1. Introduction
Close by the advancement of satellite systems with a spatial resolution beneath 1 m, there has been a significant change in the ratio between the observed object dimension and the size of pixel over the past decade, where the size of the pixels becomes very small when compared to the average size of observed object. Due to the large detail of images of high spatial resolution, the conventional pixel-based method does not fit this type of image. In this manner, specialists began refocusing from singular pixels to their group representations, i.e. objects, as the most reasonable input data for investigation. The process of identifying and distinguishing between different geographical objects was be enabled by using the OBIA.
With significant improvement in pixel size of remote sensing images with high spatial resolution compared to the observed object, the pixel-based classification technique became worthless. Over the last decade, many of the techniques used to classify images have become useless, especially with the increasing the accuracy of remote sensing images. Hence, the popularity of the new OBIA method has increased as way combining the segmentation and classification of remote sensing images. When applying the image splitting process, the image is divided into parts containing homogeneous pixels which are categorized to achieve the process of the classification purpose, which is to determine the names and characteristics of these groups on the surface of the earth. The operator attempts to determine land use classes by choosing groups of similar pixels in visual interpretation. Whereas the land use classes was determined in digital automatic classification by the means of texture, context, spectral and/or geometric, temporal information combined with statistical grouping into classes. Pixels spectral signature is used in digital pixel-based classification to assign each pixel to the most suitable class. Then again, the first step of object-based classification is gathering pixels with regular basic qualities, and after that, based on several attributes; these parts are categorized correctly and so using the OBIA method, the pixel based classification advantages and optical classification advantages are combined [1]. Therefore, the new region based segmentation of high spatial resolution images has assumed a high position and has displaced the conventional pixel based method [2]. OBIA has increased quick saleability in remote sensing zone for the most part because of the expanding accessibility of image with high resolution. Image segmentation is an essential phase in process of OBIA. Uncommonly, the quality of image segmentation affects the subsequent accuracy of image classification [3]. Various endeavors carried out for the segmentation to create techniques for the target determination of ideal parameters that are automatic or semi-automatic. The method of multiresolution segmentation is widespread utilized method of segmenting images of remote sensing that can be used in this regard [4]. In general, the segmentation anomalies called over-segmentation and undersegmentation specify the quality of segmentation goodness. Oversegmentation occurs when any of image is divided into segments with a smaller area than the portioned object, while the undersegmentation occurs when the area of the output regions of segmentation is bigger than the object reference area within the image. It is possible that in the case of oversegmentation getting object while in the case of undersegmentation it may be not get object. From the above, since it is not possible to deal correctly with the results obtained from undersegmentation, this type of segmentation must be avoided by choosing the best or close to best value of segmentation parameters. Metrics are utilized to assess the goodness of segmentation by comparing the interactions of image objects with reference objects [1]. In this research, the most broadly utilized strategies of image segmentation was applied, which is a multiresolution segmentation, on a remote sensing image obtained from VorldView-3 sensor to evaluate the segmentation technique. eCognition software was used in order to apply the segmentation process on this image and determine the most appropriate and best values for segmentation parameters. This research has been divided into sections whose contents are described as follows. At the second section of this research, a new widely used method to analyze the images of remote sensing was introduced, which is OBIA.
In Section 3, an explanation of the image segmentation technique was presented. In Section 4, the software used in this research was explained, and description of the data used and its source were presented. In Section 5, the results obtained during this study are presented. In the sixth section of this paper, a number of conclusions reached during this study were presented.
2. Object based image analysis
In the most recent decade, the OBIA has gotten a large portion of the consideration from both industry and research and has become one of the best and most widespread techniques used in the field of remote sensing image with high spatial resolution and extract object from it. The method is founded on analysis of objects analysis including of spatially adjacent pixels, rather than managing pixels. The method is characterized by the extraordinary advantage that it benefits spatial relations between the pixels, which contain a lot of information that helps in the formation of object. The basic function of OBIA is to divide the image to various segments according to their properties. This technique is a basic auxiliary technology designed to build the object, which is the basis for the analysis of images after that. Recently, OBIA has become a major field of research [2].
Truth be told, at VHR, every pixel acts as a district, which implies that increasing in identifiable objects. In general, the individual pixel is not represent a whole object in the image but can be a part of it. At this level of detail, classes can share the same spectral signatures or not. OBIA approaches endeavor to beat these challenges by gathering pixels into larger amount objects, called segments or areas. These areas permit the calculation of more informative features for example texture or shape, which can be utilized to better distinguish and depict studied area structures. Segmentation is the first fundamental step in the OBIA technique by which the basic image object is constructed, which is then used more extensively in image analysis [5]. The image object is the fundamental components in object-oriented approach. The image object can be defined as contiguous areas within the image and the object of interest differs from image object primitives. One of the most important examples in which object of interest can be expressed is the footprint of building as well as the boundaries of agricultural land. Object of interest is obtained through the process of segmenting and classifying images of remote sensing and obtaining this result is done through another intermediate step is to obtain object primitives first and the smallest image object can be considered as one pixel. Heterogeneity possibilities such as object shape and gray tones can be used to obtain object primitives. Segmentation techniques utilizing a heterogeneity depending just on prime object features can generally just convey object primitives, without reference to the concerned object. At the beginning of the classification of remote sensing images, object primitives is classified into a certain type and the classification is ready for the classification stage based on the segmentation [6]. After the image segmentation process, a resulting image was divided into objects which it can be used later in the process of image classification and the accuracy of the resulting segmentation has an effect on the classification process [7].
3. Segmentation
The beginning and use of this method is due to the 1970s [8]. The expression “segmentation” is utilized here as the synopsis of all steps that construct, develop, union, shrink or cut objects. The process of image segmentation is considered an important basic phase in the series of remote sensing science. The quality of image classification process depends mainly on the process quality of segmentation, which depends on the good choice of segmentation parameters values. A two-faced problem arises when applying image segmentation which can be explained in particular, oversegmentation is occurred when the contrast between neighboring parts is deficient and must be fused to one object, whereas undersegmentation, indicates to the presence of parts with more area than the basic object area inside the image and which must be divided into smaller parts. This refers to the inappropriate selection of segmentation parameters. Since gathering segments is easier than splitting them, so oversegmentation is less a problem than undersegmentation. In general, briefly, a good segmentation can be defined as a segmentation that results in objects that are identical to ground reference objects and do not result in undersegmentation and may result in a little oversegmentation, and the efficiency of the mathematical model of image segmentation can be judged as the model that makes the user capable of to obtain a perfect segmenting of images without the need for careful selection of the values of the segmentation parameters [9].
3.1. Segmentation algorithms
In order for the segmentation to take place, a set of components must be available to complete this process. This is the mathematical model, the data and the program used. The process of splitting the image is done on one of two levels of the following, the pixel level and the second level object level, where the image object is converted from a size to a smaller size. eCognition software gives a few distinctive ways to deal with segmentation, ranges from a simple standard model, for example chessboard, to complicated algorithm such as multiresolution algorithm. To move to new levels in object construction, a mathematical model of segmentation is used and applied based on data and information in the image layer. There are many mathematical models used in the segmentation of images, some of which are mentioned as follow: chessboard, quadtree, contrast split, multiresolution, spectral difference, multi-threshold, and contrast filter. When using the chessboard algorithm the image is divided into a matrix of square object which can be applied in the image object phase or in the pixel phase. Whereas in case of Quadtree the image is divided into quadtree lattice. If the contrast split algorithm is used, the program divides the images or image object into areas that are either bright areas or dark areas. A reasonable initial image object is determined based on pixel filters by specifying appropriate values for contrast and gradient in case of using contrast filter algorithm. Multiresolution algorithm is the mathematical model used and applied in this research and therefore will take more interest in explanation and scrutiny than others. Using this mathematical model, which is one of the most widespread models, the program can reduce the average heterogeneity of the image object. Multiresolution segmentation (MRS) is likely the wellknown algorithm for these objectives. This mathematical model is the most common mathematical models that divide the images used, which has become popular in scientific research, which can be implemented in eCognition software. When using MRS as a mathematical model to segment the images of remote sensing, the parameter that has a very large impact on this process of segmentation is the scale parameter. The accuracy of the classification process, based on the process of dividing images, depends greatly on the image object size, which is affected by the scale parameter value [11].
When using multiresolution segmentation on a set of objects for remote sensing images, this mathematical model reduces the average heterogeneity and increase homogeneity of these objects. The image object areas, as well as the classification resolution of the remote sensing images, are very much affected by the scale parameter. The image object area is large by the scale value and vice versa. The image object is small when the scale parameter is small [10].
3.1.1. Scale parameter
The scale issue, in this way, has developed as a noteworthy issue in OBIA, especially regarding studies of OBIA. It is in this manner basic to define the suitable scale value and gain optimized segmentation results. However, in many applied studies, extractions of image objects depend on the approach of a trial-and-error, and the scale parameters segmentation values are specified according to proceeding experience [12]. It is known that when a remote sensing image is segmented, at a certain value for scale parameter, the size of the image object is larger for homogeneous data, whereas the size of image object is smaller for heterogeneous data. The image objects size varies by modifying the scale parameter value.
3.1.2. Shape
The relationship between color and shape criteria influences by the shape field value. The color standard can be adjusted and set by selecting a suitable value for shape criteria. It is recognized through previous research that the maximum value of shape parameter does not exceed 0.9.
3.1.3. Compactness
The compactness of an image object can be defined by the product of the width and the length over pixels numbers.
3.2. Segmentation evaluation
Over the last several decades, numerous image segmentation techniques have been suggested. As new segmentation techniques have been suggested, different assessment techniques have been utilized to emulate new segmentation methods to previous methods [14]. The human eye is an experienced source for assessing of segmentation methods [15]. Segmentation assessment techniques can be extensively categorized as either supervised or unsupervised. In supervised approaches it is required to perform adjustment of the segmentation parameters by using reference data, so that to get a situation where there is a match between image objects and ground reference object. Supervised methods are preferable to assess the results of segmentation especially with the presence of precise ground reference data [4]. To evaluate the quality of the image segmentation process, a set of metrics is used, from which the segmented object is matched with the ground reference object, and by means of the area value of the segmented object and ground reference and the interaction between them, the evaluation of this process can be done [16]. Table 1 shows the different methods of evaluating the process of segmentation remote sensing images and the corresponding mathematical model for each method.
Table 1. Overview of the selected segmentation accuracy metrics.
| Segmentation accuracy metric | Formula |
|---|---|
| Quality rate | |
| Area fit index | |
| Over segmentation | |
| Under segmentation | |
| Root mean square |
where ARj refers to the reference total area Rj; ASi refers to the total area of corresponding terrain segments Si, and A is the area in pixels [13].
The area fit index (AFI), quality rate (QR) and root mean square (D) are global metrics that consider the whole imagery for assessment purposes. The D metric joins the oversegmentation (OS) and undersegmentation (US) metrics to assess the ‘loseness’ of the image objects to the reference objects. Utilizing US and the OS metrics is alluded to as local effectiveness because “single objects are considered”. Oversegmentation is occurred when the contrast between neighboring parts is deficient and must be fused to one object, whereas undersegmentation, indicates to the presence of parts with more area than the basic object area inside the image and which must be divided into smaller parts [4]. AFI = 0:0 and overlap is 100% for a ideal fit. Oversegmented is occurred if the overlap between image object and reference object is less than 100% and AFI > 0:0. Undersegmented is occurred if the overlap between image object and reference object is 100% and AFI < 0:0. In a few circumstances, overlap may be less than 100% and AFI < 0:0; at that point the object is oversegmented, but the largest segment is larger than the reference object [17]. The value of the metrics is between 0 and 1 and represents the proximity of the value from zero to the accuracy of the segmentation process as well as to the extent of the matching of an image object with the ground reference object. However, AFI values may be less than zero and this occurs in the case of underegmentation [13]. |
4. Data used and segmentation software (eCognition)
4.1. Data used
|
The study area of address Calle velero, Madrid – Spain covers (3439 × 1794 pixels) orthorectified Worldview-3 multispectral satellite image was utilized in order to determine the best segmentation parameters that can be used to extract buildings from satellite images (Fig. 1). WorldView-3 is the industry’s first multi-payload, high-resolution commercial satellite, super-spectral. Working at an elevation of 617 km, the resolution of WorldView-3 is 31 cm in panchromatic and 1.24 m in multispectral. 
Fig. 1. Data used – orthorectified Worldview-3 multispectral satellite image [18]. |
4.2. Segmentation software (eCognition)
|
The eCognition software is one of the first programs aimed at the process of analysis and segmentation of remote sensing images where the program was developed through definiens [6]. With the presence of the eCognition software in 2000 object-based analysis experienced a genuine blast. This product was the primary commercial software that could be utilized to implement a quality object-based analysis of multi-spectral data [1]. Trimble eCognition is developed analysis software obtainable for geospatial applications. It is intended to enhance, fast, and mechanize the interpretation of an assortment of geospatial data and empower uses to design feature extraction or change detection procedure to adjust geospatial information into geo-data. Multiresolution segmentation in the eCognition software is a Progressive from bottom to up district blending system beginning with object of one pixel [19]. |
5. Experimental analysis and results
|
Segmentation parameters should be set preceding image segmentation, and the choice of these parameters can significantly effect on the remote sensing images classification accuracy. To assess the process of segmentation of remote sensing image five segmentation goodness as shown in Table 1 was used and applied through the reference objects available on the image. Multiresolution segmentation, executed in eCognition software, was utilized for the OBIA. The segmentation depends on a district growing procedure which places seed pixels over a whole image and gathering neighboring pixels to the local seeds, on the off chance that they meet particular criteria. To define ideal or close ideal segmentation parameters used to extracting building from remote sensing images, an experimental study is implemented. In this paper, a set of segmentations was carried out utilizing the Worldview-3 image with different values for the segmentation parameters to define ideal or close ideal segmentation parameters (scale, shape, and compactness) values according to the combination values showing in Table 2. Table 2. The values of used segmentation parameters.
In this study, there were many attempts, which reached 121 sets of the three parameters scale = (10, 50, 75, 100, 125, 150, 200), shape = (0.10, 0.20, 0.30, 0.40, 0.50, 0.60, 0.70, 0.80) and compactness = (0.50, 0.60, 0.70, 0.80, 0.90) were utilized for the segmentation of the Worldview-3 image. Fig. 2 demonstrates the segmentation results in case of segmentation parameters values are 10, 0.10, and 0.50 for the scale parameters, shape, and compactness respectively. 
Fig. 2. Segmentation results for parameters values Scale = 10, Shape = 0.10, and Compactness = 0.50. The blue line in the image shows the boundary lines of every segment. Fig. 3 show the segmentation results in case of segmentation parameters values are 50, 0.50, and 0.50 for the scale parameters, shape, and compactness respectively. 
Fig. 3. Segmentation results for parameters values Scale = 50, Shape = 0.50, and Compactness = 0.50. To evaluation the quality and estimate the quality of the segmentations results for every set, a seven reference distributed building (distributed in location, form, area, texture, contrast and so on) were considered and every was visually and geometrically emulated with the segmented regions, the situation of these building as shown in Fig. 4. After several attempts were made to change the values of the segmentation parameters combinations and each time the goodness metrics of segmentation was calculated, it was discovered that the ideal segmentation parameters values are scale = 150, shape = 0.50, and compactness = 0.80 and the corresponding goodness metrics of segmentation as shown in Table 3 and the relating image for the segmentation by optimum segmentation parameters as shown in Fig. 5. The Table 4 shows the results of goodness metrics of segmentation of one attempt which the segmentation parameters values are scale = 50, shape = 0.50, and the resulting image of this segmentation and by using those segmentation parameters as described in Fig. 3. The Table 5 shows the results of one attempt which the segmentation parameters values scale = 150, shape = 0.70, and compactness = 0.50 and the resulting image of this segmentation and by using those segmentation parameters as described in Fig. 6. In the table from table # 3 to table # 5, some of the models of the calculations that have been set to assign the values of segmentation quality metrics are shown when the values of the segmentation parameters are changed and applied to image using seven buildings sample chosen as a case study. Each of the three tables consists of 8 columns, the first column represents the building number, the second column represents the real reference area of each building, the third column represents the area of the largest part that was segmented and shares the reference area in a part of the area, and the fourth to eighth columns represent different goodness metrics which is shown previously in Table 1. The value of the metrics is between 0 and 1 and represents the proximity of the value from zero to the accuracy of the segmentation process as well as to the extent of the matching of an image object with the ground reference object. 
Fig. 4. Seven building used to estimate the goodness of the segments. Table 3. The different goodness values for each building for optimum segmentation parameters scale = 150, shape = 0.50, and compactness = 0.80.
Fig. 5. Segmentation results for optimum parameters values Scale = 150, Shape = 0.50, and Compactness = 0.80. Table 4. The different goodness values for each building for segmentation parameters scale = 50, shape = 0.50, and compactness = 0.50.
Table 5. The different goodness values for each building for segmentation parameters scale = 150, shape = 0.70, and compactness = 0.50.
Fig. 6. Segmentation results for optimum parameters values Scale = 150, Shape = 0.70, and Compactness = 0.50. To evaluate the segmentations efficiency, the segmentation results were compared to reference polygons that were mostly manually delineated in the images that were analyzed (table # 3 through table # 5). A set of metrics were utilized to evaluate the spatial match between reference polygons and individual image objects of the automatically generated segmentation levels. In table #3 and table #4 the values of the evaluation parameters (QR), (AFI), and (OS) are the same even though they come using different equations as in table (1) because the status shown in Tables 3 and 4 is in the case of oversegmentation. In this case, the segmented area (As) is less than the reference area (Ar) and therefore the overlap between reference and segmented area is equal to segmented area, and union of reference area with segmented area equal reference area and therefore in this case there is equality in the three laws. Table 5 represents the status of the undersegmentation of all buildings except Building 1 and Building 4. In this case, the value of the reference area is less than the value of the segmented area, then the area fit index is less than zero and the value of oversegmentation is equal to zero. Fig. 7 shows the quality rate values at compactness 0.80 (for building_1) for different values of scale and shape parameters. From this figure the value of QR goodness doesn't change at the shape value between 0.10 and 0.20 by increasing the value of scale parameter from 97 to 150. 
Fig. 7. Quality Rate values at compactness 0.8 (for building_1). |
6. Conclusion
|
After applying the image segmentation process to the remote sensing image of the case study, a set of conclusions was reached:
|