Post-Classification Change Detection of High Resolution Satellite Images Using AdaBoost Classifier

Size: px

Start display at page:

Download "Post-Classification Change Detection of High Resolution Satellite Images Using AdaBoost Classifier"

Charlotte Lester
6 years ago
Views:

1 , pp Post-Classification Change Detection of High Resolution Satellite Images Using AdaBoost Classifier Dong-Min Woo 1 and Viet Dung Do 1 1 Department of Electronics Engineering, Myongji University San 38-2 Namdong, Cheoin-gu, Yongin, Gyeonggido, South Korea {dmwoo2000, mr.dovietdung}@gmail.com Abstract. This paper presents a post-classification approach that can achieve efficient results for change detection by using AdaBoost classifier. In the first step, a land cover of satellite image is classified in the independent fashion. For the stable classification of man-made structures, 3D features are employed. 3D co-occurrence feature is used with 2D co-occurrence feature and Harr-like feature. Adaboost classifier is implemented for the classification of a land cover into 4 classes: road, grass/bare-ground, foliage and building. In the second step, we perform the change detection for multi-temporal images. We carry out the simple comparison between terrain classification results. The experimental results show that the resulting change map efficiently reflects a newly developed area. Keywords: post-classification, AdaBoost, satellite image, co-occurrence feature, change detection 1 Introduction Change detection approaches can be divided into two categories: unsupervised and supervised methods. [1] The common unsupervised approach is by way of difference image [2]. It can be created by subtracting counterpart pixels from two given images taken at two different instances. Pixels beyond an empirical threshold are associated with changed areas, meanwhile the under threshold values are attached to unchanged areas. In practice, threshold is an arbitrary number at first and modify gradually until the change detection is recognized properly Image rationing [3] can be computed by the ratio between the two corresponding values from two co-registered images to produce difference matrix. Supervised method can be efficiently used for change detection, since it is insensitive to the change of image due to lighting variation. One of the most popular supervised methods is the post-classification comparison approach [4][5]. However, the accuracy of this approach critically depends on the result of terrain classification. In this paper, we suggest the application of AdaBoost classifier to change detection problem. The main advantage of AdaBoost algorithm is a feature selection mechanism, which can be crucial for accurate terrain classification. [6] ISSN: ASTL Copyright 2015 SERSC

2 In this paper, we use the combination of 2D and 3D co-occurrence [7] and Haarlike features [8] for feature extraction and selection. Accurately segmented temporal images lead to efficient change detection. In this context, we perform experiments on change detection in terms of the post classification comparison approach. In this paper, two remote sensing images of the same area, acquired at different times, are classified independently. Comparison between the two classification maps is used to determine the change. 2 Multi-Class AdaBoost Algorithm 2.1 AdaBoost Classifier In the original form, the AdaBoost learning algorithm is used to boost the classification performance of a simple learning algorithm. It combines a set of weak classification functions to form a stronger classifier. Each weak learner determines the optimal threshold classification function, such that the misclassified number of samples should be minimized. A weak classifier thus consists of a feature, a threshold, and a polarity indicating the direction of the inequality: { In this training process, the error rate of the current weak classifier is reduced based on the update of the previous classifier. The weights of training set determine the probability being selected for a feature and they are continuously updated via every weak classifier. 2.2 Multi-Class Implementation Multi-class problem is more practical than the binary classifications. As we mentioned before, we should distinguish the classes of building, foliage, grass/bareground, and road. And our solution is almost the same as One vs All classification. For instance, we have four classes already. We will classify three times. Firstly, we classify building with all dataset including other samples. Then we label all the buildings, and execute the classification for the rest of samples. We classify foliage from dataset which is labeled as building. Foliage s sample is labeled, and we carry out the similar way to the rest of samples. 3 Change Detection Scheme Given two satellite images taken at different times in the same location, it is expected that a change detector will be able to identify changed area and recognize types of change. In this paper, we adopt post-classification scheme. Copyright 2015 SERSC 35

For the segmentation of each temporal image, we extract 2D co-occurrence feature, and we use DEM to extract 3D co-occurrence feature.

3 In the first phase, each temporal image is classified into the classes of building, foliage, grass/bare-ground and road. Adaboost classifier performs image segmentation for each temporal image to produce a four-value segmented image. For the segmentation of each temporal image, we extract 2D co-occurrence feature, and we use DEM to extract 3D co-occurrence feature. Three types of co-occurrence features used for calculation are ASM, CON, ENT. With four angular directions, twelve 2D cooccurrence feature and twelve 3D co-occurrence features are calculated. Using extracted features, we carry out AdaBoost classification with randomly selected training samples. Only 5% of ground truth data was used in the training. In the second phase, the supervised change detection is carried out in terms of post classification comparison. The difference between segmented images is calculated for the comparison of temporal images. Then, a change map can be produced by the comparison of two segmented images. 4 Experiments For the experiments, we use IKONOS satellite images of Daejun area. From ortho image and DEM (Digital Elevation Model) with size of 2048x2048, samples of building, foliage, grass/bare-ground and road are manually collected as rectangular windows, and will be used as a ground truth for training and accuracy evaluation, as shown in Fig. 1. (a) Building (b) Foliage (c) Grass (d) Road Fig. 1. Ground truth of four classes. (a) Segmented result using AdaBoost (b) Segmented result using ANN Fig. 2. Terrain classification result. 36 Copyright 2015 SERSC

4 The segmented result is shown in Fig. 2 (a). Classification result is displayed in four gray levels so that we can visualize the classified area. ANN(Artificial Neural Network) classification is also carried out for the comparison, as shown in Fig. 2 (b). Table 1 shows classification accuracy based on the correctly classified pixels in comparison with the ground truth. We notice that the accuracy of building class is lowest among four classes. One of the main reasons for this is due to the shadow occurred by buildings. Comparison of two classifier shows that AdaBoost classifier slightly outperforms ANN with accuracy of 88.2%. However, ANN classifier comes up with inconsistent accuracy with regards to the class. In this respect, we conclude that the proposed AdaBoost classifier shows more stable performance for the terrain classification. Table 1. Classification accuracy AdaBoost Classifier ANN Classifier class number of correctly accuracy correctly classified accuracy pixels classified pixel number pixel number building 61,294 44, , foliage 75,052 67, , grass 174, , , road 66,696 51, , total 377, , , (a) New segmented image (b) Post-classification result Fig. 3. Experimental result of change detection We perform change detection experiment using multi-temporal images. Another image for the same area acquired at the different instance is used for this experiment. Fig. 3 (a) shows the segmented result of the new image acquired at the different instances. Since the size of the new image is different, only overlapped area is used for change detection experiment. Fig. 3 (b) shows the change map of post-classification comparison. In the right bottom area, we can see that 6 new structures are detected as changed area, which are indicated by three red rectangles. However, there are some false alarms in addition to the real change. Copyright 2015 SERSC 37

5 5 Conclusions In this paper, we have studied post-classification change detection method by using AdaBoost algorithm. Post-classification approach is known to be very sensitive to the accuracy of terrain classification. Since we use many features including 2D and 3D co-occurrence features and Harr-like feature, we believe that the feature selection mechanism of AdaBoost classifier improves a classification performance for terrain classification. Experimental results show that the implemented AdaBoost classifier outperforms the conventional ANN classifier. We also confirm that the proposed change detection scheme performs efficiently to detect the newly developed building. In this context, it can be concluded that the proposed method can be very useful in detecting change from multi-temporal images. Acknowledgments. This research was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (No. NRF-2012R1A1A ). Reference 1. Singh, A.: Digital change detection techniques using remotely-sensed data. In: International Journal of Remote Sensing, vol. 10, pp (1989) 2. Bruzzone, L., Prieto, D.: Automatic Analysis of the Difference Image for Unsupervised Change Detection. In: IEEE Tran. Geoscience and Remote Sensing, Vol. 38, pp (2000) 3. Coppin P., Jonckheere I., Nackaerts K., Muys B.: Digital change detection methods in ecosystem monitoring. In: International Journal of Remote Sensing, vol. 25, pp (2004) 4. Serra, P., Pons, X., Sauri, D.: Post-classification change detection with data from different sensors: Some accuracy considerations, In: International Journal of Remote Sensing, vol. 24, pp (2003) 5. Liu, H., Zhou, Q.: Accuracy analysis of remote sensing change detection by rule-based rationality evaluation with post-classification comparison. In: International Journal of Remote Sensing, vol. 25, pp (2004) 6. Zhu, J., Zou, H., Rosset, S., Hastie, T.: Muti-class AdaBoost, In: Statistics and Its Interface, Vol. 2, pp (2009) 7. Eleyan, A., Demirel,H.: Co-occurrence matrix and its statistical features as a new approach for face recognition. In: Turkish Journal of Electrical Engineering & Computer Sciences, vol. 19(1), pp (2011) 8. Lienhart, R., Maydt, J.: An Extended Set of Haar-like Features for Rapid Object Detection. In: Proceedings International Conference on IEEE Image Processing 2002, vol.1, pp. I I-903 (2002) 38 Copyright 2015 SERSC

An Improvement of the Occlusion Detection Performance in Sequential Images Using Optical Flow

, pp.247-251 http://dx.doi.org/10.14257/astl.2015.99.58 An Improvement of the Occlusion Detection Performance in Sequential Images Using Optical Flow Jin Woo Choi 1, Jae Seoung Kim 2, Taeg Kuen Whangbo