Change detection in Landsat images based on local neighbourhood information
2018; Institution of Engineering and Technology; Volume: 12; Issue: 11 Linguagem: Inglês
10.1049/iet-ipr.2018.5524
ISSN1751-9667
AutoresNeha Gupta, Gargi V. Pillai, Samit Ari,
Tópico(s)Advanced Image Fusion Techniques
ResumoIET Image ProcessingVolume 12, Issue 11 p. 2051-2058 Research Article Free Access Change detection in Landsat images based on local neighbourhood information Neha Gupta, Corresponding Author neha27brs@gmail.com orcid.org/0000-0002-2287-1259 Department of Electronics and Communication Engineering, National Institute of Technology Rourkela, Odisha, 769008 IndiaSearch for more papers by this authorGargi V. Pillai, Department of Electronics and Communication Engineering, National Institute of Technology Rourkela, Odisha, 769008 IndiaSearch for more papers by this authorSamit Ari, Department of Electronics and Communication Engineering, National Institute of Technology Rourkela, Odisha, 769008 IndiaSearch for more papers by this author Neha Gupta, Corresponding Author neha27brs@gmail.com orcid.org/0000-0002-2287-1259 Department of Electronics and Communication Engineering, National Institute of Technology Rourkela, Odisha, 769008 IndiaSearch for more papers by this authorGargi V. Pillai, Department of Electronics and Communication Engineering, National Institute of Technology Rourkela, Odisha, 769008 IndiaSearch for more papers by this authorSamit Ari, Department of Electronics and Communication Engineering, National Institute of Technology Rourkela, Odisha, 769008 IndiaSearch for more papers by this author First published: 17 September 2018 https://doi.org/10.1049/iet-ipr.2018.5524Citations: 9 AboutSectionsPDF ToolsRequest permissionExport citationAdd to favoritesTrack citation ShareShare Give accessShare full text accessShare full-text accessPlease review our Terms and Conditions of Use and check box below to share full-text version of article.I have read and accept the Wiley Online Library Terms and Conditions of UseShareable LinkUse the link below to share a full-text version of this article with your friends and colleagues. Learn more.Copy URL Share a linkShare onEmailFacebookTwitterLinked InRedditWechat Abstract In this study, a novel technique is proposed to detect the changes in bitemporal multispectral images. Utilisation of the local neighbourhood information in any image processing task may provide good noise immunity and reduces false alarms. Motivated by this, Otsu's thresholding of local information based approach is proposed in this work. It shows the effective performance in change detection of bitemporal Landsat images which suffer from different atmospheric and sunlight conditions. To get the local information around each pixel, both bitemporal images are partitioned into overlapping image blocks. Every block of the first image is concatenated with the corresponding block of the second image for each pixel position. Thus, the information of the concatenated block is considered as inter-block information. Further, Otsu's method is applied on the concatenated block for threshold calculation. Depending on the threshold, binary values are generated. Finally, binary values of both images for all bands are compared by XOR operation to detect it as the background i.e. unchanged pixel or foreground i.e. changed pixel. On the basis of majority class present in XOR output, binary change map is generated. Experiments conducted on Landsat images show that the proposed method provides better performance compared to reported techniques. 1 Introduction Change detection is the way of distinguishing differences in an object or phenomenon by observing it at various times [1]. It is useful in many applications such as land-use change, deforestation, forest fire, landscape change, urban change, forest or vegetation change, wetland change and environmental change [2]. Changes on the earth's surface can be detected using the bitemporal satellite images with various resolutions [3]. Resolutions of satellite images are varying from moderate to high, and it can be used for different applications of change detection. Bitemporal satellite images acquired at different times suffer from different atmospheric and sunlight conditions. Moreover, acquiring bitemporal ground truth data or training sample is a very difficult task for satellite images. Therefore, it is necessary to develop an unsupervised change detection technique which reduces the false decision occurring due to the effect of the aforementioned external sources. Many unsupervised reported techniques [4-11] provide change detection map either based on difference image or by analysing bitemporal images separately. A number of change detection techniques are reported based on the difference image and this difference image is created by applying algebraic operation on bitemporal images. In [4], on the basis of Bayes theory, two automatic approaches have been proposed. The first thresholding based approach assumes that pixels are independent to each other whereas in the second one, interpixel class dependency is exploited by using Markov random fields (MRF) theory. The second technique considers the spatial contextual information. In both the techniques, statistical terms are estimated with an iterative method, i.e. expectation maximisation (EM) algorithm. It is stated that MRF approach that exploits the spatial contextual information provides good performance compared to the first approach. Although many change detection methods are based on thresholding techniques [4-6], few techniques have focused the change detection based on kernels. Kernel methods produce change detection results using the non-linear transformation approach based on kernels [7-9]. In [7], the difference image is analysed in original domain as well as in kernel domain. Here, the kernel k-means (KKM) algorithm is applied for partitioning the image into two classes viz., 'change' and 'no change'. Moreover, multiple KKM clustering with local-neighbourhood information technique [8] provides strong immunity against noise due to inclusion of local-neighbourhood information into its clustering objective function. In addition, clustering-based techniques in [10] extracts the local spatial contextual feature by applying Gabor wavelet transform on difference image. This Gabor feature based two-level clustering (GaborTLC) technique produces results with less isolated spots. Although aforestated difference image based techniques provide satisfactory results, some information may be discarded during generation of difference image, e.g. it fails to consider the starting and ending pixel's location in the feature space [1, 12]. More recently, few change detection techniques have analysed the bitemporal satellite images separately to generate the binary change map. In [11], the binary descriptors for each pixel are generated by analysing the bitemporal images separately, where the local neighbourhood information are utilised around each pixel. Then, the obtained binary descriptors of bitemporal images are used to calculate the hamming distance of each pixel, and binary change map is generated by applying the Lloyd–Max's algorithm on hamming distance of each pixel. Here, this technique is represented as binary descriptor and Lloyd–Max's (BDLM) method. Another technique is called PCANet [13] in which neighbourhood patches of each pixel of both images are utilised to train the PCANet model, where PCA filters are used as convolutional filters. Taking this neighbourhood features around each pixel makes PCANet technique more robust to the speckle noise. Besides aforesaid change detection techniques, object-based change detection (OBCD) techniques have gained much attention to provide change detection of medium-resolution as well as very high-resolution satellite images. OBCD techniques, which consider the richer information in terms of texture, shape and spatial relationships with neighbouring objects, yield better performance by exploiting the contextual information against the traditional pixel-based approaches [14]. In traditional pixel-based approaches, pixel is used as an analysis unit, which produces spurious and noisy results in optical remote sensing images, whereas in OBCD techniques, these noises are filtered by using contextual or local neighbourhood information of image pixels and object-based contextual measures [15]. In addition, while using high-resolution remote sensing images, pixel-based algorithms produce less accurate change detection results due to high spectral variability within geographic objects whereas OBCD techniques, which are able to extract shape and contextual windows based on local object characteristics, reduce illumination variation, and also provide slight misregistration between multitemporal images [16]. In [17], blended high spatio-temporal resolution data is generated by object-based spatial and temporal vegetation index unmixing model (OB-STVIUM) that disaggregates MODIS NDVIs to Landsat objects using the spatial analysis. This OB-STVIUM technique reduces pseudo-changes occurring due to phenological differences. Object-based blending technique performs better compared to pixel-based blending which may have blocky appearance results [17]. Further, various methods have been discussed based on multi-temporal-object analysis which provide better results. In [18], multi-temporal-object method is realised to get multi-type land cover change in medium-resolution images where chi-square transformation is used to get a threshold to differentiate the changed object from the unchanged ones. This change detection technique, which incorporates contextual information, can provide partial changes of geographic object such as contraction, expansion and fragmentation. In [19], an unsupervised change detection technique is proposed, which uses the robust analysis unit, i.e. multi-temporal object. This technique monitors the urban growth at building level in very high resolution (VHR) imagery. By merging the features of first-order statistics of the multispectral image, ratio features as well as texture features of the near infrared (NIR) and greyscale image, this technique achieves best results even for remote sensing imagery from different sensors. Along with this, in [20], multitemporal-object based method is combined with cross-sharpening of multitemporal data. Such combination reduces the false change area detection that is occurring due to the effect of local displacements by different acquisition angles. In general, inclusion of local neighbourhood information makes change detection algorithm's performance better [4, 8, 10, 11, 13-20]. However, most of these techniques perform change detection analysis either on difference image or features obtained from two images, and this analysis uses an iterative procedure, which is applied for all pixels of the image. In this paper, determination of the changed and unchanged pixel is taken based on the analysis of local neighbourhood information of both images with iteration procedure performed on local neighbourhood information, i.e. only on block information of each pixel position. A very simple and effective approach based on Otsu's thresholding [21] of local neighbourhood information is proposed in this work. Unsupervised and non-parametric nature of Otsu's thresholding, and owing its ability to work directly on grey level, provide motivation to use it in change detection of bitemporal satellite images. Herein, overlapping square image blocks or neighbourhood information around each pixel position of both the satellite images are concatenated. This inter-block information is utilised to calculate the threshold for each pixel position by applying Otsu's thresholding method. Now, whether the centre pixel of each block in both the images belongs to same feature or different feature, this decision is taken with respect to the calculated threshold. To avoid the false detection occurring due to the external sources, another common threshold is added to each calculated Otsu's threshold for each pixel position. Further, the binary values corresponding to each pixel position of all bands are concatenated and pixels are classified as foreground or background by performing XOR operation. Finally, binary change map is created by using the majority class concept. The main contributions of this paper are as follows: (i) Both the multitemporal images analysed separately instead of analysing difference image. (ii) Threshold is calculated by taking both inter-block information and inter-image information. (iii) Iterative procedure is performed on locally, i.e. inter-block information or concatenated block only. (iv) The proposed method detects major changes by considering all the bands. Moreover, it can also be applied for different band combination or individual band for any specific thematic field of application. Experiments conducted on multispectral images show the better performance of the proposed method compared to earlier reported change detection techniques. This paper is organised as follows. Section 2 explains the Otsu's thresholding. Section 3 describes the proposed work. Experimental results with database description, qualitative and quantitative assessment are presented in Section 4. Section 5 concludes this paper. 2 Otsu's thresholding Otsu's thresholding is a non-parametric and unsupervised technique to select an automatic threshold for image segmentation [21]. In this method, a discriminant criterion is used to select the optimal threshold that will maximise the separation of the resultant classes. This algorithm works on the basis of histograms or works directly on grey level values of the image. Mathematically, Otsu's thresholding [21] is defined as follows. Given a digital image of size with L distinct intensity levels denoted as . Let denote the number of pixels with intensity i. The total number of pixels in the image is . Suppose pixels are divided into two classes and by a threshold at level k; where indicates the pixels with levels and indicates the pixels with level . In order to evaluate the threshold, the following discriminant criterion is used in Otsu's method as follows: (1)where , and represent the between-class variance, within-class variance and the total variance, respectively. Otsu's thresholding turns out to be an optimisation problem that maximises one of the above objective function for searching a threshold k. These criteria are equivalent to one another, therefore, any criterion can be used. As stated in [21], between-class is based on first-order statistics while within-class variance is based on second-order statistics, therefore, utilising is the simplest criterion to calculate the threshold k. The simplicity of this criterion also lies on that is independent of k, this is stated further. Intra-class or within-class variance is represented as (2)where and are the variances of classes and , respectively; and represent the probability of class occurrence and it is computed from the L histograms as (3) (4)where is the components of normalised histogram. Mean of the classes are computed as follows: (5) (6)The variances of the classes are given as (7) (8)The total variance [22] is represented as (9)As is constant and independent of k, the effect of changing the threshold is merely to move the contributions of the two terms, i.e. within- and between-class variances [22]. Therefore, minimising the within-class variance is same as maximising the between-class variance [22]. Hence, using the as a discriminant criterion to calculate the threshold is simple, as it consists of between-class variance and total class variance of levels. The optimal threshold that will be utilised in thresholding of image is represented here. This threshold can be seen as it maximises or equivalently maximises between-class variance [21]. The optimal threshold [21] can be evaluated as (10) 3 Proposed methodology Let us consider two coregistered satellite images and of size acquired at different times and of the same geographical region. The main objective is to obtain the binary change detection map with change and nochange information of a geographical region occurred within a time duration. Theblock diagram of the proposed method is shown in Fig. 1. The proposed method consists of the following steps: partitioning of images,concatenation of the partitioned blocks, threshold calculation, comparison of eachpixel using XOR and generation of the binary change map. Visual representation ofthe proposed method is shown in Fig. 2 using anumerical illustration. The detailed description of the proposed methodology isgiven as follows. Fig. 1Open in figure viewerPowerPoint Block diagram of the proposed method Fig. 2Open in figure viewerPowerPoint Visual representation of the proposed method by using a numericalillustration (a) Case when the corresponding pixels ofboth images are unchanged pixels. In this case the values ofcorresponding pixels will be either less or more than the calculatedthreshold, but both pixels will provide the same binary values,(b) Case when the correspondingpixels of both images are changed pixels 3.1 Partitioning of images into overlapping blocks In this step, two bitemporal satellite images and are taken. Each image is partitioned into overlapping square blocks of size . Around each pixel, a block of size is considered and this pixel is the centre pixel of considered block. Blocks around the each pixel position of both images are represented as and for the corresponding images and , respectively. 3.2 Concatenation of the blocks As stated earlier, for each pixel position, there are two blocks corresponding to images and . These blocks are concatenated to get a single block for each pixel position. The concatenated block D is represented as (11)where represents the position of each pixel of and images and is given as follows and . As D is made by overlapping blocks of both images for each pixel position, it is considered as inter-block information for further analysis. 3.3 Otsu's thresholding applied on the concatenated blocks Otsu's thresholding is applied on the obtained concatenated block for each pixel position. Here, inter-block information is utilised in the calculation of the threshold. Otsu's based threshold is represented as follows: (12)Above calculated threshold is used to classify the considered centre pixels of blocks of both images and into two classes. If these centre pixels belong to the same class, then they are assigned same binary values, otherwise, they will be assigned different binary values. Suppose centre pixels of both blocks are unchanged even then the calculated threshold may classify them into different classes due to influence of neighbouring pixels and also because of the unchanged pixels of the satellite images may not have the same values with respect to each other. To avoid this situation, one constant value as T is added to shift the threshold to a certain level where false decision can be avoided (13)where th represents the final threshold that will be used for comparison with each pixel of the images and . T is calculated as follows: (14)where and represent the mean of the images and , respectively, and are evaluated as follows: (15) (16) 3.4 Comparison of each pixel of each bitemporal image with calculated threshold Each pixel of both images is compared with calculated threshold. If the value of any pixel is greater than the threshold then this pixel will be represented as '1', otherwise as '0'. The comparison is performed as follows: (17) (18)where and represent the binary value of pixels corresponding to and images, respectively. Multispectral datasets that have multiple bands are used in this work. Therefore, binary values are evaluated for each band of the bitemporal multispectral images using the above steps. 3.5 Concatenation of binary values In this step, above evaluated binary values of each band are concatenated to get binary vector for each pixel position of each bitemporal multispectral image. The concatenated binary vectors are represented as follows: (19) (20)where and represent the binary vectors for both multispectral images, and N shows the number of considered spectral bands. represents the transpose of matrix. 3.6 XOR operation The binary vectors corresponding to each pixel position of both multispectral images are applied to XOR operation. Each bit of both binary vectors is compared. If corresponding bits of binary vectors are different, then output is '1' otherwise it is '0'. XOR output is represented as follows: (21)where X represents the XOR output corresponding to position of each pixel. For the simplicity of notation, XOR output for each position can be represented with binary vector , where v represents the element of XOR output corresponding to each band. 3.7 Binary change map In XOR output, either the number of 0‱s will be more compared to 1‱s or the number of 1‱s will be more compared to 0‱s. Depending on the maximum number of 0‱s or 1‱s, majority class is decided as '0' or '1'. The binary change map is created on the basis of the majority class present in the XOR output. This XOR output provides similarity between the binary vectors of each bitemporal image for each pixel position. If the majority class in XOR output is '0' for any pixel position, it means those pixels of both images are similar. The position of that pixel in the binary change map will be considered as an unchanged pixel. If it is '1', then that pixel will be considered as change. To find the majority class present in XOR output the total number of elements corresponding to each class is determined as follows: (22) (23)where [] is the Iverson bracket, and are the counter outputs that counts the number of ones and zeros present in XOR output, respectively, u is the element of binary vector . Generation of the binary change map is represented as follows: (24)where '1' represents the unchanged region and '0' represents the changed region. 4 Experimental results The performance of the proposed technique is evaluated on three real datasets. Optical satellite images acquired by Landsat satellite are considered to conduct the experiments. 4.1 Database description 4.1.1 Dataset I First real-world dataset consists of two multispectral images acquired by Landsat 7 EnhancedThematic Mapper Plus (ETM+) Sensor on 9 February 2001 and 21 September 2001[23]. The study area is NaturalLake, situated in Jaisalmer district, Rajasthan, India. Consideredmultispectral images consist of six spectral bands from 1–5 and 7 with aspatial resolution of 30 m. The major change is observed when the lake driedduring summer. The size of the image used for the experiment is pixels, and the bitemporal images ofdataset I are shown in Figs. 3a and b. The reference map orground truth of this dataset is shown in Fig. 3c. It is created manually by detailed visualanalysis of input images captured at two different times. Fig. 3Open in figure viewerPowerPoint Qualitative results of dataset I obtained by variouschange detection methods (a),(b) Landsat 7 EnhancedThematic Mapper Plus (ETM+) Sensor (band 4 NIR) bitemporalimages of Natural lake in Rajasthan,(c) Reference map (groundtruth),(d)–(g)EM [4], KKM [7], BDLM [11] and GaborTLC [10] methods, respectively,(h)–(k)results of proposed methods with patch sizes , , and , respectively 4.1.2 Dataset II Second real-world dataset is acquired by Landsat 5 Thematic Mapper (TM) sensor on 23 October 2009and 20 June 2010 [23]. The study areais the Dharoi Dam on the Sabarmati river situated in Gujarat, India. Thenumber of spectral bands considered in this multispectral images is sixbands from 1–5 and 7 having 30 m spatial resolution. The size of the image used for the experiment is pixels, and the bitemporal images ofdataset II are shown in Figs. 4a and b. The reference map orground truth of this dataset is shown in Fig. 4c. It is created manually by detailed visualanalysis of input images captured at two different times. Fig. 4Open in figure viewerPowerPoint Qualitative results of dataset II obtained by variouschange detection methods (a),(b) Landsat 5 Thematic Mapper(TM) sensor (band 4 NIR) bitemporal images of Dharoi dam inGujarat, (c) Reference map (groundtruth).(d)–(g)EM [4], KKM [7], BDLM [11] and GaborTLC [10] methods, respectively,(h)–(k)Results of proposed methods with patch sizes , , and , respectively 4.1.3 Dataset III This dataset is collected by Landsat 8 Operational Land Imager (OLI) sensor over an area ofYambulla State Forest, situated in Yambulla, New South Wales 2550, Australiaon 1 October 2015 and 6 February 2016 [23]. The number of spectral bands considered in thismultispectral images is seven bands from 1–7 having 30 m spatialresolution. The considered images have the size of pixels, and the bitemporal images ofdataset III are shown in Figs. 5a and b. The mainland changewas due to bushfire that has occurred across the Gold Mine road on 20December. The reference map or ground truth of this dataset is shown inFig. 5c. It iscreated manually by detailed visual analysis of input images captured at twodifferent times. Fig. 5Open in figure viewerPowerPoint Qualitative results of dataset III obtained by variouschange detection methods (a),(b) Landsat-8 OLI Sensor (band5 NIR) multitemporal images of Yambulla State Forest,(c) Reference map (groundtruth),(d)–(g)EM [4], KKM [7], BDLM [11] and GaborTLC [10] methods, respectively,(h)–(k)Results of proposed methods with patch s , , and , respectively In order to evaluate the performance of the proposed method, the qualitative and quantitative experiments are carried out on the above datasets. 4.2 Qualitative results To get the rough idea about the generated binary change detection map, visual results are given in qualitative assessment. Here, the visual results are compared with the reference map. In binary change map, black pixels are showing the 'change' region and white pixels are showing the 'no change' region. The proposed method is compared with earlier reported techniques like EM [4], KKM [7], BDLM [11] and GaborTLC [10] methods. The results are post-processed to get the refined and smoothened binary change map [24]. The visual binary change maps for dataset I, dataset II and dataset III are shown in Figs. 3d–k, 4d–k and 5d–k, respectively. Results show that the proposed method is better than EM [4], KKM [7], BDLM [11] and GaborTLC [10] methods. It is observed from visual results that EM, KKM and BDLM have more false alarms compared to the proposed method and the results with GaborTLC is less accurate than the proposed method. EM technique [4] detects the changed area heavily, which causes the more false alarms and the results appear very noisy. In KKM technique [7], information in some places is very similar and KKM technique is unable to distinguish those information. Therefore, false alarms are more in KKM method. In BDLM method [11], the bitemporal satellite images are analysed separately to generate the binary vectors. Large variations present in pixels of bitemporal images, cause the large variation in binary vectors of unchanged pixels. Sometimes, those pixels interpreted as changed pixels may result in more false alarms in BDLM technique. Utilising the inter-block information along with threshold T, the false alarms are less in the proposed method, and visual results of the proposed method are better compared to earlier reported techniques. 4.3 Quantitative results In quantitative results, generated binary change map using proposed method is compared with EM [4], KKM [7], BDLM [11] and GaborTLC [10] methods with respect to some predefined parameters [13, 25, 26]. The quantitative measures used in this work are as follows. (i) Correct classification or overall accuracy in percentage ( or ), (ii) false positives or false alarm in percentage ( or ), (iii) total error in percentage , (iv) kappa coefficient (). EM, KKM, BDLM and GaborTLC methods are implemented with default parameters given in [4, 7,10, 11], respectively. The results with quantitative measures are shownin Table 1. As shown in the table, the proposed method with patch size achieves 91.24% accuracy for dataset I,95.27% accuracy for dataset II and 92.08% for dataset III. In all cases,accuracy is better than the EM [4], KKM[7], BDLM [11] and GaborTLC [10] methods. The proposed method provides 0.7869, 0.8804 and 0.8001 askappa value for dataset I, dataset II and dataset III, respectively, which arebetter than the compared methods. The false alarm rate which shows the number ofunchanged pixels getting as changed is less in all datasets for the proposedmethod compared to EM, KKM, BDLM methods, and it is comparable with respect toGaborTLC. Moreover, the higher kappa values, and the lower values of total errorrates of the proposed method over EM, KKM, BDLM and GaborTLC methods demonstratethe effectiveness of inter-block information used in this analysis. Further,adding a common threshold T with the evaluated Otsu's thresholddecreases false alarms in the proposed method. Fig. 6 shows the qualitative resultsof the proposed method using threshold T and without usingthreshold T for all datasets. It can be observed that withoutadding T, more number of unchanged pixels are detected aschanged, which increases the false alarms. This is observed that the falsealarms increase from 1.07 to 19.78, 3.29 to 24.29 and 1.29 to 20.01% for datasetI, dataset II, and dataset III, respectively, without using thresholdT. In summary, the proposed method provides betterperformance accuracy and less false alarm rate due to the utilisation of theinter-block information and addition of a common threshold T.Furthermore, one more merit of the proposed method is that it exploitsinformation from all bands of the multispectral data using majority classconcept that results in the reduction of the false detection. Table 1. Performance measures in terms of correct classification or overallaccuracy ( or ), false Alarms (), total error and kappa value by variouschange detection methods on three datasets (Natural Lake, DharoiDam, Yambulla State Forest) Dataset Change detectionmethod , % , % , % dataset I (Natural Lake) EM [4] 80.69 25.17 19.31 0.6055 KKM [7] 80.85 8.36 19.15 0.5321 BDLM [11] 80.80 25.00 19.20 0.6074 GaborTLC [10] 87.99 0.24 12.01 0.6963 proposed 91.24 1.07 8.76 0.7869 dataset II (Dharoi Dam) EM [4] 81.42 25.38 18.58 0.6118 KKM [7] 81.95 18.11 18.05 0.5819 BDLM [11] 85.12 19.68 14.88 0.6746 GaborTLC [10] 93.09 2.57 6.91 0.8174 proposed 95.27 3.29 4.73 0.8804 dataset III (Yambulla StateForest) EM [4] 84.74 17.22 15.26 0.6651 KKM [7] 91.50 1.40 8.50 0.7848 BDLM [11] 89.96 8.43 10.04 0.7652 GaborTLC [10] 89.82 1.16 10.18 0.7368 proposed 92.08 1.29 7.92 0.8001 Fig. 6Open in figure viewerPowerPoint Qualitative results of all datasets produced by the proposedmethod for the analysis of T (a)–(c)Results with T for dataset I, dataset II anddataset III, respectively,(d)–(f)Results without T for dataset I, dataset II anddataset III, respectively The experiments are also conducted for different patch sizes where the different values ofs are taken as 3, 5, 7 and 9. Visual results with differentpatch sizes are shown in Figs. 3h–k, 4h–k and5h–k for dataset I, dataset II and datasetIII, respectively. It is observed that as patch size increases, false alarm rateincreases and also the detected change regions are getting reduced. Thequantitative results with different patch sizes are shown in Fig. 7. Here, graph is plotted for overall accuracy, false alarms and kappa () parameters corresponding to differentpatch sizes. As shown in the graph, if the patch size increases, OA and kappaparameters both decrease while FA increases. When the patch size is increased,number of pixels inside the patch is also increased. If more number of pixels ina patch are changed then centre pixel, which might be unchanged pixel, will bemisclassified as changed. Thus, unchanged areas around heavily affected changedareas are also classified as changed; this increases the FA. If more number ofpixels in a patch are unchanged then centre pixel will be misclassified asunchanged, even though it is the changed pixel; this causes the decrease inkappa value. From the qualitative and quantitative analysis, it is seen that thepatch size of provides the best performance. Fig. 7Open in figure viewerPowerPoint Performance of the proposed method for different patch sizesand datasets in terms of parameters (a) Overall accuracy (OA),(b) False alarm (FA),(c) Kappa () In order to examine the robustness of the proposed method against different level of noises,zero-mean Gaussian noise with different signal-to-noise ratios has been added to image of the bitemporal satellite images ofdataset I, dataset II and dataset III. The performance of EM, KKM, BDLM,GaborTLC and the proposed methods for different levels of noises is shown inFig. 8. The proposed method performsbetter against noise compared to earlier reported techniques as it uses localneighbourhood information and inter-image based thresholding technique. Fig. 8Open in figure viewerPowerPoint Change detection performance in terms of overall accuracy (OA)against different levels of noise (a) For dataset I,(b) For dataset II,(c) For dataset III 5 Conclusion In this paper, a novel technique based on Otsu's thresholding of local neighbourhood information is proposed. In the proposed method, inter-block information or local information along with the centre pixels of both images are used to calculate the threshold for each pixel position. For detecting the pixel as background or foreground, the thresholded pixels of both the bitemporal images are compared using the XOR operation. Then, on the basis of the majority class concept binary change map is created. The proposed technique uses the information of all the spectral bands present in the multispectral images and reduces the false alarm rate as it utilises local neighbourhood information and inter-image based thresholding technique. The proposed technique is simple and effective, and provides less false alarm rate with better accuracy and improved kappa value for all the datasets. Experiments are conducted on three Landsat image datasets and results show that the performance of the proposed technique is better compared to earlier reported techniques such as EM, KKM, BDLM and GaborTLC methods. Moreover, experimental results for noisy data show that the proposed technique performs better against noise. 6 Acknowledgment The authors thank M. Volpi for sharing their source codes for comparison studies. This Publication is an outcome of the R&D work undertaken in the project under the Visvesvaraya PhD Scheme of Ministry of Electronics & Information Technology, Government of India, being implemented by Digital India Corporation (formerly Media Lab Asia) [grant number PhD-MLA/4(13)/2015-16]. 7 References 1Singh A.: ' Review article digital change detection techniques using remotely-sensed data', Int. J. Remote Sens., 1989, 10, (6), pp. 989– 1003 Crossref Web of Science®Google Scholar 2Lu D., Mausel P., and Brondizio E. et al: ' Change detection techniques', Int. J. Remote Sens., 2004, 25, (12), pp. 2365– 2401 Crossref Web of Science®Google Scholar 3Yavariabdi A., and Kusetogullari H.: ' Change detection in multispectral landsat images using multiobjective evolutionary algorithm', IEEE Geosci. Remote Sens. Lett., 2017, 14, (3), pp. 414– 418 Crossref Web of Science®Google Scholar 4Bruzzone L., and Prieto D.F.: ' Automatic analysis of the difference image for unsupervised change detection', IEEE Trans. Geosci. Remote Sens., 2000, 38, (3), pp. 1171– 1182 Crossref Web of Science®Google Scholar 5Zanetti M., Bovolo F., and Bruzzone L.: ' Rayleigh-rice mixture parameter estimation via EM algorithm for change detection in multispectral images', IEEE Trans. Image Process., 2015, 24, (12), pp. 5004– 5016 Crossref PubMed Web of Science®Google Scholar 6Sumaiya M.N., and Kumari R.S.S.: ' Logarithmic mean-based thresholding for SAR image change detection', IEEE Geosci. Remote Sens. Lett., 2016, 13, (11), pp. 1726– 1728 Crossref Web of Science®Google Scholar 7Volpi M., Tuia D., and Camps-Valls G. et al: ' Unsupervised change detection with kernels', IEEE Geosci. Remote Sens. Lett., 2012, 9, (6), pp. 1026– 1030 Crossref Web of Science®Google Scholar 8Jia L., Li M., and Zhang P. et al: ' SAR image change detection based on multiple kernel k-means clustering with local-neighborhood information', IEEE Geosci. Remote Sens. Lett., 2016, 13, (6), pp. 856– 860 Crossref Web of Science®Google Scholar 9Jia L., Li M., and Zhang P. et al: ' Remote-sensing image change detection with fusion of multiple wavelet kernels', IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., 2016, 9, (8), pp. 3405– 3418 Crossref Web of Science®Google Scholar 10Li H.C., Celik T., and Longbotham N. et al: ' Gabor feature based unsupervised change detection of multitemporal SAR images based on two-level clustering', IEEE Geosci. Remote Sens. Lett., 2015, 12, (12), pp. 2458– 2462 Crossref Web of Science®Google Scholar 11Radoi A., and Datcu M.: ' Automatic change analysis in satellite images using binary descriptors and Lloyd–max quantization', IEEE Geosci. Remote Sens. Lett., 2015, 12, (6), pp. 1223– 1227 Crossref Web of Science®Google Scholar 12Weismiller R., Kristof S., and Scholz D. et al: ' Change detection in coastal zone environments. [by Landsat MSS data analysis]', 1977 Google Scholar 13Gao F., Dong J., and Li B. et al: ' Automatic change detection in synthetic aperture radar images based on PCANet', IEEE Geosci. Remote Sens. Lett., 2016, 13, (12), pp. 1792– 1796 Crossref Web of Science®Google Scholar 14Hussain M., Chen D., and Cheng A. et al: ' Change detection from remotely sensed images: from pixel-based to object-based approaches', ISPRS J. Photogramm. Remote Sens., 2013, 80, pp. 91– 106 Crossref Web of Science®Google Scholar 15Tewkesbury A.P., Comber A.J., and Tate N.J. et al: ' A critical synthesis of remotely sensed optical image change detection techniques', Remote Sens. Environ., 2015, 160, pp. 1– 14 Crossref Web of Science®Google Scholar 16Chen G., Hay G.J., and Carvalho L.M. et al: ' Object-based change detection', Int. J. Remote Sens., 2012, 33, (14), pp. 4434– 4457 Crossref Web of Science®Google Scholar 17Lu M., Chen J., and Tang H. et al: ' Land cover change detection by integrating object-based data blending model of Landsat and MODIS', Remote Sens. Environ., 2016, 184, pp. 374– 386 Crossref Web of Science®Google Scholar 18tong Yang X., Liu H., and Gao X.: ' Land cover changed object detection in remote sensing data with medium spatial resolution', Int. J. Appl. Earth Obs. Geoinf., 2015, 38, pp. 129– 137 Crossref CAS Google Scholar 19Leichtle T., Geiß C., and Wurm M. et al: ' Unsupervised change detection in VHR remote sensing imagery–an object-based clustering approach in a dynamic urban environment', Int. J. Appl. Earth Obs. Geoinf., 2017, 54, pp. 15– 27 Crossref Google Scholar 20Wang B., Choi S., and Byun Y. et al: ' Object-based change detection of very high resolution satellite imagery using the cross-sharpening of multitemporal data', IEEE Geosci. Remote Sens. Lett., 2015, 12, (5), pp. 1151– 1155 Crossref Web of Science®Google Scholar 21Otsu N.: ' A threshold selection method from gray-level histograms', IEEE Trans. Syst. Man Cybern., 1979, 9, (1), pp. 62– 66 Crossref Web of Science®Google Scholar 22Chaki N., Shaikh S.H., and Saeed K.: ' Exploring image binarization techniques' ( Springer, New Delhi, 2014) Crossref Google Scholar 23Available at http://earthexplorer.usgs.gov/ Google Scholar 24Huang X., Lu Q., and Zhang L. et al: ' New postprocessing methods for remote sensing image classification: a systematic study', IEEE Trans. Geosci. Remote Sens., 2014, 52, (11), pp. 7140– 7159 Crossref Web of Science®Google Scholar 25Celik T.: ' Change detection in satellite images using a genetic algorithm approach', IEEE Geosci. Remote Sens. Lett., 2010, 7, (2), pp. 386– 390 Crossref Web of Science®Google Scholar 26Banko G.: ' A review of assessing the accuracy of classifications of remotely sensed data and of methods including remote sensing data in forest inventory'. Interim Rep, Int Inst Appl Syst Anal, Vienna, Austria, 1998 Web of Science®Google Scholar Citing Literature Volume12, Issue11November 2018Pages 2051-2058 FiguresReferencesRelatedInformation
Referência(s)