A Novel Change Detection Approach for Multi-Temporal High ... - MDPI

remote sensing Article

A Novel Change Detection Approach for Multi-Temporal High-Resolution Remote Sensing Images Based on Rotation Forest and Coarse-to-Fine Uncertainty Analyses Wenqing Feng 1 ID , Haigang Sui 1, *, Jihui Tu 2 , Weiming Huang 3 and Kaimin Sun 1 ID 1

2 3

*

ID

, Chuan Xu 1

State Key Laboratory of Information Engineering in Surveying Mapping and Remote Sensing, Wuhan University, Wuhan 430079, China; [email protected] (W.F.); [email protected] (C.X.); [email protected] (K.S.) Electronics & Information School of Yangtze University, Yangtze University, Jingzhou 434023, China; [email protected] Department of Physical Geography and Ecosystem Science, Lund University, 22362 Lund, Sweden; [email protected] Correspondence: [email protected]; Tel.: +86-27-6877-8876

Received: 13 May 2018; Accepted: 20 June 2018; Published: 25 June 2018







Abstract: In the process of object-based change detection (OBCD), scale is a significant factor related to extraction and analyses of subsequent change data. To address this problem, this paper describes an object-based approach to urban area change detection (CD) using rotation forest (RoF) and coarse-to-fine uncertainty analyses of multi-temporal high-resolution remote sensing images. First, highly homogeneous objects with consistent spatial positions are identified through vector-raster integration and multi-scale fine segmentation. The multi-temporal images are stacked and segmented under the constraints of a historical land use vector map using a series of optimal segmentation scales, ranging from coarse to fine. Second, neighborhood correlation image analyses are performed to highlight pixels with high probabilities of being changed or unchanged, which can be used as a prerequisite for object-based analyses. Third, based on the coarse-to-fine segmentation and pixel-based pre-classification results, change possibilities are calculated for various objects. Furthermore, changed and unchanged objects identified at different scales are automatically selected to serve as training samples. The spectral and texture features of each object are extracted. Finally, uncertain objects are classified using the RoF classifier. Multi-scale classification results are combined using a majority voting rule to generate the final CD results. In experiments using two pairs of real high-resolution remote sensing datasets, our proposed approach outperformed existing methods in terms of CD accuracy, verifying its feasibility and effectiveness. Keywords: change detection; segmentation; scale; features; neighborhood correlation image analyses; rotation forest; majority voting

1. Introduction Change detection (CD) is an important research topic involving quantitative analyses of multi-temporal remotely sensed images to investigate changes in land cover, particularly in the contexts of urban infrastructure monitoring, urban development, and disaster assessment [1]. Along with the rapid development of remotely sensed image acquisition systems and gradual shortening of the

Remote Sens. 2018, 10, 1015; doi:10.3390/rs10071015

www.mdpi.com/journal/remotesensing

Remote Sens. 2018, 10, 1015

2 of 22

acquisition cycle, the application of CD is becoming increasingly widespread, and demand for such analyses is expanding. This demand increases the requirements and challenges for CD technology. Over the past few years, a number of methods and effective models have been proposed [2–10]. Pixel-based change detection (PBCD) methods have been widely used in low- and medium-resolution images [2]. These traditional algorithms utilize a single pixel as the analyses unit and are unable to incorporate spatial relationships with neighboring pixels, leading to isolation and the “salt and pepper” phenomenon [1,2]. To overcome this problem, some scholars have employed the level set method and its improved algorithm [3,4], Markov Random Field (MRF) [5,6], Conditional Random Field (CRF) [7,8], and other methods [9,10], all of which reduce the uncertainty of CD by effectively combining spectral and spatial information. However, the quality of the control parameters for those methods seriously affects the accuracy of the final result. Parameter setting reduces the universality of those methods. If the spatial neighborhood relationship is not defined accurately enough, a smooth transition of edge details may occur, resulting in false detections or missed detections. With improved resolution, the internal spectral difference between a particular pair of similar features gradually increases. Automated CD technology based on pixel spectral statistics cannot satisfactorily extract change information, and this problem is the main obstacle to widespread application of high-resolution remotely sensed images. The emergence of object-oriented technology for high-resolution remote sensing image analyses provides a new paradigm in that the basic unit of CD is changed from pixel to object [11]. Because the object-based change detection (OBCD) approach offers great advantages over the PBCD approach, it has received extensive attention and has been developed in recent years [12–17]. An object is defined as a single homogeneous region with similar spatial and spectral properties. Each object has characteristic features, such as its spectrum, shape, texture, and context. We can take full advantage of spectral features in combination with other features to improve accuracy of the CD process. To this end, the most commonly used methods are object-based (OB) change vector analyses (OCVA), OB correlation coefficient (OCC), and OB chi-square transformation (OCST) [18–22]. These methods perform a direct comparison of the two multispectral images under consideration. These unsupervised techniques leverage various features of the object, and incorporate them into the analyses at a later stage. These methods are suitable for some applications, such as the detection of deforestation or land-cover changes. The performance of these methods, however, relies heavily on the quality of feature selection, the allocation of feature weights, and the determination of the change threshold. Moreover, due to the difficulty of determining the segmentation scale, these methods are likely to introduce uncertainty into the CD process, thus reducing the reliability of the results. Although these methods can have good results, they are usually focused on one scale, which may not work well in more heterogeneous areas [23]. Because the landscape usually consists of different land cover types of varying size, it may not be possible to properly segment all features in a scene using a single scale [23]. To achieve better results, the segmentation scale, feature extraction, the change threshold, and many other factors must be considered. Development of the OBCD method has made tremendous progress to date. However, the results rely heavily on the determination of the segmentation scale [24]. As objects of interest exist within a scene, often within a range of different sizes, shapes and times, no single spatial resolution is sufficient to capture all their characteristics [24]. Methods to objectively evaluate segmentation results, obtain the optimal segmentation scale, and avoid the influence of visual interpretation are increasingly important. Determination of the optimal segmentation scale is directly related to subsequent extraction and analyses of change information [25]. At present, the scale parameters can only be selected based on visual interpretation or evaluation of multi-scale segmentation results, i.e., using post-segmentation evaluation to optimize scale selection [12,14,15]. With these methods, multi-scale determination is based on experience or repeated experiments. Scale selection is usually based on visual interpretation, which is inevitably subjective. Because the scope of visual observation is limited, it is difficult to obtain optimal segmentation results. In addition, some researchers have

Remote Sens. 2018, 10, 1015

3 of 22

Remote Sens. 2018, 10, x FOR PEER REVIEW

3 of 23

utilized existing geographic information system (GIS) databases and prior knowledge to accomplish high-resolution remote sensing image CDremote [26–30]. Post-classification with multi-temporal remote knowledge to accomplish high-resolution sensing image CD [26–30]. Post-classification with sensing images is one of the most popular change detection methods, providing the detailed “from-to” multi-temporal remote sensing images is one of the most popular change detection methods, change information in real applications [31,32]. However, due applications to the fact that it neglects the temporal providing the detailed “from-to” change information in real [31,32]. However, due to correlation between corresponding pixels in multi-temporal images, the post-classification approach the fact that it neglects the temporal correlation between corresponding pixels in multi-temporal usually suffers from an accumulation of misclassification errors In order to of solve this problem, images, the post-classification approach usually suffers from an[32]. accumulation misclassification some researchers have combined historical land use vector maps (HVMs) with remote sensing images errors [32]. In order to solve this problem, some researchers have combined historical land use to create a classification system, decision rules,toand used existing GIS knowledge to aid therules, CD vector maps (HVMs) with remoteset sensing images create a classification system, set decision process [26,28–30]. These studies use GIS and remote sensing tools to monitor land use and land cover and used existing GIS knowledge to aid the CD process [26,28–30]. These studies use GIS and changessensing at different The GIScover and remote toolsspatiotemporal can provide scientific remote toolsspatiotemporal to monitor landscales. use and land changessensing at different scales. procedures to analyze the pattern, rate, and trend of environmental change at all scales. Generally The GIS and remote sensing tools can provide scientific procedures to analyze the pattern, rate, and speaking, there are two change types ofatchange thatGenerally can affectspeaking, polygons there in HVMs, global change and local trend of environmental all scales. are two types of change that change. Global changes are changes to the overall category of the target polygon (e.g., region can affect polygons in HVMs, global change and local change. Global changes are changes Btofrom the impervious surface to building, and region C from farmland to impervious surface in Figure 1a,b), overall category of the target polygon (e.g., region B from impervious surface to building, and while local changes are those within a surface target polygon’s coverage (e.g., part of region A from bare region C from farmland to impervious in Figure 1a,b), while local changes are those within to impervious surface, and 1a,b). aland target polygon’s coverage (e.g.,part partofofregion regionDAfrom fromsettlement bare land to to impervious impervious surface surface,Figure and part of To effectively detect areas where local changes have occurred in an HVM, a remote sensing image region D from settlement to impervious surface Figure 1a,b). To effectively detect areas where local needs to be segmented constraints of thesensing HVM toimage separate the area a local change. changes have occurredunder in antheHVM, a remote needs to beexhibiting segmented under the Vector-rasterofintegration avoid manual settingaoflocal parameters segmentation partition constraints the HVM methods to separate thethe area exhibiting change.for Vector-raster integration and each avoid objectthe is the smallest unitofofparameters the land use Nonetheless, these methods methods manual setting forCD. segmentation partition and each have objectseveral is the drawbacks. HVMs images must havethese strict methods geometrichave registration, because theFirst, registration smallest unitFirst, of the land and use CD. Nonetheless, several drawbacks. HVMs errorimages can affect thehave CD result. Second, anregistration, HVM is produced according to the national census geography and must strict geometric because the registration error can affect the CD standard of the country being analyzed. It often reflects land use based on classification result. Second, an HVM is produced according to the national census geography standardcriteria, of the while remote images reflect the actual coverage. Transformation is necessary before country beingsensing analyzed. It often reflects land land use based on classification criteria, while remote integrating thesereflect two data Third,coverage. objects acquired through HVM are usually heterogeneous, sensing images thesources. actual land Transformation is necessary before integrating and the classification of these objects is generally complex. These flaws are likely to cause these two data sources. Third, objects acquired through HVM are usually heterogeneous, andpoor the CD results. classification of these objects is generally complex. These flaws are likely to cause poor CD results.

Figure image Figure 1. 1. Examples Examples of of global global and and local local change change with with regard regard to to the the polygons polygons in in an an HVM. HVM. (a) (a) T1 T1 image overlaid with an HVM; (b) T2 image overlaid with an HVM. overlaid with an HVM; (b) T2 image overlaid with an HVM.

To alleviate the aforementioned drawbacks, this paper incorporates the guidance of historical To alleviate the aforementioned drawbacks, this paper incorporates the guidance of historical interpretation into the OBCD process from multi-temporal high-resolution remote sensing images. interpretation into the OBCD process from multi-temporal high-resolution remote sensing images. Accordingly, an OBCD approach that combines pixel-based pre-classification of high-resolution Accordingly, an OBCD approach that combines pixel-based pre-classification of high-resolution remote sensing images using the rotation forest (RoF) classifier and HVM is proposed in this paper. remote sensing images using the rotation forest (RoF) classifier and HVM is proposed in this paper. The purpose of combining two approaches is to obtain better results [16,33–35]. In general, there are The purpose of combining two approaches is to obtain better results [16,33–35]. In general, there are two ways to undertake a combined approach: The two methods may be executed in parallel and two ways to undertake a combined approach: The two methods may be executed in parallel and then integrated, or the two methods may be performed consecutively, allowing the results of one then integrated, or the two methods may be performed consecutively, allowing the results of one method to be used as the premise for the other. Both of these strategies are employed for the purpose of achieving better results. In this paper, we perform pixel-based and OB methods consecutively. For pixel-level analyses, neighborhood correlation image analysis (NCIA) is adopted to obtain pixel-level pre-classification

Remote Sens. 2018, 10, 1015

4 of 22

method to be used as the premise for the other. Both of these strategies are employed for the purpose of achieving better results. In this paper, we perform pixel-based and OB methods consecutively. For pixel-level analyses, neighborhood correlation image analysis (NCIA) is adopted to obtain pixel-level pre-classification results, which are then used as a prerequisite for OB analyses. The HVM is used as a priori knowledge to guide selection of the optimal segmentation scales for various ground objects. RoF, a new type of machine-learning algorithm, performs better than many single predictors and integrated forecasting methods due to its high diversity of training samples and features. RoF is widely used to combine numerous weak classifiers to generate a relatively strong classifier by reducing either the bias or the variance of the individual classifiers [23,36–38]. In this paper, we combine pixel-based and OB methods with the RoF model to analyze the influences of segmentation scale, sample selection, and feature extraction on final results. The main contributions of this paper are as follows. First, we propose a pre-classification scheme based on NCIA to obtain high-accuracy labeled samples. Second, we present a coarse-to-fine uncertainty analysis model based on the combination of RoF and HVM. In this analysis, the optimal image segmentation result is obtained under the guidance of historical knowledge interpretation. RoF, combining feature extraction and classifier ensembles, is applied to binary classification. Third, an effective majority voting (MV) rule is proposed to produce the final results. The rest of this paper is organized as follows. Section 2 details the proposed method. In Section 3, the experimental results on multiple datasets are presented to show the performance of the proposed method. Section 4 provides the discussion. Finally, we conclude the paper in Section 5. 2. Methodology Among CD techniques, OBCD methods have become the most popular. However, due to the decline of spectral separability and complications such as identifying the same spectrum in different objects and/or different spectra in the same object, OBCD remains a challenging issue for several reasons. First, multi-temporal image segmentation (MTIS) is the foundation and basis of OBCD, and it seriously restricts the accuracy of OBCD. Not only do many segmentation algorithms exist that affect the resulting object geometries, but many CD techniques can affect the final results [24]. Second, change information is usually extracted at a single scale, neglecting scale set constraints. Third, objects in MTIS results are treated independently, ignoring the effects of neighboring object interactions on the change property. To address these key problems, a multi-scale and multi-level CD strategy is adopted in this paper. The overall workflow of the proposed approach is shown in Figure 2. Four aspects are investigated to improve OBCD accuracy in terms of the OBCD process, namely MTIS, training sample selection, multi-feature extraction, and coarse-to-fine uncertainty analyses. (1) MTIS is conducted using vector-raster integration and coarse-to-fine segmentation. The superimposed image is divided into homogeneous objects with eCognition software (version 8.7) [39], facilitating subsequent CD experiments. (2) Changed and unchanged objects are selected based on coarse-to-fine segmentation and pixel-level pre-classification. Uncertain objects are further classified using the trained RoF model. (3) The multi-features of each object in the images are represented by spectral and texture information of homogeneous pixels. (4) Scale sets can be considered as a collection of image sequences of different scales. CD results from different scales are usually treated independently, and thus multi-scale fusion is implemented to combine different CD results using coarse-to-fine uncertainty analyses.

Remote Sens. 2018, 10, x FOR PEER REVIEW Remote Sens. 2018, 10, 1015 Remote Sens. 2018, 10, x FOR PEER REVIEW

5 of 23 5 of 22 5 of 23

Figure 2.2.Flowchart proposedapproach. approach. Figure Flowchart of of the the proposed Figure 2. Flowchart of the proposed approach. 2.1. MTIS and Estimation of Scale Parameters 2.1. MTIS and Estimation of Scale Parameters 2.1. MTIS and Estimation of Scale Parameters Multi-resolution segmentation(MRS) (MRS) [39], [39], which which isis embedded eCognition software, is Multi-resolution segmentation embeddedin in eCognition software, is employed for image segmentation. Acquisition of coarse-to-fine sequential images is the basis of Multi-resolution segmentation (MRS) [39], which is embedded in eCognition software, employed for image segmentation. Acquisition of coarse-to-fine sequential images is the basis of scale-driven The quality of segmentation affects thesequential accuracy ofimages object is featurebasis is employed for OBCD. image Acquisition seriously of coarse-to-fine scale-driven OBCD. Thesegmentation. quality of segmentation seriously affects the accuracy of objectthe feature extraction and OBCD. Based on different image segmentation strategies, there are three primary of scale-driven OBCD. The quality of segmentation seriously affects the accuracy of object feature extraction and OBCD. Based on different image segmentation strategies, there are three primary modes of multi-temporal image segmentation at present [40,41], namely single-temporal extraction and OBCD. Based on different image segmentation strategies, there are three primary modes modes of multi-temporal image segmentation at present namely single-temporal segmentation (STS), multi-temporal separate segmentation (MTSS),[40,41], and multi-temporal combined of multi-temporal image segmentationseparate at present [40,41], namely single-temporal segmentation (STS), segmentation (STS), multi-temporal (MTSS), and multi-temporal combined segmentation (MTCS), as shown in Figure segmentation 3. Further details of the three primary modes of multi-temporal separate segmentation (MTSS), and multi-temporal combined segmentation (MTCS), segmentation (MTCS), shown in are Figure 3. Further details the and three primary modes of multi-temporal imageas segmentation described in Niemeyer et al.of(2008) Zhang et al. (2017). as shown in paper, Figure 3. Further details ofare the three primary modes ofetmulti-temporal image segmentation multi-temporal image segmentation described invia Niemeyer al.integration (2008) and Zhang et al. (2017). In this corresponding objects are extracted vector-raster and coarse-to-fine arethis described Niemeyer al.objects (2008) constraints and extracted Zhangofet HVM. al. this paper, corresponding objects are segmentation under theet boundary The In HVM, produced fromand thecoarse-to-fine National In paper,incorresponding are via(2017). vector-raster integration extracted via vector-raster integration and coarse-to-fine segmentation under the boundary constraints Census Geographic of China, is used as auxiliary data. The superimposed image is divided into segmentation under the boundary constraints of HVM. The HVM, produced from the National homogeneous objects with consistent spatial positions. of HVM.Geographic The HVM, produced the National Census Geographic of China, isimage used as Census of China,from is used as auxiliary data. The superimposed isauxiliary divided data. into The superimposed image is divided into homogeneous objects with consistent spatial positions. homogeneous objects with consistent spatial positions.

Figure 3. Multi-temporal image segmentation modes. (a) STS; (b) MTSS; (c) MTCS.

Estimation of scale parameter (ESP) [42,43] and the modified average segmentation evaluation Figure Multi-temporal segmentation modes. (a) MTSS; (c) MTCS. index (ASEI) are3. to guide theimage selection of the optimal segmentation Figure 3.used Multi-temporal image segmentation modes. (a) STS; STS; (b) (b)scales MTSS;for (c)different MTCS. ground objects. ESP, produced by Lucian et al. (2010), is a scale parameter estimation plug-in developed for Estimation of scale parameter (ESP) and the(LV) modified average segmentation eCognition software that calculates the[42,43] local variance of homogeneous image objects evaluation using Estimation of scale parameter (ESP) [42,43] and the modified average segmentation evaluation scaleused parameters. standard of the object layer is used to determine indexdifferent (ASEI) are to guideThe theaverage selection of thedeviation optimal segmentation scales for different ground indexwhether (ASEI)the aresegmentation used to guide the selection of the optimal segmentation scales for different scale is optimal. The rate of change in LV (ROC-LV) is used to indicate objects. ESP, produced by Lucian et al. (2010), is a scale parameter estimation plug-in developed for ground ESP, produced Lucian When et al. the (2010), is a scale estimation plug-in the objects. approximate scale by parameter. ROC-LV at itsparameter maximumimage value within eCognition software optimal that calculates the local variance (LV) ofis homogeneous objectsanusing developed for eCognition software that calculates the local variance (LV) of homogeneous image interval, the corresponding scale is the optimal scale. different scale parameters. The average standard deviation of the object layer is used to determine objects using different parameters. average standard deviation object layer used to However, this scale method can only beThe adapted to analyses of image data of in the a single band. Toistake whether the segmentation scale is optimal. The rate of change in LV (ROC-LV) is used to indicate full advantage of multi-spectral information, we utilize the modified ASEI to confirm the best scale determine whether the segmentation scale is optimal. The rate of change in LV (ROC-LV) is used to

the approximate optimal scale parameter. When the ROC-LV is at its maximum value within an indicate the approximate optimal scale parameter. When the ROC-LV is at its maximum value within interval, the corresponding scale is the optimal scale. an interval, the corresponding scale is the optimal scale. However, this method can only be adapted to analyses of image data in a single band. To take full advantage of multi-spectral information, we utilize the modified ASEI to confirm the best scale

Remote Sens. 2018, 10, 1015

6 of 22

However, this method can only be adapted to analyses of image data in a single band. To take full advantage of multi-spectral information, we utilize the modified ASEI to confirm the best scale parameter. The segmentation evaluation index (SEI) [44,45] in band L for an object is defined by Equation (1), where σL is the spectral standard deviation of the object, which is used as a homogeneity index. Its formula is shown in Equation (2), where ∆CL is the heterogeneity index derived from calculating the absolute value of the mean difference between the object and a neighbouring object. Its formula is shown in Equation (3), in which num is the total number of pixels within the object, CLi is the grey value of pixel i in band L, CL is the mean value of the object, l is the boundary length of the object, lk is the boundary length of the common edge with the kth adjacent object, and CLk is the mean value of the kth adjacent object: ∆CL (1) SEIL = σL s num 1 2 σL = (CLi − CL ) (2) num − 1 i∑ =1 ∆CL =

1 K lk CL − CLk ∑ l k =1

(3)

All of the formulas above are for single-band L only. For superimposed images (number of bands is 2N, where N is the band number of the single-phase image), eCognition applies different weights (w L ) to different bands to highlight the impacts of various bands on the segmentation results. Therefore, when considering a plurality of the bands for the SEI of any object, this paper modifies the calculation formula of SEI to Equation (4), and uses a weight for each band of w L = 1. This ensures that each band is equal in the segmentation process [42,43]: 2N

SEI =

∑ wL × SEIL

(4)

L =1

To compare objects at different segmentation scales, the SEI values of all objects in the study area can be averaged. Because the area (i.e., the number of pixels forming an image object) of the object may affect the segmentation results, this paper introduces the area factor, and gives a larger weight value to objects with larger areas to reduce the instability caused by object area. The modified ASEI is calculated using Equation (5): 1 Num ASEI = A j × SEIj (5) A i∑ =1 where A represents the total area of all objects, A j is the area of the j-th object, Num is the total number of objects, and SEIj is the SEI value of the j-th object. 2.2. Selection of Training Samples Neighbourhood correlation image analysis (NCIA) is a pixel-based method of CD, which is based on spectral contextual information created using correlation analyses between bi-temporal images within a specified neighborhood [46,47]. Further details of the NCIA method are described in Im et al. (2005). A rectangular (e.g., a moving window of 5 × 5 pixels) neighborhood type is used in this paper. After obtaining the difference image, the Otsu thresholding [48] algorithm is applied to classify the images into two types, changed and unchanged, and then the corresponding binary change images are generated. Sample selection is conducted based on the results of multi-scale segmentation

Remote Sens. 2018, 10, 1015

7 of 22

and pixel-level pre-classification. For the ith object Ri , the uncertainty index T is calculated based on pixel-level pre-classification results, as shown in the following Equation (6): ( T=

nc n

nc ≥ nu

− nnu

nc < nu

(6)

where nc , nu , and n are the numbers of changed, unchanged, and total pixels in object Ri , respectively. After setting the threshold Tm , we use the following equation to determine the properties of the object:

li =

   1 2   3

T < − Tm

− Tm ≤ T ≤ Tm

(7)

T > Tm

where li = 1, 2, 3 indicates the attributes of object Ri that are unchanged, uncertain, and changed, respectively, and the threshold range is Tm ∈ (0.5, 1). Changed and unchanged objects are selected as training samples for RoF, and uncertain objects are further classified. 2.3. Multi-Feature Information Extraction After sample selection, the key steps are feature extraction and analyzes. Due to the characteristics of high-resolution images, comprehensive consideration of ground objects is necessary, which is achieved by combining multiple features. It is necessary to preliminarily specify the features to be extracted according to the requirements of CD [49]. Prior to feature analyses, we filter or combine the extracted features. Finally, based on these features, change information is extracted. After fine multi-scale segmentation, eCognition can determine the features of objects by evaluating the image objects as well as their embedding in the image object hierarchy. In this paper, spectral features and texture features of objects at the same position are extracted from T1 and T2 phase images, respectively. Spectral features include the mean value, standard deviation, ratio (i.e., the amount that a given image layer contributes to the total brightness), maximum value (i.e., the value of the pixel with the maximum layer intensity value of the image object), and minimum value (i.e., the value of the pixel with the minimum layer intensity value in the image object). Haralick et al. [50] used the Grey-level Co-occurrence Matrix (GLCM) method to define 14 types of texture features. In this paper, the primary texture features are the mean value, standard deviation, contrast, entropy, homogeneity, correlation, angular second moment, and dissimilarity. The spectral and texture features are stacked, as shown in Figure 4. Table 1 lists the features extracted in this study. Each type of feature may reflect different object information from different angles, and often complement each other. After extracting the two types of features listed above, all features are normalised to the range of 0–1 [49] and combined as input RoF data for model training. Table 1. Overview of extracted features. Object Features

Feature Dimension

Tested Features (N Bands)

Spectral features

10 × N

Mean value, standard deviation, ratio, maximum value, minimum value

Texture features

16 × N

Mean value, standard deviation, contrast, entropy, homogeneity, correlation, angular second moment, and dissimilarity

Total feature dimension

26 × N

Remote Sens. 2018, 10, 1015

8 of 22

Remote Sens. 2018, 10, x FOR PEER REVIEW

8 of 23

Remote Sens. 2018, 10, x FOR PEER REVIEW

8 of 23

Figure 4. Multi-feature stacking scheme.

Figure 4. Multi-feature stacking scheme. 2.4. OBCD Based on RoF and Coarse-to-Fine Uncertaintystacking Analyses Figure 4. Multi-feature scheme.

2.4. OBCD Based on RoF and Coarse-to-Fine Uncertainty Analyses

For multi-scale image sequences, the coarse-scale object size is large, which is suitable for CD of

2.4. OBCD Based on RoF and Coarse-to-Fine Uncertainty objects with large areas. The fine-scale object size is Analyses small, andsize it isisadvantageous of small For multi-scale image sequences, the coarse-scale object large, whichfor is CD suitable for CD objects. Therefore, large objects at object a coarsesize scale, small objects a fine scale, and For multi-scale image sequences, the coarse-scale object size is large, is at suitable for CD of of objects with large detecting areas. The fine-scale is detecting small, and itwhich is advantageous for CD of then synthesising the object detection results at different scales is useful for improving the accuracy objects with large areas. The fine-scale object size is small, and it is advantageous for CD of small small objects. Therefore, detecting large objects at a coarse scale, detecting small objects at a fine scale, and reliability of the CD algorithm. Inspired by the scale, concept of “from coarse to fine, layerand by objects. Therefore, detecting objects at a coarse objects at a refine fine scale, and then synthesising the objectlarge detection results at differentdetecting scales issmall useful for improving the accuracy layer” [25], this paper proposes a coarse-to-fine uncertainty analysis method. The overall workflow then synthesising the object detection results at different scales is useful for improving the accuracy and reliability of the CD algorithm. Inspired by the concept of “from coarse to fine, refine layer by is shown in Figure 5, CD which includesInspired the following steps: and reliability of the algorithm. by thegeneral concept of “from coarse to fine, refine layer by layer” [25], this paper proposes a coarse-to-fine uncertainty analysis method. The overall workflow is layer” [25], this paper proposes a coarse-to-fine uncertainty analysis method. The overall workflow shown in Figure 5, which includes the following general steps: is shown in Figure 5, which includes the following general steps:

Figure 5. OBCD process incorporating rotation forest and coarse-to-fine uncertainty analyses.

Step 1: Uncertain object classification by RoF rotation forest and coarse-to-fine uncertainty analyses. Figure 5. OBCD process incorporating Figure 5. OBCD process incorporating rotation forest and coarse-to-fine uncertainty analyses. The RoF classifier method can successfully generate classifier ensembles based on feature Step 1: Uncertain by to RoFrandomly divide the original feature dataset into several extraction. This object paperclassification utilises RoF Step 1:subsets, Uncertain classification by carries out method feature transformation (e.g., principal component analysis, PCA) each The and RoFobject classifier canRoF successfully generate classifier ensembles based on for feature subset [23,36–38,51]. The correlations between transformed subsets are minimized. This work lays extraction. This paper utilises RoF to randomly divide the original feature dataset into several

The RoF classifier method can successfully generate classifier ensembles based on feature extraction. subsets, and carries out feature transformation (e.g., principal component PCA) forand each This paper utilises RoF to randomly divide the original feature dataset into analysis, several subsets, carries subset [23,36–38,51]. The correlations between transformed subsets are minimized. This work lays out feature transformation (e.g., principal component analysis, PCA) for each subset [23,36–38,51]. The correlations between transformed subsets are minimized. This work lays the foundation for

Remote Sens. 2018, 10,10, 1015 Remote Sens. 2018, x FOR PEER REVIEW

of 22 9 of923

the foundation for improving the accuracy of classification without changing the information in the

improving the accuracy of classification without changing the information in the original dataset original dataset by retaining all major components. The classification and regression tree (CART) by[23] retaining all major components. The classification and regression tree (CART) [23] is employed is employed as the base classifier because it is sensitive to rotation of the characteristic axis and as can the base classifier because it is sensitiveclassifier. to rotation of the characteristic axis andand can produce produce a strongly differentiating The decision tree is intuitive easy to a strongly differentiating classifier. The decision tree is intuitive and easy to understand. The training understand. The training of each base classifier is carried out in different subsets, significantly of each base classifier is carried outthe in different subsets, significantly increasing the difference withofthe increasing the difference with base classifier. This training helps to improve the accuracy base classifier. This training helps to improve the accuracy of prediction. prediction. Step 2: 2: Coarse-to-fine Step Coarse-to-fineimage imagefusion fusion Theproposed proposed approach approach makes thethe advantages of multi-scale imageimage analyses and The makesfull fulluse useof of advantages of multi-scale analyses expression, andand simultaneously usesuses multi-scale image layer object relationships to to fuse and expression, simultaneously multi-scale image layer object relationships fuse coarse-scaleCD CDresults resultswith withfine-scale fine-scale results. results. Finally, thethe final coarse-scale Finally, the theMV MVrule rule[23] [23]isisused usedtotoobtain obtain final results. CDCD results. 3. Experimentsand andResults Results 3. Experiments Dataset Description 3.1.3.1. Dataset Description verify feasibility effectiveness of proposed the proposed method, we apply the proposed ToTo verify thethe feasibility andand effectiveness of the method, we apply the proposed method method to two pairs of real high-resolution remote sensing datasets. The first experimental dataset to two pairs of real high-resolution remote sensing datasets. The first experimental dataset (DS1) (DS1) contains GF2 multi-spectral images6a,b) (Figure 6a,b) captured in 2015 2016 covering city contains GF2 multi-spectral images (Figure captured in 2015 and 2016and covering the city ofthe Liuzhou, of Liuzhou, China. GF2 is a Chinese satellite launched in August 2014, which provides China. GF2 is a Chinese satellite launched in August 2014, which provides high-resolution imagery imagery of the Earth. orbitisheight of and the its GF2 satellite isis 631 km,◦ .and of high-resolution the Earth. The orbit height of the GF2 The satellite 631 km, inclination 97.9080 Theits GF2 inclination is 97.9080°. The GF2 image comprises four spectral bands: Red (R), green (G), blue (B), image comprises four spectral bands: Red (R), green (G), blue (B), and near infrared, as well as one and near infrared, as well as one panchromatic band. The pan- and multi-spectral images are fused panchromatic band. The pan- and multi-spectral images are fused using the Pan-sharp algorithm [52] using the Pan-sharp algorithm [52] and the spatial resolution of the fused image is 0.8 m. The and the spatial resolution of the fused image is 0.8 m. The bi-temporal images used for experiments are bi-temporal images used for experiments are orthorectified and mainly include the three bands R, G, orthorectified and mainly include the three bands R, G, and B. The image contains 3749 × 3008 pixels. and B. The image contains 3749 × 3008 pixels. The vector data used in this study are an HVM of The vector data used in this study are an HVM of Liuzhou compiled in 2015 (Figure 6c). These data Liuzhou compiled in 2015 (Figure 6c). These data were produced through the National Census were produced theHVM National CensusofGeographic of are China. The after HVM and images The of the Geographic of through China. The and images the same area obtained preprocessing. same area are obtained after preprocessing. The vector data are projected using Transverse Mercator vector data are projected using Transverse Mercator projection, and the projection’s central projection, the projection’s central meridian is 111◦ E. The vector data contains 554 objects. meridianand is 111°E. The vector data contains 554 objects.

Figure 6. The first experimental dataset (DS1). (a) Image acquired on 13 April 2015; (b) image Figure 6. The first experimental dataset (DS1). (a) Image acquired on 13 April 2015; (b) image acquired acquired on 11 May 2016; (c) HVM compiled in 2015; (d) reference image. on 11 May 2016; (c) HVM compiled in 2015; (d) reference image.

The two high-resolution images in the second dataset (DS2) are also GF2 multi-spectral images The two high-resolution images inused the second dataset (DS2) areutilize also GF2 multi-spectral images (Figure 7a,b). The bi-temporal images for experiments mainly the three bands of R, G, (Figure The bi-temporal images used experiments utilizethan the three R, G, and B,7a,b). and contain 5314 × 4745 pixels. Thefor image areas aremainly much larger those bands of DS1ofand and B, and containof5314 4745types. pixels. image much larger than of (Figure DS1 and include a variety land×cover TheThe vector dataareas set forare land use contains 1540those objects 7c). From the bi-temporal images and HVMs, the main land cover types are divided into nine7c). include a variety of land cover types. The vector data set for land use contains 1540 objects (Figure categories, i.e., farmland, garden land, grassland, building, road, bare land,categories, water andi.e., From the bi-temporal imageswoodland, and HVMs, the main land cover types are divided into nine artificialwoodland, structure. The types aregrassland, in accordance withroad, the classification criteria the National farmland, garden land, building, bare land, water and of artificial structure. Census Geography of China. Among them, artificial structure mainly includes playground, The types are in accordance with the classification criteria of the National Census Geography of China. impervious surface and construction; building playground, is categorized into medium-story Among them, artificial structure mainly includes impervious surface andapartments, construction;

building is categorized into medium-story apartments, commercial building, industrial building,

Remote Sens. 2018, 10, 1015 Remote Sens. 2018, 10, x FOR PEER REVIEW

10 of 22 10 of 23

commercial industrial building, area; grassland is categorized settlement andbuilding, residential area; grassland is settlement categorizedand intoresidential artificial grassland and natural grassland; into artificial grassland and natural grassland; farmland is categorized into paddy field and dry farmland is categorized into paddy field and dry land; garden land is categorized into orchard, land; garden land is categorized into orchard, tea plantation and mulberry field. The reference tea plantation and mulberry field. The reference images shown in Figures 6d and 7d were produced images shown in Figures 6d and 7d were produced using the polygon to raster (PTR) tool built into using the polygon to raster (PTR) tool built into ESRI ArcGIS 10.1 software. Reference polygon ESRI ArcGIS 10.1 software. Reference polygon features were produced through the National features were produced through the National Census Geography of China. The black areas represent Census Geography of China. The black areas represent unchanged regions, while white regions unchanged regions, while white regions have changed. have changed.

Figure 7. The second experimental dataset (DS2). (a) Image acquired on 13 April 2015; (b) image Figure 7. The second experimental dataset (DS2). (a) Image acquired on 13 April 2015; (b) image acquired on 11 May 2016; (c) HVM in 2015; (d) reference image. acquired on 11 May 2016; (c) HVM in 2015; (d) reference image.

3.2. Evaluation Metrics

3.2. Evaluation Metrics

Evaluation of accuracy is essential to interpreting CD results and the final decision-making Evaluation of accuracy is essential to interpreting CDofresults and the final process. Four indices are used to evaluate the accuracy the final results [53]:decision-making process. Four used(FA): to evaluate the accuracy of the final [53]:incorrectly detected as having  indices False are alarms The number of unchanged pixelsresults that are changed, N FA ; the false alarm rate is defined as RFA  NFA / N0 100% , where N 0 is the total

•

False alarms (FA): The number of unchanged pixels that are incorrectly detected as having number of unchanged pixels. false The alarm rate is of defined as R NFAare /N0incorrectly × 100%, where N0 is total FA ; the(MA): FA = that  changed, Missed N alarms number changed pixels detected as the being number of unchanged pixels. unchanged, missed alarm rate is defined as RMA  NMA / N1 100% , where N1 is NMA ; the • Missed alarms (MA): The pixels. number of changed pixels that are incorrectly detected as the total number of changed unchanged, NMA ; the missed as the RMAoverall = NMA /N1 rate × 100%, where N1  being Overall error (OE): The total errorsalarm causedrate by is FAdefined and MA; alarm is calculated is the total number of changed pixels. as ROE  ( NFA  NMA ) / ( N0  N1 ) 100% . •  Overall error The total errors bymeasure FA and MA; the overall alarm ratewhich is calculated Kappa: The(OE): Kappa coefficient is a caused statistical of accuracy or agreement, reflects as ROE = ( N + N ) / ( N + N ) × 100%. 0 experimental FA MA 1 the consistency between results and ground truth data, and is expressed as • Kappa: coefficient is a statistical measure of accuracy or agreement, which reflects KappaThe  ( p0Kappa  pc ) / (1  pc ) , where p0 indicates the true consistency and pc indicates the thetheoretical consistency between experimental results and ground truth data, and is expressed as consistency. Kappa = ( p0 − pc )/(1 − pc ), where p0 indicates the true consistency and pc indicates the 3.3. Experimental Results and Analysis theoretical consistency.

To verify the feasibility and effectiveness of the proposed approach, two unsupervised pixel-based CD methods, i.e., PCA-k-means [10] and NCIA, as well as three OB methods are used verify the feasibility and effectiveness of the approach, two unsupervised pixel-based forTo comparison. The PCA-k-means method has twoproposed parameters, i.e., non-overlapping blocks H (H = in our experiments) and the[10] dimensions S (S 5 inasour experiments) of are the used eigenvector space. CD5 methods, i.e., PCA-k-means and NCIA, as=well three OB methods for comparison. OCVA [18] is employed the two unsupervised method, and extreme learning machine [54] The PCA-k-means methodashas parameters, i.e., non-overlapping blocks H (H(ELM) = 5 in our and random forest (RF) [55] are used as the supervised classifiers. The parameters involved in these experiments) and the dimensions S (S = 5 in our experiments) of the eigenvector space. OCVA [18] is methodsas are setunsupervised as in the original paper.and Theextreme implementation the ELM method available at Ref. employed the method, learning of machine (ELM) [54]isand random forest 54.[55] These are supervised applied to demonstrate the parameters advantages of the proposed approach. (RF) aremethods used as the classifiers. The involved in these methodsThe are CART set as in is adopted as the base classifier for RFmethod and RoF. The default number of treesmethods for RF isare thedecision originaltree paper. The implementation of the ELM is available at Ref. 54. These set to 100, while the number of decision trees for RoF is set to L = 50. For RoF, Xia et al. [23,38,51] applied to demonstrate the advantages of the proposed approach. The CART decision tree is adopted suggest that a small number of features per subset will increase the classification performance, as as the base classifier for RF and RoF. The default number of trees for RF is set to 100, while the number such, we set M = 10. The threshold of the uncertainty index is set to Tm = 0.55 for both experiments of decision trees for RoF is set to L = 50. For RoF, Xia et al. [23,38,51] suggest that a small number of after the comparisons.

3.3. Experimental Results and Analysis

features per subset will increase the classification performance, as such, we set M = 10. The threshold of the uncertainty index is set to Tm = 0.55 for both experiments after the comparisons.

Remote Sens. 2018, 10, 1015

11 of 22

3.3.1. Test of Scale Parameters Multi-scalar image segmentation is a fundamental step in OBCD. When using the ESP tool in DS1, the scale range is from 55–245 in steps of 5. The resulting LV and ROC-LV values vary with scale. As illustrated in Figure 8a, as the segmentation scale increases, LV values retain a trend of growth, Remote Sens. 2018, 10, x FOR PEER REVIEW 11 of 23 while ROC-LV values show the opposite trend. The change in LV from one level to another indicates 3.3.1. Test that of Scale Parameters how important scale level is for structuring information about object variability relative to the whole image. Theoretically, the peaks in ROC-LV curve the levels where LV in increases Multi-scalar image segmentation is an a fundamental step inindicate OBCD. When using the ESP tool as segments delineate their correspondents real worldLV [42]. Figurevalues 8a, the DS1, the scale range is from 55–245 in stepsin of the 5. The resulting and In ROC-LV varypeaks with in the Ascorresponding illustrated in Figure 8a, aslevels the segmentation retain a trend ROC-LVscale. curve to scale of 105, 115, scale 125, increases, 150, 175, LV 210values and potentially 240ofindicate growth, while ROC-LV values show the opposite trend. The change in LV from one level to another the meaningful scale parameters for segmentation of DS1. We selected the most obvious peaks, indicates how important that scale level is for structuring information about object variability which dominate their neighborhood, as indicators for optimal scale parameters. Such a scale is relative to the whole image. Theoretically, the peaks in an ROC-LV curve indicate the levels where generally alternative optimal Based oninsegmentation buildings, LVconsidered increases as an segments delineate their scale. correspondents the real worldresults, [42]. In industrial Figure 8a, the settlements residential better segmented a scale of 105 other peaksand in the ROC-LV areas curve are corresponding to scale at levels of 105, 115,than 125, at 150, 175, scales. 210 andFeatures 240 impervious indicate the meaningful parametersand for bare segmentation DS1. We selected the such as potentially playground, surfaces, scale construction, land areofbetter segmented at a scale most obvious peaks, which dominate their neighborhood, as indicators for optimal scale of 175, while woodlands and road areas are better segmented at 210. The object levels delineated parameters. Such a scale is generally considered an alternative optimal scale. Based on with these scale parameters matched the structures in real world for the DS1. These three scales are segmentation results, industrial buildings, settlements and residential areas are better segmented at considered suitable for rough image segmentation. Then these three scales (with floating of ±5) are a scale of 105 than at other scales. Features such as playground, impervious surfaces, construction, used to and set up theland corresponding ranges of 205–215.and This paper the modified bare are better segmented at 100–110, a scale of 170–180, 175, whileand woodlands road areasuses are better ASEI tosegmented further determine the specific optimal with scalethese within range matched (Figure the 9). structures The maximum at 210. The object levels delineated scaleeach parameters in real world for the9DS1. These179, threeand scales areinconsidered suitable for rough image values shown in Figure are 102, 213 panels a–c, respectively, whichsegmentation. can be considered Then these three scales (with floating of ±5) are used to set up the corresponding rangesoptimal of 100–110, the optimal segmentation scales of DS1. The mean object value maps at these three scales are 170–180, and 205–215. This paper uses the modified ASEI to further determine the specific optimal shown in Figure 10. scale within each range (Figure 9). The maximum values shown in Figure 9 are 102, 179, and 213 in Forpanels DS2,a–c, therespectively, scale range is from to 295 the in steps ofsegmentation 5. The maximum ROC-LV values are which can be105 considered optimal scales of DS1. The mean obtained at scales of 110, 140, 155, 165, 205, 245, and 275, as shown in Figure 8b. These values represent object value maps at these three optimal scales are shown in Figure 10. the meaningful scale are considered anmaximum alternative optimal scale For DS2, theparameters scale range isand from 105generally to 295 in steps of 5. The ROC-LV values are for the obtainedofatDS2. scales of 110, 165, 205, 245, and 275, as shown in Figure 8b.industrial These values segmentation Based on140, the 155, segmentation results, commercial buildings, buildings, represent meaningful scale are parameters and are generally considered an alternative optimal scale On the settlements, andthe residential areas segmented better at the scale of 140 than at other scales. for the segmentation of DS2. Based on the segmentation results, commercial buildings, industrial other hand, playgrounds, impervious surfaces, construction areas, and bare land are better segmented buildings, settlements, and residential areas are segmented better at the scale of 140 than at other at a scale of 205, while and roads are better segmented at 245. These three scales scales. On the othergrasslands hand, playgrounds, impervious surfaces, construction areas, and bare land are (with floatingbetter of ±5) are used to set the three corresponding ranges 135–145, 200–210, and 240–250. Then the segmented at a scale of 205, while grasslands and roads are better segmented at 245. These three scales (withtofloating of ±5)the arespecific used to optimal set the three corresponding ranges 135–145, modified ASEI is used determine scale for each range (Figure 11). 200–210, The maximum andFigure 240–250. ASEI is used to determine the specific optimal scale range values in 11Then are the 136,modified 206, and 244 in panels a–c, respectively, which canforbeeach considered the (Figure 11). The maximum values in Figure 11 are 136, 206, and 244 in panels a–c, respectively, optimal segmentation scales of DS2. The mean object value maps at these three optimal scales are which can be considered the optimal segmentation scales of DS2. The mean object value maps at shown in Figure 12. these three optimal scales are shown in Figure 12.

Figure 8. Curves showing that LV and ROC-LV vary as functions of scale. (a) DS1; (b) DS2.

Figure 8. Curves showing that LV and ROC-LV vary as functions of scale. (a) DS1; (b) DS2.

Remote Sens. 2018, 10, 1015

12 of 22

Remote Sens. 2018, 10, x FOR PEER REVIEW

12 of 23

Remote Sens. 2018, 10, x FOR PEER REVIEW

12 of 23

Remote Sens. 2018, 10, x FOR PEER REVIEW

12 of 23

Figure 9. Line chart of the ASEI index for DS1 at scale ranges of (a) 100 to 110, (b) 170 to 180, and (c)

Figure 9. Line chart of the ASEI index for DS1 at scale ranges of (a) 100 to 110, (b) 170 to 180, and (c) 205 205 to 215. Figure 9. Line chart of the ASEI index for DS1 at scale ranges of (a) 100 to 110, (b) 170 to 180, and (c) to 215. 205 to 215. Figure 9. Line chart of the ASEI index for DS1 at scale ranges of (a) 100 to 110, (b) 170 to 180, and (c) 205 to 215.

Figure 10. Mean object value maps for DS1 at various optimal scales. (a–c) Scales of 102, 179, and 213 for the acquired on 13 April scalesoptimal of 102, 179, and(a–c) 213 for the of image Figure 10. image Mean object value maps for2015; DS1 (d–f) at various scales. Scales 102, acquired 179, and Figure 10. Mean object value maps for DS1 at various optimal scales. (a–c) Scales of 102, 179, and 213 on 11for May 2016. 213 the image acquired on 13 April 2015; (d–f) scales of 102, 179, and 213 for the image acquired Figure 10. Mean object value maps for DS1 at various optimal scales. (a–c) Scales of 102, 179, and for the image acquired on 13 April 2015; (d–f) scales of 102, 179, and 213 for the image acquired on on 11 213 forMay the 2016. image acquired on 13 April 2015; (d–f) scales of 102, 179, and 213 for the image acquired 11 May 2016. on 11 May 2016.

Figure 11. Line chart of the ASEI index for DS2 at scale ranges of (a) 135 to 145, (b) 200 to 210, and (c) 240 to 250. Figure 11. Line chart of the ASEI index for DS2 at scale ranges of (a) 135 to 145, (b) 200 to 210, and (c) 240 to 250. Figure 11. Line chart of the ASEI index for DS2 at scale ranges of (a) 135 to 145, (b) 200 to 210, and (c)

Figure 11. Line chart of the ASEI index for DS2 at scale ranges of (a) 135 to 145, (b) 200 to 210, and (c) 240 240 to 250. to 250.

Remote Sens. 2018, 10, 1015

13 of 22

Remote Sens. 2018, 10, x FOR PEER REVIEW

Remote Sens. 2018, 10, x FOR PEER REVIEW

13 of 23

13 of 23

Figure 12. Mean object value maps for DS1 at different optimal scales. (a–c) Scales of 136, 206, and

Figure 12. Mean object value maps for DS1 at different optimal scales. (a–c) Scales of 136, 206, and 244 244 for the image acquired on 13 April 2015; (d–f) Scales of 136, 206, and 244 for the image acquired for the image acquired on 13 April 2015; (d–f) Scales of 136, 206, and 244 for the image acquired on on 11 May 2016. 11 May 2016. 3.3.2. Results DS1 object value maps for DS1 at different optimal scales. (a–c) Scales of 136, 206, and Figure for 12. Mean 244for for DS1 the image acquired on 13 April 2015; (d–f) Scales of 136, 206, and 244 for the image acquired 3.3.2. Results Our results are demonstrated in two ways. The CD results are displayed in figures and the on 11 May 2016. criteria are listed in tables. Figures 13 and 14 show results based on DS1.and Figure Our results are demonstrated in two ways. The the CDexperimental results are displayed in figures the 13 criteria shows CD results at the three optimal scales using RoF. The false detection phenomenon (the 3.3.2. DS1 13 and 14 show the experimental results based on DS1. Figure 13red are listed in Results tables. for Figures shows regions) is most serious at scale 102, whereas the missed detection (the yellow regions) is better at CD results at the three optimal scales using RoF. The false detection phenomenon (the red regions) Our results are demonstrated in two ways. The CD results are displayed in figures and the this scale than at the other two scales (scale 179 and scale 213) tested. As the scale increases, false tables. Figures 13 14 show the experimental resultsregions) based onis DS1. Figure is most criteria seriousareatlisted scalein102, whereas theand missed detection (the yellow better at 13 this scale detection is further reduced at scale 213, but missed detection is slightly more frequent than at the CD results at the(scale three optimal scales using RoF. The As falsethe detection phenomenon (thedetection red than other at shows thetwo other two scales 179 and scale 213) tested. scale increases, false is scales. This paper utilizes the MV approach to fuse the CD results at the three optimal regions) isatmost serious at scale 102, detection whereas the missed detection (the yellow regions) isother bettertwo at scales. further reduced scale 213, but missed is slightly more frequent than at the scales, which can effectively integrate and complement false detection or missed detection this scale than at the other two scales (scale 179 and scale 213) tested. As the scale increases, false This paper utilizes MV approach to fuse the CD results the three optimal which can phenomena atfurther a the single scale and improve accuracy the final at Figure showsscales, the results detection is reduced at scale 213, but missedofdetection isresult. slightly more14 frequent than at the of MVother using ELM, RF and RoF. Changed regions are concentrated mainly within strongly changed effectively integrate and complement false detection or missed detection phenomena at a single two scales. This paper utilizes the MV approach to fuse the CD results at the three optimal scale regions (change at final three result. scales) and obviously changed regions occurs at any two RoF. and improve the Figure 14 shows the results of(change MV ELM, RF and scales, accuracy which occurs canofeffectively integrate and complement false detection orusing missed detection scales), while subtly changed regions (change occurs only at one scale) are mostly false changes. phenomena at a concentrated single scale and mainly improve within accuracystrongly of the final result. Figure 14 shows the results Changed regions are changed regions (change occursofat three MV obviously using ELM,changed RF and RoF. Changed regions are concentrated mainly within stronglychanged changed regions scales) and regions (change occurs at any two scales), while subtly regions (change occurs at three scales) and obviously changed regions (change occurs at any two (change occurs only at one scale) are mostly false changes.

scales), while subtly changed regions (change occurs only at one scale) are mostly false changes.

Figure 13. CD results of DS1. (a) RoF (Scale 102); (b) RoF (Scale 179); (c) RoF (Scale 213). Figure 13. CD results of DS1. (a) RoF (Scale 102); (b) RoF (Scale 179); (c) RoF (Scale 213).

Figure 13. CD results of DS1. (a) RoF (Scale 102); (b) RoF (Scale 179); (c) RoF (Scale 213).

Remote Sens. 2018, 10, 1015 Remote Sens. 2018, 10, x FOR PEER REVIEW

14 of 22 14 of 23

Figure 14. CD results based on majority voting. (a) ELM-MV; (b) RF-MV; (c) RoF-MV.

Figure 14. CD results based on majority voting. (a) ELM-MV; (b) RF-MV; (c) RoF-MV. The proposed fusion strategy converts the traditional “hard detection” method into “soft detection”. Various detection results combined to attain stepwise reclassification the The proposed fusionscale strategy converts theare traditional “hard detection” method into “softof detection”. change intensity, which offers a more meaningful and flexible reference in practice than simply Various scale detection results are combined to attain stepwise reclassification of the change intensity, separating all pixels into two classes of changed and unchanged. strong changes all thatpixels which offers a more meaningful and flexible reference in practice Moreover, than simply separating occur at all three scales can be defined as the areas with maximum probability of change, while into two classes of changed and unchanged. Moreover, strong changes that occur at all three scales obvious changes that occur at two scales are viewed as having the second greatest probability. can be defined as the areas with maximum probability of change, while obvious changes that occur at Finally, the strongly and obviously changed regions are considered changed areas, whereas the two scales are viewed as having the second greatest probability. Finally, the strongly and obviously subtly changed and unchanged regions are treated as unchanged areas. changed regions are the considered changed areas, whereas theshown subtly unchanged regions Table 2 lists values used for evaluation metrics. As in changed Figure 15,and the two pixel-based are treated as are unchanged areas. methods filled many white spots due to noise. The CD accuracy of pixel-based methods is Table 2 lists the values used for evaluation metrics.features As shown in Figure 15, the two pixel-based comparatively low, mainly because only the spectral of bi-temporal images are used in these methods. Furthermore, due to the high resolution of the original images, pixel-based CD methods are filled many white spots due to noise. The CD accuracy of pixel-based methods is methods based spectralbecause statisticsonly are not to meet the requirements for change information comparatively low,onmainly theable spectral features of bi-temporal images are used in the results due obtained poor. The OCVA achieves object-level CD theseextraction methods. and Furthermore, to theare high resolution of themethod original images,better pixel-based CD methods results than the two pixel-based methods. Because the unsupervised CD method is strongly based on spectral statistics are not able to meet the requirements for change information extraction affected by the threshold, OCVA results are worse than those of the supervised RoF at the three and the results obtained are poor. The OCVA method achieves better object-level CD results than the optimal scales. The dominant change combinations include those between settlements and two pixel-based methods. Because the unsupervised CD method is strongly affected by the threshold, impervious surfaces, and between bare land and buildings. The main driver of these changes is OCVA results are than thoseinofLizhou the supervised RoF at thetothree optimal scales. The dominant urban growth worse and construction in recent years. Due the limitations of OCVA, some change combinations include those between settlements and impervious surfaces, and between changed objects are falsely detected as unchanged, while some unchanged objects are falsely bare land identified and buildings. The main these changes is urban growththrough and construction in Lizhou in as changed. Thedriver scale of constraints are fully considered implementation of uncertainty analyses. amounts uncertain data are each as scale. CD recentcoarse-to-fine years. Due to the limitations of The OCVA, someofchanged objects arereduced falsely with detected unchanged, the three optimal are combined to generate final results MV approach. whileresults someatunchanged objectsscales are falsely identified as changed. Theusing scalethe constraints are fully Compared to the RF-MV and ELM-MV methods, the kappa coefficient of RoF-MV is highest among considered through implementation of coarse-to-fine uncertainty analyses. The amounts of uncertain the three supervised classifiers. Furthermore, it should be noted that although image segmentation data are reduced with each scale. CD results at the three optimal scales are combined to generate may impair results near borders, the proposed approach not only improves CD performance but final results using the MV approach. Compared to the RF-MV and ELM-MV methods, the kappa also enhances the image boundary compared to results at a single scale. In short, the proposed coefficient of RoF-MV highest among the three supervised Furthermore, approach obtains theisbest results and greater accuracy comparedclassifiers. to other methods tested. it should be noted that although image segmentation may impair results near borders, the proposed approach not only improves CD performance but also enhances the image boundary compared to results at a single scale. In short, the proposed approach obtains the best results and greater accuracy compared to other methods tested.

Remote Sens. 2018, 10, 1015

15 of 22

Remote Sens. 2018, 10, x FOR PEER REVIEW

15 of 23

Figure 15. Comparison of different methods for DS1. (a) NCIA; (b) PCA-k-means; (c) OCVA; (d)

Figure 15. Comparison of different methods for DS1. (a) NCIA; (b) PCA-k-means; (c) OCVA; RoF-scale 102; (e) RoF-scale 179; (f) RoF-scale 213; (g) ELM-MV; (h) RF-MV; (i) RoF-MV. (d) RoF-scale 102; (e) RoF-scale 179; (f) RoF-scale 213; (g) ELM-MV; (h) RF-MV; (i) RoF-MV. Table 2. Performance comparison of different approaches for DS1.

Table 2. Performance comparison of different approaches for DS1.

Pixel-Based Method Pixel-Based Accuracy NCIA PCA-k-Means OCVA Method

NCIA Accuracy

FA (%)

PCA-k-Means

12.27

16.01

FA (%) 12.27 16.01 MA (%) 34.93 36.77 MA (%) 34.93 36.77 OE (%) OE (%) 13.40 13.40 17.04 17.04 Kappa (%)27.17 27.17 20.78 20.78 Kappa (%)

Object-Based RoF RoF RF ELM Object-Based Scale Scale 102 RoF Scale 179 MVRoF MV RF MV OCVA 213 Scale 102 Scale 179 Scale 213 MV MV 4.43 4.82 3.81 2.99 3.01 3.21 2.37 4.43 4.82 3.81 2.99 3.01 3.21 31.77 28.25 23.11 31.75 26.55 26.74 34.84 31.77 28.25 23.11 31.75 26.55 26.74 4.53 4.37 4.37 3.99 5.79 5.79 5.99 5.99 4.574.57 4.53 4.18 4.18 59.28 58.82 60.27 60.27 59.86 51.07 51.07 51.4151.41 59.28 58.82 61.4961.49

ELM MV 2.37 34.84 3.99 59.86

3.3.3. Results for DS2

3.3.3. Results for DS2

Figures 16 and 17 show the experimental results obtained for DS2. As shown in Figure 16,

Figures 16three and optimal 17 show the the experimental results obtained for As isshown in Figure among the scales, false detection phenomenon (the redDS2. regions) more serious at 16, among the136 three optimal thescales false detection red other regions) is more serious at scale than at the scales, other two (scale 206 phenomenon and scale 244);(the on the hand, the missed scale detection 136 than at the other two scales (scale 206 and scale 244); on the other hand, the missed detection phenomenon (the yellow regions) is better at that scale. As the scale increases, false phenomenon (the yellow regions) is better at that scale. As the scale increases, false detection is detection is further reduced at the scale of 244, but missed detection is slightly higher than at the other two scales. these scales of DS2 is can compensate each in the CDscales. further reduced at theTherefore, scale of 244, butthree missed detection slightly higherfor than at other the other two process. Figure 17scales showsofthe results of ELM, RF,for andeach RoFother usinginthe Strong 17 and Therefore, these three DS2 can compensate theMV CDapproach. process. Figure shows obvious changes can be set as the areas with a maximum probability of change, while subtle the results of ELM, RF, and RoF using the MV approach. Strong and obvious changes can be set changes are viewed as having secondary importance. as the areas with a maximum probability of change, while subtle changes are viewed as having secondary importance.

Remote Sens. 2018, 10, 1015

16 of 22

Remote Sens. 2018, 10, x FOR PEER REVIEW

16 of 23

Remote Sens. 2018, 10, x FOR PEER REVIEW

16 of 23

Figure 16. CD results for DS2. (a) RoF (Scale 136); (b) RoF (Scale 206); (c) RoF (Scale 244).

Figure 16. CD results for DS2. (a) RoF (Scale 136); (b) RoF (Scale 206); (c) RoF (Scale 244). Figure 16. CD results for DS2. (a) RoF (Scale 136); (b) RoF (Scale 206); (c) RoF (Scale 244).

Figure 17. CD results obtained by majority voting. (a) ELM-MV; (b) RF-MV; (c) RoF-MV. Figure 17. CD results obtained by majority voting. (a) ELM-MV; (b) RF-MV; (c) RoF-MV.

Figure 17. CD results obtained by OE, majority voting. (a) coefficient, ELM-MV; (b) (c)for RoF-MV. Four indicators, namely, FA, MA, and the kappa areRF-MV; adopted quantitative comparisons (Table 3).namely, As shown Figure OBCD methods (OCVA,for ELM, RF, and Four indicators, FA,in MA, OE, 18, andthe theresults kappaofcoefficient, are adopted quantitative RoF) indicated better performance than those of PBCD methods (NCIA and PCA-k-means). The comparisons (Table 3). As shown in Figure 18, the results of OBCD methods (OCVA, ELM, RF, and Four indicators, namely, FA, MA, OE, and the kappa coefficient, are adopted for quantitative NCIA method utilizes contextual information to make a final decision, and thus performs better RoF) indicated better performance than those of PBCD methods (NCIA and PCA-k-means). The RoF) comparisons (Table 3). As shown in Figure 18, the results of OBCD methods (OCVA, ELM, RF, and than the PCA-k-means method. The dominant land cover change combinations indicate changes NCIA method utilizes contextual information to make a final decision, and thus performs better indicated better performance than those of PBCD methods (NCIA and PCA-k-means). The NCIA between roads and impervious surfaces, bare land land cover and farmland, settlements indicate and impervious than the PCA-k-means method. The dominant change combinations changes method utilizes contextual information to make a final decision, and thus performs better than the surfaces, bare land and buildings. The unsupervised method achieved better CD result, between and roads and impervious surfaces, bare land andOCVA farmland, settlements and impervious PCA-k-means method. The dominant land cover change combinations indicate changes between roads due to theand usebare of segmented objects as the unit of analysis. thisachieved method,better changed surfaces, land and buildings. Thebasic unsupervised OCVA With method CDobjects result, and impervious surfaces, bare land and farmland, and impervious surfaces, and bare are The MV scheme obtains betterunit CDsettlements results thanWith a single scale, andchanged leads toobjects much duemore to theregular. use of segmented objects as the basic of analysis. this method, land more andmore buildings. unsupervised achieved CD result, due to the use of homogeneous regions shownobtains onOCVA the better CDmethod maps. Furthermore, the proposed RoF-MV method are regular.The The MV scheme CD results than abetter single scale, and leads to much segmented objects as the basic unit of analysis. With this method, changed objects are more regular. was superior to all other methods in terms of the kappa coefficient for DS2. Regardless of the more homogeneous regions shown on the CD maps. Furthermore, the proposed RoF-MV method computation time, it can be expected that RoF can surpass RF to some extent. The time requirement The MV bettermethods CD results than aofsingle scale,coefficient and leadsfor to DS2. muchRegardless more homogeneous wasscheme superiorobtains to all other in terms the kappa of the for ELM is on lower RF andFurthermore, RoFthat classifiers. However, final kappa coefficient of ELM isto all computation time, canmaps. be expected RoF can surpass RFthe toRoF-MV some extent. The time regions shown the itthan CD the proposed method wasrequirement superior these two ensemble learning methods. TheRegardless proposed clearly outperforms for ELMthan isinlower RF kappa and RoF classifiers. the finalapproach kappa coefficient of ELM othersmaller methods termsthan of the coefficient forHowever, DS2. of the computation time,isit can other comparison methods in both learning qualitative and quantitative analyses. smaller than these two ensemble methods. The proposed approach clearly outperforms be expected that RoF can surpass RF to some extent. The time requirement for ELM is lower than RF other comparison methods in both qualitative and quantitative analyses.

and RoF classifiers. However, the final kappa coefficient of ELM is smaller than these two ensemble learning methods. The proposed approach clearly outperforms other comparison methods in both qualitative and quantitative analyses.

Remote Sens. 2018, 10, 1015

17 of 22

Remote Sens. 2018, 10, x FOR PEER REVIEW

17 of 23

Figure 18. Comparison of different methods for DS2. (a) NCIA; (b) PCA-k-means; (c) OCVA; (d)

Figure 18. Comparison of different methods for DS2. (a) NCIA; (b) PCA-k-means; (c) OCVA; RoF-scale 136; (e) RoF-scale 206; (f) RoF-scale 244; (g) ELM-MV; (h) RF-MV; (i) RoF-MV. (d) RoF-scale 136; (e) RoF-scale 206; (f) RoF-scale 244; (g) ELM-MV; (h) RF-MV; (i) RoF-MV. Table 3. Performance comparison of different approaches for DS2.

Table 3. Performance comparison of different approaches for DS2.

Pixel-Based Method Pixel-Based Method Accuracy NCIA PCA-k-Means OCVA NCIA

PCA-k-Means

32.97

52.41 17.48 20.95 52.41 20.95 20.47 20.47

FA (%) 13.37 Accuracy MA (%) 13.3742.08 FA (%) MA (%) OE (%) 42.0816.22 OE (%)Kappa (%) 16.2232.97 Kappa (%)

17.48

OCVA

Object-Based

RoFObject-Based RoF RF ELM RoF RF Scale 136 Scale 206 Scale 244 MV RoF MV MV

4.69 Scale 5.61 136 43.82 5.61 9.41 43.82 9.41 49.08

53.71 4.69 9.57 53.71 9.57 43.78 43.78

49.08

3.74 206 Scale 47.23 3.74 8.07 47.23 8.07 52.12 52.12

3.72 3.49 Scale 2443.55 MV 46.58 45.09 3.72 42.28 3.55 7.98 7.63 46.58 6.67 42.28 7.98 56.72 6.67 52.71 54.69 52.71

56.72

4.18 MV 44.38 3.49 8.17 45.09 7.63 52.99

54.69

ELM MV 4.18 44.38 8.17 52.99

4. Discussion

4. Discussion

Scale is an important feature in the OBCD process. The accuracy of OBCD results depends

greatlyison quality of feature the MTISinand extraction methods. Contextual constraints link Scale anthe important theinformation OBCD process. The accuracy of OBCD results depends parent andquality child objects various scales. In this paper, highly homogenous objects withconstraints consistent link greatly on the of theatMTIS and information extraction methods. Contextual spatial positions are obtained using the MRS algorithm thehomogenous boundary constraints the consistent HVM. parent and child objects at various scales. In this paper,under highly objects of with The proposed approach, which is capable of delineating and analysing image ground objects at spatial positions are obtained using the MRS algorithm under the boundary constraints of the HVM. different scales based on coarse-to-fine uncertainty analyses, was successfully implemented for The proposed approach, which is capable of delineating and analysing image ground objects at different high-resolution remotely sensed images. The approach adopts RoF to classify uncertain regions of scales based on coarse-to-fine uncertainty analyses, was successfully implemented for high-resolution remotely sensed images. The approach adopts RoF to classify uncertain regions of the generated coarse-to-fine segmentation maps. Multiple CD results are combined to generate the final result

Remote Sens. 2018, 10, 1015 Remote Sens. 2018, 10, x FOR PEER REVIEW

18 of 22 18 of 23

using the MV approach. Experimental results demonstrated the effectiveness of multi-feature and coarse-to-fine image fusion in segmentation improving CD results. The proposed approach led to to acceptable the generated coarse-to-fine maps. Multiple CD results are combined generate levels the of efficiency and accuracy. Furthermore, analyses that the the accuracy is better of than final result using the MV approach.comparative Experimental resultsshowed demonstrated effectiveness other CD methods. results support proposed approach asresults. a highly efficient method to improve multi-feature andThe coarse-to-fine imagethe fusion in improving CD The proposed approach led to acceptable levelsareas of efficiency and accuracy. Furthermore, comparative analyses showed that the monitoring of urban using remotely sensed data. accuracy is betteristhan other methods. The results support as algorithms a highly Our approach based on CD a combination of pixel-based andthe OBproposed analyses. approach The hybrid efficient method to improve monitoringschemes of urbansuccessfully areas using remotely data. that combined pixeland object-based reducedsensed noisy changes, as well as the Ourspurious approach is based on a combination of pixel-based andofOB analyses. The quality hybrid of small and changes introduced by the inconsistent delineation objects. However, algorithms that combined pixeland object-based schemes successfully reduced noisy changes, as of the training samples is affected by the initial pixel-level pre-classification results. The threshold well as the small and spurious changes introduced by the inconsistent delineation of objects. uncertainty index Tm is an important parameter that can affect the final CD results. The performance However, quality threshold of the training is affected by experimental the initial pixel-level pre-classification varies with different values.samples In this paper, the two datasets dynamically adjust results. The threshold of uncertainty index Tm is an important parameter that can affect the final CD Tm within the interval (0.5, 1) at a step size of 0.1. As shown in Figure 19, among the three optimal results. The performance varies with different threshold values. In this paper, the two experimental scales for DS1, the false and overall alarm rates at scale 213 are lower than those of scale 102 and datasets dynamically adjust Tm within the interval (0.5, 1) at a step size of 0.1. As shown in Figure 179 as the threshold Tm value increases, whereas the missed alarm rate shows the opposite trend. 19, among the three optimal scales for DS1, the false and overall alarm rates at scale 213 are lower Similarly, when the threshold is within the interval (0.5, 0.85), the false and overall alarm rates at than those of scale 102 and 179 as the threshold Tm value increases, whereas the missed alarm rate scale 102 the are opposite higher than those at scales of 179 and 213, whereas rate the shows shows trend. Similarly, when the threshold is within the the missed interval alarm (0.5, 0.85), falsethe opposite pattern. When threshold T = 0.55, the missed alarm rates at the three optimal m and overall alarm rates at scale 102 are higher than those at scales of 179 and 213, whereas scales the aremissed lowest, which these three scalesWhen can compensate each in the CDrates process. alarm ratemeans showsthat the opposite pattern. threshold Tm for = 0.55, theother missed alarm at Therefore, false detection missed detection at a single can be reduced using the threethe optimal scales are or lowest, which meansphenomena that these three scales scale can compensate for each multi-scale fusion. Figure 20Therefore, shows thethe influence of this index on DS2. The false alarm rateatand overall other in the CD process. false detection or missed detection phenomena a single alarm 136 are higher than those at scale 206 244the as threshold Tmthis increases, scalerate canatbescale reduced using multi-scale fusion. Figure 20and shows influence of index onwhereas DS2. theThe missed rateand has overall the opposite trend. When the is in the at interval (0.55, false alarm alarm rate alarm rate at scale 136 arethreshold higher than those scale 206 and0.85), 244 asthe threshold Tm increases, whereas therate missed alarm rate the opposite theofthreshold false alarm rate and overall alarm at scale 244 arehas smaller those trend. ones atWhen scales 136 and is 206. in the interval (0.55, 0.85), the false alarm rate and overall alarm rate at scale 244 are smaller those When the threshold Tm = 0.55, the missed alarm rates are lowest at all three scales. Therefore, the ones at scales ofvalue 136 and the threshold m =multi-scale 0.55, the missed alarm rates are lowest at all optimal threshold is T206. 0.55 for DS2. Using T the propagation method, incorporation m = When Therefore, the optimal threshold value TmRoF = 0.55 for DS2. the multi-scale of three spatialscales. information at different scales by MRS intoisthe classifier canUsing significantly improve incorporation of spatialthe information at different scales by MRS into RoF thepropagation classifier’s method, performance, which indicates importance of spatial information. Thethe excellent classifier can significantly improve the classifier’s performance, which indicates the importance of performance of the proposed approach in our two experiments is mainly attributable to the proposed spatial information. The excellent performance of the proposed approach in our two experiments is CD scheme successfully taking advantage of spectral, texture, and spatial information at different mainly attributable to the proposed CD scheme successfully taking advantage of spectral, texture, scale levels. and spatial information at different scale levels. In the application of CD, we can combine the pixel-based pre-classification and object-based In the application of CD, we can combine the pixel-based pre-classification and object-based image analyses approaches for different purposes using RoF and HVM, to obtain final object-level image analyses approaches for different purposes using RoF and HVM, to obtain final object-level CD results. The changed objects are more regular, and the object geometries correspond to actual CD results. The changed objects are more regular, and the object geometries correspond to actual geographical features. Therefore, the combination not only exhibits the advantages of both pixel-based geographical features. Therefore, the combination not only exhibits the advantages of both and object-based also has the accuracy. pixel-based andapproaches, object-basedbut approaches, but greatest also has the greatest accuracy.

Figure 19. The influence of the uncertainty index threshold for DS1. (a) Missed alarm rate at various Figure 19. The influence of the uncertainty index threshold for DS1. (a) Missed alarm rate at scales; (b) False alarm rate at various scales; (c) Overall alarm rate at various scales. various scales; (b) False alarm rate at various scales; (c) Overall alarm rate at various scales.

Remote Sens. 2018, 10, 1015 Remote Sens. 2018, 10, x FOR PEER REVIEW

19 of 22 19 of 23

Figure 20. Influence of the uncertainty index threshold for DS2. (a) Missed alarm rate at various Figure 20. Influence of the uncertainty index threshold for DS2. (a) Missed alarm rate at various scales; scales; (b) False alarm rate at various scales; (c) Overall alarm rate at various scales. (b) False alarm rate at various scales; (c) Overall alarm rate at various scales.

5. Conclusions and Perspective

5. Conclusions and Perspective

Optimal scale selection has been a key issue affecting multi-scale segmentation of Optimal scaleremote selection has been a key issue affecting high-resolution high-resolution sensing images. Determination ofmulti-scale the optimalsegmentation segmentation of scale is directly remote sensing images. extraction Determination of theofoptimal relateda to related to subsequent and analysis change segmentation information. Inscale viewisofdirectly this problem, subsequent extraction and analysis of change information. In view of this problem, a novel novel CD approach for high-resolution remote sensing images based on leveraging RoF andCD approach for high-resolution remoteissensing images based onThis leveraging RoF andadvantage coarse-to-fine coarse-to-fine uncertainty analyses proposed in this paper. approach takes of uncertainty analyses is proposed in thisinpaper. This approach takes advantage spatial and contextual spatial and contextual information existing HVM to automatically detectofland-use changes. A coarse-to-fine CD strategy to effectively local and global changed areasCD within the is information in existing HVMistoadopted automatically detect select land-use changes. A coarse-to-fine strategy HVM. to effectively select local and global changed areas within the HVM. adopted During pixel-level pre-classification the algorithm NCIA algorithm used to consider During the the pixel-level pre-classification stage, stage, the NCIA is used to is consider neighborhood neighborhood with high probabilities of beingoreither changedInorthe information and information select pixelsand withselect high pixels probabilities of being either changed unchanged. unchanged. In the object-level stage, segmentation a series of optimal segmentation scales ranging object-level classification stage, aclassification series of optimal scales ranging from coarse to fine from coarse to fine is chosen under the boundary constraints of the HVM. These two methods can is chosen under the boundary constraints of the HVM. These two methods can effectively avoid effectively avoid subjectivity in scale selection. To obtain more evenly distributed training samples, subjectivity in scale selection. To obtain more evenly distributed training samples, the proposed the proposed sample selection strategy introduces an uncertainty index for each object. The sample selection strategy introduces an uncertainty index for each object. The uncertain objects are uncertain objects are further classified at a later stage. Multiple classifiers, each of which combines further classified at a later stage. Multiple classifiers, each of which combines RoF and a single-scale RoF and a single-scale segmentation map, are used to obtain a single-scale classification result. segmentation map, are used to obtain a single-scale classification result. Then, multiple results are Then, multiple results are integrated using the MV approach to generate the final results. Thus, integrated using the MV approach to generate the final results. Thus, spatial information from different spatial information from different segmentation scales is incorporated into the CD process of segmentation scales is incorporated into the CD process of high-resolution remote sensing images, high-resolution remote sensing images, rather than relying upon a single segmentation scale. Two rather relying upon a single segmentation scale. are Twoused pairs real high-resolution pairsthan of real high-resolution remote sensing datasets to ofverify the effectiveness remote and sensing datasets are used to verify the effectiveness and superior performance of the proposed superior performance of the proposed method. In future research, we will further extendmethod. this In future we will further extend thisto method from the supervised approach to an unsupervised methodresearch, from the supervised approach an unsupervised or semi-supervised approach. In or addition, semi-supervised approach. In addition, we will focus on automated segmentation and parameter we will focus on automated segmentation and parameter selection processes to further selection processes to further performance of the CD. improve performance of the improve CD. Acknowledgments: by primarily the National of Author Contributions:This Thisresearch researchwas wasfunded designed by Key W.F. Research and H.S.;and H.S.Development provided theProgram dataset; W.F., (No.2016YFB0502603), the National Natural Science Foundation of China Grant Nos.W.H. 41771457, J.T.,China and K.S. performed the experiments and analyzed the data; W.F. and H.S. wroteunder the manuscript. and C.X. 41471354 41601443,and andprovided Open Research Fund of technical the State guidance; Key Laboratory of Information Engineering in reviewed theand manuscript theoretical and H.S. acted as the corresponding author. Surveying, Mapping and Remote Sensing (No.16E01). W. H. was funded by China Scholarship Council and Acknowledgments: This research was funded by the National Key Research and Development Program of China Lund University (No.2016YFB0502603), the National Natural Science Foundation of China under Grant Nos. 41771457, 41471354 and 41601443, and Open Research Fund of the State Key Laboratory of Information Engineering in Surveying, Authorand Contributions: This research wasW. designed byChina W.F. and H.S.; H.S.Council provided theLund dataset; W.F., Mapping Remote Sensing (No.16E01). H. wasprimarily funded by Scholarship and University J.T., and K.S. performed the experiments and analyzed the data; W.F. and H.S. wrote the manuscript. W.H. Conflicts of Interest: The authors declare no conflict of interest. and C.X. reviewed the manuscript and provided theoretical and technical guidance; H.S. acted as the corresponding author.

Abbreviations

ofabbreviations Interest: The authors declare conflict of interest. TheConflicts following are used in thisnomanuscript:

CDAbbreviations Change detection OBCD Object-based change detection The following abbreviations are used in this manuscript: PBCD Pixel-based change detection

Remote Sens. 2018, 10, 1015

NCIA HVM RoF RF ELM IR-MAD MV MRF CRF OCVA OCC OCST GIS MTIS MRS STS MTSS MTCS ESP SEI ASEI LV ROC-LV GLCM PCA FA MA OE

20 of 22

Neighbourhood correlation image analysis Historical land use vector map Rotation forest Random forest Extreme learning machine Iteratively reweighted multivariate alteration detection Majority voting Markov random field Conditional Random Field Object-based change vector analysis Object-based correlation coefficient Object-based chi-square (χ2 ) transformation Geographic information system Multi-temporal image segmentation Multi-resolution segmentation Single-temporal segmentation (STS) Multi-temporal separate segmentation Multi-temporal combined segmentation Estimation of scale parameter Segmentation evaluation index Average segmentation evaluation index Local variance Rates of change of LV Gray-level co-occurrence matrix Principal component analysis False alarms Missed alarms Overall error

References 1. 2. 3. 4. 5. 6. 7.

8. 9. 10. 11.

Hussain, M.; Chen, D.; Cheng, A.; Wei, H.; Stanley, D. Change detection from remotely sensed images: From pixel-based to object-based approaches. ISPRS J. Photogramm. Remote Sens. 2013, 80, 91–106. [CrossRef] Bovolo, F.; Bruzzone, L. A theoretical framework for unsupervised change detection based on change vector analysis in the polar domain. IEEE Trans. Geosci. Remote Sens. 2007, 45, 218–236. [CrossRef] Hao, M.; Shi, W.; Zhang, H.; Li, C. Unsupervised change detection with expectation-maximization-based level set. IEEE Geosci. Remote Sens. Lett. 2013, 11, 210–214. [CrossRef] Cao, G.; Liu, Y.; Shang, Y. Automatic change detection in remote sensing images using level set method with neighborhood constraints. J. Appl. Remote Sens. 2014, 8. [CrossRef] Bruzzone, L.; Prieto, D. Automatic analysis of the difference image for unsupervised change detection. IEEE Trans. Geosci. Remote Sens. 2000, 38, 1171–1182. [CrossRef] Hao, M.; Zhang, H.; Shi, W.; Deng, K. Unsupervised change detection using fuzzy -means and MRF from remotely sensed images. Remote Sens. Lett. 2013, 4, 1185–1194. [CrossRef] Zhou, L.; Cao, G.; Li, Y.; Shang, Y. Change detection based on conditional random field with region connection constraints in high-resolution remote sensing images. IEEE J. Sel. Top. Appl. Earth Observ. Remote Sens. 2016, 9, 3478–3488. [CrossRef] Cao, G.; Zhou, L.; Li, Y. A new change-detection method in high-resolution remote sensing images based on a conditional random field model. Int. J. Remote Sens. 2016, 37, 1173–1189. [CrossRef] Nielsen, A. The regularized iteratively reweighted mad method for change detection in multi- and hyper-spectral data. IEEE Trans. Image Process. 2007, 16, 463–478. [CrossRef] [PubMed] Celik, T. Unsupervised change detection in satellite images using principal component analysis and k-means clustering. IEEE Geosci. Remote Sens. Lett. 2009, 6, 772–776. [CrossRef] Hazel, G. Object-level change detection in spectral imagery. IEEE Trans. Geosci. Remote Sens. 2001, 39, 553–561. [CrossRef]

Remote Sens. 2018, 10, 1015

12.

13. 14. 15. 16. 17.

18.

19. 20. 21. 22. 23.

24. 25. 26.

27.

28.

29. 30.

31. 32. 33.

21 of 22

Wang, W.J.; Zhao, Z.M.; Zhu, H.Q. Object-oriented Change Detection Method Based on Multi-scale and Multi-Feature Fusion. In Proceedings of the 2009 Urban Remote Sensing Joint Event, Shanghai, China, 20–22 May 2009; pp. 1–5. Emary, E.M.; El-Saied, K.M.; Onsi, H.M. A proposed multi-scale approach with automatic scale selection for image change detection. Egypt. J. Remote Sens. Space Sci. 2010, 13, 1–10. [CrossRef] Wang, C.; Xu, M.X.; Wang, X.; Zheng, S.; Ma, Z. Object-oriented change detection approach for high-resolution remote sensing images based on multi-scale fusion. J. Appl. Remote Sens. 2013, 7. [CrossRef] Hao, M.; Shi, W.; Zhang, H.; Wang, Q.; Deng, K. A scale-driven change detection method incorporating uncertainty analysis for remote sensing images. Remote Sens. 2016, 8. [CrossRef] Xiao, P.; Zhang, X.; Wang, D. Change detection of built-up land: A framework of combining pixel-based detection and object-based recognition. ISPRS J. Photogramm. Remote Sens. 2016, 119, 402–414. [CrossRef] Xiao, P.; Yuan, M.; Zhang, X.; Feng, X.; Guo, Y. Cosegmentation for object-based building change detection from high-resolution remotely sensed images. IEEE Trans. Geosci. Remote Sens. 2017, 55, 1587–1603. [CrossRef] Sun, K.; Chen, Y. The Application of objects change vector analysis in object-level change detection. In Proceedings of the International Conference on Computational Intelligence and Industrial Application (PACIIA), Wuhan, China, 6–7 November 2010; Volume 15, pp. 383–389. Wang, L.; Yan, L.I.; Wang, Y. Research on land use change detection based on an object-oriented change vector analysis method. Geogr. Res. 2014, 27, 74–80. Wang, Y.; Shu, N.; Gong, Y. A study of land use change detection based on high resolution remote sensing images. Remote Sens. Land Resour. 2012, 24, 43–47. Chen, Q.; Chen, Y. Multi-Feature Object-Based Change Detection Using Self-Adaptive Weight Change Vector Analysis. Remote Sens. 2016, 8. [CrossRef] Vázquezjiménez, R. Applying the chi-square transformation and automatic secant thresholding to Landsat imagery as unsupervised change detection methods. J. Appl. Remote Sens. 2017, 11. [CrossRef] Chen, J.; Xia, J.; Du, P.; Chanussot, J. Combining Rotation Forest and Multi-scale Segmentation for the Classification of Hyper-spectral Data. IEEE J. Sel. Top. Appl. Earth Observ. Remote Sens. 2016, 9, 4060–4072. [CrossRef] Gang, C.; Geoffrey, J.; Luis, M.; Michael, A. Object-based change detection. Int. J. Remote Sens. 2012, 33, 4434–4457. Zhang, Y.; Peng, D.; Huang, X. Object-Based Change Detection for VHR Images Based on Multiscale Uncertainty Analysis. IEEE Geosci. Remote Sens. Lett. 2018, 15, 13–17. [CrossRef] Alboody, A.; Sedes, F.; Inglada, J. Post-classification and spatial reasoning: New approach to change detection for updating GIS database. In Proceedings of the IEEE International Conference on Information and Communication Technologies: From Theory To Applications, Damascus, Syria, 7–11 April 2008; pp. 1–7. Yu, C.; Shen, S.; Huang, J.; Yi, Y. An object-based change detection approach using high-resolution remote sensing image and GIS data. In Proceedings of the IEEE International Conference on Image Analysis and Signal Processing, Zhejiang, China, 9–11 April 2010; pp. 565–569. Hanif, M.; Mustafa, M.; Hashim, A.; Yusof, K. Spatio-temporal change analysis of Perak river basin using remote sensing and GIS. In Proceedings of the IEEE International Conference on Space Science and Communication, Langkawi, Malaysia, 10–12 August 2015; pp. 225–230. Sofina, N.; Ehlers, M. Building Change Detection Using High Resolution Remotely Sensed Data and GIS. IEEE J. Sel. Top. Appl. Earth Observ. Remote Sens. 2017, 9, 3430–3438. [CrossRef] Ayele, G.T.; Tebeje, A.K.; Demissie, S.S.; Belete, M.A.; Jemberrie, M.A.; Teshome, W.M.; Mengistu, D.T.; Teshale, E.Z. Time Series Land Cover Mapping and Change Detection Analysis Using Geographic Information System and Remote Sensing, Northern Ethiopia. Air Soil Water Res. 2018, 11. [CrossRef] Dewan, A.; Yamaguchi, Y. Land use and land cover change in Greater Dhaka, Bangladesh: Using remote sensing to promote sustainable urbanization. Appl. Geogr. 2009, 29, 390–401. [CrossRef] Wu, C.; Du, B.; Cui, X.; Zhang, L. A post-classification change detection method based on iterative slow feature analysis and Bayesian soft fusion. Remote Sens. Environ. 2017, 199, 241–255. [CrossRef] Aguirre, J.; Seijmonsbergen, A.; Duivenvoorden, J. Optimizing land cover classification accuracy for change detection, a combined pixel-based and object-based approach in a mountainous area in Mexico. Appl. Geogr. 2012, 34, 29–37. [CrossRef]

Remote Sens. 2018, 10, 1015

34.

35. 36. 37. 38. 39.

40. 41.

42. 43. 44. 45. 46. 47. 48. 49.

50. 51. 52. 53. 54. 55.

22 of 22

Lu, J.; Li, J.; Chen, G.; Zhao, L.; Xiong, B.; Kuang, G. Improving Pixel-Based Change Detection Accuracy Using an Object-Based Approach in Multi-temporal SAR Flood Images. IEEE J. Sel. Top. Appl. Earth Observ. Remote Sens. 2015, 8, 3486–3496. [CrossRef] Feng, W.; Sui, H.; Tu, J.; Huang, W.; Sun, K. Remote Sensing Image Change Detection Based on the Combination of Pixel-level and Object-level Analysis. Acta Geod. Cartogr. Sin. 2017, 46, 1147–1155. Lu, X.; Zhang, J.; Li, T.; Zhang, Y. Hyper-spectral Image Classification Based on Semi-Supervised Rotation Forest. Remote Sens. 2017, 9. [CrossRef] Rodríguez, J.; Kuncheva, L.; Alonso, C. Rotation forest: A new classifier ensemble method. IEEE Trans. Pattern Anal. Mach. Intell. 2006, 28, 1619–1630. [CrossRef] [PubMed] Xia, J.; Du, P.; He, X.; Chanussot, J. Hyper-spectral Remote Sensing Image Classification Based on Rotation Forest. IEEE Geosci. Remote Sens. Lett. 2014, 11, 239–243. [CrossRef] Baatz, M.; Schäpe, A. Multi-resolution Segmentation-an optimization approach for high quality multi-scale image segmentation. In Proceedings of the Beiträge zum AGIT-Symposium. 2000, pp. 12–23. Available online: http://www.ecognition.com/sites/default/files/405_baatz_fp_12.pdf (accessed on 6 November 2017). Niemeyer, I.; Marpu, P.; Nussbaum, S. Change detection using object features. In Object-Based Image Analysis; Springer: Berlin/Heidelberg, Germany, 2008; pp. 185–201. Zhang, X.; Xiao, P.; Feng, X.; Yuan, M. Separate segmentation of multi-temporal high-resolution remote sensing images for object-based change detection in urban area. Remote Sens. Environ. 2017, 201, 243–255. [CrossRef] Lucian, D.; Dirk, T.; Shaun, R. ESP: A tool to estimate scale parameter for multi-resolution image segmentation of remotely sensed data. Int. J. Geogr. Inf. Sci. 2010, 24, 859–871. Drăgu¸t, L.; Csillik, O.; Eisank, C.; Tiede, D. Automated parameterisation for multi-scale image segmentation on multiple layers. ISPRS J. Photogramm. Remote Sens. 2014, 88, 119–127. [CrossRef] [PubMed] Chen, C.; Wu, G. Evaluation of optimal segmentation scale with object-oriented method in remote sensing. Remote Sens. Technol. Appl. 2011, 26, 96–102. Liu, H.; Qiu, Z.; Meng, L.; Fu, Q.; Jiang, B.; Yan, Y.; Xu, M. Study on site specific management zone of field scale based on Remote Sensing Image in black soil area. J. Remote Sens. 2017, 21, 470–478. Im, J.; Jensen, J. A change detection model based on neighborhood correlation image analysis and decision tree classification. Remote Sens. Environ. 2005, 99, 326–340. [CrossRef] Im, J. Neighborhood Correlation Image Analysis for Change Detection Using Different Spatial Resolution Imagery. Korean J. Remote Sens. 2006, 22, 337–350. Otsu, N. A threshold selection method from Gray-level. IEEE Trans. Syst. Man Cybern. 1979, 9, 62–66. [CrossRef] Tan, K.; Jin, X.; Plaza, A.; Wang, X.; Xiao, L.; Du, P. Automatic Change Detection in High-Resolution Remote Sensing Images by Using a Multiple Classifier System and Spectral–Spatial Features. IEEE J. Sel. Top. Appl. Earth Observ. Remote Sens. 2016, 9, 3439–3451. [CrossRef] Haralick, R. Texture features for image classification. IEEE Trans. Syst. Man Cybern. 1973, 3, 610–621. [CrossRef] Xia, J.; Falco, N.; Benediktsson, J.; Chanussot, J.; Du, P. Class-Separation-Based Rotation Forest for Hyperspectral Image Classification. IEEE Geosci. Remote Sens. Lett. 2016, 13, 584–588. [CrossRef] Amro, I.; Mateos, J.; Vega, M.; Molina, R.; Katsaggelos, A. A survey of classical methods and new trends in pansharpening of multispectral images. Eurasip J. Adv. Signal Process. 2016, 9, 3439–3451. [CrossRef] Hou, B.; Wang, Y.; Liu, Q. A Saliency Guided Semi-Supervised Building Change Detection Method for High Resolution Remote Sensing Images. Sensors 2016, 16. [CrossRef] [PubMed] Gao, F.; Dong, J.; Li, B.; Xu, Q.; Xie, C. Change detection from synthetic aperture radar images based on neighborhood-based ratio and extreme learning machine. J. Appl. Remote Sens. 2016, 10. [CrossRef] Wessels, K.; Bergh, F.; Roy, D.; Salmon, B.; Steenkamp, K.; Macalister, B.; Swanepoel, D.; Jewitt, D. Rapid land cover map updates using change detection and robust random forest classifiers. Remote Sens. 2016, 8. [CrossRef] © 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).