A Survey of Data Fusion in Smart City Applications

Healthcare is a crucial component in everyday life concerning medical and public practices using devices as defined by Lee and Co-authors[144, 145]. The rapid development of technology (e.g. smartphones and their in-built sensing devices such as heart rate sensors) provides more opportunities to adopt technology in healthcare applications pervasively. For telehealth application in smart city, Hossain et al. [112] have used electroencephalographic (EGG) signals and voice to monitor a specific user’s health with the support of cloud technology and doctor’s advices. In [113], work has shown to monitor elderly at home based on fuzzy fusion model using behavioral and acoustical environment data. Similarly, Noury [146] also monitors the activities and fall detection of elderly through fuzzy logic by fusing accelerometer, vibration, and orientation sensor. In [91], Marakkalage et al have used crowd-sensing data from a smartphone application (location, noise, light, etc.) and introduced sensor fusion based environment classification (SFEC) to profile elderly people for understanding their daily lifestyle. In addition, Dawar and Kehtarnavaz in [52] have implemented a Convolution Neural Network (CNN) to combine both depth camera and wearable devices to detect the transition of movements to fall. Apart from that, Hondori et al. [111] have proposed using sensor fusion between depth images and inertia to perform tele-rehab in the home. The main challenge occurs in pervasive smart healthcare data fusion is discussed in [147] as the need of a higher accuracy to improve sensing robustness against uncertainty and unreliable integration.

The concept of Smart Homes plays an important role nowadays in contemporary urban areas. According to Jiang et al. [148], the definition of a smart home provides the capability of controlling, monitoring, and accessed appliances & services through implementation of ICT. There are currently many big players in developing the smart home appliances such as Amazon, Google, Apple, IBM, Intel, Microsoft, Xiaomi, and others. The challenge faced by manufacturers are related with service integration and formulating software ontology platform. These are necessary for implementing the services through different vendors and allow for a better integration. Meanwhile in [114], physical sensors (soil moisture) and cyber (weather, traffic) have been fused to control home appliances such as alarm clock and water sprinkle. The study of user daily activity is yet another important aspect to understand urban citizen well-being. In [45], Hong et al. have combined series of life activities to understand the lifestyle pattern depends on the equally weighted sum operation and Dempster-Shafer theory. Also, similar study on the user daily activity patterns can be found in [110]. Combination of house environmental sensor (infrared, door contact, temperature, hygrometry sensor, microphone) and wearable devices (kinematic sensors) using support vector machine (SVM) can be used to identify the user activity patterns. In addition, the modeling of human behavior in a smart home [149] in order to generate learning situation models have proven the efficiency of context-aware services. In addition, smart home security is yet another study field for many researchers [150, 151, 152] due to increased usage of IoT devices in normal household. The research challenges is to develop the applications for the smart houses while retaining the privacy and security of the end user.

According to Smart Communities Guidebook [153], a smart community is described as “a geographical area ranging in size from neighborhood to a multi-county region whose residents, organizations, and governing institutions are using information technology to transform their region in significant ways”. There is only a handful of cities focus on this aspect as majority are still in the stage of transforming from facility to community welfare. First world countries such as USA, Canada, Australia, European Union, and Singapore shown in [154] have started up initiatives to create smart communities. Information fusion for smart community video surveillance system is performed in [79] to aid neighborhood in terms of security. The combination of the different modal surveillance camera provides a vast amount of visual information extraction such as video summarization for highlighting certain events. A distance learning framework is proposed in [115], which enables personalized learning to cater what is best for each individual user. It uses data fusion to understand user environment and their activities by means of hybrid data sources. Real-time community monitoring also helps to prevent emergency situations and it ensures the safety of community citizens. A good example for a smart community application in large-scale is the Social Credit System in China [155]. It is a state-owned system to collect data from both public (traffic cameras, transit data etc.) and private (online shopping, fitness trackers etc.) data sources to monitor and analyze user behaviour to generate a single ”credit score” for each person, which helps in community well-being. The techniques fuse these data sources and remains a back box to the general public. However, the effect on user privacy with the rise of “data state” remains a debate for some [156]. A mature citizen should be on alert and always responds to any potential threat, while spreading the awareness to build a safer community in the urban city.

In smart governance, managing a city is considered as a complex task as the integration of different domains and services is proven to be challenging. Transparent services integration is an example of why many governance authorities are having difficulties to sort it out. It is hard to strike a balance in developing a transparent governance policy with consideration of sensitive information. Therefore, there is only limited study materials available to the best of our knowledge. Janssen and Estevez [118] have proposed a centralized platform for cutting down government staff by shifting existing organization to rely on integration of platforms. The disaster response management is also considered as another vital element for a smart city to carry out any potential counter measurements towards disaster as shown in [157]. Apart from that, urban reporting system [56] has collected report from the city wide region on the faulty infrastructure so that immediate actions can be taken to remedy the situation. It uses cloud technology and focuses on the display of fused data report, which it also describes the location and types of infrastructure. Example of research challenges is to remove any potential fake report to prevent misuse of the reporting platform. Another example of smart governance that involves city safety can be found in [158], where it can act as an emergency aid application (light pulse on emergency through mesh network) while providing energy efficient lighting to urban area. Moreover, there are cities also working on governance platform such as New York [7], Singapore [10], Tokyo [11], Oslo [159], and others. The potential research opportunity is to propose consensus protocols within the city for better integration of services.

Urban planning plays an important role in developing the city economy by taking account of well-being of the urban residents. Traditionally in urban planning, aerial photography and statistical data sources (building size, population number, public amenities, etc.) are combined to understand the current development state of the city. The downside of such method is data sources frequently lacks of fine details, which resulting the output result is not representative. To address such issue, Cheng and Toutin [43] have combined various satellite and aerial images to generate details for the exiting urban structures. Alternately, low power sensors are capable to provide a larger coverage with lower deployment cost, which give researchers the opportunity to study different points of interest in the urban area. In [95, 119, 81], a bottom up urban planning method is implemented, where sensors are installed in a designated region to capture space utilization. From the collected data, urban planners can study public space utilization pattern using an integrated portal. Here, a hybrid processing method is proposed, where the data processing and fusion occur in different stages of data pipeline. In addition, a large variety of data sources can be used for urban planning such as physical sensors [160], photography [76, 161], or hybrid data sources [85]. Despite wide variety of data sources, human interpretation is required when it comes to make decision on a proposed urban design. The need of full automated planning system would further benefit the urban planners to combine different data sources in order to achieve a more ideal city planning.

Urban building management provides building owner a platform to understand building’s energy consumption rate while automating building resources management. It has been extensively studied in [25, 162, 163, 164] and the current trend is to optimize the building resources such as hot water systems, electrical consumption, and heating ventilation & air conditioning (HVAC). In [94], Aftab et al. have combined four different parameters to predict building occupancy to control HVAC using low-cost embedded systems. Some other works such as [97, 165, 166] also have the same objectives but using different types of data sources. The potential solution for better building management system is to rely on fusing weather, human feedback, and electricity price to fine tune the building resources in order to maximize human comfort, while minimizing the energy consumption. Apart from that, fire alarm system is considered another important features of the smart building management system. Luo and Su [117] have fused three different data sources (flame, smoke, and temperature sensor) to detect any potential fire outbreak and reduce false alarms. In addition, a notification-based system is implemented to notify the property owner and manager in case of emergency. In future, potential building safety features may include a group of robots to deal with fire hazards and double duty as building security patrols.

The main challenge of landscape monitoring in smart city is the sensing coverage of the data sources. To address such issue, two different sensing approaches have been used such as relying on mobile sensing or satellite-based data. Mobile sensing [106, 167] offers greater sensing capability by leveraging the mobility of moving objects (vehicles or humans). The mobile sensing technique provides a large spatial coverage, but it is not suitable for real-time applications unless there are multiple data sources to compensate the lack of spatial resolution concurrently. The output type of this mobile sensing includes combination of different spatial data in order to complete the data sources before proceed to data processing stage. Mobile sensing works such as [82, 168] utilized different data sources to complete spatial resolution and visualized the ambient changes across the city. The common characteristic of aforementioned works is feature extraction, which they visualize the processed features from the raw data sources. Majority of data input and output types in this domain are DAI-DAO and DAI-FEO since physical sensors are the common data sources. Using the satellite-based data sources, Shen et al. [46] have studied the UHI effect in a city using data sources collected over $26$ years. The UHI index changes are measured through the combination of Landsat and MODIS images data. Mobile sensing offers a lower deployment cost, where it sacrifice the spatial resolution given there is limited number of sensors. Also, it has a lower coverage compared to satellite data sources. In contrast, satellite data has a wider coverage of spatial resolution but it frequently needs data enhancement and lacks of finer details.

The surrounding natural resources of an urban city such as mountains and forests are considered as important assets of a city. The most common data sources in modeling the city area are satellite images, which as stated before, it requires data enhancement such as [169, 170] before using it. Therefore, prior work of data fusion  [171] was focused on improving the satellite images quality. Only until recently, the emergence of machine learning algorithms and faster computers have created new ways to extract large variety of satellite image features. For instance in [120] and [121], forest types classification have been conducted in order to understand the variety of tree species in a specific region of interest. Both methods involve region-wide data sources and classification techniques, which are used to identify the tree species based on the forest types. With a lower deployment cost, small satellite (smallsat) and nano satellite (nanosat) could improve spatial coverage to generate a better data sources. Smart city applications leveraging satellite data will also beneficial from these deployment.

With astonishing rate of garbage being generated daily, waste management for an urban city can be rather challenging. Thus, it is essential to handle the waste efficiently to improve on sustainability of a city. An example of such effort could be found in [23], where they have proposed three new aspects of a smart waste management system such as: (1) infrastructure to overlook the overall life cycle of the product, (2) new business models revolving the product life cycle for preventing any waste generation, and (3) intelligent sensor networks for waste management facilities. In [102], Catania and Ventura have combined the proximity reading and weight sensor from garbage bin to estimate the garbage capacity of a typical household. Afterwards, rubbish categories collected from user mobile devices and garbage trucks are combined to keep track of residential participation in recycling scheme. On the other hand, waste water treatment helps to manage liquid waste of urban city before discharging to river or reuse. Chang et al. [122] have combined landsat and MODIS dataset in order to trace the water pollution level of a lake. On top of that, a web portal has been deployed to visualize and monitor the water pollution region over the time. Currently, many researchers are working together to develop an efficient waste management system since there is only limited resources available on earth. The goal is to adopt the 3R (Reduce, Reuse, and Recycle) concept with the help of ICT to improve city resource sustainability.

Smart manufacturing denotes the factory that depends on ICT to optimize the manufacturing process by increasing the production throughput. In [98], De Vin et al. have proposed a simulation tool to test out the management decision support by fusing undisclosed data entries and manufacturing process events. Similar to the aforementioned approach, decision based fusion can also be seen in [173, 174], which combines different machinery sensors data and data warehouse entries. The data fusion integration also considers supply chain demand in order to further optimize the manufacturing process. The challenge in this domain is to develop a self-optimizing manufacturing process while delivering the products to meet the demand of supply chain. Therefore, smart manufacturing frequently has a high correlation with the supply chain and attempts to deliver the market needs. In addition, the robotics usage in the smart manufacturing domain is nothing new. Guo et al. [51] have proposed an anomaly detection to combat potential security aspects in the robots using sensor fusion technique such as state estimation.

The reliability and stability of the equipment and machinery is vital to all the industries to ensure smooth operation in production. Without the guarantee of smooth operation, any downtime can cost damages to reputation and also loses profit. Thus, preventive maintenance has been studied in [124, 125, 175, 176] and attempts to predict the remaining useful life (RUL) of a machine accurately. By accurately predicting the RUL, maintenance can be carried out on time to save cost only when needed. The common data fusion techniques for predicting RUL are neural network (NN) based model such as CNN and Deep NN (DNN). Please note that, common data source in this sub-domain is physical data source such as machine states, sensors readings, and related parameters. Nonetheless on the fault detection domain, machine fault detection can be found in [96, 123], where they describe the problem of fault diagnosis and apply data fusion techniques to overcome. State estimation and classification have been used to detect the current state of the machinery. The data sources share some similarity with the preventive maintenance, where lower level of data information is preferred. This yields a faster fault detection when compared to a complex data pipeline. The research challenge here is to develop a generic and a flexible maintenance system for different scale of applications adhering to the goal of accurate fault detection.

In order to produce sustainable food resources in smart city, smart farming [177, 178] has become a trend to meet the food supply demand in a smart city. There are two different sub-domains in smart farming such as land and sea agriculture. In the land agriculture aspect, planting crops using controlled environment has shed some light in fulfilling the city needs of fresh supplies. However, plant disease remains a potential threat to a highly-dense plantation crop framing. In [127], Moshou et al. have classified the plant disease infection through Self Organizing Map (SOM) by fusing the spectral reflection and fluorescence imaging data. This helps to isolate infected crops while it focuses on the production of healthy plants. Apart from that, electromagnetic induction sensors, vegetarian index, water stress level, and radiance data are combined in [179] to better determine the partition of the crop field. Similar work also can be found in [128], where Khanum et al. propose an ontology-based fuzzy logic to classify plant disease. The research gaps in this domain involve improving live stock management as well as optimizing smart farm. On the other hand, sea agriculture is responsible for supplying the seafood supplies in a city. Obtaining fresh seafood supplies in an urban city sometimes can be rather difficult due to various factors such as delivery, city location, weather, seasonal pricing, etc. Therefore, a fresh seafood supply in a city is often not guaranteed. In order to address such issue, Huang et al. [126] have provided a solution by integrating two types of cameras for seafood freshness inspection. Camera and near infrared spectroscopy are fused through PCA and use NN to classify the freshness index. The research gaps in this domain involve developing large scale fish breeding and also wide varieties of seafood product such as calm, mussels, abalone, etc. A potential solution such as smart fish breeding with IoT has been proposed in [180], where it suggests using a moving pod to breed fishes while transporting them to destination in a particular destination simultaneously.

Today, modern e-commerce platforms use multi modal data sources to reach and better understand their customers. This helps e-commerce vendors to give better product recommendations for their customers and it helps customers to make their decisions easily. Fusing customer data such as mobility, credit card purchases, and social media interactions is commonly used in modern recommender systems. In [87], Breur introduced the fusion of customer behavior data and market research data to obtain a holistic picture of the customer. Investors can leverage financial data to make investment decisions, as Hassan et al. [181] have introduced a fusion model of Hidden Markov Model (HMM), NN, and Genetic Algorithm (GA) for stock market prediction. Improving the consumer awareness is conducted in [129], by fusing real world (weather, geographical) and cyber world (Twitter, Facebook) data. The proposed system has two levels of fusion, which relies on hierarchical-based processing architecture. The data combined bottom level input and fed it into upper level for further processing to achieve its objectives.

In a smart supply chain, it often involves sources and destination tracking in order to understand the flow / processing of the objects. As discussed in [182], supply chain management and logistic are the fundamental of modern supplies on fulfilling the needs of an urban city. For instance in food supply chain, three tiers information fusion framework is proposed in [130] such as: (1) to accelerate data processing, (2) shelf life prediction, and (3) real-time supply chain planning. The proposed hierarchical information fusion architecture (HIFA) includes a process that is intelligently transforming the sensor’s data sources into usable decision-making information. Recently, combination of blockchain technology has paved a new way for revolutionizing the existing supply chain. In [107], Tian has shown the integration of blockchain and supply chain in the agri-food supply application. It aids consumers to trace the origin of food using Radio Frequency Identification (RFID) along with database or WSN. The information also includes food origin to help consumers to identify the brand authenticity and avoids consuming counterfeit products. The research gap in this sub domain concerns with the implementation of smart supply and it needs the involvement from various commercial organizations. The consensus and national regulations are also parts of the critical factors of smart supply implementation.

The advancement of transportation technology has granted accessibility for the humans to move around the globe with ease. This phenomenon has caused rapid expansion of the tourism commercial values contributed to a city side income. Since then, Internet resources such as travel blogs and recommendation systems have influenced public to venture different locations. For instance, recommendation system [57] has been developed to recommend the place to travel based on user’s information such as socioeconomic (e.g. age, education, and income) and psychological and cognitive (experience, personality, involvement, and so forth) groups. User choices are used as feedback to further fine-tune the recommendation system using Rocchio’s method. Apart from that, Miah et al. [77] have combined social media-generated big data (geo-tagged photos of tourist attraction places) to predict tourist behavioral patterns. Alternately, Viswanath et al. [131] used a smartphone based mobile application to passively track tourist location data and obtain user ratings for tourist attraction places to better understand the preferences of tourists when they visit tourist attractions. The potential research development for smart travel is to focus on using a smartphone application for improving travel experience by relying on real-time translation and augmented reality (AR) navigation.

This sub-domain aims to get the accurate position of individuals and further to provide services, such as route planning and navigation, to help them travel efficiently and comfortably, in both outdoor and indoor environments. For outdoor positioning, Global Positioning System (GPS) has been the most accurate, reliable and dominant technology since it was allowed for civilian use in 1980s [183, 184]. Less-accurate non-GPS positioning approaches, such as wifi-based localization and cell-tower triangulation, are sometimes used instead of (or together with) GPS, because they consume less energy [132, 133]. For indoor positioning, since GPS does not work well indoors, other positioning approaches have been proposed. The data collection technologies used for these approaches mainly include Wi-Fi (WLAN), inertial measurement unit (IMU), RFID tags, Bluetooth, global system for mobile communications (GSM), frequency modulation (FM), and ultra-wide band (UWB) [185, 186]. Meanwhile, multiple data sources are often fused to achieve more accurate localization results [134, 135]. Once accurate locations are obtained, either indoors or outdoors, location-based services (e.g. route planning and navigation) can be provided to end users by fusing the location sequences with other information sources such as geographic information system (GIS) data, real-time traffic data, and user preference data [136, 137, 187, 188, 189, 190]. Since the outdoor positioning and location-based services have been well developed and commercialized, the current research trend in this field is mainly focused on improving the performance (accuracy, deployment cost, and energy cost) of indoor systems and services.

Positioning systems not only enable the location-based services for individuals but also provide data sources for further monitoring and understanding human mobility in a larger and more comprehensive scale. By aggregating and analyzing (fusing) the location data of residents along with GIS data of the environment, various aspects of human mobility can be monitored and the hidden patterns can be obtained. As summarized in [100], the most common subjects of monitoring and understanding human mobility include distance and duration distributions [191], origin-destination matrices [103], individual activity-based mobility patterns [192], transportation mode identification [54], and densities and flows within a building (or a cluster of buildings) [138, 193]. Results obtained from these subjects provide clues for improving transportation system [194], urban planning [195], and communication network [196]. Typical studies in this sub-domain usually fuse one data source of people’s movement trajectories with the environment information, such as GIS data of the city or floor plan of a building. Although this type of approach has produced much deeper insights compared with traditional approach relied on survey data, there is a trend to fuse multiple data sources related to people’s movement and obtain a more comprehensive picture of human mobility [197, 198]. Moreover, social media data sources. such as Tweets, also bring in more information regarding the mobility status in cities due to the combination of spatio-temporal data and descriptive text [86, 199].

Another large part of smart mobility is the improvement of transportation systems, which mainly comes from three aspects: relieving traffic congestion, improving public transportation, and introducing new transport systems. To relieve traffic congestion, effective light control plays an important role. While existing light control systems are usually based on hand-crafted rules and do not adjust to the rapid dynamics of traffic flows, intelligent light control approaches have been proposed using different data sources, data fusion techniques, and decision making (optimization and control) algorithms [109, 200].

Challenges in this aspect mainly come from the implementation of such intelligent light control approaches. Improvement of the public transportation system is mainly conducted through the network and schedule optimization [201]. Although these two topics have been thoroughly discussed in the literature, new insights related to the public transit system (e.g. origin-destination matrices and service level obtained from big data) [139, 202] and more advanced transport modeling tools enabled by big data [203] have brought new opportunities.

Even if the existing transportation manner has been optimized, there are still problems that cannot be solved, such as last mile issue and driving accidents. Therefore, new transport systems, such as bike sharing systems and autonomous vehicle systems, are introduced. Advanced ICT and data fusion techniques are the core of the realization of these systems. For a bike sharing system, data fusion and analysis helps to understand how the system works and evaluate different operational strategies [204, 205]. As for the autonomous vehicle system, the control of an autonomous vehicle itself is a complex data fusion process, fusing various data sources about the vehicle and the road by advanced machine learning and control algorithms [140, 206]. Security plays an important role in the autonomous vehicles deployment to ensure reliability of the autonomous driving. Examples of such techniques can be found in [50, 207].

The electrical grid provides an intermediate platform for relaying the electricity from the power plant to residential and industrial area. The common goal in this sub-domain is to provide reliable and stable electricity supply with the integration of ICT, which is commonly known as smart grid. Smart grid has been extensively studied in [55, 208, 209, 210] and the goal is to address on load and demand balancing of electricity in a particular area, building, or even household. Common technique applied in this sub-domain is forecasting, and example of such application can be found in [55], which it combines the information received from residential meters and predicts the electricity consumption load. Wang et al. [88] have proposed a different approach, where the concept of multi agent systems (MAS) is used to predict building energy consumption by denoting each meter as an agent. The common goal is to use a higher information extraction technique such as prediction, where it allows grid operators to forecast the grid demand to ensure sufficient electricity load. Test bed currently is the common method for testing out the smart grid use case and has been studied in [211]. Another common research topic is security and reliability of the smart grid system. Li et al. [49] have proposed a secure state estimation, which it can be used to address single sensor or multi-sensor scenarios. Similar works addressing smart grid security also can be found in [212, 213]. On the other hand, advanced metering infrastructure (AMI) has been studied along with the smart grid to ensure the electrical metering is tamper-proof while able to accurately measure energy consumption. For instance, work in [214] uses the clustering algorithm to identify energy theft accurately while reducing potential false positives. Meanwhile, work in [105] has presented a real-time price estimation by fusing local power and global power consumption to understand real-time electric load of the grid. In future, prosumers (producer and consumer) will emerge in the smart grid market and sole distributor paradigm will be no longer valid. This scenario greatly increase the difficulty of the energy demand and load when accounting the energy as a live market

The search for clean energy resources has been an ongoing effort for many researchers in order to cut down the dependency on the fossil fuels. Therefore, the clean energy research direction mostly focuses on renewable energy, which propose to go for a green and less carbon footprint energy producing approach. Nowadays, the most common renewable energy sources emerged in the market are solar farm [101, 141] and wind power [215, 216]. Solar energy is generated based on the conversion of the sunlight into electricity, but the energy harvesting technique suffers from limited energy harvesting time. Thus, solar irradiance prediction is crucial to ensure maximum energy throughput in the solar farm within the limited time. Huang et al. [141] have proposed to use data driven algorithms such as ABB, SVM, BRT, and Lasso, in which the information from neighboring solar plants are combined to accurately predict the solar irradiance. Meanwhile in [217], Jung and Broadwater have implemented a statistical model to fuse wind speed, direction, temperature from forecast station and online measurement to determine the total power output of the wind farm. Most of the aforementioned methods focus on improving efficiency of the existing energy harvesting methodology. Future research on the clean energy relies on various data and energy sources in order to construct a high efficiency energy harvesting model.

Smart facility denotes access of physical facility that provides services to the public such as parking facility, water supply, etc. The most vital facility in a smart city would be water treatment center as clean water is an important necessity for the urban citizens. Any potential leakage or downtime of water supply in a city would be proven troublesome. Mounce et al. [104] propose a water leakage detection using classification technique, which combines all the district water meter data. Similar concept can be applied on other resources such as gas pipe leakage detection or electricity theft in smart grid. In the public facility, the wear and tear of structures can be a major issue due to the frequent rate of public usage. Hence in [99], Park et al. have combined multi-metric sensors to estimate the bridge displacement. Through this, a rough estimation of the structural health can be determined. Alternately, Khoa et al. [218] have proposed a tensor decomposition approach using the facility data sources in order to understand the facility usage details. In addition, the emergence of data centers providing various functionalities to the smart city applications such as [219, 220, 221] also one of the focuses for the ongoing efforts of smart city. There is also a few domains that is highly correlated with smart facility such as Smart Maintenance and Governance [56], where integration of a web portal is used to report potential damages.

Communication in an urban city remains an essential infrastructure for various application platforms to communicate with each other. Not all communication platforms and standards are designed equally as each of them serve different purposes. Therefore, different standards and protocols to meet varying requirements have been established. Currently, the upcoming 5G technology [28, 222] has promised to bring integration of 5G interface with support for older generation spectrum such as LTE and Wi-Fi in order to provide seamless user experience. The common data source in 5G standards is raw signal, and that is the reason why data fusion only happens at the edge level. For example, Huang et al. [27] and Rappaport et al. [222] have fused raw signals that are divided through multiple antenna during transmission. The receivers will receive multiple signal sources and reconstruct the original information being transferred. Further discussion of the energy efficient trade-off in wireless communication technology can be found in [223, 224]. In the IoT domain, wireless sensor network (WSN) is considered a common communication platform because of its wide coverage and low power consumption. WSN is built on top of nodes’ network, which is smaller than a wireless ad hoc network. Hence, multiple nodes can be combined for encoding and decoding the packets received. Kreibich et al. [47] and Luo et al. [48] have proposed approaches to improve communication between WSN focusing on the communication mechanism between nodes. The main objective is to focus on the reliability of communication channel while maintaining the coverage (from relay to sink nodes) and also low power consumption. The research significance of communication is undoubtedly a necessity in smart city as it benefits all domains leveraging communication technology. The main goal is to design efficient and reliable communication protocols to meet different requirements of applications. Alternately, low power communication is yet another goal for IoT in order to achieve long sensing operation.

Sensing coverage is one of the important factors to determine the quality of data sources. Insufficient data coverage will generate a result that is not representative, and often it implies more sensors need to be installed to increase the sensing coverage. This indirectly affects the deployment cost since more physical hardware is required to compensate the sensing coverage. Apart from that, it also affects the design of communication architecture because more physical sensors are required to transmit data, and thus potentially congests the communication platform. These factors are common obstacles for a large-scale deployment in smart city applications and getting worse when increasing the deployment scale. There are two commonly used approaches to address the aforementioned issues, which are crowdsensing and mobile sensing platform.

As shown in [225], crowdsensing is one of the most cost-efficient method as personal mobile devices such as smartphones. Smartphones offer wide variety of sensors such as vibration, magnetic field, IMU, GPS, and others. The problems with crowdsensing are related to user privacy intrusion and high battery consumption when actively collecting data. User privacy is a challenge in collecting data as regulations in many countries have been facilitated to prevent applications to collect any sensitive information. This issue will be further discussed in user privacy and security sub-section. Another problems with crowdsensing are the unavailability of geolocations information or random distribution of geographical located data. These scenarios lead to inconsistent data quality. Potential way to resolve this limitation is to collect data at a fixed time and location only when needed, where incentive is provided for valid participants. Also, the trade-off problem of the mobile sensing can be further found in [226]. Through this method, only qualified data will be included as data sources, while invalid information will be automatically filtered.

Using similar concept as crowdsensing, mobile sensing has offered the same data sensing approach but only follows designated route to collect data. The idea is to leverage the mobility of the transportation (normally public transports, cabs, and garbage trucks) to conduct data collection, where the vehicles are traveling across the city. Example of mobile sensing platform can be found in [106], where garbage trucks on duty will collect the ambient data across different parts of the city weekly. An identical concept can also be implemented with the public transport systems, since majority of them follow fixed schedules. The challenge with the mobile sensing is that spatial resolution of the data may not have a finer detail when compared to crowdsensing due to fixed data collection schedule. The main cause is due to the limited accessibility of the vehicles in certain areas (pedestrian path and residential area). Potential workaround of this limitation would be combining the mobile sensing and crowdsensing data sources to generate data that covers large area within the urban city. Services integration also plays an important role in supplying platforms alternate data sources to perform data enrichment. By simulating the different IoT services in smart city as shown in [227], potential limitation or bottlenecks of smart services can be avoided in order to design a better smart city application.

Long term data collection offers different aspects of knowledge discovery as data is able to cover more detail in a larger temporal resolution. The advancement of miniaturization has greatly reduced the power consumption of the sensors and IoT devices while maintaining the same sensing performance. As a result, combining both energy harvesting techniques and low energy devices are able to create a long self-sustaining sensing approach. This breakthrough allows physical sensors to run independently without the need of external power sources.

In order to preserve the longevity of physical sensors’ sensing capability, energy harvesting is one of the common approach in large area networks. It allows sensors to draw energy from solar energy, vibration, or temperature difference. The most widely available energy harvesting technique is solar panels and it can be easily obtained. Solar panel is affected by the presence of solar irradiance, where the energy harvested varies throughout the different time of the day. Contrast to solar farm, the goal here is to conserve as much energy, while maintaining the sensing capability of the physical sensors. The most notable influence would be the energy management architecture as well as the battery capacity and the solar panel efficiency. Apart from that, although temperature difference and sensor vibration are capable of harvesting energy but it is limited to certain use case and not suitable for general usage.

Alternately, potential replacement of the traditional energy harvesting technique is wireless power transfer. As shown in [228, 229], this method offers power to be transferred wirelessly without battery and energy harvesting module. Currently, there are different types of wireless power transfer technologies such as inductive coupling, capacitive coupling, magnetodynamics coupling, microwaves, and light-waves. Each of them has their limitation such as inductive coupling only has limited range of transferring energy. That being said, this technology is still relatively new, and it requires further investigation in order to guarantee its minimum working efficiency for smart city applications.

Other than using external power sources, low power sensing for carrying out the sensing tasks. In order to drive different smart city applications, various standards have been proposed for LPWAN, such as LoRaWAN by LoRa Alliance and NB-IoT Release 13 by 3GPP. LoRaWAN focuses on the long range IoT connectivity for industrial applications while the NB-IoT focuses on the indoor coverage, low cost, long battery life, and stable communication in high density communication channel. The main reason to use low power sensing approach is due to the high compatibility with large scale deployment relying on the low bit rate communication channel usage. However, standardization of these protocols remains a challenge in LPWAN due to the possibly of using unlicensed spectrum, where organizations may choose not to follow the agreed spectrum. In future, low power communication will ensure the long term sensing capability of physical sensors in the smart city applications and therefore will improve data sources quality.

Collecting urban residents’ data in a smart city application can be challenging due the nature of sensitive data that can be misused if poorly managed. As privacy issue has been discussed extensively by the authors in [237, 238, 116], misuse of private information may lead to catastrophic events such as information theft, or identity fraud. Currently in Europe, General Data Protection Regulation (GDPR) as discussed in [239, 240] has been proposed to better address the data privacy concern of the Internet. In other countries, there are also similar efforts to enforce data privacy protection such as Canada’s Personal Information Protection and Electronic Documents Act (PIPEDA), China’s China Data Protection Regulations (CDPR), Singapore’s Personal Data Protection Act (PDPA), Japan’s Personal Information Protection Commission (PIPC), etc. Meanwhile in USA, Health Insurance Portability and Accountability Act of 1996 (HIPAA), the Children’s Online Privacy Protection Act of 1998 (COPPA), and the Fair and Accurate Credit Transactions Act of 2003 (FACTA) have been introduced to improve with the information flow efficiency across agencies. This is also a part of the efforts to prevent sensitive information being available for unauthorized parties.

Majority of the policies and regulations emphasize on the users’ consent for collecting personal data and this can be problematic as not all platforms provide ample security for data storage. With insufficient security measurements, the data collected may be compromised, which may lead to tainted reputation and loss of public faith. For instance, Facebook and Cambridge Analytica scandal [241] has shown potential misuse of user data collected. With that in mind, potential right of accessing data sources could be revoked if the data source is not handled properly by the right person. Hence, privacy and security should be the responsibility for both platforms and users. A thorough review has been conducted in [242], which works on the IoT requirements to address privacy issues. Potential solution for the aforementioned problem is to use hybrid data fusion technique in a smart city application. The idea here is to locally fuse the sensitive information (user identity, phone number, bank account number) into generic information, before uploading to the cloud for further processing. The benefits of such approach are two-folds, which are the ability to offload computational cost and to preserve sensitive information at the physical sensor only. In addition, we can leverage machine learning approaches such as [243, 244] to generate synthetic datasets with identical data characteristic for study purpose. This eliminates the chances of private data been leaked out and encourage the openness of datasets to be studied by different researchers and data scientists. To draw a clear line between generic and sensitive data remains a debate among researchers. In future, the data fusion can be applied at the lower level to remove any potential sensitive data.

According to Kitchin [245], there are two general security concerns in the smart city applications, which are security of technology/infrastructure (data center, services, and system architecture) and data security (data generation, storage, and communication).

The security of the technology and infrastructure highly relies on the design architecture of the system being deployed. Depending on the application requirements, it varies from traditional client server architecture to decentralized architecture. The main objective is to deploy a hack-proof/exploit-less system architecture. Alternately, there are also ways of improving security of system architecture such as incentive/bounty for reporting flaws, simulating injection attacks, security assessment from third party, etc. Nowadays, the security enhancement focuses towards continuous effort as the technology has been changing rapidly. For instance, security works in [246, 247, 248] have proposed different strategies to enhance the security of the smart city application’s architecture by focusing on the common security standards/practices/protocols. This shows that as the number of smart city applications increase rapidly, system architectures implemented with the security design in mind become apparent with good practices and standard architecture design. Subsequently, regular security assessment and auditing also pave way for a safer smart city applications deployment.

Meanwhile, data security also contributes to the significant part of smart city applications ecosystem from generation, storage, and communication. The common method to combat such issue is leveraging encryption techniques, where it encodes the data so that only the authorized parties have access to it. For instance in [249], Wang et al. have introduced an attribute based encryption scheme, which it allows fine-grained access control, scalable key management, and flexible data distribution. In addition, encryption also can be used in the communication platform between IoT devices in smart city application as shown in [250, 251] to prevent information hijacking.

Despite constant effort of cyber security researchers developing new security schemes, the numbers of data breaches and cyber threats increase every year according to David et al. [252]. The main culprit of such occurrence is due to negligence of data security practices/implementation. Security often appears to be an afterthought in deployment of a smart city application. Thus, in order to combat such threat, the smart city application should comply with security standards as shown in [253] to mitigate the chances of becoming a victim.

Lately, supervised machine learning techniques focus on the DNN, where the in-depth reviews of the recent development can be found in [255, 256, 257]. Major research efforts aim to increase the explainability of the model such as NN, CNN, and DNN rather than using them as black box models. To this end, explainable AI (XAI) [258] is the new motivation for data scientists to explore the interpretable learning paradigm of the modeling in order to provide a semantic meaning behind modeling logic. This new learning process has driven three big fields in the deep learning domains, which are (1) Deep Explanation, (2) Interpretable Model, and (3) Model Induction. To develop a deep explanation on the model interpretation, the cognitive layers will act as an intermediate layer between learning and explanation layer in order to cast the learned abstractions, policies, and clusters information into an explainable format. Subsequently, the interpretable model such as Bayesian learning [259] can be built to explain the uncertainties required when developing the deep learning models to learn the choices of a learning process. Alternate approach has proposed to use subspace approximation with an adjusted bias technique [260] to build interpretable CNN, which uses feed forward design to better explain the model’s choice in allocating certain hyper-parameters. Meanwhile, model induction refers to the technique used for inferring the model’s decision and learning progress. Through a thorough understanding of the model, parameters can be fine-tuned to increase the learning optimization rate in a long-term application deployment. Hence, the search of XAI is an important milestone for the data scientists, which can be used to explain the learning process and the decision machine learning made. An example of potential use case would be trying to understand the reason behind (also known as reasoning in some literatures) the predictive maintenance decision machine learning rather than performing maintenance due to the result of predictive algorithm.

In the smart city applications, collecting the ground truth could be proven challenging due to the uncertainties and errors in the collected data sources. Hence, obtaining labels or data annotation are another problems with certain data sources. Despite the rapid development of advanced modeling tools like DNN, it still requires labels and data annotation in order to achieve objectives of extracting higher information. There are a few approaches that address the lack of labels such as manual annotation, crowd labeling, software annotation, and pattern labeling. However, manual annotation only works well with a small dataset while other approaches do not guarantee the correctness of end result. This shows a big research gap to seek a better way to label data sources accurately.

Research works such as Zhou et al. [261, 262] have attempted to fix unlabeled data by transforming them into useful features to achieve certain objectives. Traditionally, raw data is required to be preprocessed into something meaningful, but it still suffers from the need of data cleansing and amputation. The simplest method would be to solely depend on the filtering technique. However, aggressive filtering may remove large amount of raw data resulting potential loss of knowledge. Another simple solution is to increase the number of reliable data sources to be fused to create potential annotation. Increasing data sources often indicates an increment of the overall deployment cost. Alternative solution to the increased deployment cost is to use transfer learning [263], where the knowledge from existing domain can be transferred to other domain to learn from it.

The emergence of the hybrid models has become common due to wide variety of data sources available. It allows different levels of data sources (high, low, or both) to combine in order to create potential insights in a particular domain. It also helps to solve the data privacy problem along with machine learning technique, which has opened up many opportunities for researchers and data scientist to study on these big data collected. One example of the hybrid model is shown as follows: an urban planning system has different data sources as input such as human comfort factor index (environmental ambient sensors), positive urban city factor (feedback data on urban area such as greenery, surrounding amenities, recreational parks, and others), and cyber data (social media input) to design a fully automated urban planning system by fulfilling predefined criteria. The result from the data fusion needs to be explainable as discussed in the previous XAI for understanding choices made by the automation software. In this example, different tiers of data sources are fused using data sources types ($D1$, $D2$) and the result is some features. Eventually, these features will be combined to generate a potential plan for city through computation modeling ($D3$, $D4$). By joining different data sources, simulation can be used concurrently to verify the performance of urban planning system before deploying to the city. In future, implementation of the hybrid model will become a general trend due to wide availability of the data sources and processing platforms. As mentioned in the discussion, data ontology is another key factor to allow data sources to be connected from different platforms to provide knowledge for the smart city applications.