Keywords

1 Introduction

Wireless communication system is explosively growth application in urban environment. An appropriate radio propagation model is significant to predict the signal communication which is influenced by building’s shadowing, reflection, absorption and fading. Building distribution is an important factor to modify the radio propagation model in simulation environment. Building density analysis is valuable for predicting the signal propagation and visualization, and simulation parameters of radio propagation model can be modified [1,2,3,4,5]. Therefore, accuracy automation building extraction from satellite imagery is helpful to design the wireless propagation model.

The satellite imagery has many advantages, such as the macroscopic imaging, low cost of approximating in horizontal projection objects and easily to be digitized, which became more and more popular to research in computer vision [6, 7]. In recent years, many researches have been focusing on building extraction from image. These methods can be coarsely divided into three classes. The first one is extracting building by some human interaction (semi-automatic) [8,9,10,11]. Based segmentation method requires much more user workload, which is impractical for the large segmentation task. Tian [11] adopts optimal parameters (shape, spectral bands, from-related) from the given type of objects (road, building, sports filed) to process the multi-resolution segmentation on high-resolution satellite image. The experiment proves that this algorithm is capable of producing with the comparable result. However, Suggested optimal parameters are usefulness and should be preliminarily demonstrated.

In order to decrease the user workload, some methods are proposed to extract building from different types imagery information (single-view or multi-views), such as optical image and radar data [10, 12]. Rottensteiner [13] creates a 3D building model form the high-resolution lidar data. The skew error distribution function is introduced to compute the digital terrain mode by separating points of building from other object classes. After analyzing the height difference with digital surface model, building is derived by curvature-based segmentation techniques. Guo [14] locates the boundary of building by a snake-based approach from the high-resolution IKONOS satellite imagery and captures the height data by airborne laser scanning system. Teimouri [15] identifies an optimal fusion of radar and optical image. The decision-level fusion is used to assess the classification of panchromatic, multi-spectral, and radar images. Then, author applies an artificial neural network to classify those feature, and the decision level of integration for the building detection purpose.

Since multi-source image is expensive to obtain, which limited to be widely applied for ordinary users. Then, the third method according to the principle of elements acquirement (region-based or edge-based) are proposed, that could be reduced the cost of equipment. There are many methods to automatically extract the building area by this way, such as morphological-based methods [16,17,18], region and edge detection-based methods [19,20,21,22]. Huang and zhang [16, 17] rely on the morphological shadow index (MSI), and the jointly use morphological building index (MBI) for the spatial constraint of building. The dual-threshold approach of their research improves for building extraction form background without collection any training samples and process of supervised leaning. Their method was proved more accurately than the support vector machine interpretation. Jin [18] presented an automated building-extraction strategy via mathematical morphology informations (the structural, contextual, and spectral) from the high-resolution satellite imagery. This system need not provide any initial algorithm parameters, training set and pre-classification. Jiang and Zhang [19] propose a semi-automatic building extraction method to get building by edge detection and region-driven. This method result heavily relies on the image’s spectral information. Lin [20] describes a system to detect the building in aerial scenes. After Low-level segmentation, a perceptual grouping approach is used to collect fragment segments from shape properties, and the shadow is used to form and verify the hypotheses region. Gamba [21] classifies the urban area mapping separately by a neural network and an adaptive Markov random field model. Decision fusion process combines with the classification mapping to exploit objects boundaries. Lee [22] uses the classification result of Ikonos multi-spectral image and Hough transformation to detect and form the building boundary. Extraction error is mainly caused by the road class as well as the occlusion and shadow. These methods usually utilize color, space, textual, and spectral information to extract building area. In other works, the successful building extraction of these existing region-based or edge-based methods is usually based on the assumption that the source image is in high-resolution and easily distinguished form the target area. However, when the image is in low resolution and with some noise information, such as the available google or network satellite mapping with slight fog and intense illumination, these methods will be not suitable to segment building region from around the environment.

Stroke width is related to the linear feature which has applied into extracting road’s networks. Doucette [23] presents a self-supervised road classification method to detect road centerline from high resolution multi-spectral imagery. Candidate road centerline is detected by the anti-parallel-edge. It is crucial for this method that needing the sufficient number of correct roads as training samples. The author reviews a variety of techniques which identified the linear, planimetric feature form imagery based on spectral, spatial and radiometric rules. Epshtein [24] detected the text in many fonts and languages use stroke width transform (SWT). This method suggests that the SWT algorithm combined with pixel dense estimation is efficient to extract the object with same stroke (symmetry region) in natural images. Quackenbush [25] also presents a review of the system for the linear feature extraction from imagery, that regarded as significantly requirement to automation detecting interest objects ongoing research from remotely sensed imagery.

Fig. 1.
figure 1

The satellite imagery.

Full size image

In this paper, we propose a novel automatic method to extract building region from the single-view and in low-resolution satellite imagery, such as Fig. 1. The figure contains a complex environment, such as old building, playground, green landscape and waste land. Some of building is low and covered with vegetation. Our method is motivated by segmenting the symmetry and regularity architecture from this satellite mapping, and analyzed this region’s building density to modify the wireless propagation model. This study introduces SWT algorithm [24] to detect the hypothetical region, and fuses with k-means algorithm to achieve the presence of building from satellite imagery. The framework of building segmentation contains several steps. The first step uses the morphology transformation to add redundant boundary points to Canny operator [26], which benefited for calculating image pixels’ stroke width. The second step calculates images’ SWT spectral to detect the symmetry region from satellite imagery. K-means algorithm is used to obtain building region by color information. Then, the possibility building region is generated. Integrated SWT spectral with the classification result, building boundary can be totally acquired. The third step fusing second result could wipe out some interference region. Then, combined with the originate imagery mapping, Building region is grown into a fine segmentation result by a simple linear iterative clustering algorithm. For speciality of this satellite imagery, we compare our method result with K-means and Isodata clustering algorithm. The experiment results prove that our method is efficiently improvement building exaction.

The paper is organized as follows. In Sect. 2, our proposed method and the detailed of stroke width generation are introduced. In Sect. 3, experiment result and performance evaluation are presented. Finally, conclusions are given in Sect. 4.

2 Proposed Method

Our goal is to obtain the building region and evaluation its density. Therefore, this method could be performed as jointly estimating the symmetry region and color clustering problem. Intuitively, the symmetry region is applied to distinguish the character of building, and it can be detected by SWT algorithm. Besides, building always has different features with rural region in color space. We perform k-means algorithm in color clustering to removing interference. After intergrading those detected result, the simple linear iterative clustering algorithm [27] is used to maintain the building region consistency. Based on the consideration above, an unsupervised jointly framework is proposed as shown in Fig. 2. From the flowchart, we known that this method is designed simply and could be automation segmentation building region from satellite imagery.

Fig. 2.
figure 2

Flowchart of the proposed method.

Full size image

2.1 Stroke Width Generation

The stroke width of image pixel is a locate description operator based on Canny edge detector. It has been extensively applied to extract the symmetry region (letter or road) in nature scenes. Neighboring pixels with the same stroke width may be grouped together. Relied on those similar stroke width of pixels, approximately symmetry region would be clustering from image. Building’s roof has a good symmetry in geometry morphology. Therefore, focusing on building’s roof is helpful to reduce a number of interference region.

Stroke width is defined as the length of line and rough perpendicular to the orientation of stroke. Its calculation should be firstly detected image edge by Canny operator. Those symmetry pixels located in edge are found. Then, the length of pixels in line is assigned as the same stroke width. Image pixels with same width would be in high spectral distribution. Building region is mainly identified by those image pixels’ stroke width.

Fig. 3.
figure 3

Implementation pixels with stroke width using SWT algorithm.

Full size image

Intuitively, stroke width is calculated from the image edge boundary. A gradient direction \(d_p\) belongs to the pixel p on edge. If other pixel can be found from edge and signed as q, and the orientation of \(d_q\) follow the ray \(r = p + n \cdot d_p\), where \(n > 0\). The gradient direction \(d_q\) is roughly opposite to \(d_q = -d_p \pm \pi /6\). Then, the SWT spectral of output image along with the segment [p, q] is assigned the width with \(\Vert \overrightarrow{p-q}\Vert \). If a lower value already is appeared, pixel’s width would be chose the lower one as shown in Fig. 3.

If the pixel q is not existed, or \(d_p\) does not opposites with \(d_q\), the ray would be discarded. If pixels are in corner, and the average width of pixels is m, the length of pixels longer than m would be replaced by m as shown in Fig. 3.

The output of the SWT is a image, which contained the width of the stroke with each pixel as the same size of input image. Based on those theories, a detailed description of each step of the method is provided in the following section.

2.2 Building Boundary Extraction and Fusion

After calculating image pixels’ stroke width, Those symmetry region appear more intensity than others. In this section, pixels group together into building candidates via the distribution of pixels’ width. Edge detection determines the calculation scope of SWT algorithm, and is mainly influenced by the bright light and noise (shadow) in satellite imagery. Several tests prove that the more region’s edge detection, the better for the later building region extraction. Therefore, satellite imagery is necessary to be preprocessed before finding the building region.

In mathematical morphology, open operator is used to expand feature and close some of gaps in image. It followed with subtract operator could remove some of small or narrow elements, without influencing the large one. After those preprocessing, the uniformity of image background is added. Edge detection would become more complicated, so that building region would be appeared more obviously. Road and lawn region would be lower density in the distribution of SWT spectrum. The aim of this step could be to reduce several dissymmetry region.

Based on the edge detection, image’s SWT spectral is calculated. Neighbor pixels would be grouped together with the similar stroke width. For allowing more elaborate and perspective distortions of the symmetry region. If two neighboring pixels ratio does not exceed 3.0, the stroke with smoothly varying width will be grouped together. Those component region may contain building. A small set of rule is also employed by extracting the geometric character of building. The threshold of pixels’ stroke width (b) is set less than 30. After this threshold, road and lawn region (with less stroke wind) would be rejected from image.

Image’s SWT spectral obtains a lot of scattered region by Canny operator. To reduce the hypothesis building region, we classify image into two groups by K-means algorithm based on image’s color mapping. Building’s roof and road are homogenous in color and with much higher light than others. Therefore, the classification result is easily obtaining building and road from satellite imagery. We combine SWT spectral with classification result, and then the outline of building can be detected clearly. After this merged, irregular objects and mostly of roads are eliminated from the building’s region.

After above preprocessing, the hypothesis building region is brought up and scattered distribution in region. In this section, building boundary is extracted by a superpixel generation method [27]. This method generates superpixel based on K-means clustering with less memory requirement and strong performance. Superpixels contain similarly color distribution and labeled, respectively. Then, building region would be identified completely by fusing the building boundary with superpixel distribution.

Fig. 4.
figure 4

An example of this method: (a) The original satellite image. (b) The SWT mapping. (c) The result of k-means classification. (d) The result of proposed method.

Full size image

In order to analysis the advantage of proposed method, an example of subjective result is shown in Fig. 4. We can see that this area contains many complex object region, such as building, road, tree and others shadow in Fig. 4(a), it can be seen that the probability map by SWT obtains the building symmetric region, but introduces some disconnected region in Fig. 4(b). Meanwhile, the result obtained by kmeans extracts building region as consistency in color. However, it introduces many noise region such as the flat region in (Fig. 4(c)). Combined the SWT with k-means result, most of dissymmetry region is wiped out. Building region would be recognized. While, the result in (Fig. 4(d)) successfully segments building region and reduces noise region after integrating building boundary and superpixel distribution.

3 Experiment and Analysis

The experiment is tested on Google earth satellite map with 1.07 metric/pixel spatial resolution. The dataset contains 224 images are special representative to predict the signal communication as shown in Fig. 4. The building region contains many different shapes, such as square and strip. Moreover, there are many tree and grass surrounding the old building, which seriously interfered building segmentation.

Iterative self-organizing data analysis technique algorithm (Isodata) is clustering algorithm based on the k-means, which added the merged and divided of clustering result. Isodata is an adaptive method to adjust the clustering number. Isodata and K-means are basic clustering methods. For segmenting the building density of satellite image with low resolution and single information, most of building excitation methods are failure. Therefore, we use the Isodata and k-means clustering algorithm comparative with our purposed method.

Performance Evaluation. In order to demonstrate the effective of our proposed method, we use the intersection-over-union (IOU) to evaluate the segmentation performance, which is widely applied for image segmentation. The groundtruth of building region is manual annotation. IOU is defined as follows:

$$\begin{aligned} precision = \frac{A\bigcap {B}}{A\bigcup {B}} \end{aligned}$$
(1)

where A is the binary mask which denoted the segmentation region by our method. B denotes the groundtruth. The higher the iou is, the better the performance is.

Fig. 5.
figure 5

Result of building extraction for our method and compared methods. (a) The groundtruth. (b) Isodata algorithm results. (c) K-means algorithm result. (d) Our method result.

Full size image

Some subjective result and objective result are shown in the following parts. Figure 5(a) is the groundtruth of satellite imagery. Figure 5(b)–(d) show the different segmentation results of Isodata, K-means, and our method, respectively. From this figure, we can see that the result of Isodata algorithm produces much noise due to the light point in original image. Meanwhile, the result of K-means algorithm obtain much more compact building than the result of Isodata. However, it produce some false positive in the middle bottom of the image, i.e., extracted the grass region as building. The result of proposed method is better than K-means and Isodata. It could not only remove the noise point but also avoid to segment the grass region as building. Meanwhile, the objective results are shown in Table 1. It can be seen that our method performs better than Isodata and K-means algorithm. The result demonstrates the effectiveness of the proposed method. The K-means result contain lots of building region, which contained some Non-architectural area. Therefor, it has much larger rise of IOU than Isodata. The value IOU of our method is improved limitedly. However, the segmentation result show that our method proves more similarly building region. The main reason is that building region loses some part of true territory caused by in the last step fused with superpixels region. As shown in Fig. 5(d), the new building region could be partitioned completely with much higher floor.

Table 1. IOU result.
Full size table

In this paper, stroke width of image pixel is used to provide the gradient information of building boundary. K-means algorithm detects more compactness region than Isodata. So we combined image pixels’ width with the K-means classification result to discard some noise. The building region could be detected very well by our algorithm. Comparing the extraction performance with K-means algorithm, Most of roads, grassland and bare ground could be deserted. The shadow of building could add the contrast with surrounding areas. Test result shows that image with much higher building could be enhanced the performance, which influenced by the shadow. Nevertheless, some of artificial road with good symmetry is hardly eliminated completely by this method. The experiment result also shows that our method is poor performance for the image only containing road or lawn region. Form the original satellite imagery, A lots of old building with a dark brown color roof was also made to be labeled mistakenly in test.

4 Conclusions

For modifying the wire radio propagation model parameters, loaded those satellite images is used to extracted the building density from google earth with low resolution and single spectre. In this paper, we proposed a simple automation method to segment building regions epically for an arbitrarily satellite mapping. We introduce SWT algorithm to detect the symmetry region, distinguished the character of building from around region, and cluster the similar color region via K-means algorithm. Finally, fused those result with the originate imagery mapping, building region is grown into a fine segmentation result. The proposed method could extract building region automatically and doesn’t need any high-level information or others prior knowledge. Compared with the object-oriented remotely sensed building extraction, it has a more application area. Based on the value of building density, wireless transform models will be modified in further researcher.