1 Introduction

Tropical cyclones (TCs) have destructive impacts on society and economy in tropical and subtropical regions (Schmidt et al. 2010; Xiao and Xiao 2010). As the economically most developed and densely populated area on the southern coast of China, Guangdong Province has been severely impaired by TCs. For instance, the four TCs of Usagi (1319), Rammasun (1409), Hato (1713), and Mangkhut (1822) during the past decade collectively affected 15.76 million people and caused a direct economic loss of RMB 77.17 billion yuanFootnote 1 in the province. The intensity of TCs is likely to increase significantly as the global climate warms, and the disaster risk caused by TCs is projected to become more severe in the future, especially with the region’s fast population growth and economic development (Mori and Takemi 2016; Bacmeister et al. 2018). It is of great significance to accurately assess the disaster risk of TCs and rapidly predict their damages at a fine spatial scale in Guangdong Province and other coastal regions affected by TCs.

Tropical cyclone-related wind disasters cause damage to human lives, infrastructure, vegetation, and other aspects of society as TCs transit over land. These impacts include direct damages that come from the wind itself as well as the induced secondary or indirect disasters, such as storm surges, floods, heavy rainfall, landslides, and interruption to transportation, commercial activities, and so on (He et al. 2021). Most studies have carried out qualitative risk assessment for TC wind disasters based on integrated indicators and probabilistic statistical simulation (Song et al. 2020; Bloemendaal et al. 2021). Generally, disaster risk indices of TCs were established with hazard factors and elements at risk. For example, Lu et al. (2018) developed an intensity index combined with wind speed and population vulnerability to evaluate the TC wind disaster risk in Zhejiang Province. Meng et al. (2021) developed a comprehensive assessment approach that used maximum wind speed (MWS), extreme precipitation, and environmental factors to evaluate the rate of affected population in different return periods in Hainan Province. These methods merely classify the relative risk or risk severity categories, however, and they cannot reflect the potential damage to all the elements at risk. Moreover, indicators will change with regional development and environment changes, which inevitably brings challenges for rapid damage prediction and risk assessment for TC disasters.

The Intergovernmental Panel on Climate Change (IPCC) established a risk assessment approach that consists of hazard intensity, exposure, and vulnerability (IPCC 2014). The approach considers the interaction of hazard factors and elements at risk, and can effectively quantify the relationships between the TC disaster loss, hazard intensity, and environmental variables. It has been widely used in various disaster risk assessments, such as flood, drought, heatwave, and typhoon disasters (Mosquera-Machado and Dilley 2009; Nguyen et al. 2019). For instance, Zhang et al. (2021) mapped the risk of storm surges by using hazard, exposure, vulnerability, and mitigation capacity in Guangdong Province. Zhang et al. (2017) determined the risk level of TC disasters in the Yangtze River Delta.

Wind speed is the most critical hazard factor contributing to TC wind disasters. There are mainly two approaches to recover the wind speed data—numerical simulation and parametric models. Numerical models are usually used for weather and climatic forecast and regional wind field simulation (Chen et al. 2011). Although numerical models may offer better wind fields, they are not suitable for rapid simulation of wind fields due to modeling complexity and the large computational effort required. More importantly, numerical models could not capture the local MWS that was widely used in damage and loss estimate (Nellipudi et al. 2021). Parametric models, such as the Holland model (Holland 2008), have been widely applied to reconstruct wind fields from point observations due to their efficiency, simplicity, and low computation cost. The Jelesnianski model (Jelesnianski 1965), and other improved models, have been used to better reproduce the local MWS (Vickery and Wadhera 2008; Yang et al. 2021). These parametric models have been widely used in storm surge analysis (David and Cruz 2021), wind wave simulation (Ruiz-Salcines et al. 2019), and wind hazard simulation (Tan and Fang 2018). It remains rare to encounter in many previous studies the application of wind field models that determine MWS for damage prediction or risk assessment of possible TC disasters. The parametric wind field model can quickly convert the point wind speed data into a spreading grid using the key parameters of TC records, so that the potential damage caused by the wind could be rapidly estimated for TC emergency responses (Fang et al. 2020).

The recovered MWS data could be used together with historic damage records to establish a vulnerability curve of elements at risk (Kita 2017). The form of vulnerability depends on the study targets and specific conditions. The components of vulnerability can be an indicator of exposure, sensitivity, and adaptive capacity (Cutter et al. 2003). Exposure involves various aspects of society and natural environment, such as economic resources, humans, buildings, and infrastructure. The population and material assets of a TC-stricken area are the most direct parts contributing to exposure. Age, knowledge, and experiences also affect the vulnerability of a population to nature hazards. Detailed information on the spatial distribution and internal components of exposed units are crucial in determining the vulnerability curve and establishing a rapid damage prediction for TC disasters at a local scale.

The spatiotemporal analysis of wind disaster risks helps to reduce the impacts of TC disasters. Various studies have investigated the spatiotemporal variations of TC wind disaster risk at scales from regional and national to global (Hoque et al. 2019; Sajjad et al. 2019), which were much larger than those at which the disaster loss occurs. Wang et al. (2021) investigated the spatiotemporal distribution of typhoon disaster frequency and direct economic loss for 11 coastal provinces in China at the provincial scale. In large regions like Guangdong and Zhejiang Provinces, which have borne the largest losses, population, economy, and adaptive capacities can show significantly different patterns in space. Those TCs with similar intensity (for example, wind speed) can cause distinctly different disaster losses along different transit paths (Wang et al. 2016). Moreover, the MWS declines exponentially away from the TC centers, and the point wind speed data used in previous studies only offered limited information on the hazard intensity in areas away from the TC tracks. To better estimate the damage of TC wind disasters, fine grid assessment is in great demand, especially for emergency response and risk assessment.

The primary objective of this study was to develop a series of vulnerability curves at a county level using the newly established MWS grid and historic records of damage, population, and economy. Then, in a case study, those county-scale vulnerability curves were applied to make rapid damage predictions and disaster risk assessments of TCs. A fine (1-km) grid scale was used to make predictions and assessments of direct economic loss and affected population in Guangdong Province, South China. The MWS grids were generated from the TC track records by a parametric wind field model, and were used to predict the wind damage and to compute the extreme wind speeds for disaster risk assessment in various return periods.

2 Study Area and Data

This section describes the primary characteristics of the study area and those data used in this study, such as TC records, population and economy, as well as the disaster losses caused by TCs.

2.1 Study Area

Guangdong Province is located in the southern coast of China and the Western North Pacific (WNP) Ocean at 20–26°N and 109–118°E and has a tropical and subtropical monsoon climate (Fig. 1). The province consists of 21 prefecture-level cities and 122 counties (or districts). Its total population and gross domestic product (GDP) in 2020 were 126.24 million people and RMB 11.07 trillion yuan, among which 73.38 million people (58%) and 8.52 trillion yuan (77%) were concentrated in the Guangdong-Hong Kong-Macao Greater Bay Area (GBA) on the Pearl River Delta (PRD) in the central part of Guangdong Province (GPBS 2021). Guangdong Province is perennially exposed to the hazards of TCs that follow different paths: during 1949–2020 TCs followed predominantly westward (60.5%), northwestward (20.7%), and northward (18.8%) trajectories. The mean number of TCs that annually make landfall was 3.19 from 1980 to 2020, leading to direct economic losses of RMB 6.08 billion yuan and affecting 3.63 million people each year (COIN 2020; GPDNR 2020). Economic losses caused by TCs accounted for 31.17% of the annual total of natural hazard-related disaster losses (GPDNR 2020).

Fig. 1
figure 1

Land use and 21 prefecture-level cities in Guangdong Province, China and the tracks of 8 tropical cyclone (TC) events used in model validation and demonstration—Abe (9315), Hagupit (0814), Usagi (1319), Rammasun (1409), Kalmaegi (1415), Haima (1622), Merbok (1702) and Hato (1713). The number in the parentheses is the ID of the TC in the China Meteorological Administration dataset and the first two numbers represent the year of TC genesis. The two capital letters represent the 21 prefecture-level cities and two regions from the south to the north—ZJ: Zhanjiang, MM: Maoming, YJ: Yangjiang, YF: Yunfu, JM: Jiangmen, ZH: Zhuhai, MO: Macao, HK: Hong Kong, SZ: Shenzhen, ZS: Zhongshan, DG: Dongguan, FS: Foshan, GZ: Guangzhou, ZQ: Zhaoqing, QY: Qingyuan, SG: Shaoguan, HZ: Huizhou, SW: Shanwei, HY: Heyuan, JY: Jieyang, ST: Shantou, CZ: Chaozhou, and MZ: Meizhou. GBA = Guangdong-Hong Kong-Macao Greater Bay Area.

2.2 Data

Those datasets used in this study are summarized in Table 1. The best track datasets of TCs were downloaded from the Shanghai Typhoon Institute of China Meteorological Administration, including TC center’s position (latitude and longitude), 2-min maximum sustained wind speed, and sea-level pressure. Tropical cyclones are classified as tropical depression (TD: < 17.2 m/s), tropical storm (TS: 17.2–24.4 m/s), severe tropical storm (STS: 24.5–32.6 m/s), typhoon (TY: 32.7–41.4 m/s), strong typhoon (STY: 41.5–50.9 m/s), and super typhoon (SUTY: > 50.9 m/s) according to the TC definition of CMA (2012). A 300-km buffer zone was created to select the TCs that could affect Guangdong Province, and a total of 592 TCs during 1949–2020 was chosen.

Table 1 Data description and sources for tropical cyclone damage prediction and risk assessment in Guangdong Province, South China

The daily MWS data observed at meteorological stations were downloaded from the China Meteorological Data Network (CMDN 2021) and were used to validate those simulated by the parametric wind field model. The total population and GDP in each county were derived from the Guangdong Statistical Yearbook, including population, GDP, economic growth rate, and consumer price index (CPI). The 1-km grid data of population and GDP in 2015 were obtained from the Resource and Environment Science and Data Center (RESD 2017) and these data were used for calculating the losses caused by TCs. The RESD data were based on the county-level population statistics, considering the land use type, the intensity of nighttime light, and the density of residential areas (Huang et al. 2014). In addition, the 1-km grid population data provided by WorldPop (2018) in 2020 and the Guangdong Branch of China Unicom (GBCU) in 2020 were used for comparison and cross validation for the affected population by TCs. The mapping approach of WorldPop is the random forests-based asymmetric redistribution (Lloyd et al. 2017). The original population data from GBCU were discrete point data derived from the mobile phone base stations of China Unicom and were converted into 1-km grid data (Ding et al. 2022). Our analysis showed that the GBCU data have the best agreement with the county total population and with the spatial pattern of the built-up areas.

The TC disaster loss data include 1,350 records of affected counties in 21 prefecture-level cities from 1984 to 2017, and consist of disaster type, affected counties, date of disaster start and end, casualties, affected population, direct economic loss, affected area, collapsed houses, among other information. These TC disaster loss data were collected from the National Climate Center of China. Only direct economic loss and the affected population were used in this study, and both were carefully checked for outliers by comparing the MWS and the reported values in neighboring counties especially for the affected population.

3 Methodology

The methodology includes two parts. The Holland wind field model was used to compute a grid of MWS for each TC transit. The IPCC disaster risk assessment framework was adopted for damage prediction and disaster risk assessment. The disaster loss curves of affected population and GDP loss were fitted based on the annual MWS and historic disaster loss rates, and were implemented for damage prediction and disaster risk assessment.

3.1 Holland Wind Field Model

The central MWS of TCs were recorded along the TC tracks every 6 h. Those point data were directly used for disaster risk assessment of TCs in most studies, but large uncertainties exist since the MWS declines exponentially away from the TC centers. In order to reduce the uncertainty of wind damage prediction and risk assessment, a parametric wind field model was used to convert the point data of MWS along the TC tracks into grid data for better estimating the wind damage and disaster risk in this study. A 1-km grid was constructed for the parametric wind field model over Guangdong Province. The Holland gradient wind model for rapid calculation of the radial wind speed of TCs is as Eq. 1 (Holland 2008):

$${V}_{\rm r}=V\sqrt{{(R/r)}^{B}{e}^{1-{(R/r)}^{\rm B}}}$$

where Vr is the wind speed at the radial distance r from a TC center, V is the MWS near the TC center, R is the radius of maximum wind, B is a shape parameter.

Since the R values of each TCs were not provided in the best track datasets, the statistical formula was used to calculate R in Eq. 2 (Vickery and Wadhera 2008):

$$R={e}^{3.015-6.291\times {10}^{-5}\Delta {P}^{2}+0.0337\Psi }$$

where ΔP is the difference between the peripheral pressure and the central pressure, \(\Psi\) is the central latitude.

A 1-km grid of wind speed covering Guangdong Province was computed at every 6-h record for a TC, and a grid of the MWS for each TC transit was then created from all grid layers of wind speed data for subsequent damage analysis. A grid of the MWS for each year was then created to compute the extreme wind speed in a return period.

3.2 Damage Prediction and Disaster Risk Assessment

The disaster risk assessment framework adopted in the IPCC assessment report of climate change is as Eq. 3 (IPCC 2014):

$$\text{Risk}=\text{Hazard}\times \text{Exposure}\times \text{Vulnerability}$$

For the wind disaster risk assessment, the hazard factor is the MWS, the exposure is population and GDP, and the vulnerability is the disaster loss rate for the direct economic loss and the affected population.

According to the TC disaster data, the disaster losses were mainly caused by tropical storms and above and only a few records were associated with tropical depressions. Disaster losses caused by MWS below 17.2 m/s are assumed negligible in this study. The disaster loss curves were fitted based on the MWS over 17.2 m/s by an exponential function in Eq. 4:


where Y is the affected population rate or direct economy loss rate, x is the MWS, a and b are parameters to be fitted, and δ is the fitting residual.

A unique set of a and b was fitted in each county based on the historic damage data (affected population and direct economic losses), the county’s total population and GDP in the year of the disaster occurrence, and the MWS of TCs at the county seat. One set of a and b represents a loss curve for the affected population rate and another for the direct economic loss rate in a county. If the count of the damage reports in a county was smaller than seven, those damage reports were added to its neighboring counties to fit the loss curve. All curves were fitted when the count of damage data is larger than seven in a county, and then used for damage prediction and disaster risk assessment.

For damage prediction, the disaster loss curves were used to estimate the affected population and direct economic losses based on the MWS and total population and GDP at the 1-km grid for the forecast or landfall TCs in each county. The MWS was used to match the loss rate from the loss curves, and the damage was calculated by multiplying the loss rate with the total population or GDP exposure at the grid. Such a simple model is of great significance in disaster emergency response to the incoming TCs due to its easy implementation and rapid computation.

The disaster risk assessment consists of three steps—computing the extreme wind speeds (EWS) in a return period, estimating the loss rate from the loss curves with EWS, and calculating the potential loss (risk) by multiplying the loss rate with total population or GDP exposure.

The extreme wind speeds in six return periods of 10, 20, 50, 100, 200, and 500 years were calculated from the MWS in each year from 1949 to 2020 at the 1-km grid by using the Gumbel distribution in Eqs. 5 and 6 (Liu et al. 2019):

$$f\left({x};\mu ,\sigma \right)={e}^{-\frac{1}{\sigma }\left(x-\mu \right)-{e}^{-\frac{1}{\sigma }(x-\mu )}}$$
$$F({x};\mu ,\sigma )={e}^{-{e}^{-\frac{1}{\sigma }(x-\mu )}}$$

where x is the simulated MWS at the 1-km grid, \(\mu ,\sigma\) are undetermined parameters, \(f\left({x};\mu ,\sigma \right)\) is the Gumbel density function, and \(F({x};\mu ,\sigma )\) is the Gumbel distribution function. According to the Gumbel distribution, the extreme wind speed of a return period is calculated in Eq. 7:

$${W}_{\rm T}=\mu -\sigma \ln(-\ln(1-\frac{1}{T}))$$

where T is the return period, WT is the extreme wind speed in the return period T.

4 Results

The results include four parts: Validation of the simulated MWS along a TC transit; fitting and validation of the TC disaster loss curves of affected population and GDP in each county; applying the loss curves for TC damage prediction and disaster risk assessment at the 1-km grid scale.

4.1 Validation of the Simulated Wind Speed

The MWS observed at the weather stations were used to validate those simulated by the Holland model along the TC track, such as the six TCs illustrated in Fig. 2; their transit paths were marked in Fig. 1. The simulated MWS figures were comparable with the observed ones with R2 and the mean absolute error (MAE) ranging from 0.71 to 0.93 (P < 0.01 in all cases), and 3.3 to 7.5 m/s, respectively. This indicates that the MWS simulated by the parametric Holland model can accurately reflect the MWS during a TC transit. In contrast, when using the raw record of MWS at the nearest point, the R2 and MAE ranged from 0.31 to 0.54, and 10.4 to 37.1 m/s. Therefore, it was necessary to convert the raw record of the point wind speed along the TC track to a set of fine grid data for better predicting the expected damage and losses.

Fig. 2
figure 2

Comparison of the maximum wind speed (MWS) between ground observations at weather stations and the grid simulations by the Holland model of tropical cyclones in Guangdong Province, South China. MAE = mean absolute error

4.2 Tropical Cyclone Disaster Loss Curves

This section includes two parts—fitting the loss curves and validating the loss rates for the affected population and GDP losses.

4.2.1 Fitting the Loss Curves

Our study adopted the disaster loss curve as the vulnerability function with which to quantify TC damage and losses and disaster risk. A pair of loss curves for the affected population rates and the direct economic loss rates were fitted respectively for each county once the count of TC damage reports was equal to or larger than seven. Among the 122 counties, only 41 pairs of loss curves were fitted due to the limited damage reports in the inland counties far away from the coast, such as Guangning County in Zhaoqing City, Yangshan County in Qingyuan City, and Lechang County in Shaoguan City, among others. On the whole, these disaster loss curves had a good fit with the disaster loss rate, which showed that the disaster loss rate in coastal counties was higher than that in inland counties, and the disaster loss rate in coastal counties had a faster growth rate with the increase of wind speed. For brevity, only nine loss curves in the coastal areas are illustrated from south to north, that is, Xuwen, Lianjiang, Wuchuan, Yangdong, Taishan, Zhuhai, Guangzhou, Haifeng, and Jieyang (Figs. 3 and 4). The affected population rates varied from 3% to 28% with MWS ranging from 17.2 to 48.5 m/s in these nine districts and were fitted well with the exponential function (R2 = 0.69–0.87, P < 0.01) (Fig. 3).

Fig. 3
figure 3

Loss rates for the affected population fitted with maximum wind speed (MWS) by an exponential function in nine coastal districts from south to north in Guangdong Province

Fig. 4
figure 4

Direct economic loss rates in GDP fitted with maximum wind speed (MWS) by an exponential function in nine coastal districts from south to north in Guangdong Province

The direct economic loss rates varied from 0.3% to 3.0% in the nine districts and were also fitted well with the exponential function (R2 = 0.65–0.84, P < 0.01) (Fig. 4). The city of Guangzhou located at the northwestern edge of the Pearl River estuary had much smaller MWS and GDP loss rates than other districts especially in the southern part of Xuwen, Wuchuan, and so on. The loss curves at other counties that are located away from the coastline showed much lower MWS and smaller loss rates as well. The disaster loss curves showed a large spatial heterogeneity in different counties, which further confirms the significance of fine scale analysis of the vulnerability and damage estimate by using the Holland wind field model. These disaster loss curves provided the integrated vulnerability of the exposure units for the subsequent damage prediction and assessment of disaster risks caused by TCs.

4.2.2 Validating the Loss Rates

The fitted disaster loss curves were validated by the disaster loss rates of TCs Haima (1622), Kalmaegi (1415), and Hato (1713), whose disaster loss rates were set aside for validation instead of fitting. The three TCs represent the three primary paths of TCs through western, central, and eastern Guangdong Province, respectively. The loss rates of affected population and direct economic loss estimated by the loss curves demonstrate a good linear relationship (R2 = 0.83–0.91 for population, R2 = 0.81–0.93 for GDP loss, P < 0.01) with the reported loss rates for the three TCs (Fig. 5). The MAEs of the estimated direct economic loss rates (0.17%–0.34%) were smaller by one order of magnitude than those of affected population (1.18%–3.43%). This implies that the estimated loss rates by the fitted loss curves could provide a fast estimate of vulnerability for the TC affected counties.

Fig. 5
figure 5

Scatter plots of the simulated loss rates and reported ones for affected population and direct economic loss in each affected county for the three tropical cyclones. MAE = mean absolute error.

4.3 Damage Prediction

One of the primary applications for the TC disaster loss curves was to rapidly predict the damage for emergency response once the TC is formed and its moving track is forecasted. The three TCs of Kalmaegi (1415), Hato (1713), and Haima (1622) were selected to demonstrate the spatial distribution of the simulated MWS, affected population, and direct economic loss (Fig. 6). Kalmaegi, Hato, and Haima passed through and mainly affected the southwest, the central Pearl River Delta (PRD), and the northeast of Guangdong Province, respectively. The estimated values less than 30 persons/km2 for the affected population and 15 × 104 RMB/km2 for GDP loss were not shown and considered as the minimum disaster loss per unit area (uncertainty) for damage prediction and risk assessment by the wind disaster loss curves and the simulated MWS.

Fig. 6
figure 6

Spatiotemporal distribution of maximum wind speed (MWS), affected population, and direct GDP loss simulated for tropical cyclones Kalmaegi (1415), Hato (1713), and Haima (1622)

Typhoon Kalmaegi (1415) was formed in the Western Pacific Ocean on 8 September 2014, moved westward and first landed in Wenchang County in Hainan Province at 9:40 a.m. on 16 September and 3 h later made landfall at Xuwen County in the southwest of Guangdong Province (Fig. 6a). It had a MWS of 40 m/s and a minimum air pressure of 960 hPa at landfall. The MWS was mainly concentrated near the track and quickly weakened outwards. The simulation results show that about 1.16 million people were affected and the direct economic loss was RMB 3.62 billion yuan in the affected counties. In reality, the reported affected population and direct economic loss were 1.24 million people and RMB 3.12 billion yuan in the affected counties, respectively. The predicted results of affected population and direct economic loss were 6.8% lower and 16.7% higher than the reported losses (Table 2).

Table 2 Comparison of disaster losses between reports and simulations in the affected counties for tropical cyclone (TC) Kalmaegi (1415), Hato (1713), and Haima (1622)

Super typhoon Hato (1713) was formed in the Western Pacific Ocean on 20 August 2017, moved westward and struck land at Zhuhai in Guangdong Province at 12:50 on 23 August (Fig. 6d). Its MWS increased to 52 m/s just within two days after formation, and declined to 45 m/s and an air pressure of 950 hPa at landfall. It moved westward and directly struck the central PRD, causing huge damage. The predicted affected population and direct economic loss were 1.30 million people and RMB 26.74 billion yuan in the affected areas, while the reported damages were 1.11 million people and RMB 29.76 billion yuan. Most of the predicted damages were concentrated in the PRD, where a high density of population and high value of GDP are concentrated, particularly in Shenzhen, Guangzhou, and Zhuhai. The predicted result of affected population was 16.9% higher than the reported one, and the estimated direct economic loss was 10.1% lower than the reported value.

Super typhoon Haima (1622) was formed in the Western Pacific Ocean on 15 October 2016, and reached landfall at Haifeng in Shanwei in the eastern coast of Guangdong Province at 12:40 on 21 October (Fig. 6g). At contact with land, a MWS 42 m/s and an air pressure of 960 hPa was recorded. The predicted affected population and direct economic loss were 0.86 million people and RMB 3.86 billion yuan in the affected regions, whereas the reported damage and losses were 0.99 million people and RMB 3.38 billion yuan. The predicted results of affected population and direct economic loss were 13.2% lower and 13.9% higher than the reported losses, respectively. Most of the predicted damages were concentrated in the central PRD, notably in Shenzhen, Guangzhou, and Zhuhai. The loss difference between the predicted values and the reported ones was much smaller in eastern Guangdong Province, for example in Shanwei and Shantou.

Typhoons Hato and Haima had similar MWS at landfall, but the former caused much greater damage than the latter because of their different landfall locations, moving paths, and spatial distribution of exposure, such as the high population density / property value areas in Guangzhou and Shenzhen (Fig. 6, Table 2).

4.4 Disaster Risk Assessment

The disaster loss curves were used to assess the TC disaster risk according to the grid data of population, GDP, and the simulated extreme wind speed in six return periods of 10, 20, 50, 100, 200, and 500 years. The extreme wind speeds showed a decrease pattern from the coastal zone to the inland area, and from the southwestern coast to the northeastern coast in Guangdong Province (Fig. 7). In the coastal zone, the extreme wind speed increases from 35–45 to 55–65 m/s as the return period increases from 10 to 200 years. The extreme wind speed in a return period of 500 years reaches 65–75 m/s in Zhanjiang, the southwestern corner of Guangdong Province.

Fig. 7
figure 7

Spatial distribution of the extreme wind speed in six return periods of 10, 20, 50, 100, 200, and 500 years in Guangdong Province, South China

Figures 8 and 9 demonstrate the spatial distribution of the integrated TC disaster risk in six return periods on affected population and direct economic loss, which show quite similar spatial patterns and are mainly determined by the spatial distributions of population and GDP. The GBA (Guangdong-Hong Kong-Macao Greater Bay Area), especially such places as Shenzhen, Dongguan, and Zhuhai in the central Pearl River Delta, is exposed to the highest risk. In the inner delta, such as around Guangzhou and Foshan, although the extreme wind speeds are about 10 m/s lower than the coastal zone, these cities are also exposed to high TC risk. The second high-risk areas are in the northeast, in the vicinity of Shantou and Chaozhou. Another high-risk area is in the southwest, especially near Zhanjiang, Maoming, and Yangjiang. Zhanjiang has the highest extreme wind speed and high loss rates, while its disaster risk is lower than the central PRD due to its much smaller density of population and GBA. Other inland cities have relatively low risk mainly due to the fast decline of wind speed once TCs make landfalls.

Fig. 8
figure 8

Spatiotemporal distribution of affected population by tropical cyclones in six return periods in Guangdong Province, South China

Fig. 9
figure 9

Spatiotemporal distribution of direct economic (GDP) loss caused by tropical cyclones in six return periods in Guangdong Province, South China

5 Discussions

This section discusses three themes on hazard, vulnerability, and damage, that is, the MWS from point records and grid simulation, vulnerability and disaster loss curves, and the impacts of different population data.

5.1 Simulation of Grid Maximum Wind Speed

In most previous studies, the point wind speed data recorded every 6 h along TC tracks were directly used for TC disaster damage analysis and risk assessment at the state or regional scale (Sajjad et al. 2019). In order to confirm the feasibility of point wind speed along TC tracks, the MWS recorded from the neighboring TC track points were validated with those observed at the province’s weather stations. These readings showed less comparability between such point records of MWS along a TC track and the observed MWS values at the weather station in a neighboring location, with R2 and MAE ranging from 0.31 to 0.54, and 10.4 to 37.1 m/s. In contrast, the simulated 1-km grid MWS were in better agreement (R2 = 0.71 to 0.93, and MAE = 3.3 to 7.5 m/s) with ground observations at the same locations (Fig. 2). This indicates that the grid MWS simulated by the parametric Holland model can better reflect the MWS than those directly adopted from the raw record of the nearest track points, and can provide more reliable results for damage and loss prediction than those using the raw point records and collected at the regional scale (Sajjad et al. 2019; Song et al. 2020).

Generally, the simulated grid MWS were higher than those observed at the weather stations especially for MWS less than 20 m/s (Fig. 2). The higher estimates were primarily caused by the simplified parametric wind field model, which considers the terrain as a smooth plain and ignores the impact of topography and roughness on wind speed (Tan and Fang 2018). Meanwhile, the observed wind speeds at some meteorological stations in urban areas were impacted by tall buildings, which also contributed to the deviation for the simulated wind speeds.

Although the physically-based atmospheric models might provide better predictions of TC wind field (speed and direction), they are not suitable for damage estimate and near real-time disaster warning due to modeling complexity, heavy computation load, and especially underestimate to the MWS (Nellipudi et al. 2021). Alternatively, available data products also can provide wind speed at different spatiotemporal resolutions, including the reanalysis data (for example, NCEP,Footnote 2 ERA,Footnote 3 MERRAFootnote 4) and climate model data (for example, CMIP6Footnote 5). To investigate whether the wind speed datasets could reflect the MWS during TC, we used the wind speed from NCEP during Haima (1622) and Hato (1713) to compare with the observed wind speed at weather stations. The results showed large errors and poor correlation of wind speed between the NCEP and the observed one, indicating that the wind speed of NCEP may not reflect the MWS during TCs. Numerical models need more physical conditions and processes to capture the local MWS (Nellipudi et al. 2021). This might explain why few studies directly used the MWS simulated by atmospheric models for damage prediction or disaster risk assessment, and emphasizes the significance of simulating the grid MWS by parametrical wind field models.

5.2 Vulnerability and Disaster Loss Curves

Compared with the qualitative disaster assessment based on indices (Peduzzi et al. 2009; Cardona et al. 2010; Lu et al. 2018), our method could rapidly and quantitatively predict the disaster losses. Our disaster loss curves were developed at a county scale based on the average loss rates and the grid MWS at the county seat. To explore the feasibility of disaster loss curves for damage prediction, we estimated disaster losses at seven spatial scales of 1, 5, 10, 20, 30, 40, and 50 km. The results of this exploration indicate that disaster loss curves were generally feasible for the damage estimation and can best reveal the spatial distribution of disaster losses at the 1-km grid in Guangdong Province. The disaster loss curves also can be used to reconstruct TC losses for counties lack of damage records.

Compared with the comprehensive assessment approaches that combine various environmental factors (Meng et al. 2021), our method only takes the MWS as a convenient hazard factor for rapid damage prediction. The damage data were integrated statistics that included all losses caused by extreme wind, heavy rainfall, and even storm surges during a TC. We tried to separate the impact of the storm surges from the integrated statistics by dividing the study areas into two zones with elevations above 5 m and below 5 m (where storm surges may impact), but the difference was not evident. Therefore, we only took the MWS of TCs as hazard factor, while recognizing that the intensity of rainfall and storm surges are also closely related to MWS (Shi et al. 2020). The constructed TC disaster loss curves represent the overall vulnerability of the exposure units to MWS, and indirectly to the intensity of heavy rainfall and even storm surges that larger scale TCs can bring. Based on the recorded losses from storm surges and TC damage in Guangdong Province from 1995 to 2018, the average annual affected population and GDP loss caused by storm surges accounted for 63.09% and 32.83% of the TC damage and losses, respectively. In the low-lying coastal areas, storm surges might dominate the direct economic loss during a TC disaster (Wang et al. 2021) and need a separate risk analysis. Thus it is helpful to separate storm surge damages from the overall TC damages and to derive a better estimate for the wind disaster damage and risk assessment.

With the advancement of science and technology, local governments have more ability and better measures for TC wind disaster prevention and mitigation, such as warning systems, raising public awareness of disasters, disaster prevention training, and other disaster mitigation activities. Meanwhile, the capacity of disaster prevention and mitigation in China’s coastal cities is constrained by the economic development level of those cities, since cities with higher economic development levels have more ability to reduce disaster risks and losses (Huang et al. 2020). According to the marine disaster report of Guangdong Province, local governments carried out various measures on disaster prevention before the TCs landed, such as evacuating people and moving property in the potential affected area in advance. These precautions reduced TC disaster loss and partially explained the overestimate of affected people by our model in typhoon 1713-Hato (Fig. 6 and Table 2). Although our method performed well on rapid damage estimation, the influence of mitigation capacity on TC disaster loss rates and risk assessment should be considered in future studies.

5.3 Influence of Grid data on the Affected Population Estimates

One of the objectives of this study was to predict TC damage at the 1-km fine grid scale, which required high-quality grid data of the elements at risk, such as population and GDP distribution. Taking TC Hato (1713) as an example, this study compared the affected population using the same disaster loss curve and three 1-km population datasets from RESD in 2015, and WorldPop and GBCU in 2020. Overall, the spatial patterns of the affected population calculated by the three population datasets were similar, and mainly distributed in the Pearl River Delta and the northeast coastal areas in Shantou (Fig. 10). The total affected population calculated with population data from RESD, WorldPop, and GBCU was 2.88, 2.52, and 3.19 million people in the affected districts, respectively. Meanwhile, the reported number was 1.11 million for affected people in the corresponding impacted regions. The distribution of this impacted population from RESD was wider and continuous in space and omitted those in most county seats and towns, while those from WorldPop and GBCU were highly concentrated near the city centers, such as in Guangzhou and Shenzhen, and also scattered around the county seats. In particular, the affected population calculated with the GBCU data was consistent with the spatial distribution of urban areas. Although the affected people calculated with the GBCU data had the largest overestimate, we believe that the GBCU population estimate is still more accurate than the other two datasets since it was calculated with users’ mobile signal data at base stations. The higher estimation for the impacted people was likely attributed to the feasibility of the TC damage and loss curves, which were derived with the average loss rate in a county but used for damage prediction at the 1-km grid scale. The much higher density of population in the city center for the 1-km grid data than the county average resulted in overestimate of the total affected population. From this point of view, the better resolution (quality) population data of GBCU may perform worse than RESD in predicting the total affected population with the fitted damage and loss curves.

Fig. 10
figure 10

Spatial distribution of population affected by typhoon Hato (1713) calculated from three population grid datasets. Plots a, b, and c are the affected population calculated by RESD, WorldPop, and GBCU, respectively.

6 Conclusion

This study developed an approach for rapid damage prediction and disaster risk assessment at a fine 1-km grid for TCs and tested the method in Guangdong Province. This method has advantages in providing rapid and efficient calculation of hazard and vulnerability for TC-related wind disaster risk assessment with satisfactory modeling accuracy and reliability. The core components of the approach are the integrated TC disaster loss curves developed for the affected population and the direct GDP loss in each district. The disaster loss curves have a good performance in rapid TC damage prediction and risk assessment, but tend to overestimate the damage and loss especially for the affected population. This can be partially attributed to the recent improvement in TC forecasting and emergency management for disaster prevention that were not accounted for in developing the disaster loss curves based on historical loss data. The Holland parametrical wind field model could efficiently simulate grid MWS from the point wind speed data. These simulated grid MWS are comparable with the ground observations and this makes it feasible to quantitatively predict the damage and loss at the community level. The coastal areas and the Guangdong-Hong Kong-Macao Greater Bay Area part of Guangdong Province have the greatest risk of wind disasters mainly due to their high density of population and developed economy. The presented results can be used for policy design, disaster mitigation and preparedness, and public education of TC wind disaster in high risk regions in Guangdong Province. Our method has already been used to support TC warning and emergency response in Guangdong Province and also can be used in other provinces in China.