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.

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).

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, an illustration of the BSV images, with photographs showing different directions, shows that a BSV image with a 90∘ viewing angle is more appropriate for the present work because it provides a comprehensive roadside view. Hence, 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.

An effective sampling technique for generating training, validation, and test image samples is developed to detect RNBs from the large volume of BSV images collected. According to their physical shapes, the RNBs identified in this study can be categorized into the following four distinct types: upright noise barrier, top curved noise barrier, noise barrier with folded corners at the top, and large curved noise barrier, as depicted in Fig. 5. Figure 1 illustrates the different steps followed in the data preparation stage. The BSV images are classified into four tiers based on their location within the city administration hierarchy. Subsequently, the training, validation, and test sampling set are subdivided from the entire images, accounting for 60 %, 20 %, and 20 % of images, respectively. These sampling sets can be used to collect the corresponding samples and are beneficial in that they avoid the mixing of samples.

Previous investigations revealed that BSV images with RNBs are rare, accounting for less than 5 % of the sampled images. To alleviate the impact of the class imbalance problem on model training, 50 000 images are randomly selected from each city tier based on the training sampling set. These samples are labeled as positive type (i.e., image with RNB) or negative type (i.e., image without RNB) by manual visual interpretation, the details of which are shown in Fig. 6. Subsequently, the same number of positive and negative samples are maintained. Certain objects, such as tunnel inner walls, billboards, and guardrails, seem like RNBs in images, which intensifies the difficulty of deep learning, as shown in Fig. 5. Therefore, 500 images of each of these objects are added as confusing negative samples to the training samples. The ultimate training sample size is 14 484, including 6492 positive and 7992 negative samples. To generate the validation and test samples, 500 and 2500 image samples from each city tier are chosen. There are 79 positive samples and 1921 negative samples in the validation samples, while there are 350 positive samples and 9650 negative samples in the test samples. The details of the sample collection results are shown in Table 1. The labeled BSV images are available at https://doi.org/10.11888/Others.tpdc.271914 (Chen, 2021).

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 transferring 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, non-building 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://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. 1Hybrid loss=CEpimage context+2×CE(pnoise barrier),2CEp=-∑p⋅log⁡(p), where pimage context is the confidence of the image context identification, pnoise barrier is the confidence of the RNB identification, and CE(p) refers to the cross-entropy loss function (Hu et al., 2018).

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.

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 post-processing 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, 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.

To evaluate the accuracy of RNB detection, four quantitative metrics in the deep learning classification task, including overall accuracy (OA), recall, precision, and F1 score (Thomas et al., 2020) 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 F1 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 (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): 3OA=TP+TNTP+FP+TN+FN,4Precision=TPTP+FP,5Recall=TPTP+FN,6F1score=2⋅Precision⋅RecallPrecision+Recall,

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: 7RMSE=1m∑i=1m(li-li^)2, where m is the number of selected roads, li is the surveyed RNB mileage of the ith road, and li^ is the generated RNB mileage of the ith road. 8IoU=LintersectionLunion, where Lintersection is intersection mileage of the generated and surveyed RNB, and Lunion is union mileage of the generated and surveyed RNB.

Table 2 summarizes the evaluation results of the RNB identification at different city tiers based on test samples. The OA and the F1 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.

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 multi-windowed 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 performance of the proposed framework, implying the reliability of the generated RNB dataset.