Training Samples Dataset for Building
Identification in Urban Village

Liu, Y. F.^1,2  Lv, B. R.^1,3  Peng, L.^1*  Wu, T.^1,3  Liu, S.⁴

1. Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100101, China;

2. Ucastech (Beijing) Smart Co. Ltd., Beijing 100080, China;

3. University of Chinese Academy of Sciences, Beijing 100049, China;

4. Beijing Qingruanhaixin Technology Co. Ltd., Beijing 100085, China

Abstract: Identifying buildings from remote sensing imagery is an important basic methodology used in urban management. The distribution pattern of the building clusters, especially the high density of buildings and narrow streets, among others, are more critical for urban managers. Based on the remotely sensed images obtained in Google Maps, 2328 samples of building clusters in an urban village were drawn by using LabelMe software. The building information was extracted by using the Mask R-CNN, which is an example of a segmentation algorithm used in deep learning. The dataset includes: (1) original sample images (Buildingsample_pic); (2) sample segmentation results (Buildingsample_mask), and; (3) sample segmentation annotation (Buildingsample_info). The datasetconsists of 6,984 data files in three data folders, having .png and .yaml data formats. The data size is 499 MB (compressed into one file: 498 MB). The research paper related to the datasetbe published in the Proceedings of the first China Digital Earth Conference.

Keywords: urban village; building cluster; deep learning; Mask R-CNN; Proceedings of the first China Digital Earth Conference

Dataset Avialable Statement:

The dataset supporting this paper was published at: Liu, Y. F., Lv, B. R., Peng, L., et al. Training samples dataset of building identification in urban village [J/DB/OL]. Digital Journal of Global Change Data Repository, 2020. DOI: 10.3974/geodb.2020.02.16.V1.

1 Introduction

With the continuous development of urban construction and urban governance, the problem of the urban village is now of widespread concern^[1-2]. The urban village is a residential area built on the original rural collective’s land and farmers’ homestead during phases of urban expansion, in which buildings are an important part. Urban village buildings are disordered and heterogeneous settlement patterns perhaps best described as “city is not like city, village is not like village”^[3]: because of its high density of buildings, narrow streets and lanes, illegal building and other characteristics, the urban villages’ shape is diverse and is structurally complex, which has always been a contentious topic in academic research. The urban village building community is a target subject, with discernable structural characteristics, in the analysis of remotely sensed images of urban areas. Due to its unique distribution and pattern of urban village building community, it is of great significance to extract buildings based on deep learning. In recent years, with advances in artificial intelligence and deep learning techniques, many scholars have begun to research how to apply deep learning to extract buildings from such imagery. Compared with a data-driven and model-driven method, the building extraction process based on machine learning requires less prior knowledge and can achieve high extraction accuracy when using suitable samples^[4-7]. In this paper, a large and medium-sized city in northern China was selected as the sample drawing basis. By using Google Maps imagery, with a spatial resolution of 0.11 m, a total of 2,328 urban village building samples were drawn by LabelMe software. This study provides basic data for remote sensing image analysis based on deep learning, specifically the case segmentation algorithm mask R-CNN, and includes an application case of the case segmentation sample. This work has practical significance for applying artificial intelligence information extraction in urban governance.

2 Metadata of the Dataset

The metadata of the “Training Samples Dataset for Building Identification in the Urban Village” is summarized in Table 1. It includes the dataset full name, short name, authors, year of the dataset, temporal resolution, spatial resolution, data format, data size, data files, data publisher, and data sharing policy, etc.

Table 1  The metadata summary of the dataset

3 Data Development Methods

The sample of remotely sensed images were divided into target detection samples, semantic segmentation samples, and instance segmentation samples according to their specific uses^[6]. The samples used for target detection must have the location and type of the target feature labeled, i.e., by drawing the external rectangular box of the target feature and labeling its category; for semantic segmentation, its sample needs to have the outline and type of the target feature labeled, i.e., by drawing the outline of the target feature and labeling its category; for instance segmentation, its sample should have the outline and the category of the target feature marked, that is, by drawing the outline of a single object and labeling its category. Currently, the most commonly used software tools for drawing on images are LabelMe, ArcGIS, and LabelImg.

According to the remote sensing imagery and from the ground real-scene photos, this paper used LabelMe software to obtain the building instance segmentation samples from an urban villages, which were later used for deep learning by the instance segmentation

algorithm.

(1) Imagery selection

Combined with the unique distribution pattern of the urban village building community, high building density, narrow streets and lanes, an image captured via Google at a resolution of 0.11 m was selected as the image data.

(2) Image segmentation

An image was divided into the target size, which was generally an exponential square with side length of 2. The original image and their labels were subset into 512 × 512 pixels for subsequent model training. After the original image was segmented the pic file was obtained, which was the sample’s original image set Buildingsample_pic, as shown in Figure 1-a.

(3) LabelMe draws the buildings in the village in the city

Draw the outline of the building in LabelMe and mark it in the form of vbuilding*. Each building is named vbuilding1, vbuilding2, vbuilding3 in turn vbuilding*.

(4) Format conversion

According to the JSON file generated by LabelMe, the sketch sample was converted to an executable dataset format. Next, the mask images generated were sorted to derive the mask file of the instance segmentation result, which was the sample’s segmentation results (Buildingsample_mask), as shown in Figure 1-b. Instance segmentation annotation information info file, which was the sample’s segmentation annotation (Buildingsample_info) Each color shown in Figure 1-b represents a building.

Figure 1  Google image and mask of the sample

(5) Sample increase

The generated mask image (mask) and original image (pic) were flipped horizontally, vertically, and rotated to 90°, rotated to 180°, and rotated to 270° to increase the number of samples, as shown in Figure 2.

Figure 2  Schematic diagram of data expansion

4 Data Results and Validation

4.1 Data Proudcts

The dataset includes 3 data files (Table 2). A total of 2,328 urban village building samples were drawn.

Table 2  File description of the data product

4.2 Validation

The case segmentation algorithm, mask R-CNN, was used to extract building information^[10-13], and 678 urban village building samples were tested and verified. The algorithm of extracting village buildings in city by mask R-CNN is shown in Figure 3.

To evaluate the performance of the algorithm, average precision (AP) was used as the evaluation metric of experimental accuracy. AP is the area formed by the accuracy recall curve and X and Y axes, which is calculated by equation (1). The higher the AP, the better the performance of the model, and vice versa. Therefore, the calculation of AP involves the calculation of both precision (p) and recall (r). The precision refers to the ratio of TP (True Positive) to the number of all detected targets, as shown in equation (2). Recall rate refers to the ratio of TP (True Positive) to all actual target numbers, as shown in equation (3).

                                               (1)

                                                                                                                                             (2)

                                                                                                 (3)

Table 3  Evaluation metrics of target detection

Figure 5  Comparison images of the test results

The results show that the average building area of the experimental area is 75.08 m², and the average nearest-neighbor distance is 0.90 m. According to the kernel density estimation results, the building density of the studied area is 43.75%, and the green space rate is 5.12%^[14]. According to the national standard of the code for planning and design of urban residential areas of China, the area is a high-density residential area^[15].

5 Conclusion

The dataset is based on 0.11-m spatial resolution of remote sensing imagery produced by Google Maps, for which the location, outline, and type of each single building in a city’s village was marked. According to the sample, we demo an application case of case segmentation of single building in urban village. Our experimental results show the followings:

(1) The network structure of mask R-CNN has advantages in building target detection. The dataset has high practicability when using an instance segmentation algorithm mask R-CNN to extract information by deep learning. The AP of the building sample in the urban village reaches 0.66, and the highest detection accuracy AP of the single image of the building sample in the urban village reaches 0.995. Through testing and verifying 678 urban village building samples, the quality of the urban village building sample datasetis proved to be effective, which provides a sample deep learning case segmentation.

(2) The results of spatial analysis of the information extraction show that the spatial analysis can effectively reflect the distribution characteristics of small average building area, narrow streets, high density of buildings and complex building types.

The dataset provides the basic data for the use of remotely sensed images based on a deep learning algorithm to extract the buildings in urban villages. It offers good practical significance for studying the spatial distribution characteristics of urban villages and the intelligent analysis and application of urban villages’ governance.

Author Contributions

Wu, T. designed the technical route of the dataset development; Liu, Y. F. and Lu, B. R. collected and processed the sample data of urban villages; Lu, B. R. designed the models and algorithms; Lu, B. R. and Liu, S. validated the results; Peng, L. was responsible for data organization, sample types, and production process, as well as value judgment; Liu, Y. F, and Lu, B. R. wrote the manuscript.

References

[1]        Li, Z. Y., Yang, Y. C. Research progress of urban village in China [J]. Gansu Science and Technology, 2008 (7): 7–11.

[2]      Zhou, X. H. Urban village problem: an economic analysis of its formation, existence and transformation [D]. Shanghai: Fudan University, 2007.

[3]      Deng, C. Y., Wang, Y. R. A review of the research on urban villages in China [J]. Journal of Guangdong University of Administration, 2008, 20(1): 93–97.

[4]        Zhao, Y. H., Chen, G. Q., Chen, G. L., et al. Extraction of urban village buildings from multi-source big data: a case study of Tianhe district, Guangzhou city [J]. Geography and Geographic Information Science, 2018, 34 (5): 3, 13–19.

[5]      Liang, Y. D. Research on the application of UAV system in urban village reconstruction [J]. Beijing Surveying and Mapping, 2018, 32 (10): 70–73.

[6]      Mayunga, S. D. Semi-automatic building extraction in informal settlements from high-resolution satellite imagery [D]. Canada: University of New Brunswick, 2006: 1–298.

[7]      Cheng, T. Construction and application method of big data of remote sensing image sample [J]. Application of Computer System, 2017, 26(5): 43-48.

[8]      Liu, Y. F, Lv, B. R., Peng, L., et al. Training samples dataset of building identification in urban village [J/DB/OL]. Digital Journal of Global Change Data Repository, 2020. DOI: 10.3974/geodb.2020.02.16.V1.

[9]      GCdataPR Editorial Office. GCdataPR Data Sharing Policy [OL]. DOI: 10.3974/dp.policy.2014.05 (Updated 2017)

[10]   Ji, S. P., Wei, S. Q. Convolutional neural network and open source dataset method for building extraction from remote sensing images [J]. Acta Sinica Sinica, 2019, 48 (4): 50–61.

[11]   Lin, T. Y., Dollár, P., Girshick, R., et al. Feature pyramid networks for object detection. [C]. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), IEEE Computer Society, 2017.

[12]   Hirata, T., Kuremoto, T., Obayashi, M., et al. Deep belief network using reinforcement learning and its applications to time series forecasting [C]. International Conference on Neural Information Processing. Springer International Publishing, 2016.

[13]   Fu, F., Wei, J. Y., Zhang, L. N. Research on building extraction from remote sensing image based on convolution network [J]. Software Engineering, 2018, 21 (6): 8–11.

[14]   Lv, B. R., Peng, L., Wu, T., et al. Research on urban building extraction method based on deep learning convolutional neural network [J]. Earth and Environmental Science, 2020, 502: 012022.

[15]   GB 50180—93. Code for planning and design of urban residential areas [S]. Beijing: China Construction Industry Press, 2002.