State-of-the-Art Techniques for Real-Time Monitoring of Urban Flooding: A Review

Abstract

In the context of the increasing frequency of urban flooding disasters caused by extreme weather, the accurate and timely identification and monitoring of urban flood risks have become increasingly important. This article begins with a bibliometric analysis of the literature on urban flood monitoring and identification, revealing that since 2017, this area has become a global research hotspot. Subsequently, it presents a systematic review of current mainstream urban flood monitoring technologies, drawing from both traditional and emerging data sources, which are categorized into sensor-based monitoring (including contact and non-contact sensors) and big data-based monitoring (including social media data and surveillance camera data). By analyzing the advantages and disadvantages of each technology and their different research focuses, this paper points out that current research largely emphasizes more “intelligent” monitoring technologies. However, these technologies still have certain limitations, and traditional sensor monitoring techniques retain significant advantages in practical applications. Therefore, future flood risk monitoring should focus on integrating multiple data sources, fully leveraging the strengths of different data sources to achieve real-time and accurate monitoring of urban flooding.

1. Introduction

Global climate warming is frequently triggering extreme weather events. This phenomenon, combined with rapid urbanization, has led to a significant increase in surface runoff, thereby intensifying urban pluvial flooding events, which have caused substantial economic losses and casualties [1,2,3]. In 2021, the extreme storm event with an estimated return period of 1000 years in Zhengzhou, China, resulted in the most severe flooding disaster in recent Chinese history. This disaster is estimated to have caused direct economic losses exceeding 65.5 billion yuan (approximately 10 billion USD), with 292 reported fatalities and 47 missing individuals [4]. In 2022, extreme monsoon storms devastated Pakistan, affecting most of the country, causing numerous casualties, and displacing approximately 33 million people [5]. Furthermore, according to the socio-economic scenario, with a global temperature increase of 2 °C, the death toll caused by flooding is projected to increase by 50%, and direct economic losses are expected to double [6]. All these instances highlight that floods induced by extreme storms have become a severe issue that needs to be addressed and managed safely. In this context, there is an urgent need to strengthen the management of urban floodings, with the primary task being to enhance the timeliness and accuracy of flood risk warning.

Currently, flood warning systems primarily rely on two approaches: model-based forecasting and real-time monitoring. In the research community, considerable attention has been devoted to model-based forecasting [7], including physical models (such as hydrological and hydraulic models) [8,9] and black-box/data-driven models (such as machine learning models) [10]. These efforts have led to notable advancements in the accuracy and efficiency of simulations through the refinement of physical models and innovations in machine learning algorithms. Despite extensive model optimizations, significant uncertainty remains in model predictions. In urban areas with complex terrain, for instance, extreme rainfall events can alter runoff paths, leading to substantial deviations from simulated results [11]. Additionally, inaccuracies in rainfall forecasts are a critical source of model uncertainty [12]. Furthermore, both physical and black-box models require rigorous calibration and validation with empirical data before they can be reliably applied [13]. Consequently, real-time monitoring of urban flooding plays an essential role in enhancing the accuracy and effectiveness of flood warning systems.

Monitoring technology has made significant advancements over the past few years, and the sources of data have become increasingly diverse. Two main trends can be observed in terms of data sources: First, the development of new sensors and measurement devices with controlled costs. Second, monitoring methods are becoming more diversified and intelligent [14]. Traditional urban flood monitoring relies on the deployment of sensors to precisely and intuitively monitor key parameters such as water depth and flow rate. With major advancements in new sensors, wireless communication, and data platforms [15], there has been an improvement in monitoring accuracy while also controlling costs. On the other hand, with the continuous development of information technology, the concept of “smart cities” has emerged [16], and big data has become a key technology supporting the “intelligentization” of various urban sectors [14,17]. Therefore, researchers and relevant departments have begun to focus on various new sources of flood data. For instance, images obtained from social media and surveillance cameras not only provide critical information related to urban flooding but also incorporate spatiotemporal characteristics [14,18]. The analysis of such flood-related imagery necessitates the use of computer vision algorithms [19,20]. This interdisciplinary approach, while challenging for researchers, has led to significant advancements in the field. The continuous breakthroughs and applications of these technologies help enhance the identification and monitoring capabilities of urban flooding, strengthening the assessment of flood risks and disaster response capabilities. However, existing reviews on monitoring technologies tend to focus on individual categories of technology [21]. For example, Van, et al. [22] focused on IoT technologies in flood monitoring, and Wu, et al. [23] reviewed the non-contact monitoring technologies based on radar and computer vision. Given the current landscape, where traditional technologies maintain a stronghold in the market and emerging technologies are rapidly evolving, there is an urgent need for a comprehensive review of various monitoring data sources. Such a review is currently lacking.

This paper aims to summarize and analyze existing urban flood monitoring technologies from the perspective of different data sources, providing city managers and researchers with a more comprehensive technical reference. The goal is to better address urban flood challenges, enhance the city’s disaster resilience, and ensure the safety of urban residents’ lives and property. Figure 1 presents a flowchart of this review process. The current review is arranged as follows: Section 2 provides a review of literature on urban flood monitoring and identification technologies from a bibliometric perspective. Section 3 discusses a systematic review of flood monitoring technologies based on different data sources, including sensor data, big data, and multi-source data fusion. An outlook for future research and recommendations for relevant stakeholders are given in Section 4. Finally, conclusions are drawn in Section 5.

2. Bibliometric Review

2.1. Methods

Bibliometric analysis, through graphical visualization, can swiftly elucidate the developmental trajectory and emerging hotspots of a scholarly field, ultimately providing a framework for future work. This article employs CiteSpace version 6.3.R1 software for the visualization analysis of the retrieved literature data. CiteSpace is a widely used multifunctional citation visualization analysis software developed by the team of Prof. Chaomei Chen from Drexel University [24], which is designed to assist researchers in understanding the evolution of academic fields, the development of themes, and the formation of academic collaboration networks.

This paper selects the Web of Science Core Collection as the database for bibliometric analysis. The search terms are divided into three strings: “TS = (‘urban flood* OR waterlogging* OR inundation*’) AND TS = (‘water depth* OR flow rate* OR risk*’) AND TS = (‘monitor* OR sensor* OR detection*’)”. Irrelevant and non-English publications were filtered out from the search results, and ultimately, articles and reviews from the years 2000 to 2024, totaling 834, were analyzed.

2.2. Descriptive Analysis

This text provides a bibliometric analysis of literature on urban flood identification and monitoring from 2000 to 2024, using the Web of Science Core Collection. By examining the annual publication volume and citation frequency (Figure 1), this section aims to comprehensively understand the developmental trajectory of research in this field.

The bibliometric analysis reveals that between 2000 and 2024, a total of 834 research articles related to “urban flood identification and monitoring” were published, with an average citation rate of 24.53 per article. As shown in Figure 2, there was limited research in the field of urban flood identification and monitoring at the beginning of the 21st century. A significant increase in both the number of publications and citation frequency began in 2017, which may be attributed partly to the frequent urban flooding caused by extreme weather, leading to heightened attention on flood management by relevant authorities. Additionally, advancements in technology may have encouraged researchers to integrate traditional disciplines with emerging technologies, sparking numerous innovations. The academic interest in this area has continuously increased, with the year 2022 marking the peak in publication volume with 147 articles, which were cited 3704 times.

2.3. Co-Citation Analysis

Citation analysis maps the significance or popularity of publications. Due to the limitations in the number of nodes and considering the lower volume of publications prior to 2017, articles from 2017 to 2024 were selected for citation analysis. This analysis produced 281 cited articles, each represented by a node, with node size increasing with the citation count of the corresponding article. Figure 3 displays articles that were cited at least 16 times. Teng, et al. [7] lead with 34 citations, followed by Chini, et al. [25] with 21 citations, Gorelick, et al. [26] with 16 citations, and Pekel, et al. [27] also with 16 citations. It is evident that articles related to remote sensing and image recognition are frequently cited, highlighting the extensive research focus on these areas.

Figure 3. Network map of literature co-citation [7,25,26,27].

Figure 3. Network map of literature co-citation [7,25,26,27].

2.4. Country Analysis

Figure 4 illustrates the publication output from 2000 to 2024 for six countries, each with at least 50 publications. The United States leads with a total of 190 articles, followed by China, India, the United Kingdom, Australia, and Italy, with 181, 72, 67, 61, and 57 articles, respectively. The lines in the graph represent collaborations between countries, with thicker lines indicating a higher number of collaborative articles. It is evident that there is cooperation between these six countries, with the most collaborations occurring between the United Kingdom and Italy.

3. Systematic Review

Urban flood monitoring can be categorized into macroscopic and microscopic approaches. Macroscopic monitoring typically employs satellite remote sensing technology, which allows for the analysis and understanding of the extent and risk levels of urban flooding across large areas. However, due to limitations such as obstruction by buildings, this method is not suitable for precise water depth monitoring at the scale of urban streets [28,29]. Microscopic monitoring methods vary based on their data sources and can be divided into sensor-based and big data-based monitoring. Given the focus of this paper on precise monitoring of urban water levels, the following discussion will concentrate on microscopic monitoring approaches, beginning with a review of traditional sensor data and moving to the emerging field of big data analysis (Figure 5).

3.1. Sensor-Based Flood Stage Monitoring

Sensor measurement has long been the mainstream method for flood monitoring. It can monitor water level changes in real-time, providing intuitive and high-precision data that aids in the timely detection and warning of flood conditions. However, using sensors for urban flood monitoring presents several challenges, such as high costs and maintenance difficulties associated with deploying sensors over large areas. This section reviews sensors by categorizing them into contact and non-contact types based on whether they directly interact with the monitored object.

3.1.1. Contact Sensors

Contact sensors mainly include pressure sensors and electronic water gauges (Table 1). Pressure sensors calculate water column height through hydrostatic pressure, and are typically used for data collection in river water levels and drainage network operations [30]. However, static or turbulent flows can affect the measurement results, causing errors, and the pressure holes are prone to blockage, requiring regular maintenance, which is cumbersome. The hydrostatic pressure calculation process is shown in Equation (1). Electronic water gauges, such as Figure 6a, estimate water depth by utilizing the slight conductivity of water. As shown in Equation (2), when the water level changes, the length of the electrodes immersed in the water (L) also changes, thereby altering the resistance. Therefore, the working principle of the electronic water gauge is to infer the water level height by measuring the changes in current within the water [31,32]. These gauges are commonly used for monitoring water accumulation in residential areas and underpasses. Although electronic water gauges involve contact measurement, they can effectively cope with freezing, corrosion, and heat by using special material housings [33]. However, the susceptibility to displacement or damage by large debris or suspended particles in the water remains an unavoidable drawback of all contact sensors [34].

$$P=P_0+\rho gh$$

(1)

where $P$ is the total pressure (Pa) $P_0$ is the atmospheric pressure (Pa), $\rho$ is the fluid density (kg/m³), $g$ is the gravity acceleration (m/s²), and $h$ is the depth (m).

$$R={\displaystyle \frac{\rho ·L}{A}}$$

(2)

where $R$ is the measured resistance (Ω), $\rho$ is the resistivity of water (Ω·m), $L$ is the distance between electrodes (m), and $A$ is the cross-sectional area of the electrodes (m²).

Figure 6. The installation of different sensors; (a) electronic water gauges; (b) ultrasonic sensors [35].

Figure 6. The installation of different sensors; (a) electronic water gauges; (b) ultrasonic sensors [35].

Table 1. Main types of contact sensors.

Name	Operating Principle	Advantages	Disadvantages	References
Pressure sensor	Stevin’s law for hydrostatics pressure	Simple structure and low manufacturing costs	Susceptible to the conditions of installation and prone to damage	[36,37,38]
Electronic water gauge	Based on the Micro-conductivity of water	Equipped with a special material shell to effectively cope with special conditions such as freezing, rot, and heat	Easily damaged by external human intervention	[33,39,40]

Table 1. Main types of contact sensors.

Name	Operating Principle	Advantages	Disadvantages	References
Pressure sensor	Stevin’s law for hydrostatics pressure	Simple structure and low manufacturing costs	Susceptible to the conditions of installation and prone to damage	[36,37,38]
Electronic water gauge	Based on the Micro-conductivity of water	Equipped with a special material shell to effectively cope with special conditions such as freezing, rot, and heat	Easily damaged by external human intervention	[33,39,40]

3.1.2. Non-Contact Sensors

Non-contact sensing technologies primarily include ultrasonic sensors and radar sensors (Table 2) [41].

(1)

Ultrasonic sensors

Ultrasonic sensors, such as Figure 6b, use high-frequency sound wave transducers to transmit waves through the air to the water surface, determining the water depth by calculating the travel time of the sound waves. These sensors offer several advantages, including small size, low cost, non-contact operation, durability in harsh environments, long service life, and ease of operation. However, environmental factors such as temperature, humidity, and noise can introduce uncertainties in distance measurements using ultrasonic sensors [34,42]. Ultrasonic sensors have a wide range of applications, such as in rivers and reservoirs [43,44], streets [45], and outdoor low-lying areas [46]. For river and reservoir monitoring, Ranieri, et al. [34] analyzed the performance of ultrasonic and pressure sensors in river water level monitoring. They integrated two interfaces in a flood warning system corresponding to each sensor type, and the comparative analysis showed that ultrasonic sensors had smaller errors. In studies of street flooding, Silverman, et al. [45] developed a low-cost FloodNet sensor to monitor road water depth and duration, using an ultrasonic rangefinder to collect minute-by-minute distance readings between the sensor and the monitored surface. Additionally, Ni-Bin and Da-Hai [46] used the WL700 ultrasonic sensor to monitor water levels in outdoor low-lying areas, with a monitoring range of 0.2 to 6 feet. However, when the water level is too close to or submerges the transducer, it falls within a measurement blind spot, making data collection impossible. Therefore, in engineering practice, ultrasonic sensors are often combined with pressure sensors. Ultrasonic sensors are placed near manhole covers to expand the monitoring range, while pressure sensors are installed below the detection blind spots of ultrasonic sensors, away from the water surface, to reduce the likelihood of submersion and thus lower maintenance costs.

(2)

Radar sensors

The basic working principle of radar sensors is similar to that of ultrasonic sensors, both estimating water depth by measuring the range of reflected waves, as shown in Equation (3). However, they use different types of waves, with ultrasonic sensors using sound waves and radar sensors using electromagnetic waves [23]. Since radar sensors are non-imaging sensors, their accuracy is less affected by meteorological conditions that impair visibility, such as sunlight, fog, and nighttime. Compared to acoustic sensors, the propagation and attenuation of microwaves are less influenced by air temperature and humidity, resulting in a longer effective range and stronger anti-interference capability for microwave sensors [47]. Furthermore, the radar sensors can be classified into frequency modulated continuous wave (FMCW) radar and pulse radar based on the waveform of the transmitted signal. The FMCW radar sensors are more commonly used recently, and the normal pattern of mmWave FMCW is shown in Figure 2. Measuring data from radar sensors must first be processed via the Internet of Things (IoT) to calculate water depth. And then the depth results are sent to the cloud. To provide accurate depth measurements in a shorter time, innovations in local radar processing algorithms have been initiated [48,49,50]. These innovations aim to achieve high depth measurement accuracy while maintaining low computational complexity and power consumption. For example, Shui, et al. [51] proposed a cross-correlation-based algorithm called SFCC, tailored for the short-range high-resolution monitoring required by urban flood monitoring systems. This algorithm achieves a distance detection resolution of less than 5 mm while reducing computation time.

$$R={\displaystyle \frac{ct}{2}}$$

(3)

where $R$ is the distance between the sensor and the target (m), $c$ is the speed of waves (3 × 10⁸ m/s for electromagnetic waves and 340 m/s for ultrasonic waves), and t is the total time for the waves taking the round trip to the target (s).

(3)

Unconventional sensors—mobile sensing

In addition to traditional sensors, interdisciplinary research has explored using smartphones as sensors. By capturing gait characteristics of pedestrians at different flood depths, researchers can classify flood risk levels [52]. This method helps to avoid the installation and maintenance costs associated with additional equipment. Moreover, the ubiquitous nature of smartphones allows for widespread coverage across urban areas, aiding in the construction of flood risk maps and the timely execution of emergency response efforts.

Table 2. Main types of non-contact sensors.

Name	Operating Principle	Advantages	Disadvantages	References
Ultrasonic sensor	Calculated from the time between ultrasonic wave transmission and reception	No need to be submerged in water, thus causing fewer maintenance issues	Impacted by rain, snow, fog, dust, temperature, and humidity	[53,54]
Radar sensor	Similar to ultrasonic sensor, based on the principle of time travel	Large range, high precision, easy installation, almost not subject to external temperature, humidity, and other environmental conditions	Impacted by the media pressure, density, temperature, and other factors	[51]
Mobile sensing	Based on gait characteristics of pedestrians at different flood depths	No need to install or maintain additional equipment	Impacted by the surrounding environment and privacy concerns	[52]

Table 2. Main types of non-contact sensors.

Name	Operating Principle	Advantages	Disadvantages	References
Ultrasonic sensor	Calculated from the time between ultrasonic wave transmission and reception	No need to be submerged in water, thus causing fewer maintenance issues	Impacted by rain, snow, fog, dust, temperature, and humidity	[53,54]
Radar sensor	Similar to ultrasonic sensor, based on the principle of time travel	Large range, high precision, easy installation, almost not subject to external temperature, humidity, and other environmental conditions	Impacted by the media pressure, density, temperature, and other factors	[51]
Mobile sensing	Based on gait characteristics of pedestrians at different flood depths	No need to install or maintain additional equipment	Impacted by the surrounding environment and privacy concerns	[52]

3.2. Big Data-Based Flood Stage Monitoring

Due to potential blind spots in sensor deployment and the susceptibility of equipment to malfunctions, timely analysis of natural disasters such as flooding often suffers from limited data [55]. Additionally, with the rapid development of emerging technologies such as machine learning, image recognition, and semantic analysis, methods for extracting waterlogging features from social media and surveillance camera data for water level estimation have become increasingly sophisticated. This section provides a review of flood monitoring methods based on different data sources, specifically focusing on surveillance cameras and social media.

3.2.1. Surveillance Camera Data Based

More than a decade ago, video surveillance was already being used for river monitoring and flood management. However, at that time, water level readings from video images were manually obtained [56]. Although this method allowed for continuous monitoring compared to on-site observations, manually reading water levels remained a time-consuming and labor-intensive process. It was not only inefficient but also subjectively affected the accuracy of the observations, especially when image quality was low, such as when there was distortion or blurriness. In recent years, the number of traffic and security monitoring devices in cities has been steadily increasing. For example, Beijing has approximately 1.15 million cameras, with an average density of 71 cameras per square kilometer [57]. These ubiquitous surveillance devices can record the entire process of urban flooding. At the same time, computer vision, an important field of artificial intelligence, is continually advancing in its ability to process images at the pixel level and extract useful information. Its role in hydrological research is becoming increasingly significant [58]. Consequently, some studies have started to use computer vision algorithms to extract flood information from video surveillance data.

(1)

Computer vision methods for surveillance image analysis

Surveillance videos capture the process of urban flooding in the form of images, which can be defined as a computer vision task [57]. Computer vision tasks are generally categorized into five types: image classification, object detection, object tracking, semantic segmentation, and instance segmentation [59], providing technical support for extracting waterlogging information from images. Commonly used algorithms currently include convolutional neural networks (CNN), you only look once (YOLO), and transfer learning. CNNs are designed to automatically extract local features from images through convolutional layers, reduce the dimensionality of feature maps via pooling layers, and ultimately perform classification or other tasks using fully connected layers. YOLO, a real-time object detection algorithm, processes the entire image in a single forward pass (“You Only Look Once”) to predict the positions and categories of multiple objects simultaneously. Due to its high detection speed, YOLO is frequently employed in real-time water level monitoring. Transfer learning involves applying a pre-trained deep learning model to a related task, making it particularly suitable for scenarios with limited data. By fine-tuning the pre-trained model, it can rapidly adapt to the requirements of new tasks.

Currently, applications that extract flood information from surveillance videos mainly focus on two aspects: extracting waterlogged areas and estimating water levels. The latter can be approached using two primary methods [60]. The first method relies on image features and image processing algorithms to identify water levels. For example, Yu and Hahn [61] proposed a water level measurement scheme based on differential images. This scheme uses time-domain sparsely sampled images and an invariant feature index in the image to identify water levels. Sakaino [62] developed a two-step histogram method that estimates water levels by supervising the regions of interest (ROI) in consecutive frames. Vanden Boomen, et al. [63] used convolutional neural networks (CNN) to predict river water levels from river images captured by cameras. However, this method requires a separate training model for each monitoring scene and is easily affected by external dynamic factors. The second method for estimating water levels involves detecting reference objects in the images to indirectly identify water levels. The accuracy of this approach is often influenced by the type of reference objects used, making the selection of reference objects a popular research direction. Zhong, et al. [64] used the YOLOv4 object recognition model to identify pedestrians’ legs and vehicle exhaust pipes in images. After extensive data testing, they found that using vehicles as reference objects provided higher accuracy than using pedestrians, possibly because pedestrians are more flexible and less objective as reference points. Additionally, Jiang, et al. [65] employed common urban objects such as sidewalk railings, mailboxes, and traffic barrels as reference points to estimate flood depth by detecting height differences before and after flooding. Their application in a city in Hebei Province, China, demonstrated that this method offers broader applicability and enhanced accuracy, and flexibility compared to traditional object detection techniques. In addition to these reference objects, Liu, et al. [29] proposed a method combining surveillance cameras with water gauges to measure urban flood depths. They installed floating rulers in the test area and used surveillance cameras and the YOLOv5s object detection model to obtain the pixel positions of the floating rulers. Then, they used the binocular method and sub-pixel method to derive the transfer function from pixel positions to water depth, obtaining the urban flood depth. In recent years, transfer learning has gained significant attention due to its effectiveness in automatically learning image features. Jiang, et al. [57] applied a transfer learning model to extract feature vectors from urban flood video images and trained a lasso regression model using these vectors to estimate flood depth. The method’s efficacy was validated through case studies in two Chinese cities, Henan and Guizhou. This approach can be implemented across the extensive network of urban cameras, forming a robust flood monitoring system. Transfer learning features typically contain more complex information than traditional computer vision features, such as histograms of oriented gradients (HOG) and color histograms, making them effective for classification, regression, and clustering tasks [66]. Table 3 provides more examples of applications utilizing surveillance footage for flood monitoring.

(2)

Limitations

Monitoring water levels using surveillance video is an emerging research field that leverages existing surveillance infrastructure and advanced image processing technologies to enhance the efficiency of flood monitoring and management. However, this approach has certain limitations. The position and angle of surveillance cameras can restrict the field of view, affecting the accuracy and comprehensiveness of water level measurements. Factors such as image clarity, lighting conditions, and weather can also impact the accuracy of image analysis. At the same time, some areas may not have full coverage of surveillance cameras, so that comprehensive monitoring data cannot be obtained. Additionally, the data collected by surveillance cameras may involve privacy and data security concerns [67]. To address these issues, future research can focus on developing more efficient image processing algorithms and machine learning models to improve the accuracy and real-time capabilities of flood water level measurements. Integrating surveillance video technology with other flood monitoring methods, such as satellite remote sensing and hydrological models, can enhance system robustness and coverage. Furthermore, advancing data protection technologies is essential to ensure the protection of personal privacy while conducting flood monitoring.

Table 3. Applications of using surveillance cameras for flood monitoring.

Application Area	Algorithm	Application Effectiveness		References
		Advantages	Disadvantages
A city in Hebei Province, China	CNN	Low economic cost, acceptable accuracy, high spatiotemporal resolution, and wide coverage	Large relative error for shallow waterlogging	[65]
A city in Guizhou Province, China	Transfer learning	low economic cost, good real-time performance, and satisfactory accuracy	Accuracy affected by camera lens distortions, resolution, and field of view extent	[57]
2 rivers in São Paulo, Brazil	DeepLabv3 (A deep learning model based on CNN)	Robust to changes in camera viewpoints and illumination, and easy to deploy in many urban rivers	Accuracy affected by strong wind	[68]
Dongying City, Shandong Province, China	R-CNN and OpenCV	Low economic cost, high temporal and spatial resolution	Accuracy affected by the quality of the video image	[69]

Table 3. Applications of using surveillance cameras for flood monitoring.

Application Area	Algorithm	Application Effectiveness		References
		Advantages	Disadvantages
A city in Hebei Province, China	CNN	Low economic cost, acceptable accuracy, high spatiotemporal resolution, and wide coverage	Large relative error for shallow waterlogging	[65]
A city in Guizhou Province, China	Transfer learning	low economic cost, good real-time performance, and satisfactory accuracy	Accuracy affected by camera lens distortions, resolution, and field of view extent	[57]
2 rivers in São Paulo, Brazil	DeepLabv3 (A deep learning model based on CNN)	Robust to changes in camera viewpoints and illumination, and easy to deploy in many urban rivers	Accuracy affected by strong wind	[68]
Dongying City, Shandong Province, China	R-CNN and OpenCV	Low economic cost, high temporal and spatial resolution	Accuracy affected by the quality of the video image	[69]

3.2.2. Social Media Data Based

Social media platforms (such as Twitter, Facebook, and Sina Weibo) have garnered increasing attention in the field of disaster emergency management due to their rich information, near-real-time transmission pathways, and low-cost data generation [70]. For instance, during the 2007 Southern California wildfires, people turned to social media to assess the disaster event, gather real-time information, and disseminate it widely because they lacked sufficient and timely information about the specific affected areas from official sources [71]. Witnesses, acting as social “sensors”, upload real-time on-site reports (text and images) to social media, providing a feasible way to monitor detailed flood locations and water depth information [72], thereby enhancing situational awareness during disasters.

(1)

Information retrieval

Information from social media is primarily collected in bulk using web crawling technology. Web crawlers automatically browse the internet based on predefined rules, overcoming the traditional inefficiencies, high costs, and cumbersome processes associated with manual data collection and organization [73]. Given the large volume, rapid update speed, and diverse structure of social media content, processing this “big data”—specifically, extracting useful information from a vast amount of irrelevant social media data—poses a significant challenge [74]. For example, Yan, et al. [70] employed TF-IDF (term frequency-inverse document frequency) keyword extraction technology to improve dataset quality. They identified frequently occurring keywords in irrelevant articles to extract negative keywords, which were then used to filter out non-flood-related posts from the crawled data.

(2)

Image recognition

Images are the most intuitive form of information dissemination on social media. The method of inferring waterlogging from these images is similar to monitoring urban flooding using surveillance video, as both fall under computer vision tasks. However, due to the differing nature of these data sources, the research focus varies. Social media data is crowdsourced, and the high variability of such data makes extracting flood information more challenging [75]. Therefore, research in this area emphasizes image preprocessing. Convolutional neural networks (CNNs) have become a research hotspot in the field of computer vision due to their excellent feature extraction and classification capabilities. They have achieved significant results in extracting and interpreting vast amounts of social media image information and have become one of the mainstream methods in this field. For instance, Witherow, et al. [76] highlighted the differences in resolution, lighting, and environmental conditions in social media images and the complex dynamics of these environments. Consequently, they first performed a series of preprocessing operations on the images, including edge detection, image padding, and contrast adjustment, before using R-CNN to extract waterlogging information. Pereira, et al. [77] employed deep learning techniques for image classification. They used convolutional neural networks (specifically the EndseNet and EfficientNet models) to classify crowdsourced images into three categories: “no flood”, “water level below 1 m”, and “water level above 1 m”. Li, et al. [78] retrieved images related to urban flooding from social media and filtered out those without people. They then used a YOLO-based object detection model to detect human body parts from the bottom up and estimated the water depth based on the detected body parts relative to the average standard length of those body parts.

(3)

Text processing

Text, as a primary medium of information dissemination, is characterized by its rapid spread, easy retrieval, and rich content [79]. Currently, artificial intelligence methods are typically employed to classify and summarize data. Text classification, a classic problem in natural language processing, involves mapping information-bearing texts to one or more predetermined categories. The algorithms that accomplish this are known as classifiers. For example, Kankanamge, et al. [80] used a decision tree model to incorporate spatial dimensions into sentiment analysis to identify areas severely affected by floods. Bai, et al. [81] applied an improved term frequency-inverse document frequency (MTF-IDF) algorithm to efficiently classify disaster information from social media data, constructing a disaster emergency processing model (SEPM) that includes earthquake emergency information collection, vocabulary construction, and information classification, providing timely feedback for disaster relief deployment. With the development of deep learning models, text classifiers based on deep learning have shown excellent performance compared to the aforementioned shallow learning models, enabling faster and more accurate acquisition of flood information [82,83]. Traditional flood databases rely on hydrological stations, remote sensing technology, and other means (such as media reports and government reports). However, many flood events go unreported, and data in traditional databases often require long periods of manual collection and processing, resulting in delays in obtaining and disseminating information. To address this, de Bruijn, et al. [82] proposed a global flood event historical and real-time database based on social media data, aiming to detect flood events worldwide through social media information. This study used a place name parsing algorithm (TAGGS) to extract geographic location information from tweets, including countries, administrative regions, and places. Additionally, it employed a multilingual text classification algorithm based on the BERT model to filter out non-flood-related tweets. The study also analyzed the sudden increase in tweets related to specific areas to identify flood events. The proposed method can handle and analyze tweets in multiple languages, enhancing the feasibility of global flood detection. By automatically extracting and verifying flood event information from large volumes of social media data through place name parsing and machine learning models, the model’s precision and recall rates were improved. Thus, this method can monitor global flood events in real time, addressing the flood monitoring gaps in data-scarce regions to some extent. Research on using text information from social media for flood monitoring is currently limited, but many studies have used it for disaster risk analysis. For instance, Liu, et al. [84] used news reports from a Chinese website from 2008 to 2017 as key data, employing text mining, descriptive statistics, and association rule mining to extract information on the type, time, and location of natural disaster events and analyze their spatiotemporal distribution. However, social media information sometimes contains unverified rumors, which can affect the effectiveness of rescue operations and the rational allocation of resources. To address this, Mondal, et al. [85] developed a model capable of timely detecting and assessing rumors on social media following disasters, providing new perspectives and tools for information quality control in disaster response and management.

(4)

Integration of image and text information

In recent years, advancements in computer hardware have propelled the application of deep learning, making it possible to combine text and images to identify flood disasters. Simultaneous extraction and analysis of both data types allow for cross-validation and improved accuracy [86,87,88]. For example, Fan, et al. [89] proposed an integrated framework for detecting infrastructure disruptions based on three types of information embedded in social media content: images, text, and geographic maps. They applied this framework to analyze the release of two reservoirs during Hurricane Harvey in Houston, USA. By extracting and analyzing text, images, and geographic information from Twitter, they were able to spatially and temporally map key situational data and assess infrastructure damage in the flooded areas. Beyond broad risk assessment, the textual and visual data from social media also offer fine-grained flood information for urban monitoring. Yan, et al. [70] conducted a case study in Anhui Province, China, an area characterized by a complex river network, to extract flood-prone locations and water level information using social media data from Sina Weibo. This study aimed to elucidate the dynamics of urban flooding events. The researchers developed two deep learning modules designed to capture text and image data. The text analysis module synthesizes and extracts various water depth descriptions, while the image analysis module employs a water level layering technique. The decision layer then integrates these data streams to analyze the spatiotemporal distribution and patterns of the extracted information, thereby enhancing situational awareness. The study assessed urban flood risk during a three-day heavy rainfall event, demonstrating the capability of this approach to provide real-time inundation information. The location data were detailed to the level of neighborhoods, roads, and intersections, while the water depth information offered precise descriptions of the flood extent in affected areas.

(5)

Limitations

While existing research has made some progress in utilizing social media for flood monitoring, this approach still has certain limitations. For instance, the distribution of social media users (such as Twitter) is geographically biased, with higher user densities in regions like North America and Europe compared to certain areas in Africa and Asia. This disparity can affect the detection frequency and accuracy of flood events, particularly in regions where social media usage is low. Therefore, it is essential to choose social media platforms according to local conditions to ensure sufficient and usable data. Additionally, the accuracy of social media data depends on users’ subjective descriptions, which may contain errors, exaggerations, or information unrelated to floods. Obtaining accurate and continuous flood water level data for specific locations remains a challenge [64]. Furthermore, most existing studies have focused on large-scale urban floods, such as riverine floods or flash floods, with relatively few studies on waterlogging at the street block scale within cities. In summary, using social media data for flood monitoring has great potential but also presents certain limitations. Future research needs to address these issues and conduct further studies based on practical needs.

3.3. Multi-Source Integration for Flood Stage Monitoring

As the diversity of data sources for flood monitoring increases, the limitations of using a single data source to collect flood information are becoming more apparent [90]. Researchers are beginning to explore the potential of integrating multiple data sources for monitoring, particularly in time-sensitive major disaster situations where using multiple data sources can enhance the timeliness and accuracy of emergency management. Additionally, scientifically and effectively integrating various types of data to study urban storm and flood disasters and achieving complementary advantages in temporal and spatial scales are pressing theoretical and technical issues that need to be addressed [91].

(1)

Social media and remote sensing

Remote sensing imagery has been one of the primary data sources for extracting flood extents in early flood disaster research. It offers high spatial resolution, wide coverage, accurate data, strong data integration capabilities, and cost-effectiveness. However, due to its broad coverage, remote sensing monitoring cannot provide precise local waterlogging information. Additionally, its accuracy and timeliness are affected by urbanization and weather changes [92]. On the other hand, social media data, as an emerging source for flood monitoring, contains rich waterlogging information and offers significant advantages such as timeliness, broad temporal scale, rich information, and easy access. However, compared to data provided by professional monitoring platforms, information on social media is more scattered and incomplete, significantly impacting its application in urban flood monitoring [93]. Therefore, integrating social media data with remote sensing data on the feature dimension can fully exploit the advantages of multi-source data, compensating for the limitations of single data sources, and achieving efficient and accurate acquisition of waterlogging information [86,94], thereby enabling accurate real-time monitoring of urban flooding. For example, Fohringer, et al. [95] supplemented remote sensing data with waterlogging information extracted from social media images for real-time monitoring of the extent and depth of water. Jongman, et al. [96] achieved near-real-time monitoring of flood location, timing, causes, and impacts through a comprehensive analysis of satellite and Twitter data. Huang, et al. [97] combined social media data with remote sensing imagery, overcoming the limitations of traditional flood inundation probability calculation methods. However, integrating different data sources often relies on the subjective determination of model variable weights by users. To address this issue, Rosser, et al. [98] developed a Bayesian statistical model to quantitatively assess the contribution of each data source and used evidence weighting analysis to estimate the probability of flood inundation. Duan, et al. [99] utilized neural networks and other techniques to integrate and extract text and remote sensing data, constructing objective weight models for weight calculation, thereby avoiding subjective bias.

(2)

Social media and hydrological modeling

In addition to being integrated with remote sensing data, social media data can also be combined with hydrological data. For instance, Restrepo-Estrada, et al. [100] used social media data to calibrate urban flood models, improving the accuracy of flood simulations and enhancing the performance of flood warning systems. Similarly, Chen and Wang [101] used flood information extracted from social media text in combination with emergency management cloud data to establish a more precise warning mechanism in the context of floods in Taiwan. Shoyama, et al. [102], using the 2019 typhoon in eastern Japan as an example, conducted a comparative analysis of the temporal changes in the number of flood-related tweets and flood monitoring data. They found that the surge in tweet numbers was closely related to the occurrence of disaster events. Based on the correlation characteristics between tweets, rainfall, and water level data, they improved the flood disaster warning mechanism.

4. Future Perspectives and Recommendations

As discussed in Section 3.3, the integration of multiple data sources for urban flood monitoring is becoming an increasingly important trend as research advances. However, current studies remain somewhat limited in their application of integrated data sources. Therefore, future research should explore alternative forms of data integration to further enhance flood monitoring capabilities. Additionally, it is crucial for future studies to address the challenges associated with data interpretation, particularly those arising from the incompatibility of different data sources and the uncertainties introduced by varying data origins.

Furthermore, given the gap between current research and practical application, we recommend that urban planners, policymakers, and relevant stakeholders place greater emphasis on emerging monitoring technologies. Strengthening collaboration between these stakeholders and researchers will be essential to facilitate the practical implementation and translation of research findings into real-world applications. Furthermore, addressing flood issues in border regions requires the timely acquisition of upstream water level and rainfall data, which is of paramount importance. To this end, establishing interconnectivity with the monitoring networks of neighboring countries or adjacent regions is essential to ensure the prompt transmission and sharing of critical information. Such cross-border cooperation can significantly enhance flood defense and response capabilities in border areas, thereby mitigating the risks and potential damages associated with floods [103].

5. Conclusions

In the context of increasingly frequent urban flooding, timely and accurate flood monitoring is crucial for addressing flood risks and mitigating the impact of disasters. This paper first uses the Web of Science Core Collection as the database and Citespace as the analysis software to conduct a bibliometric analysis of 834 articles related to urban flood monitoring from the past twenty-five years (2000 to 2024). The analysis reveals a continuous upward trend in research within this field, with explosive growth starting in 2017. Subsequently, this paper systematically reviews methods for monitoring urban flood water depth from various data sources, with a particular focus on traditional sensor-based methods and intelligent big data-driven approaches, highlighting the research hotspots, advantages, and limitations of each method. Finally, the paper provides an outlook on future research directions and offers targeted recommendations for relevant stakeholders. The key points are as follows:

• Sensor measurement has long been a primary method for flood monitoring, providing real-time, high-precision water level data. Sensors can be broadly categorized into contact and non-contact types. Contact sensors provide accurate and intuitive monitoring but require frequent maintenance, while non-contact sensors are more durable but their accuracy can be affected by environmental factors. To mitigate these limitations, a combination of different sensor types, often integrated with IoT technologies, is commonly used. However, the challenges of high costs and maintenance difficulties remain significant.
• With the advent of the big data era, the application of social media data, surveillance camera data, and various integrated data sources in flood monitoring has gradually attracted the attention of researchers. The analysis of these data typically involves artificial intelligence methods, such as computer vision and text analysis, which have demonstrated significant potential in both research and practical applications. Although substantial progress has been made in optimizing algorithms, challenges persist in areas such as data acquisition, information filtering, data analysis, and ensuring data reliability. Future research should focus on addressing these challenges to achieve more accurate and efficient urban flood monitoring.
• As the diversity of data sources for flood monitoring increases, the limitations of using a single data source to collect flood information are becoming more apparent. Researchers are beginning to explore integrating multiple data sources, particularly in major disaster situations, to enhance the timeliness and accuracy of emergency management. Current studies primarily focus on the integration of social media data with remote sensing and hydrological data. Moving forward, it is crucial to fully leverage the rich, complementary, and comprehensive information provided by various data sources while addressing challenges related to data interpretation and the incompatibility of data sources.

In summary, this paper deepens the understanding of urban flood monitoring technologies and provides actionable recommendations for researchers and urban managers, contributing to the improvement of urban flood management and the promotion of its intelligent transformation.