Journal of Electrical and Computer Engineering

Volume 2017 (2017), Article ID 3123967, 8 pages

https://doi.org/10.1155/2017/3123967

## Statistical Similarity Based Change Detection for Multitemporal Remote Sensing Images

Computer Science & Engineering Department, Rajshahi University of Engineering & Technology, Rajshahi 6204, Bangladesh

Correspondence should be addressed to Mumu Aktar; moc.liamg@teur.umum

Received 11 March 2017; Revised 23 May 2017; Accepted 14 June 2017; Published 24 July 2017

Academic Editor: Huan Xie

Copyright © 2017 Mumu Aktar et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

#### Abstract

Change detection (CD) of any surface using multitemporal remote sensing images is an important research topic since up-to-date information about earth surface is of great value. Abrupt changes are occurring in different earth surfaces due to natural disasters or man-made activities which cause damage to that place. Therefore, it is necessary to observe the changes for taking necessary steps to recover the subsequent damage. This paper is concerned with this issue and analyzes statistical similarity measure to perform CD using remote sensing images of the same scene taken at two different dates. A variation of normalized mutual information (NMI) as a similarity measure has been developed here using sliding window of different sizes. In sliding window approach, pixels’ local neighborhood plays a significant role in computing the similarity compared to the whole image. Thus the insignificant global characteristics containing noise and sparse samples can be avoided when evaluating the probability density function. Therefore, NMI with different window sizes is proposed here to identify changes using multitemporal data. Experiments have been carried out using two separate multitemporal remote sensing images captured one year apart and one month apart, respectively. Experimental analysis reveals that the proposed technique can detect up to 97.71% of changes which outperforms the traditional approaches.

#### 1. Introduction

Change detection (CD) plays an important role in earth observation since a lot of changes occurring on different areas of earth surface cause severe damage to that area. So, it is very important to identify changes occurring on that area to take steps for subsequent recovery from damage. Most of the damage is caused due to natural disasters like flooding, rainfall, and droughts and also by man-made activities and thus it is necessary to observe the areas on regular intervals for detecting that change. Multitemporal remote sensing images taken at different times provide the periodic repeated look of an area which is very necessary for regular observation of earth surface.

The spatial, radiometric, spectral, and temporal resolution of an image has been improved due to recent advances in satellite imaging that has made earth observations much easier [1]. Multispectral remote sensing image is a valuable source of satellite data providing necessary information by each band, that is why in this research two separate multispectral datasets of two different dates have been chosen to identify changes. CD is applied to different tasks including land cover change detection [2, 3], building change detection [4, 5], vegetation change detection [6], wetland monitoring [7], and so on. Generally, CD detection can be categorized in two ways: binary change detection and multiclass change detection. Binary change detection focuses on change versus no change whereas multiclass change detection uses some supervised or unsupervised approaches for classifying multiple classes and further identifies changes of each class separately. This paper is concerned with binary change detection. A number of different techniques are used for binary change detection such as image differencing [8, 9], log-ratio operator [10], Kullback Leibler divergence [11], regression analysis [12], vegetation index differencing [13, 14], and principal component analysis [1, 15].

Ground truth data is needed for supervised approach being used for change detection. When this ground truth or prior knowledge is unavailable then threshold selection is very important to take decision about change/no change. An adaptive threshold can be used for generating a binary change detection map. Change detection of multisensor images at pixel level is verified using manual threshold which faced false alarms because only the mean value is considered for threshold selection [3]. Here in this research a difference image is generated using image differencing method which requires a threshold to take decision about changes. OTSU method for threshold selection [16, 17] has been used here since it maximizes variance between changed and unchanged pixels for generating a binary change detection map.

Since threshold used in image differencing method depends on radiometric differences of pixel values, singly it is not always enough to determine changes between two times’ images since they can provide false detection due to registration errors, that is why a new approach is proposed here based on statistical similarity to determine the percentage of changes along with image differencing. In [3] a number of similarity measure techniques such as distance to independence, normalized standard deviation or Woods criterion, correlation ratio, mutual information (MI), and cluster reward algorithm (CRA) have been applied on land cover change detection for multisensor remotely sensed images and the performances have been compared in case of two separate fixed window sizes. Mean value of similarity image has been used there as a threshold for decision about change/no change of any pixel. Although result of the statistical measures has been normalized, optimal dimension of the estimated windows has not been investigated. Land cover change detection has been performed in [2] considering sliding window by transforming spectral values into local spectrum trend using curve fitting and raster encoding techniques. This approach is time consuming and even difficult when calculating the local trend spectrum for different window sizes. Multisensor remote sensing image change detection using mutual information, maximal information coefficient, and distance correlation measure based similarity measures has been developed in [18] using sliding window approach with fixed window size of . This approach used manual threshold for interpreting change and no change regions and generated false positives. Selection of threshold is a difficult task there as it is justified manually.

Among a number of similarity measures, MI has been chosen here for change detection purpose because of its many advantages. MI is a nonparametric approach that does not require any assumption about the shape of the distribution of input variables and able to measure both the linear and nonlinear relationships among input variables [19]. A lot of papers considered MI as a similarity measure for change detection task including multivariate statistical model [20] which concentrated on homogeneous and heterogeneous sensors, SAR image change detection [21] which considered different kind of changes, canonical information analysis [22] for image change detection, multicontextual mutual information data as an improved form of image spatial mutual information [23] for SAR image change detection, and temporal behavior of multichannel scene characterization for change detection [24].

But the value of MI is affected by the entropies of the input variables and the calculated MI without being normalized cannot measure similarity effectively, so the MI measure can be improved by normalizing it to a specified range such as [19], that is why MI being normalized (NMI) has been analyzed here in this research work as a similarity measure to detect changes. But when just traditional NMI is applied on the whole image then a pixels’ marginal and joint probability density functions (PDFs) are affected with the globally spread information that can be noisy and has less significance on them. But if sliding window with varying window sizes is considered then pixels’ local neighborhood is used for calculating PDF which is the most significant one. The window which is too small or too large affects the CD results [2]. So different window sizes have been considered here in this research to find out which one is more efficient for better change detection result. Thus to significantly perform change detection on multitemporal remote sensing images sliding window based normalized mutual information (SWB-NMI) with varying window sizes has been considered here to statistically detect changes using 1-Imagesimilarity(**X**,** Y**) concept. Experimental analysis has been performed using remotely sensed multispectral imagery of Canberra region of Australia with multitemporal datasets of two separate dates.

The rest of the paper is organized as follows. Section 2 provides a clear concept of proposed change detection method. The proposed methodology for multispectral images’ change detection is presented in Section 3. In Section 4 experimental analysis and performance evaluation of the proposed method have been performed. Finally, a conclusion is given in Section 5.

#### 2. Proposed Change Detection Method

##### 2.1. OTSU Threshold Selection Method

Threshold selection is a key concept used in unsupervised approaches for generating a binary change detection map. OTSU’s method is a popular approach which maximizes variance between changed and unchanged pixels to get an optimal threshold [17]. The OTSU algorithm assumes the image containing two classes of pixels as foreground pixels and background pixels; it then calculates the optimum threshold separating the two classes to minimize their combined spread (intraclass variance), , or equivalently (because the sum of pairwise squared distances is constant), to maximize their interclass variance, [16]. The optimal threshold, , can be obtained by maximizing the following criteria with respect to [17]:where and are the count and mean value of elements in difference image, which are less than , and and are the count and mean value of elements in difference image, which are no less than .

Finally the decision about change/no change can be defined as [17]

##### 2.2. NMI as a Statistical Similarity Measure

MI of two discrete random variables** X** and** Y** can be defined as [19] using Shannon entropy theorem [25]:where , , and are the joint and marginal probability distribution functions of** X** and** Y,** respectively. MI between these two variables is zero if they are independent of each other and will be higher if they are similar to some degree. If the MI value is normalized to the range between 0-1 then the effect of input variables’ entropy can be removed and thus MI can be used as an improved similarity measure. Therefore, NMI has been applied here to identify percentage of changes between two images** X** and** Y **in which one is the reference and the other is the target image. The NMI is defined as [19]where is the entropy of the variable** X** which represents the amount of information held in and the same representation in case of .

##### 2.3. SWB-NMI as a Statistical Similarity Measure

Figure 1 is an example showing the basic concept of how sliding window can be moved over an image to apply NMI on each local window. An efficient way for remote sensing image change detection is to analyze the image’s local geographical layout for its similarity measure. For this purpose the sliding window approach can be considered where a change indicator is applied separately on each window. Thus the statistical properties such as joint and marginal probabilities of the pixels can be calculated from the local neighborhood information which are most significant ones. Different window sizes are considered here to achieve the relevant information. Window sizes such as , , , , , , and are used instead of full image to avoid the global statistics of pixels which may contain irrelevant information. Thus SWB-NMI has been chosen in this research with an expectation of detecting changes locally and the initial results are promising. Since a large window results noise and small window do not contain enough spatial contextual information [2], varying sized windows have been applied to the experiment here to find an optimal window size and the performances have been compared with traditional approaches stated in [2, 3].