Search

Purpose: Bone age (BA) is needed to assess developmental status and growth disorders. We evaluated the clinical performance of a deep-learning-based BA software to estimate the chronological age (CA) of healthy Korean children. Methods: This retrospective study included 371 healthy children (217 boys, 154 girls), aged between 4 and 17 years, who visited the Department of Pediatrics for health check-ups between January 2017 and December 2018. A total of 553 left-hand radiographs from 371 healthy Korean children were evaluated using a commercial deep-learning-based BA software (BoneAge, Vuno, Seoul, Korea). The clinical performance of the deep learning (DL) software was determined using the concordance rate and Bland-Altman analysis via comparison with the CA. Results: A 2-sample t-test (P<0.001) and Fisher exact test (P=0.011) showed a significant difference between the normal CA and the BA estimated by the DL software. There was good correlation between the 2 variables (r=0.96, P<0.001); however, the root mean square error was 15.4 months. With a 12-month cutoff, the concordance rate was 58.8%. The Bland-Altman plot showed that the DL software tended to underestimate the BA compared with the CA, especially in children under the age of 8.3 years. Conclusion: The DL-based BA software showed a low concordance rate and a tendency to underestimate the BA in healthy Korean children. [ABSTRACT FROM AUTHOR]

Background: This study aimed to assess the error range of cephalometric measurements based on the landmarks detected using cascaded CNNs and determine how horizontal and vertical positional errors of individual landmarks affect lateral cephalometric measurements. Methods: In total, 120 lateral cephalograms were obtained consecutively from patients (mean age, 32.5 ± 11.6) who visited the Asan Medical Center, Seoul, Korea, for orthodontic treatment between 2019 and 2021. An automated lateral cephalometric analysis model previously developed from a nationwide multi‐centre database was used to digitize the lateral cephalograms. The horizontal and vertical landmark position error attributable to the AI model was defined as the distance between the landmark identified by the human and that identified by the AI model on the x‐ and y‐axes. The differences between the cephalometric measurements based on the landmarks identified by the AI model vs those identified by the human examiner were assessed. The association between the lateral cephalometric measurements and the positioning errors in the landmarks comprising the cephalometric measurement was assessed. Results: The mean difference in the angular and linear measurements based on AI vs human landmark localization was.99 ± 1.05°, and.80 ±.82 mm, respectively. Significant differences between the measurements derived from AI‐based and human localization were observed for all cephalometric variables except SNA, pog‐Nperp, facial angle, SN‐GoGn, FMA, Bjork sum, U1‐SN, U1‐FH, IMPA, L1‐NB (angular) and interincisal angle. Conclusions: The errors in landmark positions, especially those that define reference planes, may significantly affect cephalometric measurements. The possibility of errors generated by automated lateral cephalometric analysis systems should be considered when using such systems for orthodontic diagnoses. [ABSTRACT FROM AUTHOR]

The triage process in emergency departments (EDs) relies on the subjective assessment of medical practitioners, making it unreliable in certain aspects. There is a need for a more accurate and objective algorithm to determine the urgency of patients. This paper explores the application of advanced data-synthesis algorithms, machine learning (ML) algorithms, and ensemble models to predict patient mortality. Patients predicted to be at risk of mortality are in a highly critical condition, signifying an urgent need for immediate medical intervention. This paper aims to determine the most effective method for predicting mortality by enhancing the F1 score while maintaining high area under the receiver operating characteristic curve (AUC) score. This study used a dataset of 7325 patients who visited the Yonsei Severance Hospital's ED, located in Seoul, South Korea. The patients were divided into two groups: patients who deceased in the ED and patients who didn't. Various data-synthesis techniques, such as SMOTE, ADASYN, CTGAN, TVAE, CopulaGAN, and Gaussian Copula, were deployed to generate synthetic patient data. Twenty two ML models were then utilized, including tree-based algorithms like Decision tree, AdaBoost, LightGBM, CatBoost, XGBoost, NGBoost, TabNet, which are deep neural network algorithms, and statistical algorithms such as Support Vector Machine, Logistic Regression, Random Forest, k-nearest neighbors, and Gaussian Naive Bayes, as well as Ensemble Models which use the results from the ML models. Based on 21 patient information features used in the pandemic influenza triage algorithm (PITA), the models explained previously were applied to aim for the prediction of patient mortality. In evaluating ML algorithms using an imbalanced medical dataset, conventional metrics like accuracy scores or AUC can be misleading. This paper emphasizes the importance of using the F1 score as the primary performance measure, focusing on recall and specificity in detecting patient mortality. The highest-ranked model for predicting mortality utilized the Gaussian Copula data-synthesis technique and the CatBoost classifier, achieving an AUC of 0.9731 and an F1 score of 0.7059. These findings highlight the effectiveness of machine learning algorithms and data-synthesis techniques in improving the prediction performance of mortality in EDs. [ABSTRACT FROM AUTHOR]

In this study, a combined convolutional neural network for the diagnosis of three benign skin tumors was designed, and its effectiveness was verified through quantitative and statistical analysis. To this end, 698 sonographic images were taken and diagnosed at the Department of Dermatology at Severance Hospital in Seoul, Korea, between 10 November 2017 and 17 January 2020. Through an empirical process, a convolutional neural network combining two structures, which consist of a residual structure and an attention-gated structure, was designed. Five-fold cross-validation was applied, and the train set for each fold was augmented by the Fast AutoAugment technique. As a result of training, for three benign skin tumors, an average accuracy of 95.87%, an average sensitivity of 90.10%, and an average specificity of 96.23% were derived. Also, through statistical analysis using a class activation map and physicians' findings, it was found that the judgment criteria of physicians and the trained combined convolutional neural network were similar. This study suggests that the model designed and trained in this study can be a diagnostic aid to assist physicians and enable more efficient and accurate diagnoses. [ABSTRACT FROM AUTHOR]

Artificial intelligence (AI) has demonstrated its ability to complete complex tasks in various fields. In urban studies, AI technology has been utilized in some limited domains, such as control of traffic and air quality. This study uses AI to better understand diverse urban studies data through a novel approach that uses a convolutional neural network (CNN). In this study, a building outline in the form of a two-dimensional image is used with its corresponding metadata to test the applicability of CNN in reading urban data. MobileNet, a high-efficiency CNN model, is trained to predict the location of restaurants in each building in Seoul, Korea. Consequently, using only 2D image data, the model satisfactorily predicts the locations of restaurants (AUC = 0.732); the model with 2D images and their metadata has higher performance but has an overfitting problem. In addition, the model using only 2D image data accurately predicts the regional distribution of restaurants and shows some typical urban forms with restaurants. The proposed model has several technical limitations but shows the potential to provide a further understanding of urban settings. [ABSTRACT FROM AUTHOR]

Colposcopy is the gold standard diagnostic tool for identifying cervical lesions. However, the accuracy of colposcopies depends on the proficiency of the colposcopist. Machine learning algorithms using an artificial intelligence (AI) system can quickly process large amounts of data and have been successfully applied in several clinical situations. This study evaluated the feasibility of an AI system as an assistive tool for diagnosing high-grade cervical intraepithelial neoplasia lesions compared to the human interpretation of cervical images. This two-centered, crossover, double-blind, randomized controlled trial included 886 randomly selected images. Four colposcopists (two proficient and two inexperienced) independently evaluated cervical images, once with and the other time without the aid of the Cerviray AI® system (AIDOT, Seoul, Republic of Korea). The AI aid demonstrated improved areas under the curve on the localization receiver-operating characteristic curve compared with the colposcopy impressions of colposcopists (difference 0.12, 95% confidence interval, 0.10–0.14, p < 0.001). Sensitivity and specificity also improved when using the AI system (89.18% vs. 71.33%; p < 0.001, 96.68% vs. 92.16%; p < 0.001, respectively). Additionally, the classification accuracy rate improved with the aid of AI (86.40% vs. 75.45%; p < 0.001). Overall, the AI system could be used as an assistive diagnostic tool for both proficient and inexperienced colposcopists in cervical cancer screenings to estimate the impression and location of pathologic lesions. Further utilization of this system could help inexperienced colposcopists confirm where to perform a biopsy to diagnose high-grade lesions. [ABSTRACT FROM AUTHOR]

Traffic congestion is a significant problem that adversely affects the economy, environment, and public health in urban areas worldwide. One promising solution is to forecast road-level congestion levels in the short-term and long-term, enabling commuters to avoid congested areas and allowing traffic agencies to take appropriate action. In this study, we propose a hybrid deep neural network algorithm based on High-Resolution Network (HRNet) and ConvLSTM decoder for 10, 30, and 60-min traffic congestion prediction. Our model utilizes the HRNet's multi-scale feature extraction capability to capture rich spatial features from a sequence of past traffic input images. The ConvLSTM module learns temporal information from each HRNet multi-scale output and aggregates all feature maps to generate accurate traffic forecasts. Our experiments demonstrate that the proposed model can efficiently and effectively learn both spatial and temporal relationships for traffic congestion and outperforms four other state-of-the-art architectures (PredNet, UNet, ConvLSTM, and Autoencoder) in terms of accuracy, precision, and recall. A case study was conducted on the dataset from Seoul, South Korea. [ABSTRACT FROM AUTHOR]

Owing to the recent increase in mobile banking customers, studies exploring self-service channels and customer responses as distribution channels in the retail banking industry are also rapidly expanding. Moreover, with the emergence of big data and a series of artificial intelligence (AI) technologies, customer pattern analysis using deep learning, insurance fraud prevention, software development and various types of blockchain-based FinTech technologies, offline banks are disappearing. Accordingly, many commercial banks are attempting to find technological alternatives. However, maintaining a profitable bank branch is a crucial factor in the relationship between service quality and customer satisfaction because excellent service quality prevents existing customers from leaving. This study sought to quantitatively prove the change in customer service quality and profit due to the introduction of technology in the financial industry. We microscopically compared the effects between bank branch closures and changes in net profit using a time-series analysis. Specifically, we quantitatively analyzed actual customer attrition behavior with a time-series analysis across the three quarters before and after the closure of 88 branches of major commercial banks in South Korea in the Seoul metropolitan area and nearby cities. The findings prove that branch closures and multi-channel effects in the financial sector are gradually being resolved through immediate technology acceptance, contrary to popular concern. [ABSTRACT FROM AUTHOR]

The Bukchon area in Seoul boasts a high density of Hanok, the traditional Korean architecture representing the region. Because the Hanok facade plays a vital role in the streetscape formation, we must record it in terms of social, cultural, historical, artistic, and scenic values. However, recording the facade of an existing Hanok building through drawing or image information is time consuming and labor intensive, and therefore costly. Further, its digital conversion is inherently difficult. This study proposes the use of deep learning to identify the form elements that comprise the Hanok facade. Three-dimensional modeling was performed on 405 well-preserved Hanok facades in the region, and 2808 items of image data were created under similar conditions and at differing angles. Labeling was performed on the shape elements of the Hanok facade, and a methodology was established to identify the facade elements using MASK R-CNN. The type of roof, windows, the lower part of the outer wall, and the design were identified with high accuracy. [ABSTRACT FROM AUTHOR]

As social network service usage is rapidly surging in our daily life, social network service data plays a crucial role in identifying region of attractions and analyzing tourism destination image. In recent years, the computer vision technology is just beginning to be applied in the tourism field through the transfer learning of a deep learning model. However, the pre-trained models have limitations of properly classifying the photos with the unique landscape or specific elements of the tourism destination. With the purpose of going beyond these limitations, we generated a tourists' photo classification reflecting regional characteristics and developed a deep learning model to classify photos according to this classification. Through the analysis of 168,216 Flickr photos, we analyzed the tourism destination image of Seoul. Key findings are that (1) tourists prefer to enjoy local food, to visit authentic traditional palaces, and to see inherent cityscape which can be uniquely enjoyed in Seoul, (2) tourist attractive factors differ by region of attractions, (3) tourist preferences differ by continent. This study has novelty in that it develops a tourist's photo classification suitable for regional characteristics and analyzes tourism destination image by classifying photos using an artificial intelligence technology. [ABSTRACT FROM AUTHOR]

Ensemble deep learning methods have demonstrated significant improvements in forecasting the solar panel power generation using historical time-series data. Although many studies have used ensemble deep learning methods with various data partitioning strategies, most have only focused on improving the predictive methods by associating several different models or combining hyperparameters and interactions. In this study, we contend that we can enhance the precision of power generation forecasting by identifying a suitable data partition strategy and establishing the ideal number of partitions and subset sizes. Thus, we propose a feasibility study of the influence of data partition strategies on ensemble deep learning. We selected five time-series data partitioning strategies—window, shuffle, pyramid, vertical, and seasonal—that allow us to identify different characteristics and features in the time-series data. We conducted various experiments on two sources of solar panel datasets collected in Seoul and Gyeongju, South Korea. Additionally, LSTM-based bagging ensemble models were applied to combine the advantages of several single LSTM models. The experimental results reveal that the data partition strategies positively influence the forecasting of power generation. Specifically, the results demonstrate that ensemble models with data partition strategies outperform single LSTM models by approximately 4–11% in terms of the coefficient of determination (R2) score. [ABSTRACT FROM AUTHOR]

Many forms of air pollution increase as science and technology rapidly advance. In particular, fine dust harms the human body, causing or worsening heart and lung-related diseases. In this study, the level of fine dust in Seoul after 8 h is predicted to prevent health damage in advance. We construct a dataset by combining two modalities (i.e., numerical and image data) for accurate prediction. In addition, we propose a multimodal deep learning model combining a Long Short Term Memory (LSTM) and Convolutional Neural Network (CNN). An LSTM AutoEncoder is chosen as a model for numerical time series data processing and basic CNN. A Visual Geometry Group Neural Network (VGGNet) (VGG16, VGG19) is also chosen as a CNN model for image processing to compare performance differences according to network depth. The VGGNet is a standard deep CNN architecture with multiple layers. Our multimodal deep learning model using two modalities (i.e., numerical and image data) showed better performance than a single deep learning model using only one modality (numerical data). Specifically, the performance improved up to 14.16% when the VGG19 model, which has a deeper network, was used rather than the VGG16 model. [ABSTRACT FROM AUTHOR]

As urbanization progresses, many studies about the analysis and prediction of land-use change and urban sprawl have been conducted recently. As the sprawl phenomenon progresses rapidly, the urban expansion phenomenon became uncontrolled and it has affected negatively on the city's environment and transportation finally. So, it is essential to identify lands likely to be urbanized in the future because it aids in establishing land use plans and policies pre-acting the negative impact of spatially urban expansion the sprawl by determining factors affecting the urban sprawl. Previous studies based on statistical models are limited to identifying determining factors, so the prediction performance is low compared to deep learning. On the other hand, existing studies using machine learning and deep learning overlook selecting specific region-focused variables. Therefore, this study aims to analyze and predict changes in the Seoul Metropolitan Area's sprawl in Korea using the Convolutional Long Short-Term Memory Network (ConvLSTM) with factors at the city scale and neighboring factors at the local scale in the Seoul Metropolitan Area (SMA). ConvLSTM is a type of combination model: combining Recurrent Neural Network(RNN) and Convolutional Neural Network(CNN). This study showed that ConvLSTM with factors at the city and neighboring factors at the local scale predicted the urbanized land. The determinants contain population and roads ratio at the city scale, and neighboring urban lands, distance to the nearest subway stations, slope, and elevation at the local scale. The results reveal that predicted urban lands in 2020 increase over the entire region. In particular, the expected urban lands in 2020 increase by reducing farmlands in the southern part of the SMA. It is consistent with the trend of urbanized lands from 1980 to 2010. In addition, urbanization occurs in areas adjacent to Seoul due to the well-established urban infrastructure. The results of this study can be used as evidence to establish sustainable land use plans and regulations in the future. [ABSTRACT FROM AUTHOR]

Summary: Dust, or particulate matter (PM2.5), is among the most harmful pollutants negatively affecting human health. Predicting indoor PM2.5 concentrations is essential to achieve acceptable indoor air quality. This study aims to investigate data‐driven models to accurately predict PM 2.5 pollution. Notably, a comparative study has been conducted between twenty‐one machine learning and deep learning models to predict PM2.5 levels. Specifically, we investigate the performance of machine learning and deep learning models to predict ambient PM2.5 concentrations based on other ambient pollutants, including SO2$$ {}_2 $$, NO2$$ {}_2 $$, O3$$ {}_3 $$, CO, and PM10. Here, we applied Bayesian optimization to optimally tune hyperparameters of the Gaussian process regression with different kernels and ensemble learning models (i.e., boosted trees and bagged trees) and investigated their prediction performance. Furthermore, to further enhance the forecasting performance of the investigated models, dynamic information has been incorporated by introducing lagged measurements in the construction of the considered models. Results show a significant improvement in the prediction performance when considering dynamic information from past data. Moreover, three methods, namely, random forest (RF), decision tree, and extreme gradient boosting, are applied to assess variables contribution and revealed that lagged PM2.5 data contribute significantly to the prediction performance and enables the construction of parsimonious models. Hourly concentration levels of ambient air pollution from the air quality monitoring network located in Seoul are employed to verify the prediction effectiveness of the studied models. Six measurements of effectiveness are used for assessing the prediction quality. Results showed that deep learning models are more efficient than the other investigated machine learning models (i.e., SVR, GPR, bagged and boosted trees, RF, and XGBoost). Also, the results showed that the bidirectional long short term memory (BiLSTM) and bidirectional gated recurrent units (BiGRU) networks produce higher performance than the investigated machine learning models (i.e., SVR, GPR, bagged and boosted trees, RF, and XGBoost) and deep learning models (i.e., LSTM, GRU, and convolutional neural network). [ABSTRACT FROM AUTHOR]

Tree species and canopy structural profile ('tree profile') are among the most critical environmental factors in determining urban ecosystem services such as climate and air quality control from urban trees. To accurately characterize a tree profile, the tree diameter, height, crown width, and height to the lowest live branch must be all measured, which is an expensive and time-consuming procedure. Recent advances in artificial intelligence aids to efficiently and accurately measure the aforementioned tree profile parameters. This can be particularly helpful if spatially extensive and accurate street-level images provided by Google ('streetview') or Kakao ('roadview') are utilized. We focused on street trees in Seoul, the capital city of South Korea, and suggested a novel approach to create a tree profile and inventory based on deep learning algorithms. We classified urban tree species using the YOLO (You Only Look Once), one of the most popular deep learning object detection algorithms, which provides an uncomplicated method of creating datasets with custom classes. We further utilized semantic segmentation algorithm and graphical analysis to estimate tree profile parameters by determining the relative location of the interface of tree and ground surface. We evaluated the performance of the model by comparing the estimated tree heights, diameters, and locations from the model with the field measurements as ground truth. The results are promising and demonstrate the potential of the method for creating urban street tree profile inventory. In terms of tree species classification, the method showed the mean average precision (mAP) of 0.564. When we used the ideal tree images, the method also reported the normalized root mean squared error (NRMSE) for the tree height, diameter at breast height (DBH), and distances from the camera to the trees as 0.24, 0.44, and 0.41. [ABSTRACT FROM AUTHOR]

This study suggests a natural language processing framework for collecting, analyzing, and, visualizing online natural language data, consisting of a web crawler for data collection, tokenizer for text preprocessing, Word2vec for word embedding, and deep-learning long short-term memory networks for sentiment classification. The framework was exemplified on online brokerage platforms in New York City and Seoul. The visualized framework-driven results showed regional similarities and differences between the cities. The proposed approach provides a way to gather big data, not through surveys or interviews. The framework-driven analysis may provide descriptive precursors to explore how laypersons experience built environments and city spaces. [ABSTRACT FROM AUTHOR]

Deep learning approaches are widely employed for forecasting short-term travel demand to respond to real-time demand. Although it is critical for demand forecasting to be evenly distributed in the spatial and temporal views to support real-time mobility service operations, in related studies, the predictive performance of models has been evaluated only in terms of aggregated errors. Therefore, the present study was conducted to investigate the distribution of errors to explore spatiotemporal correlations. Six deep learning models with the same architecture, except for the base module, consisting of three stacked layers, were constructed. These models were used to forecast demands for a station-based bike-sharing service in Seoul, South Korea. To attain our goals, global and local Moran's I of the errors was introduced to evaluate the spatial and temporal performances of the deep learning approaches. The results showed that the model with convolutional long short-term memory layers, which are effective at predicting spatiotemporal data, outperformed the other models in terms of aggregated performance. However, the global Moran's I of the errors in the model reflects spatial dependency over the regions. This suggests that the best predictive performance of the model does not necessarily imply that it performs well in demand forecasting in all regions. Furthermore, cluster and outlier analyses of the errors indicated that excessive or insufficient predictions were clustered or dispersed throughout the regions. These results can be used to enhance the model by introducing the spatial correlation index into the loss function or by incorporating additional features for handling spatial correlations. [ABSTRACT FROM AUTHOR]

Purpose: This study aimed to evaluate the performance of transfer learning in a deep convolutional neural network for classifying implant fixtures. Materials and Methods: Periapical radiographs of implant fixtures obtained using the Superline (Dentium Co. Ltd., Seoul, Korea), TS III (Osstem Implant Co. Ltd., Seoul, Korea), and Bone Level Implant (Institut Straumann AG, Basel, Switzerland) systems were selected from patients who underwent dental implant treatment. All 355 implant fixtures comprised the total dataset and were annotated with the name of the system. The total dataset was split into a training dataset and a test dataset at a ratio of 8 to 2, respectively. YOLOv3 (You Only Look Once version 3, available at https://pjreddie.com/darknet/yolo/), a deep convolutional neural network that has been pretrained with a large image dataset of objects, was used to train the model to classify fixtures in periapical images, in a process called transfer learning. This network was trained with the training dataset for 100, 200, and 300 epochs. Using the test dataset, the performance of the network was evaluated in terms of sensitivity, specificity, and accuracy. Results: When YOLOv3 was trained for 200 epochs, the sensitivity, specificity, accuracy, and confidence score were the highest for all systems, with overall results of 94.4%, 97.9%, 96.7%, and 0.75, respectively. The network showed the best performance in classifying Bone Level Implant fixtures, with 100.0% sensitivity, specificity, and accuracy. Conclusion: Through transfer learning, high performance could be achieved with YOLOv3, even using a small amount of data. [ABSTRACT FROM AUTHOR]

Road condition and quality are critical road maintenance and risk reduction factors. Most existing road monitoring systems include regular on-site surveys and maintenance. However, major roads in urban areas are generally complicated and have heavy traffic during the daytime, so such field investigation can be significantly limited. Moreover, any road work at nighttime can be risky and dangerous and incur excessive expenses. Based on a review of existing systems for monitoring road conditions, this study focuses on overcoming two unsolved challenges: the capacity of the monitoring range and the avoidance techniques to ensure traffic is not hindered. To solve these challenges, these paper proposes an integrated road monitoring system called Car-free Street Mapping (CfSM) using unmanned aerial vehicles (UAV), aerial mapping cameras, and deep learning (DL) algorithms. The use of the aerial mapping camera mounted on the UAV is to widen the monitoring viewing range, and general-purpose drones are used in this study rather than expensive special equipment. Since the drone-taken images include many passing vehicles that conceal the road surface from the camera vision, the DL model was applied to detect the vehicles and their shadows and then remove them from the images. To train the DL model, two image datasets were used: publicly available cars overhead with context (COWC) images and orthoimages additionally taken for the project to further improve the accuracy. The two datasets consist of 298,623 labeled objects on 9,331 images in total. The tests resulted in a mean average precision (mAP) of 89.57% for trucks, 95.77% for passenger vehicles, and 76.51% for buses. Finally, the object-removed images were composited into one whole car-free image. The CfSM was applied to two areas in Yeouido and Sangam-dong, Seoul, Korea. The car-free images in both regions show a spatial resolution of 10 mm and can be used for various purposes such as road maintenance and management and autonomous vehicle roadmaps. [ABSTRACT FROM AUTHOR]

Identifying Mycobacterium using matrix-assisted laser desorption/ionization time-of-flight mass spectrometry (MALDI-TOF MS) is challenging. We evaluated the performance of MALDI-TOF MS in identifying nontuberculous mycobacteria (NTM) using the ASTA MicroIDSys system (ASTA Inc., Suwon, Korea) with the MycoDB v1.95s and upgraded MycoDB v2.0-beta databases. We tested 124 NTM isolates collected from Ogawa medium at Severance Hospital, Seoul, Korea, between January and April 2019. MicroIDSys scores were categorized into three groups: ≥140, reliable identification; 130–139, ambiguous identification; and <130, invalid identification. To validate the results, we used the reverse blot hybridization assay (Molecutech REBA MycoID, YD Diagnostics Corp., Korea). Initial analysis using MycoDB v1.95s resulted in 26.6% (33/124) reliable, 43.5% (54/124) ambiguous, and 29.8% (37/124) invalid identifications. Re-analysis using the upgraded MycoDB v2.0-beta database resulted in 94.4% (117/124) reliable, 4.0% (5/124) ambiguous, and 1.6% invalid (2/124) identifications. The percentage of reliable identifications that matched with the reference increased from 26.6% (33/124) with MycoDB v1.95s to 93.5% (116/124) with MycoDB v2.0-beta. The upgraded databases enable substantially improved NTM identification through deep learning in the inference algorithm and by considering more axes in the correlation analysis. MALDI-TOF MS using the upgraded database unambiguously identified most NTM species. Our study lays a foundation for applying MALDITOF MS for the simple and rapid identification of NTM isolated from solid media. [ABSTRACT FROM AUTHOR]

Traditional accident prediction models have been mostly designed with statistical analysis that finds and analyzes the causal relationships between traffic accidents and a variety of human, road geometry, and environmental factors. However, these statistical methods have limitations in that they are based on assumptions about data distribution and function type. Therefore, this study suggests an accident prediction model using deep learning. This newly suggested risk prediction model is for predicting risk by reflecting static features of the road, such its length and the speed limit on it, and dynamic features of the road, such as traffic volume when driving on it, and the altitude and azimuth of the sun. For this purpose, 4470 accident cases, collected over 5 months from August to December 2018 in Seoul—the most complex, high-traffic, and accident-prone city in Korea—were analyzed. As a result of testing the model using such data, it was found to have an accuracy of 75% and recall of 81%. Based on testing results for the suggested risk prediction model, a system was developed to guide not only accident-prone regions predicted using statistical data but to also guide a risk level for the road. This level of risk is estimated based upon each given situation, so it can change even for the same road. This guide system can be used to provide a level of risk for each road segment and region but also to improve roads with recommendations, such as installation of safety features. In addition, it could support a mobile system that provides a driver with the optimized driving path for safety. [ABSTRACT FROM AUTHOR]

With the rise of social media platforms, tourists tend to share their experiences in the form of texts, photos, and videos on social media. These user-generated contents (UGC) play an important role in shaping tourism destination images (TDI) and directly affect the decision-making process of tourists. Among UGCs, photos represent tourists' visual preferences for a specific area. Paying attention to the value of photos, several studies have attempted to analyze them using deep learning technology. However, the research methods that analyze tourism photos using recent deep learning technology have a limitation in that they cannot properly classify unique photos appearing in specific tourist attractions with predetermined photo categories such as Places365 or ImageNet dataset or it takes a lot of time and effort to build a separate training dataset to train the model and to generate a tourism photo classification category according to a specific tourist destination. The purpose of this study is to propose a method of automatically classifying tourist photos by tourist attractions by applying the methods of the image feature vector clustering and the deep learning model. To this end, first, we collected photos attached to reviews posted by foreign tourists on TripAdvisor. Second, we embedded individual images as 512-dimensional feature vectors using the VGG16 network pre-trained with Places365 and reduced them to two dimensions with t-SNE(t-Distributed Stochastic Neighbor Embedding). Then, clusters were extracted through HDBSCAN(Hierarchical Clustering and Density-Based Spatial Clustering of Applications with Noise) analysis and set as a regional image category. Finally, the Siamese Network was applied to remove noise photos within the cluster and classify photos according to the category. In addition, this study attempts to confirm the validity of the proposed method by applying it to two representative tourist attractions such as 'Gyeongbokgung Palace' and 'Insadong' in Seoul. As a result, it was possible to identify which visual elements of tourist attractions are attractive to tourists. This method has the advantages in that it is not necessary to create a classification category in advance, it is possible to flexibly extract categories for each tourist destination, and it is able to improve classification performance even with a rather small volume of a dataset. [ABSTRACT FROM AUTHOR]

An adversary can use SSH communication as a route for information leakage or hacking. Many studies have focused on TCP header analysis to detect encrypted communication. However, SSH detection using TCP header analysis is limited when changing TCP port information or modifying components of the SSH protocol. Various machine-learning (ML) techniques have been introduced to enhance network traffic classification by analyzing TCP headers. Most ML-based traffic classification research has analyzed network packet flows. However, because of the complex structures and the various implementations of the TCP protocol, a lot of time and resources are required for the recombination of network packet flows. This paper presents a novel contribution to overcome the problems of network packet analysis that employs web proxy session logs, which do not require the recombination of packets to prepare a dataset for analysis. Moreover, we propose a hybrid predictive model that is useful for web proxy session log analysis. In the modeling process, we collected the web proxy logs from an actual network of ICT companies (more than 10,000 employees, Seoul, South Korea) and used the random forest and decision tree algorithms for the supervised learning. The detection rate (DR) for the training dataset was 99.9%, which is similar to or higher than that of other studies using ML and deep learning. Using the dataset of DARPA99, we proved that the DR and FPR for our proposed model were better than those achieved by Alshammari et al.'s model. We expect that the proposed predictive model can be used to block illegal attempts at SSH communication over HTTP CONNECT by changing the destination port and to detect novel illegal communication protocols. [ABSTRACT FROM AUTHOR]

Nitrous acid (HONO), one of the reactive nitrogen oxides (NOy), plays an important role in the formation of ozone (O3) and fine aerosols (PM2.5) in the urban atmosphere. In this study, a simulation model of Reactive Nitrogen species using Deep neural network model (RND) was constructed to calculate the HONO mixing ratios through a deep learning technique using measured variables. A Python-based Deep Neural Network (DNN) was trained, validated, and tested with HONO measurement data obtained in Seoul during the warm months from 2016 to 2019. A k-fold cross validation and test results confirmed the performance of RND v1.0 with an Index Of Agreement (IOA) of 0.79 ~ 0.89 and a Mean Absolute Error (MAE) of 0.21 ~ 0.31 ppbv. The RNDV1.0 adequately represents the main characteristics of HONO and thus, RND v1.0 is proposed as a supplementary model for calculating the HONO mixing ratio in a high-NOx environment. [ABSTRACT FROM AUTHOR]

Objective: To investigate the accuracy of automated identification of cephalometric landmarks using the cascade convolutional neural networks (CNN) on lateral cephalograms acquired from nationwide multi-centres.Settings and Sample Population: A total of 3150 lateral cephalograms were acquired from 10 university hospitals in South Korea for training.Materials and Methods: We evaluated the accuracy of the developed model with independent 100 lateral cephalograms as an external validation. Two orthodontists independently identified the anatomic landmarks of the test data set using the V-ceph software (version 8.0, Osstem, Seoul, Korea). The mean positions of the landmarks identified by two orthodontists were regarded as the gold standard. The performance of the CNN model was evaluated by calculating the mean absolute distance between the gold standard and the automatically detected positions. Factors associated with the detection accuracy for landmarks were analysed using the linear regression models.Results: The mean inter-examiner difference was 1.31 ± 1.13 mm. The overall automated detection error was 1.36 ± 0.98 mm. The mean detection error for each landmark ranged between 0.46 ± 0.37 mm (maxillary incisor crown tip) and 2.09 ± 1.91 mm (distal root tip of the mandibular first molar). A significant difference in the detection accuracy among cephalograms was noted according to hospital (P = .011), sensor type (P < .01), and cephalography machine model (P < .01).Conclusion: The automated cephalometric landmark detection model may aid in preliminary screening for patient diagnosis and mid-treatment assessment, independent of the type of the radiography machines tested. [ABSTRACT FROM AUTHOR]

This paper proposes a deep learning model that integrates a convolutional neural network and a gated recurrent unit with groups of neighboring stations to accurately predict PM2.5 concentrations at 25 stations in Seoul, South Korea. The deep learning model uses observations obtained from one Korea Meteorological Administration (KMA) station, 25 National Institute of Environmental Research (NIER) stations, and 28 automatic weather stations (AWSs) throughout Seoul. To train the deep learning model, we use all available meteorological and air quality data observed between 2015 and 2017. With the trained model, we predict PM2.5 concentrations at all 25 NIER stations in Seoul for 2018. This study also proposes a geographical polygon group model that determines the optimal number of neighboring NIER stations required to increase the accuracy of PM2.5 concentration predictions at the target station. Comparing the model measures for each of the 25 monitoring sites in 2018, we find that the geographical polygon group model achieves an index of agreement of 0.82–0.89 and a Pearson correlation coefficient of 0.70–0.83. Compared to the method using only meteorological and air quality data from one target station (average IOA = 0.77) to predict PM2.5 concentrations at the 25 stations in Seoul, the proposed method using geographical correlation-based neighboring NIER stations as polygonal groups (average IOA = 0.85) improves the PM2.5 prediction accuracy by an average of about 10%. This approach, based on deep learning, can be updated to predict air pollution or air quality indices up to several days in advance. [ABSTRACT FROM AUTHOR]

Direct microscopic examination with potassium hydroxide is generally used as a screening method for diagnosing superficial fungal infections. Although this type of examination is faster than other diagnostic methods, it can still be time-consuming to evaluate a complete sample; additionally, it possesses the disadvantage of inconsistent reliability as the accuracy of the reading may differ depending on the performer's skill. This study aims at detecting hyphae more quickly, conveniently, and consistently through deep learning using images obtained from microscopy used in real-world practice. An object detection convolutional neural network, YOLO v4, was trained on microscopy images with magnifications of 100×, 40×, and (100+40)×. The study was conducted at the Department of Dermatology at Veterans Health Service Medical Center, Seoul, Korea between January 1, 2019 and December 31, 2019, using 3,707 images (1,255 images for training, 1,645 images for testing). The average precision was used to evaluate the accuracy of object detection. Precision recall curve analysis was performed for the hyphal location determination, and receiver operating characteristic curve analysis was performed on the image classification. The F1 score, sensitivity, and specificity values were used as measures of the overall performance. The sensitivity and specificity were, respectively, 95.2% and 100% in the 100× data model, and 99% and 86.6% in the 40× data model; the sensitivity and specificity in the combined (100+40)× data model were 93.2% and 89%, respectively. The performance of our model had high sensitivity and specificity, indicating that hyphae can be detected with reliable accuracy. Thus, our deep learning-based autodetection model can detect hyphae in microscopic images obtained from real-world practice. We aim to develop an automatic hyphae detection system that can be utilized in real-world practice through continuous research. [ABSTRACT FROM AUTHOR]

Ground-penetrating radar (GPR) is widely used for detecting buried underground object. Deep learning technique is recently being adopted into this field thanks to its powerful image classification capacity. However, it uses only GPR B-scan images, although multichannel GPR device can provide more informative three-dimensional (3D) data for underground object. In this study, a novel deep learning-based underground object classification method is proposed by using two-dimensional (2D) grid image which consists of several B-scan and C-scan images. Spatial information of an underground object can be well represented in the 2D grid image. The 2D grid images are then used to train deep convolutional neural networks. The proposed method is experimentally validated by field data collected from urban roads in Seoul, South Korea. The performance is also compared to a conventional method which uses only B-scan images. The proposed method successfully classifies cavity, pipe, manhole and subsoils background having very small false-positive errors. [ABSTRACT FROM AUTHOR]

Purpose: Prior studies on the application of deep-learning techniques have focused on enhancing computation algorithms. However, the amount of data is also a key element when attempting to achieve a goal using a quantitative approach, which is often underestimated in practice. The problem of sparse sales data is well known in the valuation of commercial properties. This study aims to expand the limited data available to exploit the capability inherent in deep learning techniques. Design/methodology/approach: The deep learning approach is used. Seoul, the capital of South Korea is selected as a case study area. Second, data augmentation is performed for properties with low trade volume in the market using a variational autoencoder (VAE), which is a generative deep learning technique. Third, the generated samples are added into the original dataset of commercial properties to alleviate data insufficiency. Finally, the accuracy of the price estimation is analyzed for the original and augmented datasets to assess the model performance. Findings: The results using the sales datasets of commercial properties in Seoul, South Korea as a case study show that the augmented dataset by a VAE consistently shows higher accuracy of price estimation for all 30 trials, and the capabilities inherent in deep learning techniques can be fully exploited, promoting the rapid adoption of artificial intelligence skills in the real estate industry. Originality/value: Although deep learning-based algorithms are gaining popularity, they are likely to show limited performance when data are insufficient. This study suggests an alternative approach to overcome the lack of data problem in property valuation. [ABSTRACT FROM AUTHOR]

This paper presents a plan for preparing for changes in the business environment by analyzing and predicting business district data in Seoul. The COVID-19 pandemic and economic crisis caused by inflation have led to an increase in store closures and a decrease in sales, which has had a significant impact on commercial districts. The number of stores and sales are critical factors that directly affect the business environment and can help prepare for changes. This study conducted correlation analysis to extract factors related to the commercial district's environment in Seoul and estimated the number of stores and sales based on these factors. Using the Kendaltau correlation coefficient, the study found that existing population and working population were the most influential factors. Linear regression, tensor decomposition, Factorization Machine, and deep neural network models were used to estimate the number of stores and sales, with the deep neural network model showing the best performance in RMSE and evaluation indicators. This study also predicted the number of stores and sales of the service industry in a specific area using the population prediction results of the neural prophet model. The study's findings can help identify commercial district information and predict the number of stores and sales based on location, industry, and influencing factors, contributing to the revitalization of commercial districts. [ABSTRACT FROM AUTHOR]

Interest in the economic effects of walkable environments is growing as the multidimensional benefits of walkability can be capitalized into surrounding property values. Commercial properties may be more sensitive to the economic effects of walkability compared to other types of properties due to their distinct attributes, but not many studies have investigated the relationships between walkable environments and commercial property values. Additionally, previous studies identified walkable environments on a neighborhood-scale, not at a meso- and micro-scale. This study fills these gaps by examining the effects of multifaceted walkable environments on commercial property values in Seoul, Korea. Based on sales transaction data between 2017 and 2019, we employed the hedonic price model to clarify the relationships between walkable environments and commercial property values. This study estimated walkable environments using semantic segmentation based on Google Street View images. Furthermore, we accounted for commercial submarkets to identify how the impacts of walkability vary across areas stratified by living population density. The results show that various walkable environments positively affect nearby commercial property values. Additionally, these effects vary across commercial submarkets' heterogeneity. Our findings provide a variety of suggestions for encouraging the creation of walkable environments and increasing the economic benefits of commercial properties. • We explore the effects of walkable environments on commercial property values. • This study Identifies walkable environments on macro-, meso-, and micro-scales. • Using deep learning semantic segmentation, we estimate meso-scale streetscapes. • Walkable environments positively affect neighboring commercial property values. • Impacts of walkable environments vary across commercial submarkets. [ABSTRACT FROM AUTHOR]

High-density memory devices are essential to sustain growth in information technology (IT). Furthermore, brain-inspired computing devices are the future of IT businesses such as artificial intelligence, deep learning, and big data. Herein, we propose a facile and hierarchical nickel cobaltite (NCO) quasi-hexagonal nanosheet-based memristive device for multilevel resistive switching (RS) and synaptic learning applications. Electrical measurements of the Pt/NCO/Pt device show the electroforming free pinched hysteresis loops at different voltages, suggesting the multilevel RS capability of the device. The detailed memristive properties of the device were calculated using the time-dependent current–voltage data. The two-valued charge-flux properties indicate the memristive and multilevel RS characteristics of the device. Interestingly, the Pt/NCO/Pt memristive device shows a compliance current (CC)-dependent RS property; compliance-free RS was observed from 10−2 to 10−4 A, and the compliance effect dominated in the range of 10−5–10−6 A. In CC control mode, the device demonstrated three resistance states during endurance and retention measurements. In addition, the device was successful in mimicking biological synaptic properties such as potentiation-depression- and spike-timing-dependent plasticity rules. The results of the present investigation demonstrated that solution-processable NCO nanosheets are potential switching materials for high-density memory and brain-inspired computing applications. Neuromorphic computing: Nickel cobaltite nanosheets for memory devices A simple technique for producing nanostructured oxides shows promise for developing neuromorphic computing systems. Tae Geun Kim from Korea University in Seoul, South Korea, and co-workers report that nickel cobaltite, a low-cost material being considered for non-volatile memory devices, can be synthesized as porous nanosheets through a co-precipitation method. The team demonstrated that when stacked a few hundred nanometers thick, these sheets had properties ideal for 'resistive switching', an effect that stores data by transforming insulators into conductors using metal filaments. Characterizations revealed that the material's morphology proved ideal for fine-tuning filament growth, enabling the memory cell to switch between three conductive states. Because the nanosheet devices retained memories of their switching history, they were capable of mimicking neural network behavior, such as the potentiation, depression and spike-timing-dependent plasticity processes associated with synapses. A resistive switching device is fabricated using nanostructured nickel cobaltite for high-density data storage and synaptic learning applications. The active switching layer of the device consists of quasi-hexagonal porous nanosheets that enable smooth charge transport. The device shows voltage tunable and forming-free resistive switching effect and non-ideal memristive properties. The rational design of the device helps to show controlled multilevel resistive switching property and thereby switch between three conductive states. The analog switching of the device helps to mimic specific neural network behavior, such as the potentiation, depression and four spike-timing-dependent plasticity rules. [ABSTRACT FROM AUTHOR]

Purpose: This study evaluated the performance of a commercially available deep-learning algorithm (DLA) (Insight CXR, Lunit, Seoul, South Korea) for referable thoracic abnormalities on chest X-ray (CXR) using a consecutively collected multicenter health screening cohort. Methods and materials: A consecutive health screening cohort of participants who underwent both CXR and chest computed tomography (CT) within 1 month was retrospectively collected from three institutions' health care clinics (n = 5,887). Referable thoracic abnormalities were defined as any radiologic findings requiring further diagnostic evaluation or management, including DLA-target lesions of nodule/mass, consolidation, or pneumothorax. We evaluated the diagnostic performance of the DLA for referable thoracic abnormalities using the area under the receiver operating characteristic (ROC) curve (AUC), sensitivity, and specificity using ground truth based on chest CT (CT-GT). In addition, for CT-GT-positive cases, three independent radiologist readings were performed on CXR and clear visible (when more than two radiologists called) and visible (at least one radiologist called) abnormalities were defined as CXR-GTs (clear visible CXR-GT and visible CXR-GT, respectively) to evaluate the performance of the DLA. Results: Among 5,887 subjects (4,329 males; mean age 54±11 years), referable thoracic abnormalities were found in 618 (10.5%) based on CT-GT. DLA-target lesions were observed in 223 (4.0%), nodule/mass in 202 (3.4%), consolidation in 31 (0.5%), pneumothorax in one 1 (<0.1%), and DLA-non-target lesions in 409 (6.9%). For referable thoracic abnormalities based on CT-GT, the DLA showed an AUC of 0.771 (95% confidence interval [CI], 0.751–0.791), a sensitivity of 69.6%, and a specificity of 74.0%. Based on CXR-GT, the prevalence of referable thoracic abnormalities decreased, with visible and clear visible abnormalities found in 405 (6.9%) and 227 (3.9%) cases, respectively. The performance of the DLA increased significantly when using CXR-GTs, with an AUC of 0.839 (95% CI, 0.829–0.848), a sensitivity of 82.7%, and s specificity of 73.2% based on visible CXR-GT and an AUC of 0.872 (95% CI, 0.863–0.880, P <0.001 for the AUC comparison of GT-CT vs. clear visible CXR-GT), a sensitivity of 83.3%, and a specificity of 78.8% based on clear visible CXR-GT. Conclusion: The DLA provided fair-to-good stand-alone performance for the detection of referable thoracic abnormalities in a multicenter consecutive health screening cohort. The DLA showed varied performance according to the different methods of ground truth. [ABSTRACT FROM AUTHOR]

To illustrate what is inside the so-called black box of deep learning models (DLMs) so that clinicians can have greater confidence in the conclusions of artificial intelligence by evaluating adversarial explanation on its ability to explain the rationale of DLM decisions for glaucoma and glaucoma-related findings. Adversarial explanation generates adversarial examples (AEs), or images that have been changed to gain or lose pathologic characteristic-specific traits, to explain the DLM's rationale. Evaluation of explanation methods for DLMs. Health screening participants (n = 1653) at the Seoul National University Hospital Health Promotion Center, Seoul, Republic of Korea. We trained DLMs for referable glaucoma (RG), increased cup-to-disc ratio (ICDR), disc rim narrowing (DRN), and retinal nerve fiber layer defect (RNFLD) using 6430 retinal fundus images. Surveys consisting of explanations using AE and gradient-weighted class activation mapping (GradCAM), a conventional heatmap-based explanation method, were generated for 400 pathologic and healthy patient eyes. For each method, board-trained glaucoma specialists rated location explainability, the ability to pinpoint decision-relevant areas in the image, and rationale explainability, the ability to inform the user on the model's reasoning for the decision based on pathologic features. Scores were compared by paired Wilcoxon signed-rank test. Area under the receiver operating characteristic curve (AUC), sensitivities, and specificities of DLMs; visualization of clinical pathologic changes of AEs; and survey scores for locational and rationale explainability. The AUCs were 0.90, 0.99, 0.95, and 0.79 and sensitivities were 0.79, 1.00, 0.82, and 0.55 at 0.90 specificity for RG, ICDR, DRN, and RNFLD DLMs, respectively. Generated AEs showed valid clinical feature changes, and survey results for location explainability were 3.94 ± 1.33 and 2.55 ± 1.24 using AEs and GradCAMs, respectively, of a possible maximum score of 5 points. The scores for rationale explainability were 3.97 ± 1.31 and 2.10 ± 1.25 for AEs and GradCAM, respectively. Adversarial example provided significantly better explainability than GradCAM. Adversarial explanation increased the explainability over GradCAM, a conventional heatmap-based explanation method. Adversarial explanation may help medical professionals understand more clearly the rationale of DLMs when using them for clinical decisions. [ABSTRACT FROM AUTHOR]

Ground penetrating radar (GPR) is a promising non-destructive evaluation technique for detecting buried underground objects in urban area. Deep learning technique is recently being applied into this field to automate the GPR data interpretation. However, there is no proper technique that can reflect the uniqueness of urban road pavements. In this study, an underground object detection technique suitable for urban road pavement is proposed by using a statistically determined threshold amplitude and a large amount of GPR B-scan image libraries. An automated thresholding technique is newly developed based on the statistical distribution of GPR data. Deep learning technique is then applied to the reconstructed GPR data to detect underground objects in urban area. The proposed method is experimentally validated by field data collected on urban roads in Seoul, South Korea. In addition, its application possibility is also tested with full-size GPR data. The proposed method successfully emphasises the feature of underground objects and classifies hyperbola, manhole cover, layer interface and subsoil background. [ABSTRACT FROM AUTHOR]

Although deep learning-based valuation models are spreading throughout the real estate industry following the artificial intelligence boom, property owners and investors continue to doubt the accuracy of the results. In this study, we specify a neural network for predicting house prices. We suggest a standard feed-forward network with two hidden layers, and show that it is sufficiently reasonable to apply its prediction to real-world projects such as property valuation. In addition, we propose a Bayesian neural network for describing uncertainty in house price predictions while providing a means to quantify uncertainty for each prediction. We choose Gangnam-gu, Seoul for the analysis, and predict house prices in the area using both networks. Although the Bayesian neural network did not perform better than the conventional network, it could provide a tool to measure the uncertainty inherent in predicted prices. The findings of this study show that a Bayesian approach can model uncertainty in property valuation, thereby promoting the adoption of deep learning tools in the real estate industry. [ABSTRACT FROM AUTHOR]

Ground-penetrating radar is a typical sensor system for analyzing underground facilities such as pipelines and rebars. The technique also can be used to detect an underground cavity, which is a potential sign of urban sinkholes. Multichannel ground-penetrating radar devices are widely used to detect underground cavities thanks to the capacity of informative three-dimensional data. Nevertheless, the three-dimensional ground-penetrating radar data interpretation is unclear and complicated when recognizing underground cavities because similar ground-penetrating radar data reflected from different underground objects are often mixed with the cavities. As it is prevalently known that the deep learning algorithm-based techniques are powerful at image classification, deep learning-based techniques for underground object detection techniques using two-dimensional GPR (ground-penetrating radar) radargrams have been researched upon in recent years. However, spatial information of underground objects can be characterized better in three-dimensional ground-penetrating radar voxel data than in two-dimensional ground-penetrating radar images. Therefore, in this study, a novel underground object classification technique is proposed by applying deep three-dimensional convolutional neural network on three-dimensional ground-penetrating radar data. First, a deep convolutional neural network architecture was developed using three-dimensional convolutional networks for recognizing spatial underground objects such as, pipe, cavity, manhole, and subsoil. The framework of applying the three-dimensional convolutional neural network into three-dimensional ground-penetrating radar data was then proposed and experimentally validated using real three-dimensional ground-penetrating radar data. In order to do that, three-dimensional ground-penetrating radar block data were used to train the developed three-dimensional convolutional neural network and to classify unclassified three-dimensional ground-penetrating radar data collected from urban roads in Seoul, South Korea. The validation results revealed that four underground objects (pipe, cavity, manhole, and subsoil) are successfully classified, and the average classification accuracy was 97%. In addition, a false alarm was rarely indicated. [ABSTRACT FROM AUTHOR]

This study uses a deep learning approach to forecast ozone concentrations over Seoul, South Korea for 2017. We use a deep convolutional neural network (CNN). We apply this method to predict the hourly ozone concentration on each day for the entire year using several predictors from the previous day, including the wind fields, temperature, relative humidity, pressure, and precipitation, along with in situ ozone and NO2 concentrations. We refer to a history of all observed parameters between 2014 and 2016 for training the predictive models. Model-measurement comparisons for the 25 monitoring sites for the year 2017 report average indices of agreement (IOA) of 0.84–0.89 and a Pearson correlation coefficient of 0.74–0.81, indicating reasonable performance for the CNN forecasting model. Although the CNN model successfully captures daily trends as well as yearly high and low variations of the ozone concentrations, it notably underpredicts high ozone peaks during the summer. The forecasting results are generally more accurate for the stations located in the southern regions of the Han River, the result of more stable topographical and meteorological conditions. Furthermore, through two separate daytime and nighttime forecasts, we find that the monthly IOA of the CNN model is 0.05–0.30 higher during the daytime, resulting from the unavailability of some of the input parameters during the nighttime. While the CNN model can predict the next 24 h of ozone concentrations within less than a minute, we identify several limitations of deep learning models for real-time air quality forecasting for further improvement. [ABSTRACT FROM AUTHOR]

Three-dimensional ground penetrating radar data are often ambiguous and complex to interpret when attempting to detect only underground cavities because ground penetrating radar reflections from various underground objects can appear like those from cavities. In this study, we tackle the issue of ambiguity by proposing a system based on deep convolutional neural networks, which is capable of autonomous underground cavity detection beneath urban roads using three-dimensional ground penetrating radar data. First, a basis pursuit-based background filtering algorithm is developed to enhance the visibility of underground objects. The deep convolutional neural network is then established and applied to automatically classify underground objects using the filtered three-dimensional ground penetrating radar data as represented by three types of images: A-, B-, and C-scans. In this study, we utilize a novel two-dimensional grid image consisting of several B- and C-scan images. Cavity, pipe, manhole, and intact features extracted from in situ three-dimensional ground penetrating radar data are used to train the convolutional neural network. The proposed technique is experimentally validated using real three-dimensional ground penetrating radar data obtained from urban roads in Seoul, South Korea. [ABSTRACT FROM AUTHOR]

Solving the supply–demand imbalance is the most crucial issue for stable implementation of a public bike-sharing system. This gap can be reduced by increasing the accuracy of demand prediction by considering spatial and temporal properties of bike demand. However, only a few attempts have been made to account for both features simultaneously. Therefore, we propose a prediction framework based on graph convolutional networks. Our framework reflects not only spatial dependencies among stations, but also various temporal patterns over different periods. Additionally, we consider the influence of global variables, such as weather and weekday/weekend to reflect non-station-level changes. We compare our framework to other baseline models using the data from Seoul’s bike-sharing system. Results show that our approach has better performance than existing prediction models. [ABSTRACT FROM AUTHOR]