Abstract
The high dimensionality of the modern remote sensing data of construction land makes it complicated to extract image data. This paper proposes a dimensionality reduction and extraction strategy for the remote sensing data of construction land, with the aid of building information modeling (BIM) and geographical information system (GIS). Firstly, the BIM was employed to reduce the size of the remote sensing data of construction land and to obtain the information of each element. Next, the remote sensing data of construction land were parsed, and the key BIM elements were extracted through semantic filtering. In addition, the remote sensing data were converted into a triangulated irregular network (TIN), which can be processed by the geographical information system (GIS). In the end, random projection was utilized to reduce the dimensionality and compress the remote sensing data, and realize the data extraction. Experimental results show that our approach can compress and extract the information from construction land images in the remote sensing data with a high accuracy.
1. Introduction
With the rapid development of society, urban construction requires better management, standardization, refinement, scientific planning, and low-carbon environmental protection development. This poses a huge challenge to scientific research and engineering technology. It is impossible to satisfy the high requirements of construction land date and information with a single technical means.
In recent years, building information modeling (BIM) has developed quickly as a technique applicable to construction and related industries [1–3]. The technique boasts rich and consistent semantic information, as well as refined geometric structure. For example, BIM can integrate the information of each stage of engineering construction into the BIM model and display the detail information of buildings through the model. It is in line with the goal of refined management of urban buildings. The BIM can realize the intelligence and collaboration of building management, support the real-time sharing of massive data in buildings [4, 5], and meet the various requirements of building planning and approval.
Three-dimensional (3D) geographical information system (GIS) is an important technique to combine and display the spatial structure of buildings and the terrain. The technique facilitates the micro and macro management of multiscale data of buildings and provides technical support for the approval and investigation of construction land [6–8].
The information related to urban construction land can be exchanged through multiple fields. The multifield information exchange is an important aspect of information sharing and collaboration in the development and construction of smart cities. The urban construction land can be presented in digital form using the BIM [9] to thoroughly display the rich semantic information of the construction land, including physical properties, geometry, and regularity.
Remote sensing images are featured by rich information, complex features, and clear details. Feature compression and extraction is an important way to analyze massive remote sensing data [10]. As a spatial database technique, the GIS can store a huge amount of 3D spatial data and manage and visualize the massive data effectively [11, 12]. Thus, it plays an important role in the management and analysis of urban construction land. Remote sensing images and data are the main information sources and core components of the GIS. With the aid of the GIS, it is possible to effectively analyze and manage spatial data, and enhance the utilization value of images and data [13].
Under the BIM, this paper proposes a dimensionality reduction and extraction approach for remote sensing data of construction land. The BIM was adopted to process a massive amount of urban construction land. The processed data were converted into the GIS software. Experimental results show that the data on urban construction land were effectively extracted through our approach of dimensionality reduction and extraction. The extracted construction land data can be used to support urban renewal and transformation planning, land acquisition planning, construction land approval, economic and social development planning, digital city (smart city) construction, etc.
2. Methodology
2.1. Flow
The BIM was applied to construction land, aiming to support building planning and approval with complete and accurate data. Specifically, the BIM was integrated with the GIS [14–18]. The collected remote sensing data were parsed by the BIM and compressed to extract urban construction land. The flow of the compression and extraction scheme of urban construction land is explained as follows (see Figure 1):
The target data are imported to the BIM model. The data size is reduced to obtain the element information required for construction land. After a thorough analysis on the data of construction land, the key elements of the BIM are extracted, and the postextraction data are converted to the GIS model. Then, the remote sensing data are dimensionally reduced through random projection and extracted with the aid of visual word graph.
2.2. Key Element Extraction and Data Conversion
The BIM involves building settings and spatial interfaces, which cover the whole process from building design, construction, to management. Depending on the different research directions, the BIM model was simplified and used to process the remote sensing data. After filtering the relevant information, the key elements of construction land were obtained. Figure 2 shows the flow of key element extraction and data conversion.
In the past, the application of BIM to construction land extraction has the disadvantage of data redundancy, which complicates the construction land approval [19]. Here, semantic filtering is implemented according to the specific requirements of construction land, before extracting the land. Different data filter banks were adopted to filter out the redundant semantics of construction land [20].
In remote sensing images, the main land uses include high-density construction areas, low-density construction areas, new construction areas, mountains, forests, shrubs, farmlands, lakes, seas, and rivers. According to the type of the construction land, the words irrelevant to the construction land were selected as semantic filter banks, such as mountains, forests, shrubs, farmlands, lakes, seas, and rivers. Only the semantics related to construction land were preserved.
Next, the internal elements of the BIM model were combined with the constraints of semantic filter banks to obtain the key information from the semantic filter banks. After that, the BIM entity model was converted into a triangulated irregular network (TIN), which can be processed by the GIS. During the conversion, the coordinate conversion matrix of the GIS model was obtained, and the contours of construction land were mapped to the global coordinate system of the GIS model.
The identity (ID) value was employed to provide data, including the semantic information and geometric information related to construction land, for data management of the GIS model. The BIM relies on coordinates to locate construction land. When massive data were imported to the GIS model, the specific coordinates were determined by comparing each Revit measuring point with actual measuring point. In this way, the construction land, which is extracted from massive remote sensing data after being compressed by the BIM, can be converted to the correct positions in the GIS model coordinate system.
2.3. Dimensionality Reduction and Extraction
The dimensionality reduction and extraction of remote sensing data consists of two parts: dimensionality reduction and feature extraction. The former was realized through random projection, and the latter was achieved with the aid of visual word graph.
2.3.1. Dimensionality Reduction
Let I be a remote sensing image containing n wave bands; b × b be the size of the image window extracted from each band of the image according to the ranking of pixels. The pixels can be ranked by the following rule:where q0.0 is the central pixel; .
All the grayscale values on the rectangular ring centering on the b × b window are ranked by formula (1). After ranking, these values are connected in series to form the original eigenvector .
Then, spatial information is integrated with the spectrum, and the eigenvector containing n bands are prepared into a new eigenvector :where is the random projection matrix, with and . The subspace of compressed features encompasses all compressed eigenvectors.
The size parameter of the random projection matrix should conform to the Johnson–Lindenstrauss (JL) lemma. According to the compressed sensing theory, this parameter must have the restricted isometry property.
2.3.2. Feature Extraction Based on Visual Word Graph
In the visual word graph, the texture primitives are defined as the central words at all pixels and the words in the u × u neighborhood of each pixel. Here, the visual word graph is utilized to extract the construction land from the original remote sensing image and treat it as the texture feature of the pixel of interest. The texture information-containing texture primitives thus obtained are of the size b × b. They can reflect the properties of texture primitives with high quality [21]. In the visual word graph, the words in the u × u neighborhood manifest the global spatial information, as well as the spatial information of the class of each central pixel. The multiscale remote sensing data can be illustrated by reflecting the problem features with two windows. With the aid of visual word graph, the global texture features of construction land can be extracted in three steps: dictionary learning, word encoding, and feature extraction.(1)Dictionary learning. The k-means clustering (KMC) is applied to compress the feature subspace. With Euclidean distance as the similarity metric, the training samples are clustered. The cluster center is the dictionary of the corresponding cluster. Let C and K be the number of sample classes and the number of cluster centers, respectively. From the different types of dictionaries, the final compressed texture dictionary can be obtained, with the size of CK.(2)Word encoding. The texture dictionary S is established by the e nearest neighbor algorithm. Then, the Euclidean distance from each word in S to the texture primitives in S is computed, and each texture primitive is numbered with the code of the nearest word. Next, the codes of the words in the neighborhood and the central words are compiled into the texture primitive li,j corresponding to each word in the visual word graph, with the central pixel being denoted as x0,0.(3)Feature extraction. The statistical features of the remote sensing data are represented by the word histogram of the visual word graph, where is the number of appearances of word i in the visual word graph. The spatial information of words in the visual word graph is added to improve the extraction accuracy [22]. The spatial distribution information in the graph is represented by the second-order moments of different words relative to the central pixel, where , with and being the distance to the central pixel and the mean distance to the central pixel, respectively.
The second-order moment information and histogram information are fused to obtain the final texture metric :
3. Experiments and Result Analysis
3.1. Experimental Plan
This section intends to verify the effectiveness of the proposed BIM-based approach for dimensionality reduction and extraction of remote sensing data of construction land. The region in the east of Dazu District, Chongqing, China was selected as the experimental area. The proposed approach was programmed in Java on a computer running on Windows XP. Figure 3 shows the original remote sensing image of the experimental area.
3.2. Result Analysis
Figure 4 shows the construction land extracted from the original remote sensing image by the proposed approach. It can be seen that our approach can effectively extract and compress the construction land in the remote sensing image.
As shown in Figure 3, the experimental region mainly covers seven land uses: high-density construction areas, low-density construction areas, new construction areas, mountains, forests, shrubs, and farmlands. In total, there are three kinds of construction land, and four kinds of nonconstruction land. The extraction accuracy of our approach on each land use was tested to verify its effectiveness in compression and extraction. In addition, our approach was compared with deep learning and partial least square (PLS) regression (Figure 5). It can be observed that our approach extracted different land uses with an accuracy greater than 99%, much higher than that (<98.5%) of deep learning and PLS.
In the remote sensing image, the buildings are mainly multistory residential buildings, factory buildings, public buildings, single-story new houses, single-story old houses, and high-rise residential buildings. The above three approaches were separately adopted to extract the six kinds of buildings, and the number of extracted buildings was compared with the actual number of buildings (Table 1). The comparison shows that the number of buildings extracted by our approach was very close to the actual number, while that by deep learning and PLS was far from the actual number. Again, our approach was found to be highly accurate in extraction.
Next, eight buildings were randomly selected from the remote sensing image. The above three approaches were separately adopted to extract the mean circumference ratio and mean area ratio of each building. The extraction results in Figure 6 show that the construction land extracted by our approach had a much greater mean circumference ratio and mean area ratio that extracted by deep learning and PLS. Hence, our approach can extract the details of construction land with a high accuracy.
(a)
(b)
Furthermore, 10 areas were randomly selected from the remote sensing image. The above three approaches were separately adopted to extract the buildings from each area. The two errors of each approach were counted: mistaking building for nonbuilding (BFNB) and mistaking nonbuilding for building (NBFB). The results in Table 2 show that our approach had much smaller BFNB and NBFB than deep learning and PLS, evidence to the high extraction accuracy of our approach for construction land.
Finally, the above three approaches were separately adopted for feature extraction of the original remote sensing image seven times. The single-to-noise ratios (SNRs) of the construction land extracted by each approach were counted. The results in Figure 7 show that the construction land extracted by our approach from the remote sensing image had much higher SNRs than that extracted by deep learning and PLS. This means our approach can compress and extract construction land, while preserving the useful information in the image. After being processed by our approach, the images maintained high SNRs, which demonstrate the high extraction performance of our approach.
4. Conclusions
This paper applies the BIM to compress and extract the remote sensing data of construction land. After analyzing the necessary component information of the BIM, the authors filtered out the information unrelated to construction land and extracted construction land features through random projection and visual word graph. In this research, the data completeness of the BIM is fully utilized to maximize the extraction accuracy of construction land and to provide a technical basis for the planning and approval of urban buildings. Experimental results show that our approach can extract features at ease and differentiate between ground objects effectively.
Data Availability
The data used to support the findings of this study are available from the corresponding author upon request.
Conflicts of Interest
The authors declare that they have no conflicts of interest.