5.3. c p = DD j 1 < j < p (5.3) A(k 1,k 2 ) = A p (k 1,k 2 )+ c p (5.4) This linear simple additive reconstruction formula is the unique convenience of trous algorithm along with its translation invariance and both account for its ease of interpretation. As a consequence of these two properties this algorithm is often employed in object detection Wavelet Filtering Algorithm The steps involved in the wavelet filtering algorithm are as follows: 1. Choose the resolution level p for the wavelet decomposition of the image as 0 j log 2 N, where N is the spatial resolution of the image. 2. Apply series of convolutions along rows and then columns of image matrix using G(n) of the haar filter (db1) to get the filtered image c p. 3. Implementation proceeds with the subtraction of original image from the smoothened image. 4. If j p, go to step 2. Else go to step 5. 82

11 5. Road Extraction from Satellite Images 5. Obtain the output image A p = A c p for G Є db Watershed Segmentation using Extended Minima Transform The next stage is segmenting the image using watershed transform. The different steps involved in the process are described below: Image Gradient The Sobel operator (Sobel edge emphasizing filter) (Gonzalez & Woods, 2008) is used to estimate the image gradient. Each pixel of a gradient image measures the change in intensity of that same point in the original image, in a given direction. To get the full range of direction, gradient images in the x and y directions are computed. If we define A p as the source image, and G x and G y are two images which at each point contain the horizontal and vertical derivative approximations, the computations are as follows (eqn.5.5): Gx = * Ap and Gy= * A p (5.5) where * here denotes the 2-dimensional convolution operation. At each point in the image, the resulting gradient approximations can be combined to give the gradient magnitude, using eqn G = Gx + Gy (5.6) To remove the undesired small gradients, a threshold T is applied to the gradient image, and values smaller than T are set to zero. The threshold T is selected as a standard value, such as T = K max (M), where max (M) is the maximum of the gradient magnitudes, and K is a constant (0 < K < 1) Extended Minima Transform The extended minima transform (Soille, P., 1999) is the regional minima of h- minima transform. The h minima transform suppresses all the minima in the intensity image whose depth is less than h, where h is a scalar. Regional minima are connected components of pixels with the same intensity value t, whose external 83

12 Study & Analysis of Satellite Images for the Extraction of Structural Features boundary pixels all have a value greater than t. The output of regional minima is a binary image with the same size as the original image, in which the pixels with value 1 represents the regional minima in the original image. This also sets all other pixel values to 0. Comparing the extended minima, it can be verified that when h is increased, the area of some objects increase and some other objects disappear. In some cases, the images are noisy and present low contrast, making the choice of the parameter h critical. Considering the union of the pixel sets of the regions obtained by the extended minima when h varies between 1 and k, if k is small, we obtain small regions centred on regional minima of the image. If k increases, the regions grow and can be merged. Based on this, a suitable value is determined for h Segmentation based on Morphological Watersheds A well known morphological approach to segmentation, the watershed algorithm, is generally applied to the gradients of the image. The gradient image can be considered as topography with boundaries between regions, known as ridges. Over time, the watershed transformation has been established to be a very useful and powerful tool for image segmentation. It is the first algorithmic approach invented from the field of topography (Band, L. E., 1986). The segments correspond to the individual regions identified in the image. Every pixel in the image is assigned to the catchment basin corresponding to a regional minimum. Numerous morphological watershed segmentation techniques have been proposed based on immersion, flooding and rainfalling. In this work, the morphological rainfalling watershed technique (Stanislav, L. S., 2000) was chosen as it is computationally faster than the other techniques. Segmentation by morphological watershed normally suffers from the problem of over-segmentation, especially if the image is corrupted with different kinds of noises during acquisition, transmission and storage which happen to be very common for the case of satellite images.to avoid over-segmentation while applying watershed transform compute the extended minima transform of the gradient image and impose the regional minima on the gradient image Image Reconstruction using Inverse Discrete Wavelet Transform Reconstructed image is obtained by performing the inverse DWT (IDWT) on the watershed segmented image with the help of reconstruction filters (fig.5.14). To 84

13 5. Road Extraction from Satellite Images prevent noise from being introduced back into the unsampled image during inverse transform, all detail images are set to zero, except those whose position correspond to the watershed lines of the image. The IDWT is the summation of the products of the decomposed coefficients and wavelet function at respective scales. Figure 5.14 Discrete Wavelet Reconstruction of the Image Figure 5.15: Reconstructed Image The procedure of imposing the selective minima on the gradient image followed by application of watershed transform and image projection to the next layer has been repeated till we obtained the full resolution segmented original image at high-resolution. The IDWT is obtained by the summation of the products of the decomposed coefficients and wavelet function at respective scales. Fig shows the result at this stage for the degraded satellite image in fig.5.1. As seen from the result, the degraded effects are eliminated in this step which can be given to any intelligent system for the extraction of desired features Fuzzy Inference System For a dense urban area a fully automatic and yet accurate road extraction is impossible without using any intelligent computational techniques. Here this goal is 85

14 Study & Analysis of Satellite Images for the Extraction of Structural Features achieved by carefully setting rules for a fuzzy inference system after examining a number of images taken from different satellites. Whatever be the resolution of the satellite image, whether it is noisy or blurred, road pixels have some unique characteristics compared to other features in the scene, which can be utilized for constructing rules for the fuzzy inference system (Tuncer, O., 2007). The characteristics of the designed system are given in table 5.1: Table 5.1: Fuzzy System Characteristics Type AND Method OR Method Defuzzification Method Implication Method Aggregation Method Mamdani Min Max Centroid Min Max Since road pixel intensity values are of medium range their mean will be average and standard deviation is always low. Therefore mean and standard deviation are chosen as the two linguistic variables for the fuzzy system. Most of the roads are linear in nature and so hough transform (Duda, R. O. et al., 1972) is chosen as the third variable. The output linguistic variables are assigned as road, road unlikely and not road. The linguistic variables generated using MATLAB are shown in fig Figure 5.16 Fuzzy Linguistic Variables: (a) Mean (b) Standard Deviation (c) Hough (d) Road Status: Road, Road Unlikely, Not Road 86

15 5. Road Extraction from Satellite Images After evaluating a number of satellite images including high & low resolutions, degraded, blurred and noisy images from their mean and standard deviation, 11 rules are developed for the accurate classification of road pixels. The rules are given in table 5.2. Table 5.2: Developed Fuzzy Rules No. Fuzzy Rules 1 If (mean is low) and (standard deviation is low) and (hough is not line) then (output is not road) 2 If (mean is low) and ( standard deviation is low) and (hough is line) then (output is road unlikely) 3 If (mean is low) and ( standard deviation is high) and (hough is not line) then (output is not road) 4 If (mean is low) and ( standard deviation is high) and (hough is line) then (output is not road) If (mean is average) and ( standard deviation is low) and (hough is not line) then (output is not 5 road) 6 If (mean is average) and ( standard deviation is low) and (hough is line) then (output is road) If (mean is average) and ( standard deviation is high) and (hough is not line) then (output is not 7 road) If (mean is average) and ( standard deviation is high) and (hough is line) then (output is road 8 unlikely) 9 If (mean is high) and ( standard deviation is low) and (hough is not line) then (output is not road) 10 If (mean is high) and ( standard deviation is low) and (hough is line) then (output is road unlikely) If (mean is high) and ( standard deviation is high) and (hough is line) then (output is road 11 unlikely) The output of the fuzzy system is defuzzified using centroid method. If d i is the fuzzified value obtained having a membership value µ i, the defuzzified output is given by eqn n μ d i 1 i i y = = n μ i = 1 i Finally all pixels are classified as white or black depending on whether the defuzzified value exceeds or not a threshold value which is set as Results and Discussions The multistep algorithm is evaluated with different satellite images of varying features and varying resolutions. Output for all types of test images are compared with the hand drawn reference. The results of high resolution RGB test images (fig.5.2 and fig.5.3) are also compared with the method proposed by Tuncer, O. (2007). The extracted road maps are given in fig It is seen that for high 87 (5.7)

16 Study & Analysis of Satellite Images for the Extraction of Structural Features resolution satellite images as well as for degraded satellite images the method proposed in our work is superior as shown in the results given in fig.5.17 & table 5.3. Accuracy assessment is determined by calculating detection rate, false alarm rate and quality percentage (Jin, H. et al., 2012) given in the following equations (5.8, 5.9 and 5.10): TP Detection Rate = TP + FN (5.8) FP False Alarm Rate = TP + FP (5.9) TP Quality Percentage = 100 TP + FP + FN (5.10) where TP (true positive) is the number of road pixels correctly identified, FN (false negative) is the number of road pixels identified as non- road, FP (false positive) is the number of non-road pixels identified as road pixels. The quality percentage is estimated for both the affected and unaffected images. It is proved that competitive results are obtained even for degraded images. The value of the above measures for all the test images are given in table 5.3. Figure 5.17: Road Extraction Results: (a) Road Map of fig. 5.1 (b) Hand Drawn Reference (c) Road Map of fig. 5.2 (d) Road Map of fig. 5.2 by the Method Proposed in Tuncer, O. (2007) 88

17 5. Road Extraction from Satellite Images (e) Road Map of fig. 5.4 (f) Hand Drawn Reference (g) Road Map of fig. 5.3 (h) Road Map of fig. 5.3 by the Method Proposed in Tuncer, O. (2007) (i) Road Map of fig. 5.5 (j) Hand Drawn Reference (k) Road Map of fig. 5.6 (l) Hand Drawn Reference 89

18 Study & Analysis of Satellite Images for the Extraction of Structural Features Table 5.3: Accuracy Assessment Reference Images Detection Rate False alarm Rate Quality percentage Fig.5.1- Degraded Image Fig.5.2- High Resolution (1m) RGB Image [Proposed method] Fig.5.2- High Resolution (1m) RGB Image [Method proposed in Tuncer, O., 2007] Fig.5.3- High Resolution(60cm) RGB Image [Proposed method] Fig.5.3- High Resolution (60cm) RGB Image [Method proposed in Tuncer, O., 2007] Fig.5.4- Noisy Satellite Image Fig.5.5- Blurred Satellite Image Fig.5.6- Low Resolution Satellite Image From the quantitative analysis (table 5.3) and comparison with the hand drawn reference (fig.5.17), it can be seen that almost all the road segments are detected with acceptable accuracy in spite of the degraded effects and poor quality. Though 50% speckle noise is added with the original high resolution image, the quality percentage is only reduced to 3.5% when compared with the noise unaffected image. By applying a guassian blur of radius 3 pixels, as given in table 5.3 the quality percentage is reduced around 1% only. For the low resolution image, though the road segments are not vivid clearly even in the actual scene, the method is able to identify all the road segments. But because of poor quality of the original image, lot of miss detections are there and so the quality percentage is somewhat low. For high resolution images, comparing with the existing technique of road extraction by Tuncer, O. (2007) it is proved quantitatively as seen from the results given in table 5.3 and qualitatively from visual appearance as shown in fig.5.17 that our method is more efficient. 90

19 5. Road Extraction from Satellite Images 5.6 Applicability of soft computing methods in Multi Spectral Satellite Images As seen from the results, fuzzy inference system works well if the road features are somewhat distinguishable with the surroundings irrespective of the high curvature. But as seen from the road network given in fig.5.6, if the road features are not clearly identified, using a simple fuzzy inference system will not be satisfied. Combining neural networks, fuzzy clustering and genetic algorithms which is explained in Mokhtarzade, M. et al. (2008) will give good results for such cases but the computation time is much higher. Therefore, if road features are identifiable, we don t want to go for such time consuming algorithm. Here texture parameters like Energy, Entropy, Contrast, Homogeneity and spectral parameters having R, G, B values are given to the neural network input. Before giving inputs to the classifiers, masking vegetation and water regions (refer section of chapter 2) will give better results as the intensity properties of water and road segments are somewhat similar and if a pixel is identified as that in vegetation area, it will never become a road pixel. Also the computation time is not much increased by doing this extra pre-processing. The inferred results are given in table 5.4. But, for this we require multispectral images of the same region to get the NIR values. The PS-MS level 3A test image given in fig. 5.18(a) has vegetation areas (appears as bright red) which is not much distinguishable from the surroundings. The image is a satellite view of Kathipara Flyover which is an important road junction in Chennai, India. The image is acquired by WorldView-2 earth observation satellite owned by Digital Globe obtained from Satellite Imaging Corporation (refer section 3.4). The image is of 1.8m resolution taken on 5 th March, Results given in fig show that though spatial domain processing and intelligent techniques can be applied for accurate detection, regions having high NDVI values can be masked to increase the detection efficiency. Here accuracy assessment is done by calculating RCC, BCC and RMSE values Mokhtarzade, M. et al. (2008). RCC and BCC, stand for Road/Background Detection Correctness Coefficient respectively, are the average of correct response for road and background detection by comparing the hand drawn reference image. Regarding the difference between the actual response and its expected response 91

20 Study & Analysis of Satellite Images for the Extraction of Structural Features (0 for background and 1 for road pixels) as the error values, the Root Mean Square Error (RMSE) can be computed. As seen from the results given in table 5.4, the RMSE value is improved by masking pixels having higher NDVI values without consuming much time. Figure (a) PS-MS Satellite Image (1.8m) (b) Hand drawn reference Figure 5.19: (a) Result using FIS Alone (b) Result of 5.19(a) After Masking Vegetation Areas using NDVI (c) Result of the Method Proposed by Mokhtarzade, M. et al. (2008) (d) Result of 5.19 (c) after Masking Vegetation Areas using NDVI 92

21 5. Road Extraction from Satellite Images Table 5.4: Comparison of Different Methods Reference Image - PSMS satellite image [Fig. 5.18(a)] RCC BCC RMSE Computation Time (Sec) Using FIS alone - Fig.5.19(a) Using FIS alone after masking vegetation areas using NDVI - Fig.5.19(b) Using the method proposed in Mokhtarzade, M. et al. (2008) - Fig.5.19(c) Using the method proposed in Mokhtarzade, M. et al. (2008) after masking vegetation areas using NDVI - Fig.5.19(d) Summary A fully automatic road extraction algorithm for both high and comparatively low resolution satellite images and degraded satellite images are discussed. The method is compared with the method proposed by Tuncer, O. (2007). A brief investigation into the applicability of soft computing methods in multi spectral images is also given and a comparative study is done with the method proposed by Mokhtarzade, M. et al. (2008). Both quantitative and qualitative analysis and comparison with the hand drawn reference is done in each and every method discussed in this study to validate the performance of our algorithm. 93

22