Search

This paper contributes to the ongoing debate on the nature and characteristics of the volatility transmission channels of major crash events in international stock markets between 03 July 1997 and 09 March 2021. Using dynamic conditional correlations (DCC) for conditional correlations and volatility clustering, GARCH-BEKK for the direction of transmission of disturbances, and the Diebold-Yilmaz spillover index for the level of volatility contagion, the paper finds that the climbs in external shock transmissions have long-lasting impacts in domestic markets due to the contagion effect during crisis periods. The findings also reveal that the heavier magnitude of financial stress is transmitted between Asian countries via the Hong Kong stock market. Additionally, the degree of volatility spillovers between advanced and emerging equity markets is smaller compared to the pure spillovers between advanced markets or emerging markets, offering a window of opportunity for international market participants in terms of portfolio diversification and risk management applications. Furthermore, the study introduces a novel early warning system created by integrating DCC correlations with a state-of-the-art deep learning model to predict the global financial crisis and COVID-19 crisis. The experimental analysis of long short-term memory network finds evidence of contagion risk by verifying bursts in volatility spillovers and generating signals with high accuracy before the 12-month crisis period. This provides supplementary information that contributes to the decision-making process of practitioners, as well as offering indicative evidence that facilitates the assessment of market vulnerability for policymakers. [ABSTRACT FROM AUTHOR]

Large-scale assessment of window views is demanded for precise housing valuation and quantified evidence for improving the built environment, especially in high-rise, high-density cities. However, the absence of a semantic segmentation dataset of window views forbids an accurate pixel-level assessment. This paper presents a City Information Model (CIM)-generated Window View (CIM-WV) dataset comprising 2,000 annotated images collected in the high-rise, high-density urban areas of Hong Kong. The CIM-WV includes seven semantic labels, i.e., building, sky, vegetation, road, waterbody, vehicle, and terrain. Experimental results of training a well-known deep learning (DL) model, DeepLab V3+ , on CIM-WV, achieved a high performance (per-class Intersection over Union (IoU) ≥ 86.23%) on segmenting major landscape elements, i.e., building, sky, vegetation, and waterbody, and consistently outperformed the transfer learning on a popular real-world street view dataset, Cityscapes. The DeepLab V3+ model trained on CIM-WV was robust (mIoU ≥ 72.09%) in Hong Kong Island and Kowloon Peninsula, and enhanced the semantic segmentation accuracy of real-world and Google Earth CIM-generated window view images. The contribution of this paper is three-fold. CIM-WV is the first public CIM-generated photorealistic window view dataset with rich semantics. Secondly, comparative analysis shows a more accurate window view assessment using DL from CIM-WV than deep transfer learning from ground-level views. Last, for urban researchers and practitioners, our publicly accessible DL models trained on CIM-WV enable novel multi-source window view-based urban applications including precise real estate valuation, improvement of built environment, and window view-related urban analytics. [ABSTRACT FROM AUTHOR]

Numerous news or event pictures are taken and shared on the internet every day that have abundant information worth being mined, but only a small fraction of them are geotagged. The visual content of the news image hints at clues of the geographical location because they are usually taken at the site of the incident, which provides a prerequisite for geo-localization. This paper proposes an automated pipeline based on deep learning for the geo-localization of news pictures in a large-scale urban environment using geotagged street view images as a reference dataset. The approach obtains location information by constructing an attention-based feature extraction network. Then, the image features are aggregated, and the candidate street view image results are retrieved by the selective matching kernel function. Finally, the coordinates of the news images are estimated by the kernel density prediction method. The pipeline is tested in the news pictures in Hong Kong. In the comparison experiments, the proposed pipeline shows stable performance and generalizability in the large-scale urban environment. In addition, the performance analysis of components in the pipeline shows the ability to recognize localization features of partial areas in pictures and the effectiveness of the proposed solution in news picture geo-localization. [ABSTRACT FROM AUTHOR]

This paper develops a cost-effective sequential route recommender system to provide real-time routing recommendations for vacant taxis searching for the next passenger. We propose a prediction-and-optimization framework to recommend the searching route that maximizes the expected profit of the next successful passenger pickup based on the dynamic taxi demand-supply distribution. Specifically, this system features a deep learning-based predictor that dynamically predicts the passenger pickup probability on a road segment and a recursive searching algorithm that recommends the optimal searching route. The predictor integrates a graph convolution network (GCN) to capture the spatial distribution and a long short-term memory (LSTM) to capture the temporal dynamics of taxi demand and supply. The GCN-LSTM model can accurately predict the pickup probability on a road segment with the consideration of potential taxi oversupply. Then, the dynamic distribution of pickup probability is fed into the route optimization algorithm to recommend the optimal searching routes sequentially as route inquiries emerge in the system. The recursion tree-based route optimization algorithm can significantly reduce the computational time and provide the optimal routes within seconds for real-time implementation. Finally, extensive experiments using Beijing Taxi GPS data demonstrate the effectiveness and efficiency of the proposed recommender system. History: Accepted by Ram Ramesh, Area Editor for Data Science and Machine Learning. Funding: This work was partially supported by the Hong Kong Research Grants Council [Grants CityU 21500220, CityU 11504322] and the National Natural Science Foundation of China [Grant 72201222]. Supplemental Material: The software that supports the findings of this study is available within the paper and its Supplemental Information (https://pubsonline.informs.org/doi/suppl/10.1287/ijoc.2021.0112) as well as from the IJOC GitHub software repository (https://github.com/INFORMSJoC/2021.0112). The complete IJOC Software and Data Repository is available at https://informsjoc.github.io/. [ABSTRACT FROM AUTHOR]

One of education's primary goals is to cultivate citizens who want to contribute to society. However, surprisingly little research has been conducted on how students' desire to contribute to society is related to crucial learning-related outcomes. The aim of this study was to examine how the desire to contribute to one's society, which we call societal motivation for shorthand, is associated with optimal learning-related outcomes such as self-regulated learning, deep learning, and achievement. The sample included 8,773 secondary school students from Hong Kong. The mean age of the students was 13.28 (SD = 1.09) years old. Students were asked to respond to self-reported surveys and answer achievement tests across two-time points, one year apart. Structural equation modeling was used to analyze the data. Results indicated that Time 1 societal motivation positively predicted Time 2 self-regulated learning and deep learning. These associations held despite controlling for auto-regressive effects and other relevant covariates such as Time 1 goals (mastery, performance, social, and extrinsic) and demographic variables. Our results showed that societal motivation is associated with optimal learning-related outcomes. This paper has theoretical implications for educational research by showing that societal motivation is an important yet neglected aspect of student motivation. [ABSTRACT FROM AUTHOR]

Artificial intelligence (AI) and deep learning techniques are playing an increasing role in astronomy to deal with the data avalanche. Here we describe an application for finding resolved planetary nebulae (PNe) in crowded, wide-field, narrow-band Hα survey imagery in the Galactic plane, to test and facilitate more objective, reproducible, efficient and reliable trawls for them. PNe are important for studying the late-stage stellar evolution of low-mass to intermediate-mass stars. However, the confirmed ∼3800 Galactic PNe fall far short of the numbers expected. Traditional visual searching for resolved PNe is time-consuming because of the large data size and areal coverage of modern astronomical surveys. The training and validation data set of our algorithm was built with the INT Photometric Hα Survey (IPHAS) and true PNe from the Hong Kong/AAO/Strasbourg Hα (HASH) data base. Our algorithm correctly identified 444 PNe in the validation set of 454 PNe, with only 16 explicable 'false' positives, achieving a precision rate of 96.5 per cent and a recall rate of 97.8 per cent. After transfer learning, it was then applied to the VST Photometric Hα Survey of the Southern Galactic plane and bulge (VPHAS+), examining 979 out of 2284 survey fields, each covering 1° × 1°. It returned ∼20 000 detections, including 2637 known PNe and other kinds of catalogued non-PNe. A total of 815 new high-quality PNe candidates were found, 31 of which were selected as top-quality targets for optical spectroscopic follow-up. We found that 74 per cent of them are true, likely, and possible PNe. Representative preliminary confirmatory spectroscopy results are presented here to demonstrate the effectiveness of our techniques, with full details to be given in our forthcoming paper. [ABSTRACT FROM AUTHOR]

This paper shows that investments based on deep learning signals extract profitability from difficult-to-arbitrage stocks and during high limits-to-arbitrage market states. In particular, excluding microcaps, distressed stocks, or episodes of high market volatility considerably attenuates profitability. Machine learning-based performance further deteriorates in the presence of reasonable trading costs because of high turnover and extreme positions in the tangency portfolio implied by the pricing kernel. Despite their opaque nature, machine learning methods successfully identify mispriced stocks consistent with most anomalies. Beyond economic restrictions, deep learning signals are profitable in long positions and recent years and command low downside risk. This paper was accepted by Kay Giesecke, finance. Funding: D. Avramov acknowledges the Israel Science Foundation (Grant 288/18) for financial support. S. Cheng acknowledges the General Research Fund of the Research Grants Council of Hong Kong [Project 14502318] for financial support. Supplemental Material: The data files and online appendix are available at https://doi.org/10.1287/mnsc.2022.4449. [ABSTRACT FROM AUTHOR]

Travel time estimation plays an essential role in the high-granular traffic control and management of urban roads with distinct lane-changing conditions among lanes. However, little attention has been given to the estimation of distributions of travel times among different lanes and different vehicle types in addition to their expected values. This paper proposes a new method for estimating lane-based travel time distributions with consideration of different vehicle types through matching low-resolution vehicle video images taken by conventional traffic surveillance cameras. The vehicle type classification is based on vehicle sizes and deep learning features extracted by densely connected convolutional neural networks, and the vehicle re-identification is conducted through a lane-based bipartite graph matching technique. A case study is carried out on a congested urban road in Hong Kong. Results show that the proposed method performs well in estimating the lane-level travel time distributions by vehicle type which can be very helpful for various lane-based and vehicle type-specific traffic management schemes. [ABSTRACT FROM AUTHOR]

Rock classification provides vital information to geosciences and geological engineering practices. Reaping the benefits of the advances of computer vision-based deep learning artificial intelligence (AI) technology, this study aims to develop a next-generation convolutional neural network (CNN) to perform automatic rock classification. Two major challenging issues have been particularly addressed. First, most of the previous rock classifications are simply transfer learning of CNNs that are trained by life-like scenarios. Second, classifying rock types with similar textures leads to severe overfitting of CNNs. In this study, a novel CNN called HKUDES_Net is proposed and implemented to classify seven common Hong Kong rock types, namely fine-grained granite, medium-grained granite, coarse-grained granite, coarse ash tuff, fine ash tuff, feldsparphyric rhyolite, and granodiorite. With the aid of dynamic expansion, squeeze and excitation, and other strategies, HKUDES_Net can classify rock types with similar texture patterns/colors but different grain sizes. As compared with the other ten landmark CNNs and seven feature-based algorithms, HKUDES_Net has the best performance in precision (90.9%), recall (90.4%), and f1-score (90.5%). By implementing the alerting level, which restricts the training loss hovering above a small constant and prevents the validation loss from rising, the overfitting has been efficiently eliminated. The proposal and implementation of HKUDES_Net highlight the value of interdisciplinary research and will continuously pave the way for better coupling AI and geosciences. Highlights: A next-generation convolutional neural network (CNN) called HKUDES_Net is developed to perform automatic rock classification. HKUDES_Net can classify rock types with similar texture patterns/colors but different grain sizes. By implementing the alerting level in HKUDES_Net, the overfitting has been efficiently eliminated. Plain Language Summary: Rock classification provides vital information to geosciences and geological engineering practices. However, the traditional way of rock classification is to a certain degree experience-based. Computer vision-based artificial intelligence (AI), especially convolutional neural network (CNN), a sub-branch of AI, may help people automatically classify rock types with the hope to train AI models (CNNs), which can instantly and precisely perceive rock images similar to how human beings observe rocks. In other words, by putting a rock image into a CNN, the CNN will read the information from the image and tell people what rock type in the image is. This study develops and implements a next-generation CNN called HKUDES_Net to classify seven different rock types. Adopting different computational strategies, HKUDES_Net can classify rock types of different grain sizes with similar texture patterns/colors, which is challenging for other CNNs. As compared with the other landmark CNNs and machine learning algorithms, HKUDES_Net has much a better prediction accuracy performance. The proposal and implementation of HKUDES_Net exemplify an interdisciplinary research study, and we are confident that this paper will be instrumental to pave the way for further coupling AI and geosciences. [ABSTRACT FROM AUTHOR]

Nowadays, event forecasting in Twitter can be considered an essential, significant and difficult issue. Maximum conventional methods are focusing on temporal events like sports or elections. These methods do not calculate the spatial features too their correlation analysis. Hence, this paper proposes an Improved Deep Belief Neural Network (iDBNN) for civil unrest event forecasting in twitter data. This proposed method is utilized to forecast the future event with the consideration of the tweets. The proposed method is designed with three phases named as pre-processing phase, feature extraction phase, and civil unrest event forecasting. Initially, the proposed method is used to train the Hong Kong Protest event 2019 tweet data for forecasting events. In the pre-processing phase, removal of special symbol, removal of URL, username removal, tokenization and stop word removal are done. After that, the essential features such as domain weight, event weight, textual similarity, spatial similarity, temporal similarity, and Relative Document-Term Frequency Difference (RDTFD) are extracted and then applied for training the proposed model. To empower the training phase of proposed iDBNN method, the Jellyfish Algorithm is utilized to select optimal weight parameter coefficients of DBNN for training the model parameters. The projected technique is authenticated by statistical capacities and compared with the conventional methods such as Hidden Markov Model (HMM) and Random Forest (RF) respectively. Comparing with other traditional methods, the proposed model shows better performance in terms of prediction and processing time. The iDBNN model shows 91% prediction accuracy that is much higher than the traditional DBNN. [ABSTRACT FROM AUTHOR]

Mapping of highly dynamic changes in land use and land cover (LULC) can be hindered by various cloudy conditions with optical satellite images. These conditions result in discontinuities in high-temporal-density LULC mapping. In this paper, we developed an integrated time series mapping method to enhance the LULC mapping accuracy and frequency in cloud-prone areas by incorporating spectral-indices-fused deep models and time series reconstruction techniques. The proposed method first reconstructed cloud-contaminated pixels through time series filtering, during which the cloud masks initialized by a deep model were refined and updated during the reconstruction process. Then, the reconstructed time series images were fed into a spectral-indices-fused deep model trained on samples collected worldwide for classification. Finally, post-classification processing, including spatio-temporal majority filtering and time series refinement considering land–water interactions, was conducted to enhance the LULC mapping accuracy and consistency. We applied the proposed method to the cloud- and rain-prone Pearl River Delta (i.e., Guangdong–Hong Kong–Macao Greater Bay Area, GBA) and used time series Sentinel-2 images as the experimental data. The proposed method enabled seamless LULC mapping at a temporal frequency of 2–5 days, and the production of 10 m resolution annual LULC products in the GBA. The assessment yielded a mean overall accuracy of 87.01% for annual mapping in the four consecutive years of 2019–2022 and outperformed existing mainstream LULC products, including ESA WorldCover (83.98%), Esri Land Cover (85.26%), and Google Dynamic World (85.06%). Our assessment also reveals significant variations in LULC mapping accuracies with different cloud masks, thus underscoring their critical role in time series LULC mapping. The proposed method has the potential to generate seamless and near real-time maps for other regions in the world by using deep models trained on datasets collected globally. This method can provide high-quality LULC data sets at different time intervals for various land and water dynamics in cloud- and rain-prone regions. Notwithstanding the difficulties of obtaining high-quality LULC maps in cloud-prone areas, this paper provides a novel approach for the mapping of LULC dynamics and the provision of reliable annual LULC products. • Novel LULC mapping via indices-fused deep models and time series reconstruction. • High-temporal-density 10 m seamless LULC mapping in cloudy and rainy areas (PRD). • Overall mapping accuracy yielded up to 87.01%, outperforming existing LULC products. • Potential to generate seamless near real-time maps and high-quality LULC data sets. [ABSTRACT FROM AUTHOR]

In Hong Kong, merchandise trade statistics are compiled based on the commodity information given on the trade declarations submitted by traders. Due to the complexity of the standardised commodity classification system (i.e. Hong Kong Harmonized System, or HKHS in short), there are often reporting errors, especially in the commodity codes and quantities. With around 20 million declarations received annually, the availability of this big data source motivates us to adopt deep learning techniques to detect the reporting errors. This paper proposes a mechanism consisting of three deep learning models for checking the commodity code, quantity and value, which offers an end-to-end solution to data quality assurance for declarations. The results show that the proposed mechanism could enhance the accuracy of error detection, which is conducive to improving the quality of trade statistics. With the use of text analytics techniques, the mechanism could fully utilise free-text commodity descriptions declared by traders to check the accuracy of the declared information comprehensively. It also overcomes some limitations of the traditional rule-based models. The whole study demonstrates the potential of using deep learning approach in quality assurance of existing statistical systems for official statistics. [ABSTRACT FROM AUTHOR]

Short-term origin–destination (OD) flow prediction is vital for operations planning, control, and management in urban railway systems. While the entry and exit passenger demand prediction problem has been studied in various studies, the OD passenger flow prediction problem receives much less attention. One key challenge for short-term OD flow prediction is the partial observability of the OD flow information due to trips having not been completed at a certain time interval. This paper develops a novel deep learning architecture for the OD flow prediction in urban railway systems and examines various mechanisms for data representation and for dealing with partial information. The deep learning framework consists of three main components, including multiple LSTM networks with an attention mechanism capturing short/long-term temporal dependencies, a temporally shifted graph matrix for spatiotemporal correlations, and a reconstruction mechanism for partial OD flow observations. The model is validated using smart card data from Hong Kong's Mass Transit Railway (MTR) system and compared with state-of-the-art prediction models. Experiments are designed to examine the characteristics of the proposed approach and its various components. The results show the superior performance (accuracy and robustness) of the proposed model and also the importance of partial observations of OD flow information in improving prediction performance. In terms of data representation, predicting the deviation of OD flows performs consistently better than predicting OD flows directly. [ABSTRACT FROM AUTHOR]

The dissemination of fake news on social media platforms is an issue of considerable interest, as it can be used to misinform people or lead them astray, which is particularly concerning when it comes to political events. The recent event of Hong Kong protests triggered an outburst of fake news posts that were identified on Twitter, which were then promptly removed and compiled into datasets to promote research. These datasets focusing on linguistic content were used in previous work to classify between tweets spreading fake and real news using traditional machine learning algorithms (Zervopoulos et al., in: IFIP international conference on artificial intelligence applications and innovations, Springer, Berlin, 2020). In this paper, the experimentation process on the previously constructed dataset is extended using deep learning algorithms along with a diverse set of input features, ranging from raw text to handcrafted features. Experiments showed that the deep learning algorithms outperformed the traditional approaches, reaching scores as high as 99.3% F1 Score, with the multilingual state-of-the-art model XLM-RoBERTa outperforming other algorithms using raw untranslated text. The combination of both traditional and deep learning algorithms allows for increased performance through the latter, while also gaining insight regarding tweet structure from the interpretability of the former. [ABSTRACT FROM AUTHOR]

Tourism demand forecasting is a critical process in the planning of tourism utilities. In recent years, Internet search indices have been popularly used as tourism demand indicators. However, due to the complex relationship between tourism demand and the search index and the vast amounts of search engine data, the traditional econometric and artificial intelligence models could not be enough to complete the prediction task. Under this scenario, this paper proposes a novel two-step method to improve tourism demand prediction accuracy. Firstly, a double-boosting algorithm is proposed to select the keywords and their lags from the potential relevant high-dimensional search queries. Second, the ensemble Support Vector Regression (SVR) based Deep belief Network (DBN) approach is adopted to capture the possible non-linear relationship and to improve the forecasting performance through deep learning combination. The empirical results demonstrate that this procedure significantly outperforms other benchmark models when forecasting monthly Hong Kong tourist arrivals. [ABSTRACT FROM AUTHOR]

• This paper focuses on daily natural gas consumption forecasting. • A novel hybrid model is developed for daily natural gas consumption forecasting. • The most recent datasets of four cities in three climate zones are collected. • An improved singular spectrum analysis is first proposed. In daily natural gas consumption forecasting, the accuracy of forecasting models is vulnerably affected by the noise data in historical time series. Singular spectrum analysis (SSA) is often introduced into hybrid models for denoising. However, as a deterministic-based algorithm, SSA does not give good results when a time series is contaminated with a high noise level. Considering this fact, this paper proposes an improved SSA (ISSA) that modifies the determination method of subseries selection in the reconstruction stage of SSA. Combining ISSA with long short-term memory (LSTM), a novel hybrid model, ISSA-LSTM, is thus developed. Additionally, for validating the robustness and superiority of ISSA-LSTM, the historical datasets of four representative cities located in three climate zones are collected as the training and testing datasets, and a comparison of ISSA-LSTM with five advanced models is performed. The results reveal that SSA would generate negative values when time series close to zero and the contribution of SSA in improving the forecasting accuracy of LSTM is insignificant. In contrast, ISSA avoids generating negative values and reduces the mean absolute range normalized error (MARNE) of LSTM by a range of 0.86–11.86%. Among the models, ISSA-LSTM achieves the best performance and its MARNEs for London (temperate zone), Melbourne (subtropical zone), Karditsa (subtropical zone), and Hong Kong (tropical zone) are 4.68%, 5.72%, 5.76%, and 14.10%, respectively. The MARNE of the tropical city is higher than that of others, which is caused by the complex natural gas consumption pattern of itself. [ABSTRACT FROM AUTHOR]

Squatter structures have been a serious threat to human safety and health for a long time. And monitoring their changes is important to facilitate government management of squatters. However, existing methods are still not automatic, accurate and fast enough to meet the actual needs of practical applications. In this paper, we propose a novel deep learning-based method to detect squatter structure changes from bi-temporal remotely sensed (RS) images and digital surface models (DSMs). The proposed convolutional neural network (CNN) takes the advantages of the spectral information from high resolution image and the height information from the DSM, so as to detect changes more accurately in type and height of squatter structures. Moreover, we create a data set for deep learning model training, covering a variety of squatter structures in Hong Kong. Compared with three existing representative methods, Our model performs the best, with Kappa of 0.6786 and 0.6458 in the detection results of the two test regions, respectively, which indicates that it has application potential. [ABSTRACT FROM AUTHOR]

Purpose: Our purpose was to develop a deep learning-based computed tomography (CT) perfusion mapping (DL-CTPM) method that synthesizes lung perfusion images from CT images.Methods and Materials: This paper presents a retrospective analysis of the pulmonary technetium-99m-labeled macroaggregated albumin single-photon emission CT (SPECT)/CT scans obtained from 73 patients at Queen Mary Hospital in Hong Kong in 2019. The left and right lung scans were separated to double the size of the data set to 146. A 3-dimensional attention residual neural network was constructed to extract textural features from the CT images and reconstruct corresponding functional images. Eighty-four samples were randomly selected for training and cross-validation, and the remaining 62 were used for model testing in terms of voxel-wise agreement and function-wise concordance. To assess the voxel-wise agreement, the Spearman's correlation coefficient (R) and structural similarity index measure between the images predicted by the DL-CTPM and the corresponding SPECT perfusion images were computed to assess the statistical and perceptual image similarities, respectively. To assess the function-wise concordance, the Dice similarity coefficient (DSC) was computed to determine the similarity of the low/high functional lung volumes.Results: The evaluation of the voxel-wise agreement showed a moderate-to-high voxel value correlation (0.6733 ± 0.1728) and high structural similarity (0.7635 ± 0.0697) between the SPECT and DL-CTPM predicted perfusions. The evaluation of the function-wise concordance obtained an average DSC value of 0.8183 ± 0.0752 for high-functional lungs (range, 0.5819-0.9255) and 0.6501 ± 0.1061 for low-functional lungs (range, 0.2405-0.8212). Ninety-four percent of the test cases demonstrated high concordance (DSC >0.7) between the high-functional volumes contoured from the predicted and ground-truth perfusions.Conclusions: We developed a novel DL-CTPM method for estimating perfusion-based lung functional images from the CT domain using a 3-dimensional attention residual neural network, which yielded moderate-to-high voxel-wise approximations of lung perfusion. To further contextualize these results toward future clinical application, a multi-institutional large-cohort study is warranted. [ABSTRACT FROM AUTHOR]

This paper reports a new knowledge-driven engineered feature learning approach in response to the Global AI Challenge for Building E&M Facilities held by the Electrical and Mechanical Service Department (EMSD) of the Hong Kong SAR. The results were awarded with a Grand Prize by the competition organizer. A dynamically engineered multi-modal feature learning (DEMMFL) method is proposed for predicting the cooling load of two office buildings. The DEMMFL model is estimated with the Lasso-ridge regression and compared with other well-known methods such as the Lasso. The novel approach applies control system knowledge to engineer useful features and explore load patterns for multi-mode modeling. Deep learning methods including LSTM, GRU, and AutoGluon are implemented for automated machine learning and tested in parallel to compare the performance of the proposed model with existing methods. The proposed model is demonstrated to predict long-term cooling load most accurately using engineered features from weather information only. • A new dynamic feature engineering method for office building cooling load prediction. • Slow and fast dynamic features engineered from available weather data only. • Control-knowledge based feature to learn on-hours dynamic responses. • Multi-modal approach with eight day-type features to learn occupancy loads. • Deep recurrent learning and AutoML methods compared to the proposed method. [ABSTRACT FROM AUTHOR]

Coastal ship berth planning and management necessitate accurate identification of ship position and size, yet this is complicated by the presence of diverse ship types and mixed distribution patterns. Traditional survey-based and position-tracking ship monitoring methods have limitations in reflecting the precise spatial information of coastal ships and their surroundings. This paper introduces a novel aerial visual data-driven methodology to estimate the berthing capacity of sheltered space for stationary ships in geographically dispersed water areas. To measure the berthing capacity, image data from an unmanned aerial vehicle (UAV) survey is collected to reflect the spatial information of all types of ships. A deep learning-based computer vision algorithm is used to automatically detect, identify, and classify ships in sample images. Additionally, we introduce the occupancy factor to empirically quantify the spatial proportional correlation between the ship size and the requisite berthing capacity. A validation test of the proposed method is conducted across all sheltered spaces in Hong Kong, and the UAV survey yields over 7,000 informative images of all types of local ships. The ship identification algorithm achieves an average identification rate of 98.6%, demonstrating higher accuracy and reduced labor and time consumption than observational survey and manual counting methods. Findings indicate that the suggested sheltered space for individual ships is 7.75 times the ship size. The approach provides greater flexibility in estimating berthing capacity for mixed berthing forms in extensive or geographically dispersed water bodies, thereby having the potential to support the design of berths, anchorage areas, and typhoon shelter facilities in coastal regions. [ABSTRACT FROM AUTHOR]

In densely built urban areas such as Hong Kong, the positive effect of urban trees is to help maintain high environmental and social sustainability for the city while unmanaged trees lead to negative effects such as accidents, outbreaks of pests and diseases. The public awareness of urban tree population has been increasing and preserving all the benefits offered by trees, a continuous monitoring concept would be required. In this work, an efficient 3D map system for tree inventory in Hong Kong is presented to the based on automated tree detection from publicly available Google street view (GSV) panorama images. First, Convolutional Neural Networks (CNNs) based object detector and classifier – YOLOv3 with pretrained model is adopted to learn GSV images to detect tree objects. GSV depth image has been utilized to decode depth values of each GSV panorama image and will provide accurate information to calculate the tree geographic position. A "field of view" filter was designed to remove duplicated tree detection within the overlapped areas followed by spatial clustering applied to further increase the tree localization accuracy. The average distance between the detected trees and ground truth data was achieved within 3 meters for selected roads used for the experiment. Second, a 3D Map platform prototype for facilitating the urban tree monitoring and management was developed. Currently, there is no true 3D platform for interpreting the results of tree records in Hong Kong city areas. With the help of webGL technology, contemporary browsers are able to show 3D buildings, terrain and other scene components together with the obtained tree records in an open source 3D GIS platform, the level of visualization is enhanced as all the detected trees are placed on the 3D digital terrain model. Consequently, it is easy for end-users to know the actual position of the trees and their distribution. [ABSTRACT FROM AUTHOR]

Multipath and non-light-of-sight (NLOS) reception greatly deteriorate the GNSS positioning accuracy in urban areas. Currently, most of the attempts using machine learning in this area focus on signal status prediction (LOS, multipath, NLOS). This paper exploits the capability of deep learning in multipath/NLOS mitigation. A new input feature, the single-differenced residual map is proposed, which has a high correlation with the user location and is very effective in multipath/NLOS mitigation. Combining the domain knowledge in GNSS, features of residual maps from different satellites are extracted by the proposed network and generate the heat map to indicate the user location. The proposed network can significantly improve the positioning accuracy of 84% of the epochs in the dense urban to 5-meter level. In addition, our network has a superior generalization ability, reducing the error of 90% of the epochs to 7 m level in a new scenario. • A deep learning network structure to directly estimate urban GNSS positioning. • A new GNSS feature, single-differenced residual maps, is proposed to improve the positioning accuracy of the neural network. • The proposed network is validated and tested using real GNSS raw data collected in urban canyons at Hong Kong. [ABSTRACT FROM AUTHOR]

BACKGROUND: The protective effect of urban greenery from adverse heat impacts remains inconclusive. Existing inconsistent findings could be attributed to the different estimation techniques used. OBJECTIVES: We investigated how effect modifications of urban greenery on heat–mortality associations vary when using different greenery measurements reflecting overhead-view and eye-level urban greenery. METHODS: We collected meteorological and daily mortality data for 286 territory planning units between 2005 and 2018 in Hong Kong. Three greenery measurements were extracted for each unit: a) the normalized difference vegetation index (NDVI) from Landsat remote sensing images, b) the percentage of greenspace based on land use data, and c) eye-level street greenery from street view images via a deep learning technique. Time-series analyses were performed using the case time series design with a linear interaction between the temperature term and each of the three greenery measurements. Effect modifications were also estimated for different age groups, sex categories, and cause-specific diseases. RESULTS: Higher mortality risks were associated with both moderate and extreme heat, with relative risks (RRs) of 1.022 (95% CI: 1.000, 1.044) and 1.045 (95% CI: 1.013, 1.079) at the 90th and 99th percentiles of temperatures relative to the minimum mortality temperature (MMT). Lower RRs were observed in greener areas whichever of the three greenery measurements was used, but the disparity of RRs between areas with low and high levels of urban greenery was more apparent when using eye-level street greenery as the index at high temperatures (99th percentile relative to MMT), with RRs for low and high levels of greenery, respectively, of 1.096 (95% CI: 1.035, 1.161) and 0.985 (95% CI: 0.920, 1.055) for NDVI (푝=0.0193), 1.068 (95% CI: 1.021, 1.117) and 0.990 (95% CI: 0.906, 1.081) for the percentage of greenspace (푝=0.1338), and 1.103 (95% CI: 1.034, 1.177) and 0.943 (95% CI: 0.841, 1.057) for eye-level street greenery (푝 =0.0186). Health discrepancies remained for nonaccidental mortality and cardiorespiratory diseases and were more apparent for older adults (≥ 65 years of age) and females. DISCUSSION: This study provides new evidence that eye-level street greenery shows stronger associations with reduced heat–mortality risks compared with overhead-view greenery based on NDVI and percentage of greenspace. The effect modification of urban greenery tends to be amplified as temperatures rise and are more apparent in older adults and females. Heat mitigation strategies and health interventions, in particular with regard to accessible and visible greenery, are needed for helping heat-sensitive subpopulation groups in coping with extreme heat. [ABSTRACT FROM AUTHOR]

• A real-time flood simulation method based on cross-scale correlation has been proposed. • Super-resolution deep learning models can downscale flow depth and velocity. • The deep learning model can be trained with limited assumed rainfall scenarios. • The proposed method simulates pluvial flooding process under arbitrary rainstorms. • The analysis is 2690 times faster, with errors of 0.082 m for depth and 0.088 m/s for velocity. Real-time information on flooding extent, severity, and duration is necessary for effective metropolitan flood emergency management. Existing pluvial flood analysis methods are unable to simulate real-time regional flooding processes under spatiotemporally varying rainstorms. This paper presents a deep learning-enabled super-resolution hydrodynamic flood analysis method to simulate the real-time pluvial flooding process over a large area under spatiotemporally varying rainstorms. Compared with existing flood downscaling techniques, which are limited to flow depth, the proposed method produces high-resolution flow depth and velocity predictions, providing more comprehensive information for flood emergency management. The proposed method adopts a coarse-grid hydrodynamic model to generate a low-resolution flood map time series, which is subsequently converted to high-resolution flood maps by a deep learning model. The deep learning model can be trained using a limited number of assumed rainfall scenarios, which greatly reduces data preparation effort. The proposed method is applied to a complex terrain of 352 km2 in Hong Kong that covers both mountainous and urban areas. Results show that the proposed method simulates the spatiotemporal variations of flood depth and velocity with root mean square errors as low as 0.082 m and 0.088 m/s, respectively, and correlation coefficients of 0.962 and 0.921, respectively. The computation time for a 48-h rainfall event in the study area is less than 30 s, which is 2690 times faster than the direct fine-grid hydrodynamic analysis. The deep learning-enabled super-resolution hydrodynamic flood analysis method provides a promising computational tool for emergency flood risk management. [ABSTRACT FROM AUTHOR]

Short-term wind speed predication is of great significance for scholars (e.g., understanding wind profiles), practitioners (e.g., building energy management), regulators (e.g., urban microclimate regulation), and even the general public. Current wind speed forecasting methods either generate sparse predictions or occur high cost. This paper reports a novel, inexpensive framework to forecast urban local dense wind speed. The central tenet is a convolutional long short-term memory (ConvLSTM) and LSTM combinatorial deep learning model to learn the features of input historical weather image series coupled with spatial-temporal correlations. The model was trained and tested using Hong Kong datasets. The feasibility and effectiveness of the proposed model are verified and compared with parallel models under different criteria, including mean absolute error (MAE), root mean square error (RMSE) and R-squared (R2). The experimental results show that: (1) the proposed ConvLSTM-LSTM deep learning model can effectively forecast wind speed regardless of location; (2) the overall MAE, RMSE, and R2 value of the proposed model are improved by 14.84%, 15.04%, and 7.51%, respectively, compared to the ConvLSTM-full connected (ConvLSTM-FC) model, and by 22.12%, 22.80%, and 12.24%, respectively, compared to the convolutional neural network-LSTM (CNN-LSTM) model; and (3) compared with parallel models, the proposed model has better performance in predicting wind speed series with large amplitude variations and rapid frequency changes. • A novel framework and a deep learning model are proposed for dense wind speed forecasting. • The limitations of wind speed prediction intervals and locations are addressed. • The spatial-temporal correlations of weather image sequences are learnt by ConvLSTM-LSTM model. • The proposed model is compared with the parallel model at different prediction step sizes. [ABSTRACT FROM AUTHOR]

Public open space (POS) is essential to urban areas. Assessing them usually requires tedious approaches such as fieldwork and manual processes. Street View Imagery (SVI) and Computer Vision (CV) have been adopted in some urban environment research, bringing fine granularity and human perspective. However, limited aspects have been subject in these studies, and SVI and CV have not been used for holistic POS assessment. This research introduces a novel approach of employing them in conjunction with traditionally used geospatial and remote sensing data for automating POS assessment and doing so extensively. Indicators from both subjective and objective perspectives are developed, and CV algorithms are adopted for retrieving visual features. In a case study spanning 800 POS in Hong Kong and Singapore, a method is designed to predict both subjective and objective scores. The results demonstrate the perceptual models achieved acceptable to high accuracy scores, and suggest that SVI reflects different aspects of POS compared to previous approaches. The paper concludes that SVI can be adopted in POS assessment as a new instrument, extending their research scope to rarely considered off-road areas, and contributing with a new approach for the design and allocation of POS in urban planning. • Comprehensive assessment using street view imagery and computer vision • Uncovering the usability of rarely used street-level images beyond streetscapes • Developing and comparing objective and subjective perspectives for assessment • Combination of multi-sourced indicators reduces the bias of single data source. • Imagery adds unparalleled insights to traditionally used GIS or remote sensing data. [ABSTRACT FROM AUTHOR]

Deep learning plays a highly essential role in the domain of remote sensing change detection (CD) due to its high efficiency. From some existing methods, we can observe that the fusion of information at each scale is quite vital for the accuracy of the CD results, especially for the common problems of pseudo-change and the difficult detection of change edges in the CD task. With this in mind, we propose a New Fusion network with Dual-branch Encoder and Triple-branch Decoder (DETDNet) that follows a codec structure as a whole, where the encoder adopts a siamese Res2Net-50 structure to extract the local features of the bitemporal images. As for the decoder in previous works, they usually employed a single branch, and this approach only preserved the fusion features of the encoder's bitemporal images. Distinguished from these approaches, we adopt the triple-branch architecture in the decoder for the first time. The triple-branch structure preserves not only the dual-branch features from the encoder in the left and right branches, respectively, to learn the effective and powerful individual features of each temporal image but also the fusion features from the encoder in the middle branch. The middle branch utilizes triple-branch aggregation (TA) to realize the feature interaction of the three branches in the decoder, which enhances the integrated features and provides abundant and supplementary bitemporal feature information to improve the CD performance. The triple-branch architecture of the decoder ensures that the respective features of the bitemporal images as well as their fused features are preserved, making the feature extraction more integrated. In addition, the three branches employ a multiscale feature extraction module (MFE) per layer to extract multiscale contextual information and enhance the feature representation capability of the CD. We conducted comparison experiments on the BCDD, LEVIR-CD, and SYSU-CD datasets, which were created in New Zealand, the USA, and Hong Kong, respectively. The data were preprocessed to contain 7434, 10,192, and 20,000 image pairs, respectively. The experimental results show that DETDNet achieves F1 scores of 92.7%, 90.99%, and 81.13%, respectively, which shows better results compared to some recent works, which means that the model is more robust. In addition, the lower FP and FN indicate lower error and misdetection rates. Moreover, from the analysis of the experimental results, compared with some existing methods, the problem of pseudo-changes and the difficulty of detecting small change areas is better solved. [ABSTRACT FROM AUTHOR]

The aim of this study was to examine the synergistic effects of prosocial and mastery goals on learning-related outcomes including deep learning, self-regulated learning, and academic achievement. We recruited 8,773 secondary school students from Hong Kong to participate in the study. A longitudinal study design was used. Goals were measured at Time 1, while learning-related outcomes were measured at both Time 1 and 2. Structural equation modeling with latent interaction analysis was conducted. Results partially confirmed our synergistic effects hypothesis. Prosocial goals were associated with optimal learning-related outcomes such as deep learning and self-regulated learning only when mastery goals were also high. However, achievement was only predicted by mastery goals. These effects held despite controlling for the baseline variables. This study provides empirical evidence for the possible synergies between prosocial and mastery goals. A better understanding of student learning might be achieved when both mastery and prosocial motivation are examined. [ABSTRACT FROM AUTHOR]

This article presents a proof‐of‐concept designer‐in‐the‐loop schematic map drawing tool, based on the marriage of two approaches—manual and automated, which provides the technical interactivity of drawing tools between the user and the computer. We focus on concentric circle maps as opposed to the commonly used orthogonal mode representation, which is suggested by previous studies that it could promote better network learning. In comparison with existing methods, the proposed method is more compatible with the framework of effective map design from psychological and aesthetic perspectives, and a range of options can be provided in conjunction with users' preferences. We evaluated our approach on a set of iterations with case studies of Hong Kong metro with a group of three co‐authors from the fields of geography, transport engineering, and education. [ABSTRACT FROM AUTHOR]

Background: The heterogeneity of COVID-19 spread dynamics is determined by complex spatiotemporal transmission patterns at a fine scale, especially in densely populated regions. In this study, we aim to discover such fine-scale transmission patterns via deep learning. Methods: We introduce the notion of TransCode to characterize fine-scale spatiotemporal transmission patterns of COVID-19 caused by metapopulation mobility and contact behaviors. First, in Hong Kong, China, we construct the mobility trajectories of confirmed cases using their visiting records. Then we estimate the transmissibility of individual cases in different locations based on their temporal infectiousness distribution. Integrating the spatial and temporal information, we represent the TransCode via spatiotemporal transmission networks. Further, we propose a deep transfer learning model to adapt the TransCode of Hong Kong, China to achieve fine-scale transmission characterization and risk prediction in six densely populated metropolises: New York City, San Francisco, Toronto, London, Berlin, and Tokyo, where fine-scale data are limited. All the data used in this study are publicly available. Results: The TransCode of Hong Kong, China derived from the spatial transmission information and temporal infectiousness distribution of individual cases reveals the transmission patterns (e.g., the imported and exported transmission intensities) at the district and constituency levels during different COVID-19 outbreaks waves. By adapting the TransCode of Hong Kong, China to other data-limited densely populated metropolises, the proposed method outperforms other representative methods by more than 10% in terms of the prediction accuracy of the disease dynamics (i.e., the trend of case numbers), and the fine-scale spatiotemporal transmission patterns in these metropolises could also be well captured due to some shared intrinsically common patterns of human mobility and contact behaviors at the metapopulation level. Conclusions: The fine-scale transmission patterns due to the metapopulation level mobility (e.g., travel across different districts) and contact behaviors (e.g., gathering in social-economic centers) are one of the main contributors to the rapid spread of the virus. Characterization of the fine-scale transmission patterns using the TransCode will facilitate the development of tailor-made intervention strategies to effectively contain disease transmission in the targeted regions. [ABSTRACT FROM AUTHOR]

Previous evidence suggests that temperature is associated with the number of emergency department (ED) visits. A predictive system for ED visits, which takes local temperature into account, is therefore needed. This study aimed to compare the predictive performance of various machine learning methods with traditional statistical methods based on temperature variables and develop a daily ED attendance rate predictive model for Hong Kong. We analyzed ED utilization among Hong Kong older adults in May to September from 2000 to 2016. A total of 103 potential predictors were derived from 1- to 14-day lag of ED attendance rate and meteorological and air quality indicators and 0-day lag of holiday indicator and month and day of week indicators. LASSO regression was used to identify the most predictive temperature variables. Decision tree regressor, support vector machine (SVM) regressor, and random forest regressor were trained on the selected optimal predictor combination. Deep neural network (DNN) and gated recurrent unit (GRU) models were performed on the extended predictor combination for the previous 14-day horizon. Maximum ambient temperature was identified as a better predictor in its own value than as an indicator defined by the cutoff. GRU achieved the best predictive accuracy. Deep learning methods, especially the GRU model, outperformed conventional machine learning methods and traditional statistical methods. [ABSTRACT FROM AUTHOR]

Deep learning has garnered increasing attention in many research fields. However, prior research seldom focused on tourists' perception prediction and prescription towards tourism destinations in turbulence. This study attempts to fill the gap by investigating Mainland Chinese tourists' perception of a turbulent Hong Kong society through deep learning‐based sentiment analytics. This incorporates a convolutional neural network (CNN) model for sentiment prediction and feature frequency analysis for sentiment prescription for 52,950 Chinese travel microblogs about Hong Kong. Results show that the CNN‐based deep learning approach can obtain improved sentiment predictive performance with minimal domain knowledge and human effort. The trend of Mainland Chinese tourists' (MCTs') sentiments about Hong Kong and the unique sentiment features revealed by our approach can provide practitioners with new insights into design customised tourism marketing and development strategies. The MCTs' shared mentalities emerged from sentiment features may help to enhance our theoretical knowledge about tourists' perception in turbulent tourism markets. [ABSTRACT FROM AUTHOR]

Adequate empirical antimicrobial coverage is instrumental in clinical management of community-onset Enterobacteriaceae bacteraemia in areas with high ESBL prevalence, while balancing the risk of carbapenem overuse and emergence of carbapenem-resistant organisms. It is unknown whether machine learning offers additional advantages to conventional statistical methods in prediction of ESBL production. To develop a validated model to predict ESBL production in Enterobacteriaceae causing community-onset bacteraemia. 5625 patients with community-onset bacteraemia caused by Escherichia coli, Klebsiella species and Proteus mirabilis during 1 January 2015–31 December 2019 from three regional hospitals in Hong Kong were included in the analysis, after exclusion of blood cultures obtained beyond 48 h of admission. The prevalence of ESBL-producing Enterobacteriaceae was 23.7% (1335/5625). Deep neural network and other machine learning algorithms were compared against conventional statistical model via multivariable logistic regression. Primary outcomes compared consisted of predictive model area under curve of receiver-operator characteristic curve (AUC), and macro-averaged F1 score. Secondary outcomes included sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Deep neural network yielded an AUC of 0.761 (95% CI 0.725–0.797) and F1 score of 0.661 (95% CI 0.633–0.689), which was superior to logistic regression (AUC 0.667 (95% CI 0.627–0.707), F1 score 0.596 (95% CI 0.567–0.625)). Deep neural network had a specificity of 91.5%, sensitivity of 37.5%, NPV of 82.5%, and PPV of 57.9%. Deep neural network is superior to logistic regression in predicting ESBL production in Enterobacteriaceae causing community-onset bacteraemia in high-ESBL prevalence area. Machine learning offers clinical utility in guiding judicious empirical antibiotics use. [ABSTRACT FROM AUTHOR]

Continuous sign language recognition (CSLR) using different types of sensors to precisely recognize sign language in real time is a very challenging but important research direction in sensor technology. Many previous methods are vision-based, with computationally intensive algorithms to process a large number of image/video frames possibly contaminated with noises, which can result in a large translation delay. On the other hand, gesture-based CSLR relying on hand movement data captured on wearable devices may require less computation resources and translation time. Thus, it is more efficient to provide instant translation during real-world communication. However, the insufficient amount of information provided by the wearable sensors often affect the overall performance of this system. To tackle this issue, we propose a bidirectional long short-term memory (BLSTM)-based multi-feature framework for conducting gesture-based CSLR precisely with two smart watches. In this framework, multiple sets of input features are extracted from the collected gesture data to provide a diverse spectrum of valuable information to the underlying BLSTM model for CSLR. To demonstrate the effectiveness of the proposed framework, we test it on an extremely challenging and radically new dataset of Hong Kong sign language (HKSL), in which hand movement data are collected from 6 individual signers for 50 different sentences. The experimental results reveal that the proposed framework attains a much lower word error rate compared with other existing machine learning or deep learning approaches for gesture-based CSLR. Based on this framework, we further propose a portable sign language collection and translation platform, which can simplify the procedure of collecting gesture-based sign language dataset and recognize sign language through smart watch data in real time, in order to break the communication barrier for the sign language users. [ABSTRACT FROM AUTHOR]

Background: The early detection of carotid wall plaque is recommended in the prevention of cardiovascular disease (CVD) in moderate-risk patients. Previous techniques for B-mode carotid atherosclerotic wall plaque segmentation used artificial intelligence (AI) methods on monoethnic databases, where training and testing are from the "same" ethnic group ("Seen AI"). Therefore, the versatility of the system is questionable. This is the first study of its kind that uses the "Unseen AI" paradigm where training and testing are from "different" ethnic groups. We hypothesized that deep learning (DL) models should perform in 10% proximity between "Unseen AI" and "Seen AI". Methodology: Two cohorts from multi-ethnic groups (330 Japanese and 300 Hong Kong (HK)) were used for the validation of our hypothesis. We used a four-layered UNet architecture for the segmentation of the atherosclerotic wall with low plaque. "Unseen AI" (training: Japanese, testing: HK or vice versa) and "Seen AI" experiments (single ethnicity or mixed ethnicity) were performed. Evaluation was conducted by measuring the wall plaque area. Statistical tests were conducted for its stability and reliability. Results: When using the UNet DL architecture, the "Unseen AI" pair one (Training: 330 Japanese and Testing: 300 HK), the mean accuracy, dice-similarity, and correlation-coefficient were 98.55, 78.38, and 0.80 (p < 0.0001), respectively, while for "Unseen AI" pair two (Training: 300 HK and Testing: 330 Japanese), these were 98.67, 82.49, and 0.87 (p < 0.0001), respectively. Using "Seen AI", the same parameters were 99.01, 86.89 and 0.92 (p < 0.0001), respectively. Conclusion: We demonstrated that "Unseen AI" was in close proximity (<10%) to "Seen AI", validating our DL model for low atherosclerotic wall plaque segmentation. The online system runs < 1 s. [ABSTRACT FROM AUTHOR]

Time series forecasting technology and related applications for stock price forecasting are gradually receiving attention. These approaches can be a great help in making decisions based on historical information to predict possible future situations. This research aims at establishing forecasting models with deep learning technology for share price prediction in the logistics industry. The historical share price data of five logistics companies in Hong Kong were collected and trained with various time series forecasting algorithms. Based on the Mean Absolute Percentage Error (MAPE) results, we adopted Long Short-Term Memory (LSTM) as the methodology to further predict share price. The proposed LSTM model was trained with different hyperparameters and validated by the Root Mean Square Error (RMSE). In this study, we found various optimal parameters for the proposed LSTM model for six different logistics stocks in Hong Kong, and the best RMSE result was 0.43%. Finally, we can forecast economic recessions through the prediction of the stocks, using the LSTM model. [ABSTRACT FROM AUTHOR]