Vectorized dataset of roadside noise barriers in China using street view imagery

. Roadside noise barriers (RNBs) are important urban infrastructures to ensure that cities remain live-able. However, the absence of accurate and large-scale geospatial data on RNBs has impeded the increasing


Introduction
In recent years, several studies have documented the substantial impact of traffic noise problems in cities (Apparicio et al., 2016;Begou et al., 2020).Roadside noise barriers (RNBs) are a vital urban infrastructure that contribute significantly to mitigate undesirable traffic noise in communities (Abdulkareem et al., 2021;Ning et al., 2010).Additionally, RNBs contribute to the development of sustainable cities in many ways.For example, with the emphasis on new energy, RNBs are being used to install solar photovoltaic panels, thereby increasing the utility of new energy sources (Gu et al., 2012;Zhong et al., 2021).The reasonable presence of RNBs also enables the airflow in the urban canyon region to be adjusted, thereby improving the roadside air quality (Huang et al., 2021;Zhao et al., 2021).Because of the importance of RNBs in building sustainable cities, the demand for RNBs has increased alongside traffic growth in recent decades (Den Boer and Schroten, 2007;Oltean-Dumbrava and Miah, 2016).There are bottomup benefits from establishing an accurate and standardized large-scale RNB dataset with detailed geospatial information about RNBs, including their mileage, location, and distribution (Liu et al., 2020;Wang and Wang, 2021).Specifically, precise RNB locations enable traffic departments to effectively manage and maintain this type of infrastructure (Sainju and Jiang, 2020), urban research can simulate dynamic cities based on accurate RNB geospatial information (Wang and Wang, 2021;Zhao et al., 2017), and governments can rely on the RNB maps to examine urban layouts and create green and sustainable cities (Song et al., 2021;Song and Wu, 2021).
Over the past few years, extensive geospatial databases have been established to store data on many aspects of urban infrastructure (Griffiths and Boehm, 2019;Perkins and Xiang, 2006).However, the sharing and exchange of RNB data in these databases are restricted, and the data only cover a limited geographic area (Wang et al., 2019;K. Zhang et al., 2022).These challenges to data acquisition are because databases have to adhere to various standards related to geographic data (e.g., file format and geographic coordination reference; Lafia et al., 2018).On the other hand, the RNB data are often created and updated manually through road inspections and investigations which are costly and time consuming, especially on a large scale (Potvin et al., 2019;Ranasinghe et al., 2019).The RNB geospatial dataset must be generated, and kept up to date, as soon as possible using alternative, efficient methods.
Street view imagery is georeferenced data densely covering the road network of cities.As a new geospatial data source, it provides depictions of real-world surroundings, in-cluding natural landscapes and the built environment, and enables users to recognize physical objects, urban dynamics features, and geographic scenes on a large scale (Zhang et al., 2018).In addition, as part of the data sharing movement, an increasing number of community-based organizations and corporations, such as Baidu Maps, Tencent Maps, and Google Maps, are regularly generating and updating open-access street view imagery (Qin et al., 2020;Zhang et al., 2019).Such big data bring great prospects for acquiring urban infrastructure information (e.g., RNBs), with benefits such as broad coverage, a rapid update speed, and low acquisition costs (Kang et al., 2020).However, manual interpretation is a tedious task, and conventional computer vision algorithms struggle when confronted with large amounts of data and complex image features (Zhang et al., 2018).
With the advancement of computing hardware and frameworks, deep learning methods now have an increased capacity for extracting semantic features from a large amount of data (Lecun et al., 2015;Liu et al., 2022).The emerging approaches are increasingly being used to interpret physical objects and detect interior patterns from Earth observation data (Z.Zhang et al., 2022;Qian et al., 2022).Meanwhile, image classification based on deep learning has been used to identify RNBs using street view imagery (Zhong et al., 2021).However, for the purposes of identifying RNBs, prior geographic knowledge, which is essential, is frequently overlooked, such as the fact that RNBs are frequently located between roads and densely populated regions (e.g., residential, educational, and medical areas; Arenas, 2008;Wang et al., 2018;K. Zhang et al., 2022).In recent years, a new framework of data-driven research based on geospatial artificial intelligence (GeoAI) and machine learning has resulted in multiple notable improvements in the discovery of geographic scene knowledge (Goodchild and Li, 2021;Li, 2020).When empirical and prior spatial information are included into deep learning approaches, they can help to develop a more holistic understanding of a research subject and mitigate the effects of data scarcity or representational bias (Janowicz et al., 2019;Qian et al., 2020).As a result, it is possible to enhance the effectiveness of deep learning methods in identifying RNBs by incorporating some prior geographic knowledge from street view imagery.Additionally, Wolpert and Macready (1997) introduced the "no free lunch" theory, demonstrating that a single model must pay for some accuracy by degrading its generalizability.This is acceptable, as it is challenging to construct a perfect solution for all scenarios using a single model, particularly when dealing with vast volumes of data and large-scale areas (Wang and Li, 2021).
The purpose of this study is to build an accurate and nationwide vectorized RNB dataset utilizing Baidu Street View (BSV) imagery.To improve the performance for the detection of RNBs, this work proposes a GeoAI framework.Concretely, an ensemble of convolutional neural networks incorporating image context information (IC-CNNs) is developed, which considers the prior geographic knowledge contained in street view images.Subsequently, a post-processing method is applied to generate the vectorized RNB dataset based on the identified RNB locations.Last, the RNB dataset quality is quantitatively evaluated from two perspectives, i.e., the detection accuracy and the completeness and positional accuracy.The main contributions of this study can be summarized as follows: 1.This study provides the first reliable and nationwide vectorized RNB dataset in China and provides labeled BSV images which can be used as a benchmark dataset.
2. A GeoAI framework is presented for the processing of numerous BSV images in order to generate the RNB mapping and for the comprehensively evaluation of the generated results.
3. This study presents multiple IC-CNNs based on prior geographic knowledge and an ensemble learning strategy to achieve high-performance object identification from street view imagery.
The remainder of this paper is organized as follows.Section 2 briefly describes the data and methods used to generate and evaluate the RNB dataset.Section 3 presents the results of the RNB mapping and an evaluation and analysis for the RNB dataset.Section 4 discusses the capability of proposed methods, as well as the challenges and limitations of this work.
The last section provides the conclusions of this study.

The GeoAI framework
The GeoAI framework's workflow is divided into three stages: data preparation, modeling, and evaluation, as shown in Fig. 1.To begin with, BSV images are gathered during the data preparation stage using OpenStreetMap (OSM) road data and the BSV application programming interface (API).Subsequently, BSV images are used to generate various samples for modeling and evaluation.During the modeling stage, deep learning approaches are used to detect RNBs from the BSV imagery.Using the vectorization post-processing method, the identified and scattered RNB locations are subsequently processed into a vectorized dataset.During the evaluation stage, the quality of the created dataset is quantitatively assessed in two aspects, i.e., the detection accuracy and completeness and positional accuracy.

Data preparation
There are three types of data are acquired for this study, i.e., the road networks, administrative boundary, and street view imagery.Afterwards, training, validation, and test samples are collected based on these data.The data from Taiwan Province are scarce.

Road networks
The road networks were downloaded from OSM (https: //www.openstreetmap.org/,last access: 16 May 2021) in May 2021, which are polyline-based and include a variety of road types, including motorways, trunk roads, primary roads, and secondary roads.According to previous findings, the quality of OSM road networks in China is high in terms of completeness and positional accuracy (Liu and Long, 2015).
In addition, RNBs have a high probability of being installed on motorways and trunk roads (K.Zhang et al., 2022).Therefore, given the expense of acquiring and computing BSV images, in this study, samples on motorways and trunk roads are only considered for downloading BSV images.Figure 2a depicts the spatial distribution of these two types of roads.

Administrative boundary
The city boundary was acquired from http://bzdt.ch.mnr.gov.cn/ (last access: 21 April 2021) in April 2021.According to the urban management hierarchy established by the Chinese government, cities in China are divided into four tiers (Guan and Rowe, 2018;Jia et al., 2020), including municipalities, sub-provincial cities, and prefecture-level cities, and their locations are shown in Fig. 2b.Specifically, tier 1 is centrally administered cities and municipalities.Tier 2 is primarily sub-provincial cities, whereas tier 3 is province capitals and large prefecture-level cities. Tier 4 is ordinary prefecture cities. Cities with varying administrative levels have varying authorities over resource allocation and jurisdiction (Guan et al., 2018).

Street view imagery
With their high-resolution and detailed information on Chinese streets, BSV images are of comparable quality to Google Street View images, which are not available in China (H.Zhou et al., 2019).Numerous sample points along OSM roads are collected, and the BSV API is utilized to obtain street view images at those locations.Following the work of K. Zhang et al. (2022), a sampling interval of around 25 m is utilized to account for the tradeoff between data granularity and the expense of downloading imagery.As a result, the total number of sample points is 24 871 839.As shown in Fig. 3   to identify the RNBs along the corresponding roadside, BSV images with a 90 • viewing angle are acquired.Owing to the absence of BSV images on a few road segments in a particular year, these will be supplemented in adjacent years.
Additionally, the BSV sensors may be obstructed by some vehicles or other surrounding objects.These issues are resolved through the use of multitemporal BSV images (ones from 2013 to 2021 are downloaded in this study).A total of 6 008 674 BSV images are downloaded with a size of 500 pixels × 400 pixels, and their spatial locations are shown in Fig. 2c. Figure 4 depicts the spatial distribution of the number of BSV images in China, with the eastern region and higher city tiers having a greater number of BSV images.Previous investigations revealed that BSV images with RNBs are rare, accounting for less than 5 % of the samhttps://doi.org/10.5194/essd-14-4057-2022 Earth Syst.Sci.Data, 14, 4057-4076, 2022    RNBs are widely placed on the roadside in densely populated regions, such as residential areas and educational and government institutions, as previously described in other studies (Arenas, 2008;Wang et al., 2018;K. Zhang et al., 2022).Therefore, based on this prior geographic knowledge, an IC-CNN that leverages the context information contained in BSV images is developed which aims at enhancing the RNB detection accuracy.Figure 7 illustrates the construction of IC-CNN, which adopts the ResNet architecture (He et al., 2016).In this workflow, prior geographic knowledge is incorporated into the neural network by means of transfer-ring learning.Initially, 500 samples are randomly selected from positive and negative training samples in each tier.
There are three context labels added, depending on the context of these BSV images, i.e., building dominated, nonbuilding dominated, and uncertain (unable to judge the background of the BSV image because it is obscured by objects), as shown in Fig. 6.The context labels are interpreted by semantic segmentation models released by the MIT Computer Vision team (B.Zhou et al., 2019).Besides the sky and ground objects, images are judged to be building dominated if the ratio of building objects is the most; otherwise, they are evaluated to be non-building dominated.Additionally, the uncertain type is classified by a visual interpretation of whether the background environment in the image is obscured.These labeled images are available at https: https://doi.org/10.5194/essd-14-4057-2022Earth Syst.Sci.Data, 14, 4057-4076, 2022 //doi.org/10.11888/Others.tpdc.271914(Chen, 2021).Next, 4000 samples with image context labels are used to train the IC-CNN on a preliminary basis, using hybrid loss to optimize the parameters in IC-CNN for image context and RNB identification, as formulated in Eq. ( 1).After the network has converged, the IC-CNN's classifier is replaced with a binary classification, and all the training samples are supplemented to fine-tune and intensively train the network.
Hybrid loss = CE p image context + 2 × CE(p noise barrier ), ( 1) where p image context is the confidence of the image context identification, p noise barrier is the confidence of the RNB identification, and CE(p) refers to the cross-entropy loss function (Hu et al., 2018).

Ensemble learning strategy
Owing to the high cost of labeling and the restricted quantity of trained samples, an ensemble learning strategy for enhancing RNB detection accuracy is utilized in this study based on the "no free lunch" theory (Wolpert and Macready, 1997).In an ensemble learning domain, the effective strategy to boost performance is to integrate the numerous high-variance models together (Cao et al., 2020).Therefore, this study integrates four IC-CNNs, and their convolutional layers are chosen from the ResNet family (He et al., 2016;Zagoruyko and Komodakis, 2016), including ResNet101, ResNet152, Wide ResNet50, and Wide ResNet101.The integration of the four IC-CNNs with varying capacities for feature extraction can make a significant contribution to achieving high detection accuracy.

Vectorization post-processing
After performing a detection run by an ensemble of IC-CNNs, the identified and scattered RNB locations are connected to create a vectorized RNB dataset by a postprocessing technique, which is based on the spatial neighbor relationship between samples.Specifically, if adjacent sample images of the same road contain RNB objects, their locations will be connected.Furthermore, the findings of Sainju and Jiang (2020) demonstrated that the "near objects are more related" principle (Tobler, 1970(Tobler, , 2004) holds true when using street view imagery to detect objects at the urban scale.Therefore, in this study, given the likelihood of RNB misidentification, if a sample image is flanked by images containing RNBs in the same road, it will be considered as a positive type to minimize the impact of misidentification.

Metrics for detection accuracy
To evaluate the accuracy of RNB detection, four quantitative metrics in the deep learning classification task, including overall accuracy (OA), recall, precision, and F 1 score (Thomas et al., 2020)

Metrics for completeness and positional accuracy
To quantitatively evaluate the completeness and positional accuracy of generated RNBs, two quantitative metrics, including the root mean squared error (RMSE) and the intersection over union (IoU) are adopted (Rezatofighi et al., 2019).To calculate these metrics, numerous roads are selected from various cities and are surveyed manually as ground truths based on BSV imagery.Based on the mileage deviation and overlap relationship between the generated and surveyed RNBs, RMSE and IoU are calculated following Eqs.( 7) and ( 8), respectively: where m is the number of selected roads, l i is the surveyed RNB mileage of the ith road, and li is the generated RNB mileage of the ith road.
where L intersection is intersection mileage of the generated and surveyed RNB, and L union is union mileage of the generated and surveyed RNB.

Implementation configuration
Several techniques to enhance the performance of the model throughout the training and inference stages are employed in this study.Data augmentation techniques such as random resized cropping and random horizontal flipping are utilized to increase the data volume and decrease model bias error.The model parameters are optimized using the cosine annealing learning rate scheduler (Bhattacharyya et al., 2021) and AdamW optimizer (Loshchilov and Hutter, 2017).Long training and inference resized tuning (Touvron et al., 2019) are employed to improve the model's performance.Finally, an ensemble of models identifies RNBs based on the voting mechanism.

RNB mapping result
The final RNBs dataset is available at https://doi.org/10.11888/Others.tpdc.271914(Chen, 2021).Details of the BSV image identification results are shown in Appendix A, and details of RNB mileage by city in China are shown in Appendix B, with the total RNB mileage of 2667.02km and the average RNB mileage for each city tier of 102.39 km (±117.83km), 66.36 km (±18.70 km), 22.19 km (±12.52 km), and 1.12 km (±0.42 km), respectively.The quantitative results suggest that there are substantial variations between the different city tiers.Tiers 1 and 2 contain a major portion of the total RNB mileage compared with the other city tiers; moreover, confidence intervals show that the higher the city tier, the greater the difference in the level of RNB construction in that city tier.The reason for these variances is that the unique urban administration system in China mandates lower-tier cities to rigidly follow the leadership of higher-tier cities (Ma, 2005;Zhao et al., 2003).
Higher-tier cities are rapidly increasing in size and occupying considerable resources, while lower-tier cities are developing slowly (Au and Henderson, 2006;Lin, 2002).The spatial distribution of RNB mileage among cities is further depicted in Fig. 8, where blank areas indicate the absence of RNBs or BSV images (there are 17 cities lacking BSV images, as shown in Appendix B). Figure 8 suggests that RNBs in eastern China are more densely distributed and have longer mileage.To a certain extent, it shows that the statistics correlate with the development of Chinese cities, implying that higher-tier cities have a high probability of covering and updating BSV imagery or laying down RNBs.
After analyzing the generated RNB dataset from a national scale, three cities with the highest RNB mileage in each tier are selected to analyze the citywide mapping results, as shown in Fig. 9.The figure shows that RNBs are generally clustered in the central areas of these cities.For example, the RNBs in Shanghai are mainly clustered on the third ring road, while those in Beijing are mainly clustered on the sixth ring road.As a result, when combined with the planned layout and actual mapping of RNB distribution, the generated RNB dataset can partially reflect the rationality of urban infrastructure planning and layout.

RNB detection accuracy
Table 2 summarizes the evaluation results of the RNB identification at different city tiers based on test samples.The OA and the F 1 score for the overall city tiers are 98.61 % and 81.44 %, respectively.However, the accuracy is greater for higher-tier cities than for lower-tier cities.This may be attributed to the fact that cities with lower tiers appear to have less RNB infrastructure, resulting in a more severe class imbalance problem for deep learning methods, which impacts the training and generalization of the model.Therefore, the results indicate that, prior to using this dataset, an assessment of the influence of regional quality differences on specific applications is required.

RNB completeness and positional accuracy
To evaluate the completeness and positional accuracy of the RNB dataset, approximately 254.45 km of roads are selected from different city tiers and manually surveyed using the BSV imagery.Appendix C summarizes the detailed quantitative differences between generated and surveyed RNBs in terms of mileage deviation and level of overlap.The overall RMSE for the mileage deviation is 0.08 km, and the IoU for the overlay level is 88.08 % ± 2.95 %.The results shows that the generated and surveyed RNBs are highly consistent in terms of mileage and distribution, demonstrating the high completeness and positional accuracy of the generated RNB dataset.Moreover, as illustrated in Fig. 10, a visual comparison between surveyed and generated RNBs on various roads depicts that the generated and surveyed RNBs on the road are overall consistent in terms of mapping.However, several validated points demonstrated that the proposed deep learning approach incorrectly recognized small RNB objects in the images, such as validated points IV, II, and III on Beijing's Jingmen Highway, Zhengzhou's Longhai Expressway, and Wenzhou's Ouhai Boulevard, respectively.Additionally, several objects that looked similar to RNBs, such as multiwindowed buildings, are misclassified as a positive type, for example, point IV on Wenzhou's Ouhai Boulevard and points II and III on Nantong's Binjiang Bridge.Despite these misclassifications, most of the validated points demonstrated a high accuracy of the RNB prediction and the high perfor-mance of the proposed framework, implying the reliability of the generated RNB dataset.

Model capability
An ablation study is conducted to demonstrate the quality of the generated dataset and validate the effectiveness of developed methods (Table 3).As shown in Table 4, the combination of proposed strategies achieves the highest performance.The ablation results illustrate that the effectiveness of proposed strategies, including integrating image context information into CNN, adding confusing negative samples, and using an ensemble learning strategy.Additionally, Fig. 11 depicts the areas of the IC-CNN's attention, revealing that IC-CNNs not only have a capacity for focusing on RNB objects in BSV images but also have a sense of their surroundings.The results suggest the reliability of the generated dataset and partially decipher the "black box" of deep learning to explain the high performance of the developed methods.Notably, this study successfully achieves incorporating some of the prior geographic knowledge into the deep learning method.RNB detection accuracy can be increased further by combining more comprehensive knowledge of geographic scenes from BSV images into deep learning network, such as various geographic elements and processes and the associated construction theory (Lü et al., 2018).

Limitations and future work
This study has several limitations in the process of dataset generation which can be grouped into three categories, namely data source, ground scenario, and modeling.https://doi.org/10.5194/essd-14-4057-2022 Earth Syst.Sci.Data, 14, 4057-4076, 2022 Due to the economic status, topographical conditions, or government policies, not all Chinese cities are covered by BSV imagery, with data not being available for 17 cities (Deng et al., 2021;Du et al., 2020).In addition, challenges owing to overexposure or obstruction of the sensors by vehicles hinder the capturing of a complete street scene.As a result, the natural characteristics of the data source can have certain impacts on the accuracy of the RNB dataset.
The road/traffic environment is often complex.Concretely, BSV sensors can detect RNBs on distant highways or other lanes, and it may result in some mistakes during RNB detection and mapping.However, the likelihood of this occurring is small (about 4 % of RNB samples) by sampling investigation.
This study implicitly presupposes that BSV images are independent and identically distributed.As shown in Fig. 9, the developed GeoAI framework can achieve a high performance in continuous RNB mapping.However, the spatial autocorrelation effect in BSV images is overlooked, as BSV images taken along the same road network path frequently resemble the adjacent one (Sainju and Jiang, 2020).
Moreover, there are some uncertainties in cities with short mileage RNBs which may be generated due to misidentification.A manual survey is performed to verify the confidence level of these cities.Table 5 shows the quantitative results, which indicate that the shorter the RNBs, the lower the confidence level.In addition, the results show that the confidence level is lowest for cities with RNBs of less than 0.2 km, so further validation is needed when applying them in specific applications.
In the future, to address the data shortage issue, more data sources, such as Google Maps and Tencent Maps, will be used.Additionally, approaches for photogrammetry and image scene understanding techniques will be developed to tackle the complex ground scenario.Finally, end-to-end deep learning algorithms will be constantly enhanced by the addition of more powerful units and structures to account for spatial autocorrelation in street view imagery.

Data availability
The road networks come from OSM (https: //www.openstreetmap.org/,OpenStreetMap contributors, 2021), a collaborative project dedicated to providing many types of freely editable geographic data for the world.City boundaries can be obtained from http://bzdt.ch.mnr.gov.cn/(Map Technical Review Center, Ministry of Natural Resources, China, 2021).In addi-tion, BSV images can be downloaded by using BSV API (https://api.map.baidu.com/panorama/v2?key=parameters,Baidu Maps, 2022).Finally, the generated RNB dataset, labeled BSV image benchmark, and RNB detection results are available to the public at https://doi.org/10.11888/Others.tpdc.271914(Chen, 2021).Specifically, the generated RNB dataset is grouped by city level, with attributes of city tier, city name, province, and RNB mileage; the image labels are documented in a * .csvfile for a benchmark of the image context, and all images are categorized into specific folders.The RNB detection results include meta information of all BSV samples, such as longitude, latitude, city name, city tier, timing of imaging, and detection label (0 presents non-RNB type, while 1 presents RNB type).The mileage in RNB dataset is calculated with the Albers equal-area conical projection.

Code availability
The codes of deep learning approaches in this study are available at https://doi.org/10.11888/Others.tpdc.271914(Chen, 2021) and https://github.com/ChanceQZ/NoiseBarrierIdentification (last access: 20 January 2022 ).Python 3 packages such as PyTorch, NumPy, and OpenCV are used to develop the code.The vectorization post-processing procedure is performed in the ArcGIS Pro platform.

Conclusion
This study presents the first nationwide vectorized dataset of RNB and the benchmark dataset of the labeled BSV images in China using BSV imagery and a GeoAI framework.In this study, based on prior geographic knowledge in BSV imagery, RNB samples are identified based on deep learning approaches, and the vectorized RNB dataset is subsequently constructed using the vectorization post-processing procedure.The created RNB dataset is evaluated from two perspectives, i.e., the detection accuracy and the completeness and positional accuracy.The four quantitative metrics, OA, recall, precision, and F 1 score, with values of 98.61 %, 87.14 %, 76.44 %, and 81.44 %, illustrate high accuracy of the model in RNB detection.The level of mileage deviation and overlay between the generated and surveyed RNBs are further determined via a manual survey of around 254.45 km of roads in various cities, with the RMSE of 0.08 km and IoU of 88.08 % ± 2.95 % revealing that the created and surveyed RNBs are consistent and reliable.
The intended applications for the two datasets are diverse.In terms of the vectorized dataset of RNBs, urban studies can benefit from accurate information of RNB mileage, location, and distribution.For example, the regional energy potential of solar photovoltaic panels on RNB can be estimated, finer 3D urban models are able to be developed, and the sustainability of urban layouts can be evaluated.On the other hand, the benchmark dataset of labeled BSV images may contribute to multiple other research and applications related to RNBs identification, such as developing advanced deep learning algorithms and fine-tuning existing computer vision models to detect RNBs more accurately and exploring the further relationship between the RNB locations and surrounding environment.

Figure 1 .
Figure 1.The flowchart of the GeoAI framework to generate the vectorized RNB dataset.

Figure 2 .
Figure 2.There are three data sources used in this study.OSM road network data (a), the Chinese administrative boundary, with four city tiers (b), and the spatial distribution of the BSV image locations (c) are shown.Road networks are from © OpenStreetMap contributors (2022) and distributed under the Open Data Commons Open Database License (ODbL) v1.0.

Figure 4 .
Figure 4. Zonal statistics of the number of BSV images in China.

Figure 5 .
Figure 5. Illustration of BSV samples, including four typical types of RNBs based on physical shapes (a) and three types of confusing negative samples which look like RNBs (b) (BSV images are from © Baidu Maps, 2022).

Figure 6 .
Figure 6.The flowchart of BSV image labeling.

Figure 7 .
Figure 7.The construction of the convolutional neural network incorporating image background information (BSV images are from © Baidu Maps, 2022).

Figure 8 .
Figure 8. Zonal statistics of RNB mileage in China.The blank areas indicate no RNBs or a lack of BSV images.

(
TN), true positive (TP), and false positive (FP) images is calculated.True positive means that the prediction and ground truth of images are both positive.Conversely, false negative means the predictions are negative while the ground truths are positive.The four metrics are calculated based on the following (Thomas et al., 2020): OA = TP + TN TP + FP + TN + FN ,

Figure 9 .
Figure 9. Distribution of RNBs in several representative cities (base maps are from Esri).

Table 1 .
Details of sample collection results.
are analyzed.Due to the class imbalance problem in BSV images, OA is susceptible to being affected by a large number of sample types in this study (i.e., negative type sample).In comparison, precision and recall can concentrate on positive type samples.The F 1 score is the most comprehensive of these metrics because it considers both precision and recall.After detecting the RNBs in BSV images, the number of false negative (FN), true negative

Table 2 .
Evaluation results of RNB identification in different city tiers.The evaluation results of every city tier are calculated using the test samples of the corresponding city tier, while the overall evaluation results are calculated using the entire test samples.

Table 3 .
Ablation study design.The ablation study combines the four strategies used in this study to illustrate their effectiveness.

Table 4 .
Quantitative results of ablation.The ablation results show that the proposed methods have the highest RNB detection accuracy.The bold values indicate the highest value in each metric.

Table 5 .
Confidence assessment in the mapping accuracy for cities with low-mileage RNBs.

Table A1 .
Details of the BSV image identification results.

Table A2 .
Identification of a confusion matrix based on test samples.