2Department of Applied Physics, University of Granada, Granada 18071, ..... By means of the scalar invariant of octonions, we find that both electromagnetic.
Progress In Electromagnetics Research Symposium Proceedings, Marrakesh, Morocco, Mar. 20–23, 2011 1057
Use of the Neural Net for Road Extraction from Satellite Images, Application in the City of Laghouat (Algria) F. Benkouider1 , L. Hamami2 , and A. Abdellaoui3 1
2
Electrical Engineering Department, Amar Telidji University of Laghouat, Algeria Electrical and Computer Engineering Department, National Polytechnic School, Algeria 3 Department of Geography, URF Letters and Human Sciences, Paris XII, France
Abstract— Road extraction in urban areas has been an important task for generating Geographic Information Systems (GIS). In recent years, mainly, the rapid development of urban areas makes it urgent to provide up-to-date road maps. The timely road information is very useful for decision-makers in urban planning, traffic management and car navigation fields, etc. The satellite image is characterized by its big quantity of rich and varied information and constitutes a source of data for generating roads maps by automatic overseeing techniques according to their importance. The interpretation of this one requires mostly, a treatment based on a set of techniques: shaping information, filtering, segmentation and classification, etc. We present in this paper a method based on neuronal net strategy for extracting road networks based on the spectral characteristics of the pixel in satellite image. Normalized spectral information in a window (3 × 3) around each pixel are used, as 9 red, 9 green and 9 blue (corresponding respectively to red, green and near infra red channels) constitute the input vector of 27 neurons. The origin of the motivation is the homogeneity of roads in high-resolution satellite images, since homogeneity is a characteristic that can be recognized with respect to neighbor pixels, and their spectral information. The system output is represented by one normalized neuron representing the road or not-road characteristics. As the neuronal network requires a large coded data bases in their training stage, we have used a set of road net manually drown using special software. The image used in this application concerns an HVR SPOT image acquired on March 26, 2007 (10 m of resolution) over Laghouat (Algeria), an oasis city located 400 km south of Algiers (Algeria). We have obtained very accurate results with less than 0.022 for the MSE. A set of various applications are presented. 1. INTRODUCTION
The satellite image is characterized by its big quantity of rich and varied information; it constitutes a source of data for generating the road maps by automatic techniques according to their importance. Road extraction in urban areas has been an important task for generating geographic information systems GIS). Nowadays, we are experiencing an explosion in the amount of satellite image data, which provides us with abundant data and also brings challenges to the road extraction task at the same time. The conventional road extraction methods by manual processes are time consuming and tedious, and cannot meet the increasing requirement for such tremendous data. However, automatic extraction of urban roads from high resolution remote sensing imagery is still a challenging problem in digital Photogrammetry and computer vision, the main reason is that the diverse road surfaces and the complex surrounding environments such as trees, vehicles and shadows induced by high buildings make the urban roads take on different textures and gray levels in images. Many researchers have been conducted for this purpose. Using specific operators such as Duda operator for finding linear structures are based on a score function which takes into account the homogeneity and contrast [1]; several road detectors have been applied, the blind operators such as Top to Hat Form (THF), derived from mathematical morphology and designed to extract peak intensity in the image spot [2]; the THF is not very selective and gives noisy results. Followed by several other works in [3, 4]. However, most of them focus on extracting roads in rural or open areas. By contrast, the efforts made for urban road extraction are relatively few [5, 6]. Some works focuse on automatic road extraction in urban area from high resolution satellite images using based machine learning approach [7, 8]. The other semi-automatic methods such as active contour (or snake) [9, 10] and dynamic programming [11] have been the subject of several studies. One can find an excellent survey paper in [12] on road network extraction using the organizing maps applied to classified images. Multilayer neural networks applied in particular for IKONOS images based on the RGB spectral characteristics is presented in [13], those involving texture analysis, fuzzy clustering and genetic algorithms has been treated in [14, 15]. In the present research, road extraction is performed on Spot XS images, using artificial neuronal network algorithms; with a new input
PIERS Proceedings, Marrakesh, MOROCCO, March 20–23, 2011
1058
Table 1: Colour composite. Channel
Wave length
XS3 (Spot 1 to 3) – B3 (Spot 4 to 5) XS2 (Spot 1 `a 3) – B2 (Spot4 to 5) XS1 (Spot 1 `a 3) – B1 (Spot 4 to 5)
0.78 to 0.89 micrometers 0.61 to 0.68 micrometers 0.50 to 0.59 micrometers
Storage level in the standard composite colour Red Green Blue
and architecture (RGB). The format of a window as input is related to that road can be shown as elongated homogenous region with different contrast from the background. We present a different application of the proposed model, and the correspondent MSE error and Kappa coefficient. We use at the end mathematic morphology operations to extract road sides. 2. METHODOLOGY
Road detection can be considered as the first step in road extraction, it is the process of assigning a value to each pixel that can be used as criteria of road and not-road pixels. The problem of road detection from high-resolution satellite images is performed using: 1) artificial neuronal network and 2) pixel spectral characteristics specially the red, green and near infra red channels. The origin of the motivation is the homogeneity of roads in high-resolution satellite images, since homogeneity is a characteristic that can be recognized with respect to neighbor pixels, and their spectral information. In our work, the following hypothesis and input requirements have been considered: A. Spectral characteristics: The RGB bands of satellite image are calqued on our visual perception, it uses three basic colors: the Red (λ = 700 nm), the green (λ = 546 nm) and the blue (λ = 435.8 nm). In our work we have used the so called Standard Colour Composite (SCC), Table 1. B. Road appearance: In the RGB standard color composite roads appears in blue. In some cases building appear in blue colour too, this problem appears in classification methods when roads are classified as a building. C. Road homogeneity: Road networks in high resolution satellite images are presented as elongated homogenous areas having a distinctive brightness pattern compared to their surroundings. 3. PRINCIPLE OF THE ARTIFICIAL NEURAL NETWORK
Artificial Neuronal networks are made up of simple processing units called neurons which are usually organized into layers with full or partial connections. The principal task associated with a neuron is to receive the activation values from its neighbors, compute an output based on its weighted input parameters and send that output to its neighbors. ANNs have already been used in few instances in photogrametry and satellite image processing. In our method they are applied as sophisticated pattern classifiers. We have chosen a feed-forward backpropagation neural network, which is one of the most frequently implemented network types. Learning (training) neural networks is a time-consuming task. For its efficiency, the Back Propagation learning algorithm which is an iterative gradient decent algorithm, was used. It is designed to minimize error function expressed in Equation (1): E=
1 XN (Dj − OjM )2 j=1 2
(1)
where Dj and Oj are the desired input and the current response of the neurons j in the input layer, respectively and N the number of neurons in the output layer. The iterative method, corrections to weight parameters are computed and added to the previous values as illustrated below: ∂E ∆wi,j = −η (2) ∂w ∆w (t + 1) = i,j ∆w + α∆w (t) i,j
i,j
i,j
Progress In Electromagnetics Research Symposium Proceedings, Marrakesh, Morocco, Mar. 20–23, 2011 1059
In this equation, wi,j is weight parameter between neurons i and j, η a positive constant that controls the amount of adjustment and is called learning rate, a α momentum factor that can take on values between 0 and 1 and “t” denotes the iteration number. The parameter α can be called smoothing or stabilizing factor as it smoothes the rapid changes between the weights. 4. ARTIFICIAL NEURAL NETWORKS FOR ROAD EXTRACTION
Road detection from satellite images can be considered as a classification process in which pixels are divided into road and background classes. A backpropagation neural network (BNN) with one hidden layer is used. Normalized spectral information in a window of (3 × 3) around each pixel of RGB images are used, as 9 red neighbours pixels, 9 green neighbours pixels and 9 blue neighbours pixels to constitute the input vector of 27 neurons. The output layer consists of one neuron that represents the networks output by a number between 0 and 1 as not road and road pixel, respectively, Figure 1. As the neuronal network requires a large coded data bases in the training stage, a set of 100 road net manually drown using special software are chosen as training set in learning stage. It is recommended to have representative pixels of all present objects in the training set. 5. RESULT AND INTERPRETATION
The combination of both 27 input parameters made the network powerful in the detection of road and background, reducing also the request hidden layer and size iteration time. We used as a first step a set of, 10, 15 and 20 neurons in the hidden layer to test the performance of the neural network and we found that a layer of 20 neurons in the hidden layer is sufficient to improve the network’s ability in both road and background detection. Figure 2 (a1 ), (a2 ) shows sections of a satellite image SPOT XS (2007) from the city of Laghouat (Algeria). They were considered as original image and are applied to the input of the neural network. (b1 ), (b2 ) are their respective produced road map used in accuracy assessment. The results of the neural network for road detection are shown by image (c1 ), (c3 ). For accuracy assessment, we consider two parameters: the mean square error (MSE) and the Kappa coefficient, Table 2. The MSE, proved to be the most reliable parameter to be used as termination condition, improve the accuracy of the results. The Kappa coefficient, the overall accuracy parameter, is calculated by the same way as classification methods. The proposed ANN has presented no over-trained problem. To improve extracted roads, we propose the application of two morphological operations to the ANN extracted roads: A. Morphological erosion is applied to the gray scale image extracted by the road ANN model
Figure 1: Proposed neural net for road extraction.
PIERS Proceedings, Marrakesh, MOROCCO, March 20–23, 2011
1060
(a 1)
(b1)
(c 1)
(d1)
(e1)
(a2)
(b2)
(c2)
(d2)
(e2)
Figure 2: Result of Proposed ANN for Road detection: (a1 ), (a2 ) Sections of Spot Image (XS, 2007) of the city of Laghouat (Algeria). (b1 ), (b2 ) Manually produced maps. (c1 ), (c2 ) Extracted road by ANN model. (c1 ) and (c2 ) Road extracted by ANN model. (d1 ) and (d2 ) Road after gray scale erosion. (e1 ) and (e2 ) binarization by threshod. Table 2: Accuracy assignment. Image (a1 ) (b1 )
MSE 0.0206 0.0211
Kappa Coefficient 86.2% 84.5%
with a chosen structuring element in order to smooth roads without modifying its wideness. Erosion generally decreases the size of objects and removes small anomalies by subtracting objects with a radius smaller than the structuring element. The result of this algorithm applied to images (c1 ) and (c2 ) are respectively shown in (d1 ) and (d2 ) of Figure 2. B. The obtained images are then transformed to binary image by threshold; the result is shown in images (e1 ) and (e2 ). 6. CONCLUSIONS
In this paper, a method based on neuronal net strategy for detecting road net in the city of Laghouat (Algeria) is presented. This method is based on spectral characteristics of the pixel from satellite image. As the neuronal net require a large coded data bases in their training stage, we have used a set of road net manually drown using special software. We have obtained very accurate results with less than 0.022 for the MSE. A set of applications are presented. This approach is distinguished from previous works by the choice and the structure of the multilayer neural network input, which is mainly based on the spectral characteristics of the pixel intensity level in the three channels (red , green and blue) and the neighbors of the considered pixel, that influence greatly the quality of output (extracted road network image). These results are very accurate since the method can extract the road despite the resolution of the image (10 m). It is important to notice that, automatic extraction of urban roads from remote sensing imagery is still a challenging problem in digital Photogrammetry and computer vision. The main reason is that the diverse road surfaces and the complex surrounding environments such as trees, vehicles and shadows induced by high buildings make the urban roads take on different textures and gray levels in images. We propose some morphological operations in order to obtain the road cartography: grayscale erosion is applied to the extracted road by the proposed ANN system, followed by a binarization process. For further work we propose using recalled images and geometric characteristics of road for network’s training in order to improve network’s ability in road detection.
Progress In Electromagnetics Research Symposium Proceedings, Marrakesh, Morocco, Mar. 20–23, 2011 1061 ACKNOWLEDGMENT
Our thanks to Professor Andr´e Ozer, Director of the Geomorphology and Geography Laboratory, University of Liege (Belgium) and Doctor Marc Salmon, to provide us the image data needed to develop this work. REFERENCES
1. Fisher, M. A., J. M. Tenenbaum, and H. C. Wolf, “Detection of road and linear structures in low-resolution areal imagery using a multisource knowledge integration technique,” CGIP, No. 15, 201–223, 1982. 2. Roux, M., “Recalage images SPOT/carte routi`ere,” International Archives of Photogrammetry and Remote Sensing, Vol. 29, Tome B3, Commission III, 384–391, 1992. 3. Zhang, C., S. Murai, and E. Baltsairas, “Road network detection by mathematical morphology,” ISPRS Workshop on 3D Geospatial Data Production: Meeting Application Requirements, 185–200, Paris, 1999. 4. Mohammadzadeh, A., A. Tavakoli, and M. J. Valadan Zoej, “Automaticlinear feature extraction of Iranian roads from high resolution multi-spectral satellite imagery,” 20th ISPRS Congress, 764–768, Istanbul, Turkey, 2004. 5. Hinz, S., A. Baumgartner, C. Steger, H. Mayer, W. Eckstein, H. Ebner, and B. Radig, “Road extraction in rural and urban areas, semantic modeling for the acquisition of topographic information from images and maps,” SM-ATI’99, 7–27, M¨ unchen, 1999. 6. Geman, D. and B. Jedynak, “An active testing model for tracking roads in satellite images,” IEEE Trans. PAMI, Vol. 18, No. 1, 1–14, January 1996. 7. Narayanan, P. J., et al., “Extraction of main road high satellite images by machine learning,” ACCV 2006, LNCS 3851, 236–245, 2006. 8. Sowmya, A. and S. Singh, “Rail: Extraction road segments from areal image using machine learning,” Proc. ICML 99 Workshop on Learning in Vision, 8–18, 1999. 9. Marikhu, R., M. N. Dailey, S. Makhanov, and K. Honda, “A family of quadratic snakes for road extraction,” ACCV 2007, Part 1, LNCS 4843, 85–94, 2007. 10. Hinz, S. and A. Baumgartner, “Automatic extraction of urban road networks from multi-view aerial imagery,” ISPRS Journal of Photogrammetry & Remote Sensing, Vol. 58, 83–98, 2003. 11. Dal Poz, A. P. and G. M. do Vale, “Dynamic programming approach for semi-automated road extraction from medium- and high-resolution images,” ISPRS Archives, Vol. 34, Part 3/W8, 17–19, Munich, September 2003. 12. Mokhtarzade, M., H. Ebadi, and M. J. V. Zoej, “Optimisation of road detection from high resolution satellite images using texture parameters in neuronal network classifiers,” Canadian Journal of Remote Sensing, Vol. 33, No. 6, 484–491. 13. Mokhtarzade, M. and M. J. V. Zoej, “Road detection from high-resolution satellite images using artificial neural networks,” International Journal of Applied Earth Observation and Geoinformation, Vol. 9, No. 1, 32–40, 2007. 14. Mokhtarzade, M., M. J. V. Zoej, and H. Ebadi, “Automatic road extraction from high resolution satellite images using neuronal net work, texture analysis, fuzzy clustering, Genetic algorithm,” The International archive of Photogrammetry, Remote Sensing and Spacial Information Sciences, Vol. 37, Part B3b, Beijing, 2008. 15. Mohammadzadeh, A., A. Tavakoli, and M. J. V. Zoej, “Road extraction based on fuzzy logic and mathematical morphology from pan sharpened IKONOS images,” The Photogrammetric Record, Vol. 21, No. 113, 44–60, 2006. 16. Civco, D. L. and Y. Waug, “Classification of multi-spectral, multitemporal multi-source spatial data using artificial neural networks,” Proceeding of the ASPRS 1994 Annual Convection, 123– 133, Reno, NV, USA, 1994.
