Introduction

Floods have become more frequent and intensified under global climate change1,2,3,4,5. To cope with these trends, IPCC’s Sixth Assessment Report(AR6) makes it clear that future work should adhere to the development of climate resilience to adapt to climate change for broader benefits6. Cities are crucial to practical implication for extreme flood adaptation4,7,8,9,10.

Cities have been seeking practical and effective solutions to adapt to the damage caused by floods11,12,13,14,15, which remains a huge challenge. The current adaptation intervention to flood could pose a challenge for cities, as various limitations and adaptation gaps16. Therefore, the enhancement of technical adaptation interventions to flooding is urgently needed. The study collates technical adaptive interventions for flooding. Conventional methods of flood protection, such as levees, are increasingly insufficient as standalone solutions under climate change. To enhance resilience, future-proof planning must integrate adaptive, non-structural measures alongside traditional infrastructure to address unpredictable flood risks. A multi-pronged approach that focuses on long-term, sustainable solutions was recommended to improve the resilience of these cities to flooding17. Enhancing resilience was widely proposed and highlighted to flooding1,4,18,19, which is crucial for flood adaptation1,20,21. Several technology-related strategies exist to improve resilience to floods. Technology-related strategies against flooding can be summarized into pre-flood technology and post-flood technology. The pre-flood technology strategies include developing appropriate urban infrastructure4,22 to reduce the impact of extreme rainfall (e.g., by integrating sustainable green roofs and making sidewalks permeable), buildings that are in dynamic harmony with natural systems, and the prediction of flood risk in downstream urban areas17. Previous research and adaptive intervention mostly focused on pre-flood technical strategies. However, some emerging post-flood information technologies [such as Unmanned Aerial Vehicle (UAV)23, Tencent Life-saving documents (LSD)24 and Autonavi Map(AMap)25, see Figs. S4–7] played a groundbreaking and crucial role in a 1000-year-return extreme flood in China. Novel and disruptive technology applications have the potential to fill the adaptation gap to extreme flood26.

While new technologies have been advocated as adaptive strategies for extreme climates26, the persistent adaptation gap requires in-depth analysis of the impact of technologies to fill the adaptation gap16. The adaptation gap varies across city sectors. Therefore, cities need to figure out where adaptation interventions for these new technologies are insufficient or sufficient for the city’s sectors, in order to adjust technological strategies to favor those sectors that are more in need of enhanced intervention. The current adaptation intervention to flood could pose a challenge for cities, as various limitations and adaptation gaps16. Understanding the role of different technologies for different city sectors is urgently needed to reduce the adaptation gap in practice. Quantifying and comparing the direct and indirect roles of technology applications are prerequisites for developing appropriate technology mitigation strategies in the future. However, existing studies have not yet evaluated the mitigation effects of technological applications as extreme flood adaptation measures, nor their heterogeneous impacts across urban sectors.

Here, this work addresses this gap and provides a breakthrough assessment of the effects of new technological applications as extreme flood adaptation through modeling their underlying mechanism using a novel framework of Text-mining based Structural Equation Modeling (TSEM, Fig. 1). Social media has been an instant source of information for natural disasters, such as urban floods, throughout the world. A study in Pakistan proposes an evaluation framework to integrate disorganized online public opinion about urban flood disasters by analyzing emotional and conceptual characteristics27. In Great Britain, social media, particularly Twitter, is increasingly used to improve resilience during extreme weather events/emergency management situations27. The study introduced a TSEM framework, which integrates text mining techniques3,28,29 with structural equation modeling (SEM)30,31 to evaluate the role of technologies by examining their cross-relationships with perceived flood losses. The framework captured temporal changes in the multi-sectoral adaptation gap through public perception and flood loss characteristics. In our TSEM model, human-perceived flood losses are designed as core indicators as they provide more insight into the consequences of a disaster throughout the recovery process. This information is particularly helpful for efforts to build resilience for city sectors that most need improved adaptation interventions32. The mitigation effects of various technologies across city sectors are quantified, including the interaction between technologies and the direct and indirect links between technologies and flood perception. Intuitively, comparing the benefits of these technologies for post-flood recovery in each sector facilitates the deployment of appropriate technical adaptive strategies to reduce the city sector-specific adaptation gap.

Fig. 1: Framework of this study.
figure 1

a Schematic of novel technology applications for flood loss perception in post-flood response. b TSEM framework overview. TSEM framework developed human-perceived multisectoral flood losses and human-perceived Technology Adoption indexes (TAIs) from social media and revealed the effects of technology in multiple sectors against flooding. The multisectoral temporal corpus and four technical temporal corpora were filtered from the original microblog data. The MFLPs, including Casualty Loss Perception (CALP), Transport Loss Perception (TRLP), Communication Loss Perception (CMLP), Water and Electricity supply Loss Perception (WELP), and Building Loss Perception (BDLP), were estimated in five sensitive sectors on Weibo. Technology Adoption Indexes (TAIs, e.g., UAV, LSD, Amap, and DW) were constructed via hourly technology-related word frequency statistics. Structural equation models were built to reveal the relationship between technological adoption and flood loss perception in various sectors. TSEM (Fig. 1b) incorporates two core modules: text data processing module and structural equation modeling module. Text processing module completed the construction of the multisectoral perceived flood loss index, four technology adoption indexes, and Rescue Effort Effectiveness (REE). REE is quantified by pre-trained BERT and a multi-layer perceptron (MLP). By analyzing the topics of four technological corpora, seven hypotheses are proposed. Then, in the SEM module, these indexes and hypotheses obtained from the text module are used for structural equation modeling. The appropriate models were selected by comparing the performance indicators (e.g., Chi-square, GFI, and RMSEA) of the models to identify the best-fitting models. After determining the optimal model, the SEM results were used to compare and evaluate technological effects across sectors. The TSEM process is detailed in the Methods section.

Full size image

This research investigated the “7.20 extreme rainstorm” event, which hit two major cities (a 1000-year flood in Zhengzhou and a 60-year flood in Xinxiang) in Henan Province, China, between July 17th and 23rd, 2021. A total of 3,586,895 flood-related Weibo posts were extracted from Henan Province during the flood event. Regarding the four most concerned technologies (i.e., UAV, LSD, Amap, and DW) in Weibo, the first three are emerging information technologies, while DW is a traditional technology. This study developed Multisectoral Flood Loss Perception Indexes (MFLPs), which included five sectors: casualty transportation, communications, buildings, and water and electricity supply. Human-perceived technology adoption indexes (TAIs) were developed to reflect the intensity of technological implementation. The SEM analysis revealed significant causal relationships between major flood loss perceptions (MFLPs) and key drivers, including Rescue Effort effectiveness (REE), rainfall intensity, and four TAIs. The study quantitatively assessed the impacts of technological applications across five critical sectors during flood events. This study highlighted the significance of emerging information technologies for extreme flood adaptation.

Results

Multisectoral flood loss perceptions (MFLPs)

Using text-based word frequency statistics, this work formulated MFLPs based on Weibo to assess perceived flood consequences throughout post-flood responses. The MFLPs, including CALP, TRLP, CMLP, WELP, and BDLP, were estimated in five sensitive sectors by Weibo: casualties, transportation, communication, water and electricity supply, and building. The cities of Zhengzhou and Xinxiang in Henan experienced the most severe floods, with the maximum hourly rainfall in these regions serving as the rainfall indicator (see Fig. S1). While maximum rainfall is not the same as flooding, as flooding is also affected by land surface conditions such as land cover and intermittent precipitation, maximum rainfall is still the main cause of flooding. There are lead-lag relationships between MFLPs and maximum rainfall. So, the study conducted a time lag correlation analysis between the maximum rainfall and flood loss perceptions in five different sectors. Assuming there is a causality between the intensity of rainfall and MFLPs, the Granger causality test is performed. The results of the Granger causality test reveal a significant 3-h time lag between synthetic flood loss perception and maximum rainfall (p = 0.025, F = 3.157, see Table S5). Furthermore, there are differences in lag order between multisectoral loss perception and maximum rainfall, which reveals that the sensitivity of human perception varies from sector to sector. Specifically, citizens are most sensitive to water and electricity supply (lag time:1 h, p = 0.004, F = 8.299), followed by human communication loss (lag time:2 h, p = 0.000, F = 19.281), building damage (lag time:2 h, p = 0.027, F = 3.642), casualties (lag time:2 h, p = 0.009, F = 3.157), and traffic loss (lag time:6 h, p = 0.001, F = 3.655).

In Fig. 2, prior to the extreme rainstorm, both the maximum rainfall curve and the MFLP curves remained close to zero. This reflects the low level of negative comments on various aspects of social media before the onset of the rainstorm. As the rainstorm emerges, the maximum rainfall curve exhibits an increasing trend, accelerating rapidly until it reaches its peak. Simultaneously, the MFLPs curve also shows a rising trend. Once the maximum rainfall curve reaches its peak, it begins to decline, whereas the MFLP curves continue to rise rapidly, reaching their zenith. Subsequently, the maximum rainfall curve fluctuates within a relatively small amplitude, gradually declining and returning to zero. In contrast, the MFLP curves exhibit strong and long-lasting volatility, as the flooding’s damage persists even after the extreme rainfall, requiring cities and individuals to confront and manage the destructive consequences.

Fig. 2: Hourly human-perceived multisectoral flood loss perception index and max rainfall.
figure 2

a Time series of hourly synthetical flood loss perception and maximum rainfall. b Time series of hourly communication loss perception and maximum rainfall. c Time series of hourly casualty loss perception and maximum rainfall. d Time series of hourly transport loss perception and maximum rainfall. e Time series of hourly flood loss perception for water and electricity supply and maximum rainfall. f Time series of hourly building loss perception and maximum rainfall. Maximum rainfall refers to the maximum hourly rainfall in Zhengzhou and Xinxiang Cities.

Full size image

Human-perceived technology adoption indexes

The most discussed was LSD, followed by UAV, AMap, and DW on Weibo (see Fig. S2). To capture the human-perceived implementation intensity of technologies, the study constructed the TAIs via hourly technological word frequency statistics (see Fig. 3), which are presented as time series. In Fig. 3d, the peak of the DW curve appears later than the other technologies, which reflects the recent attraction of public attention for DW, resulting from the belated arrival of DW in Zhengzhou City. UAV is the most prominent rescue technology on Weibo, but its duration is shorter. LSD, by contrast, is the longest-lasting rescue technique. The LSD curve also exhibits prolonged fluctuations after reaching its peak, possibly due to long-term availability and durable supply for LSD in comparison to UAV and DW.

Fig. 3: Text frequency of four popular technologies.
figure 3

a The UAV indicator represents the hourly frequency of discussions about UAVs on Weibo. b The LSD indicator represents the hourly frequency of discussions about LSD on Weibo. c The AMap indicator represents the hourly frequency of discussions about AMap on Weibo. d The DW indicator represents the hourly frequency of discussions about DW on Weibo.

Full size image

Mechanisms of new technological applications

Several hypotheses (see “Associations between technologies and perceived flood losses from corpus analysis” in Method) were established, and possible causal paths were tested and compared (Supplementary Figs. 8–12) according to our empirical SEM results. Specifically, the study has examined the causal path in which MFLP is indirectly influenced by UAV through other technologies as a mediator. Moreover, the study confirmed that technologies affect the changes in perceived flood losses through REE as a mediator. Chi-square, P-value, Goodness of Fit Index (GFI), and root mean square error of approximation (RMSEA) were considered performance evaluation indexes for the model structure. By comparing the performance metrics of these possible causal paths, the study determines the most appropriate structural path. Eventually, the most explanatory models in this study were identified (Supplementary Tables 6–10). Remarkably, UAV plays a pivotal and significant role, both directly and indirectly, in shaping the change of flood loss perceptions. These findings have been further elucidated through the SEMs (see Fig. 4).

  1. 1)

    Casualty loss perception (CALP).

    The findings from the CALP model indicate that CALP is significantly mitigated by both the utilization of UAVs (coef = −0.364, 95% CI: [−0.456, −0.272], p = 0.000) and the use of LSD (coef = −0.302, 95% CI: [−0.412, −0.193], p = 0.000). Furthermore, UAVs also play a significant indirect role by facilitating the application of LSD (coef = 0.309, 95% CI: [0.187, 0.432], p = 0.000) and enhancing REE (coef = 0.078, 95% CI: [−0.007, 0.163], p = 0.073). LSD also has an indirect mitigating effect on CALP by promoting REE (mediator) (coef = 0.577, 95% CI: [0.499, 0.655], p = 0.000).

  2. 2)

    Communication loss perception (CMLP).

    In terms of CMLP, UAV exhibited a strong and significant effect on restoring communication (coef = −0.608, 95% CI: [−0.697, −0.519], p = 0.000). Additionally, UAVs played an indirect role in communication perception by facilitating the use of LSD (mediator, coef = 0.308, 95% CI: [0.186, 0.430], p = 0.000) and enhancing REE (coef = 0.078, 95% CI: [−0.007, 0.163], p = 0.073). A pathway involving multiple mediators (i.e., UAV→LSD→REE→CMLP) reveals how UAVs further facilitate REE by facilitating the application of LSD, which ultimately alleviates CMLP. Overall, UAV had the strongest mitigating effect, followed by LSD and REE.

  3. 3)

    Water and electricity supply Loss Perception (WELP). UAV and LSD showed significant mitigation effects (coef = −0.470, 95% CI: [−0.564, −0.376], p = 0.000; coef = −0.169, 95% CI: [−0.280, −0.057], p = 0.003) on promoting the recovery process of water and electricity supply. Additionally, UAVs played an indirect role in WELP by facilitating the use of LSD (coef=0.308, 95% CI: [0.186, 0.430], p = 0.000). REE, as a mediator, moderates the effect of UAV on CMLP and the effect of LSD on CMLP. UAVs have broader effects on the overall structural effects network. UAVs influence WELP through UAV→LSD→REE→WELP pathway through two mediators. Overall, for WELP, UAVs, LSD, and REE have played an important mitigating role.

  4. 4)

    Traffic loss perception (TRLP). For transport loss, novel technologies such as UAV (coef = −0.423, 95% CI: [−0.553, −0.294], p = 0.000) and AMap (coef = −0.176, 95% CI: [−0.299, −0.052], p = 0.005) exhibit the most robust and statistically significant direct effects in mitigating negative consequences, in comparison to DW (indirect effect of 0.027). As far as UAV is concerned, it affects the change of TRLP through multiple paths, which involves REE and AMap, two mediators, because UAV can promote the application of AMap (coef = 0.775, 95% CI: [0.681, 0.868], p = 0.000). As a mediator in this structural network, REE mediated the influence of two technologies on TRLP changes, including DW and AMap. Overall, in terms of the mitigation effects of TRLP, the total effects of the three technologies are ranked as UAV (0.694), AMap (0.350), and DW (0.027).

  5. 5)

    Building loss perception (BDLP). For the BDLP, the use of UAVs has a strong and significant positive effect in reducing the perception of building loss (coefficient = −0.626, 95% CI: [−0.732, −0.521], p = 0.000). DW also exhibits a significant positive effect in mitigating perceived building loss (coefficient = −0.209, 95% CI: [−0.281, −0.137], p = 0.000). Nevertheless, the SEM analysis reveals that UAV has a greater mitigation effect than DW in BDLP. However, both technologies are significant and indispensable in relieving BDLP. As a mediator in this structural network, REE mediated the influence of two technologies on BDLP changes, including UAV and DW. Overall, in terms of the mitigation effects of BDLP, the total effects of the two technologies are ranked as UAV (0.738) and DW (0.231).

    Fig. 4: Optimal structural equation modeling through performance screening.
    figure 4

    a The driving mechanism of technological applications to the perceived loss of human casualties. b The driving mechanism of technological applications for changes in perceived traffic losses. c The driving mechanism of technological applications for changes in perceived communication losses. d The driving mechanism of technological strategies for changes in the perceived loss of water and electricity supply. e The driving mechanism of technological applications for changes in perceived building losses. f Total effects of technological applications for multisectoral loss perception and the total effects of rescue efforts. Here, UAV, LSD, Amap, and DW referred to text frequency indices for the four techniques. Significant paths are shown in light red if positive or in blue if negative.

    Full size image

Discussion

While a growing literature supports the notion that disruptive technologies are crucial as extreme flooding adaptation4, research on technological applications as part of this adaptation and their underlying mechanism is currently lacking. This study, for the first time, contributes to a modeling method of the cross-relationships between techniques and perceived flood losses by inferencing from large-scale temporal social media data33,34 to during extreme floods. Using human-perceived flood losses as opposed to actual asset loss as a metric to assess disaster impacts provides more insight into the consequences of a disaster throughout the recovery process35. Understanding how to develop and integrate appropriate technical strategies and apply them to urban flooding to provide practical and systematic guidance is key to adaptation and resilience. TSEM provides important advances in assessing the implications of technological strategies for post-flood recovery across different city sectors.

Based on the analysis of the model results, the work reports the effects of new technology applications as an adaptation to extreme flooding and their variability across sectors. The results reveal high heterogeneity in technology strategy effects across different city sectors, mainly due to the characteristics of the existence of technology. Understanding the cross-sectoral heterogeneity and its underly mechanisms of technological applications can help urban emergency departments deploy appropriate technology-driven post-flood adaptation strategies. The work highlights that the current adaptation intervention to flood could pose a challenge for cities as various limitations and adaptation gaps16. The application of new technologies in cities can help fill the adaptation gap. Our findings contribute to advancing discussions on new technological applications as extreme flood adaptation to cope with more frequent flooding and enhance flood resilience17,36.

Heterogeneous technological effects across different city sectors

The four technology effects on post-flood recovery are evaluated through cross-sectoral comparison (Fig. 5a). Novel information technologies (i.e., AMap, LSD, and UAVs) presented more powerful benefits than traditional technology (i.e., DW). UAV had a strong mitigating effect on the perceived loss of traffic, buildings, and communications, followed by water and electricity supply and casualties. The high mitigating effect of UAVs on perceived losses in all sectors (greater than 0.5) indicated the central strategic position of UAVs in rescue efforts, which drove the application of other technologies and facilitated relief efforts, as it provided emergency communication and disaster monitoring. LSD mainly has a greater mitigation significance for casualties, water and electricity supply, and communication. It is the most effective technological strategy for perceived casualty loss, other than UAV. AMap had a strong restorative effect on perceived traffic losses but little effect on perceived losses in other sectors. DW has a strong mitigation effect on perceived building losses and some mitigating effect on perceived traffic losses, but has little effect on perceived losses in other sectors.

Fig. 5: The effects of four technologies.
figure 5

a Cross-sectoral comparisons of the effects of four post-flood technological strategies. Figure 5a shows the total mitigation effects of UAV, LSD, Amap, and DW on MFLPs. The values of these total effects include the sum of direct and indirect effects based on structural equation models. b The relative importance of four technologies in promoting REE (mediator) across five urban sectors. This is obtained by comparing the path strength of the technical effect propagating through REE. c The relative importance of four technologies across five urban sectors. Here, the effect values are converted into percentages to obtain relative importance.

Full size image

Our findings highlight REE as the most central mediator in the overall mechanism of the models. Technology indirectly influences the recovery of city sectors through the REE pathway. Furthermore, UAVs also serve as an important mediator, helping to drive the application of LSD, Amap and DW. This work report the relative importance of REE path intensity of technical effects across five city sectors as it reflects the importance of technological applications to driving REE across different city sectors (see Fig. 5b). For the three city sectors of casualties, communications, and water and electricity supply, LSD had the strongest contribution (88.1–88.3%) to the effectiveness of rescue efforts, followed by UAVs (11.7–11.9%). Amap (86.6%) was the most effective for REE in the city transportation sector, followed by DW (13.4%). UAV (77.3%) greatly contributed to the effectiveness of rescue efforts in the city building sector, followed by DW (22.7%).

The study discusses the relative importance of technical effects across five city sectors. For the three city sectors of casualties, communication, and water and electricity supply, the three adaptive technologies have roughly similar effects on their recovery, with UAV (43.0–62.2%) and LSD (19.5–37.1%) ranking in importance compared with REE (17.3–28.5%). In contrast, the transportation sector and the construction sector show inconsistent patterns of technological effects. In terms of the transportation sector, in order of importance, UAV (45.6%) ranked first, followed by REE (29.3%), AMap (23.1%) and DW (1.8%). As for the importance of technology and rescue efforts in the city building sector, UAV (57.4%) ranked first, followed by REE (24.7%), and finally DW (18.0%). Overall, for casualties, communications, and water and electricity supplies, these findings underscore the importance of UAVs and LSD, as the benefits of one alone exceed those provided by REE. UAVs occupy the most dominant position in all city sectors, compared to other technologies and REE.

Practical implications

Firstly, monitoring real-time multisectoral perceived flood losses from online review data is critical. This contributes to identifying sectors that are in urgent need of mitigation based on public perception. Understanding the differences in the online concern and sensitivity of different sectors to infer which aspects of the strategy are the most urgent and important will help urban management departments to formulate appropriate strategies and programs according to the significance of the sectors in flood emergency management. Public perception is crucial for management and decision-making37. For example, residents are most sensitive to the shortage of water and electricity supply, so the government should give priority to this sector. Secondly, monitoring the intensity of human-perceived technology adoption is an important means of recognizing the extent to which technical strategies are effective in the post-flood recovery process. Lastly, policies should integrate the emerging information technology strategies into the flood adaptation framework. Adaptation gaps occur when adaptation efforts are limited by some factors and fall short of expectations16. Existing adaptation gaps should be considered systematically, especially where they may lead to flood adaptation limits. Therefore, it is particularly important to distinguish the adaptation efforts of urban sectors to highlight which sectors have adaptation gaps. Policies should provide sector-targeted relief efforts on the basis of a full understanding of the recovery effects of adaptive technology strategies on different sectors. Understanding how to develop and integrate appropriate technological strategies and apply them to urban flooding to provide practical and systematic guidance is the current key to improving urban post-flood resilience. Policymakers can draw upon the empirical evidence presented in this study to regulate the deployment of technology-driven flood adaptive strategies toward more urgent sectors based on MFLPs and flood situations. This study argued that technological applications for flood adaptation are critical to fill the adaptation gap.

The implications of this research extend beyond the specific case study of Henan flooding, resonating with urban areas worldwide facing similar challenges. By showcasing the effectiveness of technology-driven adaptive strategies, the study offers valuable lessons that can be applied in diverse geographic contexts. Although influenced by diverse geographical and socio-technical factors, the adoption of technologies varies across countries. However, the characteristics of these emerging technologies remain consistent, offering valuable insights even in different geographical and institutional contexts. However, the same technology may produce divergent effects in different geographical and institutional settings, requiring deeper analysis. Empirical evidence of technology adoption contributed to understanding the relationship between technological strategies and climate resilience, and provides practical guidance for reducing the flood adaptation gap in flood-prone urban areas. In light of the forthcoming AR7 IPCC report to cope with climate change impacts, the empirical insights from this study contributed to the AR7’s recommendations for climate adaptation enhancement, which provides policymakers with the latest scientific evidence on the benefits of technology strategies for flood adaptation across city sectors. For example, this study provides a concrete approach to assessing the impact of technology and the adaptation gap based on public perception. Our research provides an assessment of technological support for mitigating the effects of climate change to facilitate effective implementation. The proposed TSEM approach is valid in modeling the cross relations between technology and perceived disaster loss. The results of this research provide insights for the flood adaptation planning of technology readiness, considering different flood damage scenarios with resource constraints.

Limitations and future directions

Several limitations of this study should be mentioned. First, while the study leverages a social dataset and advanced text modeling techniques, it primarily focuses on Henan Province, which may not fully capture higher resolution spatial scales (e.g., street-level data). The reliance on aggregate data may obscure intra-city variations, such as differences between urban cores and suburban areas. Second, the work lacks comparisons of different cities. Thirdly, the Weibo dataset inherently excludes populations without internet access, such as the elderly or non-social media users. Fourth, the weighting scheme adopted for REE construction is acknowledged to be empirically derived, representing a potential area for methodological refinement. This limitation suggests that the sample may not adequately represent vulnerable groups—including older adults, individuals with disabilities, and low-income populations—who are less likely to engage in online platforms. Despite these limitations, the study has significant strengths. It is the first to investigate the effects of post-flood technological strategies and their variability across city sectors. Future research should be conducted in several aspects. Firstly, online reviews of flood events from different regions worldwide should be collected to investigate the similarities and differences of urban multisectoral flood perception and technology adoption under various city backgrounds. Extensive, multi-regional data studies will provide a more comprehensive understanding of the significance of technology strategies. Secondly, global technology strategies need to be collated to build the inventory of mitigation effects of technology strategies to deploy efficient and targeted technology strategies. This study provides a methodological framework for evaluating post-flood technological adaptation strategies, offering insights for optimizing global applications of emerging flood resilience technologies.

Methods

Overview of the study area, background, and emerging information technologies

China, as the second-largest economy in the world, exhibits diverse geographic and climatic features. Among the natural disasters commonly experienced in China, flooding poses the most severe threat. The Yellow River Basin, which ranks as the second largest river in China, boasts a distinct river valley topography. Climate change has resulted in increased rainfall and frequent storm floods in the northern and central regions of Henan province, where the Yellow River meanders. As a consequence of this persistent and heavy precipitation, multiple floods occurred in the cities of Zhengzhou, Kaifeng, and Xinxiang, situated in the north-central part of Henan province (see Fig. S3), during the years 2004, 2009, 2010, 2012, 2016, and 2021. Severe rainstorms and devastating flooding hit Henan Province from 17 to 23 July 2021. In Henan, 14.786 million people were affected, and 398 people died, which accounts for USD 18.9 billion in losses. Zhengzhou in Henan experienced a 1000-year return period flood with the influence of Typhoon In-Fa and the Western Pacific subtropical high. The flood affected a total of 14.786 million people in 150 counties (cities and districts) of Henan Province, with a death or missing toll of 398, which included 380 individuals from Zhengzhou alone, amounting to 95.5% of the province.

This study discusses three novel technologies (UAV, LSD, and AMap) and one traditional technology (DW). The utilization of technological strategies has the potential to significantly contribute to the resilience of cities in managing flood emergencies26,38. Various examples of such technologies and strategies were employed during the flooding, including UAVs, the “Dragon Water” drainage device (DW), LSDs, and the AMap application, among others.

(1). UAV. The Wing Loong UAV (Fig. S4) was widely recognized as a prominent novel technology utilized in response to the devastating flood that occurred in Henan province on 21 July 2021. In Mihe Township, Gongyi City, several villages experienced severe communication disruptions. To address this issue, the Emergency Management Department employed the Wing Loong UAV aerial emergency communication platform. This platform facilitated long-term, stable, and continuous mobile signal coverage over an approximate area of 50 square kilometers by utilizing a mobile public network base station carried by the UAV. The UAV’s exceptional flight range, endurance, carrying capacity, and environmental adaptability enabled field surveys to be conducted even under extreme conditions, such as power and network outages and circuit breaks.

(2). LSD. A special document called “Information for Rescue Workers” is an online form for non-governmental rescue organizations to collect and update rescue information on Tencent documents. This online form served as a vital tool for non-governmental organizations to gather rescue information and was hosted on Tencent documents39 (Fig. S5). LSD was guaranteed 24-h access. Its decentralized capabilities and cloud-based nature enable people who are not connected to each other to efficiently connect and reallocate resources to those in need.

(3). DW. Another noteworthy feature is that the “Dragon Water” drainage device (Fig. S6), as a traditional technology, has the capability to drain water up to one meter deep and cover an area of 3000 square meters within an hour.

(4). AMap. Autonavi Map (AMap) is China’s largest map navigation App. The AMap (Fig. S7), introduced a channel feature to facilitate mutual assistance and share flood information in Henan Province. Local users can access this channel by clicking on the “waterlogging” icon displayed on the map. Through this feature, users can conveniently find contact details of rescue teams, seek information on nearby shelters and even post requests for assistance.

Data

Social media data in the study is sourced from Sina Public Opinion Platform. Sina Public Opinion Pass is a big data service platform for public opinion of government and enterprises developed by Shanghai Midu Information Technology Co., LTD. The study collected a total of 3,586,895 flood-related tweets from Weibo, China’s largest social media platform, with IP addresses in Henan Province, from July 12 to August 10, 2021. The collected tweet dataset comprehensively covers the duration of Henan Province’s 7.20 extreme rainfall event, encompassing two major urban areas: Zhengzhou (a 1000-year return period flood) and Xinxiang (a 60-year return period flood) from July 17 to 23, 2021. Hourly precipitation records from 16 meteorological stations in these cities, spanning July 12 to August 10, 2021, were obtained from the China National Meteorological Science Data Center.

Human-perceived technology Adoption indexes (TAIs) and Multisectoral flood loss perceptions (MFLPs)

Social media data is used to monitor disaster situations as it provides near-real-time public perception29,40,41,42. Extracting online human flood perception from Weibo is a practical approach to exploring and analyzing flood disasters40,41,43. Human-perceived disaster losses have become a vital indicator for assessing the devastating consequences of floods. To measure flood perception across multiple sectors, the Weibo dataset was divided into 720 hourly intervals spanning from July 10th to August 10th. For each sector, the hourly count of urban flood disaster discussions was calculated. These counts were then transformed into logarithmic values, serving as flood perception indices for multiple sectors. Specifically, these indices include the Transport Loss Perception Index (TRLP), Water and Electricity Loss Perception Index (WELP), Communication Loss Perception Index (CMLP), Building Loss Perception Index (BDLP), and Casualty Loss Perception Index (CALP). The flood loss perception indicators for these five sectors were constructed through frequency statistics based on related terms, as detailed in Supplementary Text S2. Indicators of technical adoption intensity were constructed based on hourly word frequency statistics for four technology-related terms. The human-perceived technology adoption index is constructed to serve as a proxy indicator of technology adoption intensity.

Text-based indicator of rescue effort effectiveness (REE)

Rescue efforts include various efforts from the government, enterprises, society, etc. The study first manually annotated 9000 tweets from the 3.58 million microblog data, including 3000 tweets mentioning that rescue efforts had reached recovery goals, 3000 tweets involving rescue efforts, and 3000 tweets not involving rescue efforts. To measure the intensity of the rescue effort and its effectiveness in the tweets, the work used a pre-trained language model BERT-base-Chinese of text-embedding to transform the text semantic information into computable high-dimensional word vectors to represent the meaning of sentences44. Text embedding models have become an emerging computational approach in the social sciences, as word vectors in well-trained embeddings represent semantic information corresponding to how people understand them45. The study converted 9000 tagged tweets into 768-dimensional sentence vectors through the pre-trained Bert model and then trained the classification model using a multi-layer perceptron classifier. Nine thousand tweets were randomly scrambled, with 90% of them as training set data and 10% as test set data. The classification results showed an accuracy of 87% and an AUC of 96.75% (see Table S20 for details). This work applied this trained model to 3.58 million tweets for text classification, and every tweet was classified. The comments were divided into time intervals for each hour to obtain statistical indicators for each hour. The work uses online comment data to construct a macro text-based indicator of REE hourly.

$${e}_{t}=\left\{{{{\rm{tweets}}}}\; {{{\rm{related}}}}\; {{{\rm{to}}}}\,{{{\rm{rescue}}}}\; {{{\rm{efforts}}}}\; {{{\rm{at}}}}\,t\,{{{\rm{hour}}}}\right\}$$
(1)
$${r}_{t}=\left\{{{{\rm{tweets}}}}\; {{{\rm{related}}}}\; {{{\rm{to}}}}\; {{{\rm{rescue}}}}\; {{{\rm{efforts}}}}\,{{{\rm{that}}}}\,{{{\rm{achiev}}}}{{{\rm{ing}}}}\,{{{\rm{recovery}}}}\; {{{\rm{effects}}}}\; {{{\rm{at}}}}\,t\,{{{\rm{hour}}}}\right\}$$
(2)
$${{{{\rm{R}}}}{{{\rm{E}}}}{{{\rm{E}}}}}_{t}={w}_{1}* {{{\rm{Count}}}}\left({e}_{t}\right)+{w}_{2}* {{{\rm{Count}}}}\left({r}_{t}\right)$$
(3)

Here, \({w}_{1}\) is the weight of the tweets related to rescue efforts, \({w}_{2}\) is the weight of the tweets related to rescue efforts that achieve recovery effects, and \(t\) is for a specific hour. For rescue efforts, tweets about rescue efforts that have achieved recovery should be given a higher weight than normal rescue efforts. The weight parameters were set as w₁ = 1 and w₂ = 2.

LDA (latent Dirichlet allocation) modeling and interconnectivity of topics

Four corpora of l technologies were extracted from the original Weibo comment dataset. To analyze the usefulness of technology during the Henan flooding, the study applied Latent Dirichlet Allocation (LDA) modeling to each corpus separately. LDA is a topic modeling technique that assigns a probability distribution of topics to each document in a corpus. The four extracted corpora consist of discussions on UAV, LSD, AMap, and Dragon Water (DW). By extracting topics from the Tech dataset related to urban flood events, this work provided insights into the impact of novel technologies in flood management. For this purpose, this work employed the widely used LDA technique to identify underlying topics in the corpus of flooding-related reviews. Using LDA, this work obtained the vocabulary for each topic and the frequency of vocabulary occurrence. Specifically, the topic modeling was performed using the Gensim Python library, implementing LDA with 1000 iterations across five predefined topics. Further information on the LDA modeling process can be found in Text S3. Association inferences were further derived by evaluating interconnections between topic-specific keywords.

Associations between technologies and perceived flood losses from corpus analysis

Based on technological corpus analysis, radar plots (Fig. 6e) showed that LSD and UAV influence communication perception, casualty perception, and water and electricity supply perception. In contrast, AMap primarily influences the perception of casualties and traffic.

Fig. 6: Technological corpus analysis and hypothetical path.
figure 6

a Ranking and interconnectivity of the top 30 topic words of UAV technological corpus. Nodes represent topic words, and the edges represent the co-occurrence of topic words in the corresponding corpus. The thickness of edges indicates interconnectedness between topic words. b Ranking and interconnectivity of the top 30 topic words of LSD technological corpus. c Ranking and interconnectivity of the top 30 topic words of AMap technological corpus. d Ranking and interconnectivity of the top 30 topic words of DW technological corpus. e Radar plots for multisector statistics of flood loss perception reflected by topic word. Radar plots were separated into five dimensions, specifically communication perception (CMP), casualty perception (CAP), building perception (BDP), transport perception (TRP), and water and electricity supply perception (WEP). The significance of four technologies for flood perception in five subsectors was evaluated by categorizing topic words into five flood perceptions. f Hypothetical path structure based on H1-7. Here, \(\Delta {{{\rm{MFLP}}}}\) s represent change in multisectoral flood loss perception.

Full size image

The relevance of UAV to communication perception

Figure 6a indicates that communication-related topic words carry significant weight in the UAV corpus, with examples like “information” (0.016), “Cell phone” (0.022), “telephone” (0.011), and “communication” (0.003). The interconnectivity of UAV topics demonstrated a clear association with mobile communication-related words, such as “telephone”, “information”, and “cellphone”. However, UAV has the potential to promote the recovery of multiple sectors as their disaster monitoring and emergency communication network. Hypothesis 1 (see Table1) is proposed.

Table 1 Mitigate effects hypotheses
Full size table

The relevance of LSD to the perception of casualties and communications

Possible mitigating effects between LSD and casualty, communication, as well as power supply perception, emerged based on relevant topic words in Fig. 6b, including examples such as “Be trapped” (0.005), “Seek assist” (0.029), “Rescue” (0.041), “power supply” (0.005), “Hydrops” (0.005), “information” (0.024), and “telephone” (0.019). Hypothesis 2 is proposed.

The relevance of AMap to traffic perception

AMap potentially influences traffic perception, as revealed by relevant topic words like “refuge” (0.008), “map” (0.013), and “Icon” (0.006) (Fig. 6c). Hypothesis 3 is proposed.

The relevance of DW to traffic and building perception

DW may potentially mitigate traffic and building loss perception, as indicated by relevant subject words such as “rescue” (0.049), “subway” (0.018), “tunnel” (0.010), and “drainage” (0.007) (Fig. 6d). Hypothesis 4 is proposed.

UAV support LSD applications

Figure 4b shows the interconnectedness between LSD topics and UAVs, highlighting a significant correlation with specific topic words (e.g., “information” and “rescue”). Further, multiple significant indirect links were discovered, i.e., “UAV”- “cell phone”- “information”- “rescue”. Based on the analytical results, it is inferred that UAVs facilitate the deployment of LSD. In fact, UAV provides emergency communications, which is a prerequisite for the adoption of LSD and AMap. UAV provides disaster monitoring to facilitate deploying appropriate rescue efforts, so Hypothesis 5 is proposed.

The perceived flood losses were continuously affected by the intensity of rainfall

Considering that there exists a Granger causality relationship between precipitation and MFLPs (The maximum lag order is 6). Hypothesis 6 is proposed.

Technology strategies promote overall flood rescue efforts

Evidence shows that the word “rescue” is present in all four technology corpora. Technology adoption is part of the rescue effort, while they also promote the overall rescue effort. LSD, through its advantages of distributed collaborative interaction, promotes the dissemination of important information to facilitate the timely allocation of social rescue resources. UAV facilitates other social relief efforts through disaster monitoring and provide an emergency communication network. So, H7 is proposed.

Based on these hypotheses, several possible causal paths(H1-7) were integrated into the connected causal structure (Fig. 6f).

Structural equation modeling (path analysis)

SEM was first used to test the significance of pathways from technological strategies to changes in human-perceived flood loss(outcome) via the REE (mediator). An SEM was separately formulated for each sector and each technology strategy (UAV, AMap, LSD, and DW) to assess the direct and indirect effects of technology strategy on perceived flood loss. The SEMs were constructed using the semopy package in Python, with regression coefficients estimated for all specified paths. Based on text-based inference and hypothesized effects, prior models were established to guide this evaluation. All indicators were transformed, including variational mode decomposition and discount transformation (Supplementary Text S4, 5), to extract major trends to ensure the robustness of time series data. All variables were normalized by mean and standard deviation across all individuals before model fitting to ensure that the magnitude of the regression coefficient is comparable across models.

Model performance comparison

The optimal model was determined through comparative evaluation of overall effects across all specified models. The overall goodness of fit for the SEMs was assessed using several criteria: a chi-square test with P > 0.05, a GFI greater than 0.90, and an RMSEA below 0.08. Additionally, multiple possible causal paths were identified through the improvement percentage of GFI and the chi-square difference test. However, the study acknowledges that the identified mechanisms may not encompass all potential pathways, and the causal models may not fully capture the process’s systemic complexity. The current study was constrained by both underdeveloped theoretical frameworks and limited data availability, restricting the modeling capacity for intermediate pathway mechanisms. Nevertheless, the robustness of the model results was systematically tested.

Model robustness

Potential deviations in indicator measurement are acknowledged in this study. For instance, there could be differences in the perception index due to the incomplete and unfair representation of the population on social media or the biases in the REE. To assess the robustness of the SEM, a comprehensive uncertainty analysis incorporating bootstrap resampling methods was implemented. Specifically, this work employed Monte Carlo simulations with a maximum permitted error of 12% for REE indicators and 5% for all other model indices. This conservative approach ensured rigorous testing of model stability under measurement uncertainty conditions. Using 1,000 bootstrap samples, this work performed iterative model fitting to generate empirical sampling distributions for all parameter estimates. Confidence intervals (95%) were derived using percentile methods, with statistical significance determined by the proportion of bootstrap estimates that exceeded or fell below zero46. The bootstrap results (Supplementary Tables S21–S25) demonstrate remarkable consistency with our original findings, showing that all initially significant paths retained their statistical significance (p < 0.01), and parameter estimates exhibited minimal variation across resamples. This means that SEM is robust.

The process of TSEM

Specifically, in the initial and subsequent phases, the original Weibo text and rainfall data were collected and preprocessed. Detailed information on the data retrieval and preprocessing can be found in Text S1. From the original microblog dataset, multisectoral flood corpus and four technology corpora were extracted and sorted by applying predefined relevant terms (refer to Table S4). Text mining techniques were then employed to construct comprehensive MFLPs (Text S2) and TAIs. LDA modeling was utilized to highlight the significance of topic words (Text S3). Furthermore, potential influences of novel technologies across sectors were inferred through topic interconnectivity analysis. Then, word frequency analysis was conducted to ascertain the implementation intensity of applied technologies (i.e., UAV, LSD, AMap, and DW), and corresponding indices were developed for in-depth analysis of their perceived frequency. Multisectoral flood loss perception indices were developed to uncover patterns in the perception of flood disasters across different sectors. To extract valuable components from the data, two data transformations, including VMD (Variational mode decomposition, see Text S4) and discount transformations (Text S5), were performed to adjust the indexes. Finally, a variety of SEMs were constructed, incorporating prior knowledge assumptions of text inference, and the optimal model was selected based on the evaluation criteria.