Detecting Land Cover Changes by extracting features and using SVM supervised classification ABSTRACT Mohammad Mahdi Mohebali MSc (RS & GIS) Shahid Beheshti Student mo.mohebali@gmail.com Ali Akbar Matkan, Alireza Shakiba, Mohammad Hajeb Department of RS & GIS, University of Shahid Beheshti, Tehran In change detection studies Support Vector Machine classification method has given reliable results. In this paper the aim is to detect land cover changes using SVM classification from Google Earth images that have high spatial and temporal resolution. Firstly, images of the twenty second zone of Tehran municipality was extracted from Google Earth so as to be used as raw images. In order to get the maximum accuracy for the outputs, several features were extracted from the images. Sensitivity analysis was then performed on the outputs in order to compare the accuracy of the result. The sensitivity analysis has indicated notable changes classifying raw images stacked with features such as saturation stretched bands and also different textures with different parameters. Out of the results the most accurate maps were chosen in order to detect land cover changes in the case study area. Moreover the maps can be further studied in order for city planners to develop a sustainable urban plan. Key Words: Change Detection, SVM, Feature Extraction, Google Earth Images 1-Introduction There are many approaches in Land Cover change detections; traditional techniques such as land survey and remote sensing techniques using satellite images.the disadvantages for the traditional land survey approach are that it is time-consuming, costly and it is very human dependent. Using remote sensing techniques is very beneficial due to the temporal and spatial resolution that it can provide and also it is semi-automated [1]. In remotely sensed data change detection studies Support Vector Machine (SVM) classification method that is derived from statistical learning theory is very reliable. SVM Classification method often gives better results when case study area consist urban regions due to complexity of urban developed areas [2].Detecting changes in an urban scale requires high resolution imageries.[3] Google Earth Images high spatial and temporal resolution has a high potential in remote sensing field. Due to the complexity of images consisting urban areas and the correlation between different bands of these images, they are needed to be enhanced in order to get better results. Performing classification according to the value of each pixel does not provide acceptable results. By adding texture features, enhancing the image color and eliminating the correlation between bands the classification performed provides higher accuracy [4]. Feature extraction techniques usually change the outcome of classification. Also using texture filters would enhance the result of classifications [5]. These textures add different features to the image. Many studies in the field of image processing using texture cooccurrence, has been done. The results in various studies indicates changes in classification accuracy in the assessment of urban development process [6]. Although these textures change the outcome of the classifications but they do not always improve the result. The output should be compare with each other in order to find the optimum result. Sensitivity analysis on the
features would help in identifying critical features that would improve the result so as to yield the highest accuracy possible. In this paper, the aim of the research was to integrate google earth images with remote sensing techniques to produce maps with the highest accuracy possible in order to detect land cover changes. 2-Material and Method a. The case study area is located at North West of Tehran. The area extent shown with red box in Figure (1). For the purpose of this paper the images of two years (2013, 2009) of the twenty second zone of Tehran municipality was exported from Google Earth. Figure (2) and Figure (3) were then used as raw images for this study. FIG 1. Tehran Municipality Zonings FIG 2. 2004 Google Earth Image from case study area. FIG 3. 2014 Google Earth Image from case study area.
b. Image enhancing techniques was used in order to get the optimum result from these raw images. Saturation Stretch (SS) enhances the color of an input image by producing output bands that have more saturated colors. The input data are transformed from red, green, and blue (RGB) space to hue, saturation, and value (HSV) space. A Gaussian stretch is performed on the saturation band so the data fills the entire saturation range. The HSV data are then automatically transformed back to RGB space. c. Co-occurrence texture filter were applied on the raw images. These eight textures were the result of applying this filter; mean, variance, homogeneity, contrast, dissimilarity, entropy, second moment, and correlation. In order to test the potential of using co-occurrence filter for classification different window processing size (3 by 3, 5 by 5, 7 by 7, 9 by 9 and 11 by 11) was used. d. To classify the raw images SVM method were used with optimum parameters that were obtained empirically (these parameters were constant true the entire process in order to test the sensitivity of extracted features). e. A sensitivity analysis was then performed using extracted features with the raw images so as to choose the accurate map. Confusion matrix was used to get the accuracy of a classification result by comparing a classification result with ground truth information. Ground truth data that was digitized from the raw images is presented in Figure (4). (Red color are areas covered by soil, green color stands for vegetation cover and yellow color are urban areas). FIG 4.Methodology flowchart f. Change detection statistics used to accumulate a detailed tabulation of changes between two classification images. To perform change detection between years 2013 and 2004, the post-classification comparison change detection technique was employed. For applying this technique the highest accuracy possible outputs chosen from classification maps to be compared. The classes of two maps will be compared with the other classes in this technique. The development of classification outputs in order to apply change detection has been presented in Figure (5).
Google Earth Images Ground Truth Data 2004 2013 Feature Extraction Texture Filter Saturation Stretch Training Data ------------------ Parameter Selection SVM Classification Method Output for 2004 Output for 2013 Sensitivity Analysis Change Detection Statistics FIG 5.Methodology flowchart
3-Results The outputs with highest accuracy (for the years 2004 and 2013 the highest accuracy achieved were 96.7% and 92% respectively) using SVM supervised classification for two years (2004, 2013) are shown in Figure (6) and Figure (7) (yellow color represents road networks and urban areas, green color stands for vegetation cover, red color shows undeveloped areas). FIG 6.The output map for 2004 FIG 7.The output map for 2013(blue color represents artificial lake) The output RGB bands from applying saturation is shown in Figure (8). FIG 8. Enhanced 2004 raw image using saturation stretch.
The result of post-classification comparison change detection is presented in Table (1) (urban class consists of road networks and buildings and water class represents artificial lake). This table illustrates land cover changes had happened in the case study area. Table 1. Change Detection Statistics Area (Square km) Vegetation Urban Soil Row Total Class Total Water 0.00 0.03 1.17 1.19 1.19 Vegetation 0.88 0.35 0.38 1.61 1.61 Urban 0.97 9.10 15.58 25.64 25.64 Soil 0.45 0.81 12.33 13.59 13.59 Class Total 2.30 10.28 29.45 0.00 0.00 Class Changes 1.42 1.18 17.13 0.00 0.00 Image Difference -0.69 15.36-15.86 0.00 0.00 In the Figure (9) and Figure (10) the overall accuracy of comparison by sensitivity analysis are presented. 97 96 95 94 93 92 91 Sensitivity Analysis Raw Images Saturation Stretch Texture 5*5 Texture 7*7 Texture 9*9 Texture 11*11 FIG 9. Overall accuracy comparison for between 2004 outputs (texture number*number stands for texture feature created with number*number window processing size). 96.75 96.7 96.65 96.6 96.55 96.5 96.45 96.4 96.35 96.3 96.25 Sensitivity Analysis Texture 7*7 Texture 7*7[-] FIG 10. Overall accuracy comparison between 2004 outputs ( [-] stands for stacked images excluding entropy and second moment)
4-Conclusion This research indicates that using features would create notable difference in accuracy of output maps. The most accurate maps for years of 2013 and 2004 was 92 % and 96.7 % respectively. Without using texture filters the outputs accuracy would drop almost 3%. As it has been presented on Figure (6) and Figure (7) the changes in land cover are noticeable. Most of these changes were in the soil class which was almost 16km 2 in ten years, this indicates rapid urban growth in this zone. Table (1) also describes reduction in vegetation cover in this zone due to urbanization. It can be realized from Figure (9) that window processing sizes can change the result of the classification outputs. The reason as it has been presented in Figure (11) is that when window size is changed from 3*3 window size to 7*7 window size the complexity between features (e.g. buildings) becomes lessened and when we move from 7*7 window size to 11*11 window size the edge between the features (for e.g. between soil and buildings) becomes unnoticeable. FIG 11.Variance textures with different window size form 2004 raw image (from left to right first 3*3, 5*5, 7*7, 9*9, 11*11) It can be understood from Figure (12) that as the more complex and noisy the raw images gets the raw image stacked with saturation stretched band and also texture feature would give better result. 98 96 94 92 90 88 86 Sensitivity Analysis Texture 7*7(2013) Texture 7*7+SS+RI(2013) Texture 7*7(2004) Texture 7*7+SS+RI(2004) FIG 12. Overall accuracy comparison between 2013 outputs The outcome of this paper shows the potential of integrating google earth images with remote sensing techniques. Furthermore Image fusion of these maps with different satellite images (for e.g. thermal images) and with other feature layers would give better understanding and help city planners to develop sustainable plans.
References 1. Singh, Ashbindu. "Review Article Digital change detection techniques using remotely-sensed data." International journal of remote sensing 10.6 (1989): 989-1003. 2. Mountrakis, Giorgos, Jungho Im, and Caesar Ogole. "Support vector machines in remote sensing: A review." ISPRS Journal of Photogrammetry and Remote Sensing 66.3 (2011): 247-259. 3. Nnam, Victor CHUKWUEMEKA, and Godwin UCHECHUKWU Nnam. "Assessment of Urban Development Planning using Supervised Classification of remotely sensed Imageries and GIS, A Case study of Independence Layout (Part of), Enugu, Nigeria." 4. Zhang, Yun. "Optimisation of building detection in satellite images by combining multispectral classification and texture filtering." ISPRS Journal of Photogrammetry and Remote Sensing 54.1 (1999): 50-60. 5. Kishore Das, D., K. Gopal Rao, and A. Prakash. "Improvement of effective spatial resolution of thermal infrared data for urban landuse classification." Remote Sensing and Data Fusion over Urban Areas, IEEE/ISPRS Joint Workshop 2001. IEEE, 2001. 6. Anys, H., A. Bannari, D. C. He, and D. Morin, 1994. Texture analysis for the mapping of urban areas using airborne MEIS-II images, Proceedings of the First International Airborne Remote Sensing Conference and Exhibition, Strasbourg, France, Vol. 3, pp. 231-245.