Optimization of CMIP6 Precipitation Projection Based on Bayesian Model Averaging Approach and Future Urban Precipitation Risk Assessment: A Case Study of Shanghai

Abstract

Urban flooding, intensified by climate change, poses significant threats to sustainable development, necessitating accurate precipitation projections for effective risk management. This study utilized Bayesian Model Averaging (BMA) to optimize CMIP6 multi-model ensemble precipitation projections for Shanghai, integrating Delta statistical downscaling with observational data to enhance spatial accuracy and reduce uncertainty. After downscaling, RMSE values of daily precipitation for individual models range from 10.158 to 12.512, with correlation coefficients between −0.009 and 0.0047. The BMA exhibits an RMSE of 8.105 and a correlation coefficient of 0.056, demonstrating better accuracy compared to individual models. The BMA-weighted projections, coupled with Soil Conservation Service Curve Number (SCS-CN) hydrological model and drainage capacity constraints, reveal spatiotemporal flood risk patterns under Shared Socioeconomic Pathway (SSP) 245 and SSP585 scenarios. Key findings indicate that while SSP245 shows stable extreme precipitation intensity, SSP585 drives substantial increases—particularly for 50-year and 100-year return periods, with late 21st century maximums rising by 24.9% and 32.6%, respectively, compared to mid-century. Spatially, flood risk concentrates in peripheral districts due to higher precipitation exposure and average drainage capacity, contrasting with the lower-risk central urban core. This study establishes a watershed-based risk assessment framework linking climate projections directly to urban drainage planning, proposing differentiated strategies: green infrastructure for runoff reduction in high-risk areas, drainage system integration for vulnerable suburbs, and ecological restoration for coastal zones. This integrated methodology provides a replicable approach for climate-resilient urban flood management, demonstrating that effective adaptation requires scenario-specific spatial targeting.

1. Introduction

Urban flood disasters are one of the most significant challenges faced by current urban development, causing substantial negative impacts on the socioeconomic fabric and urban services, such as disruptions to transportation, communication, water supply, and power systems, and damages to the urban infrastructure [1,2]. Such damages pose a serious threat to the safety and development of cities. Urbanization-induced changes in the surface cover and morphology are the main internal causes of urban flooding [3,4], while climate change is the main external factor [5,6]. The increasingly severe impact of climate change increases the urban flood risk faced by cities and is an important issue to be addressed by current and future urban flood management strategies. Climate change significantly affects the urban water cycle [7,8], the probability of precipitation events, and the spatiotemporal pattern of rainfall, thereby directly leading to changes in the frequency and volume of surface runoff and floods. The Intergovernmental Panel on Climate Change’s Sixth Assessment Report (IPCC AR6) indicates that the Earth’s surface temperature has risen by 1.1 °C since 1850–1900, and global temperatures are projected increase by 1.5 °C or more over the next 20 years, heightening the risk of climate hazards [9]. Therefore, analyzing the future rainfall risks in cities is of great significance for the construction of sustainable cities and the improvement of residents’ quality of life.

Evaluating urban flood risk under a changing climate and devising adaptive planning relies on reliable characterizations of present rainfall regimes and credible projections of future precipitation. The spatial resolution and fidelity of such projections directly condition flood safety thresholds and management options [10,11]. Global Climate Models (GCMs), advanced within the Coupled Model Intercomparison Project (CMIP), serve as the primary instruments for reconstructing past climates and exploring future trajectories [12,13]. Yet their coarse grids necessitate downscaling to resolve hydrologically relevant scales [14,15]. Among available techniques, statistical approaches are less computationally demanding than dynamical schemes; therefore, they are widely adopted [16,17]. Single-site statistical downscaling, however, yields only point-specific or loosely connected scenarios; therefore, multi-site frameworks are increasingly employed to capture the spatial heterogeneity of precipitation change and its hydrological consequences [18,19,20].

To date, six rounds of model comparison projects have been completed, with the seventh phase, Coupled Model Intercomparison Project Phase 7 (CMIP7), currently underway, and estimated to be completed in 2028. Compared to CMIP6, CMIP7 offers several improvements and advantages. CMIP7 enhances model resolution, providing more detailed simulations of regional climate processes. This increased resolution can better capture observed trends and extreme events, such as tropical cyclones and droughts. [21]. Most research on the impact of climate change on regional hydrological conditions and rainfall patterns is based on the fifth phase, Coupled Model Intercomparison Project Phase 5 (CMIP5) [12]. On a larger regional scale, Rao et al. [22] used 32 climate models from CMIP5 to study the spatiotemporal distribution of precipitation in northern China, showing that most models performed adequately in spatial distribution but overestimated precipitation levels in northern China. Ta et al. [23] assessed the accuracy of precipitation simulation in the Central Asian region by 37 GCMs from CMIP5, finding that six models could accurately simulate precipitation conditions in the study area. At the watershed scale, researchers have used CMIP5 data and models to evaluate the impact of climate change on runoff processes in the Hinkson Creek Watershed in the United States and the runoff response in the Murrumbidgee River Basin in Australia [24,25]. At the urban scale, Zahmatkesh et al. [26] used rainfall predictions from CMIP5 to assess the effectiveness of Low Impact Development (LID) implementation in mitigating urban flooding in New York City. Zhu et al. [27] evaluated the adaptability of urban flood control in Shanghai under the Representative Concentration Pathway (RCP) 4.5 and RCP8.5 scenarios of CMIP5, simulating urban flood scenarios in the central urban area of Shanghai under extreme rainfall conditions and analyzing the impact of climate change on Shanghai’s flood control.

More recently, many studies indicated that compared to CMIP5, CMIP6 has made significant strides in enhancing the accuracy of precipitation simulation for the Chinese region [28,29]. CMIP6 has shown marked improvement in simulating extreme rainfall, largely due to the reduction of dry biases, thereby better capturing extreme climate events [30]. Additionally, CMIP6 has advanced the simulation of the spatial distribution and temporal changes of precipitation compared to CMIP5 [31,32].

Current CMIP6-based precipitation studies predominantly focus on regional and watershed scales. Rivera and Arnould [33] utilized CMIP6 to simulate the spatiotemporal distribution of precipitation in southwestern South America, whereas Wang et al. [13] and Yue et al. [34] simulated future rainfall in the Han and Yangtze basins, respectively. Although these analyses confirm the improved accuracy of CMIP6 at watershed scales, notable deficiencies persist. Li et al. [35], evaluating 18 CMIP6 models over the Yangtze River Basin, observed a systematic overestimation of the projected upward trend; conversely, Patel et al. [36] documented instances of both positive and negative biases in Indian precipitation projections.

Utilizing Global Climate Model (GCM) data to forecast future climate constitutes a principal approach in climate change-related research. However, individual GCMs lack universality and are not suitable for direct application in regional climate simulation and prediction studies. Consequently, researchers have developed numerous hybrid modeling methods and model optimization algorithms to enhance the accuracy of predictions for extreme temperature and precipitation, such as the CEEMD-BiLSTM model [37], the CPSO-LSTM model [38], the orthogonal opposition-based-learning Yin–Yang-pair optimization algorithm [39], and the Arctic Puffin Optimization algorithm [40], among others. These models have improved prediction stability and the R² of prediction results while reducing the RMSE or MSE. In terms of precipitation prediction accuracy based on CMIP6, Bayesian Model Averaging (BMA) significantly enhances prediction accuracy and reduces uncertainty [41,42]. Therefore, this study uses Bayesian methods to process CMIP6 data and further predict future precipitation.

These studies provide guidance and reference value for formulating urban flood management strategies and planning corresponding responses. Although lots of research utilizes CMIP5 and CMIP6 data, studies that apply CMIP5 or CMIP6 data at the urban scale remain limited. At this scale, the majority of research directly employs regional rainfall predictions from CMIP5 or CMIP6, often in conjunction with hydrological models for analysis. This approach typically omits the downscaling necessary to capture the spatial variability of precipitation across different urban areas. As a result, the spatiotemporal distribution pattern of rainfall is inaccurately represented, failing to fully leverage the enhanced precipitation simulation accuracy offered by CMIP6.

Therefore, by leveraging CMIP6 multi-model data processed through Bayesian Model Averaging, we can investigate urban-scale flood risks under climate change. Furthermore, we propose a method for integrating model data with observations from urban meteorological stations to forecast urban precipitation volumes. This integrated approach enables more accurate hydrological model simulations of urban flood risk patterns. Based on the simulation results, corresponding planning strategies can be formulated, which not only help to construct a higher urban flood safety framework considering climate change but also provide guidance for strengthening water management and enhancing urban resilience during China’s urbanization process.

2. Materials and Methods

2.1. Study Area

Shanghai, situated in eastern China, lies on the western coast of the Pacific Ocean. Shanghai occupies the low-lying alluvial plain of the Yangtze River Delta (Figure 1). Mean elevation is 2.19 m, peaking at 103.7 m on Dajinshan Island. A humid subtropical monsoon regime delivers 1173.4 mm/year of rainfall, half to three-fifths of which falls between June and August. Eleven national weather stations (National Meteorological Observing Stations) are operational—Fengxian, Songjiang, Qingpu, Jinshan, Pudong, Nanhui, Xujiahui, Chongming, Baoshan, Minhang, and Jiading—though the recently commissioned Jiading site lacks historical records. Consequently, the analysis draws on the remaining ten stations. For the purposes of this study, the territory excluding Chongming and smaller outlying islands is designated the Shanghai mainland, while the contiguous districts of Yangpu, Hongkou, Zhabei, Putuo, Changning, Jing’an, and Xuhui are collectively regarded as the central urban core.

2.2. Data Source

The data for this study encompasses meteorological, geographical, and CMIP6 global model data. Daily precipitation from June to August (1965–2015) at 0.1 mm resolution for ten Shanghai stations were obtained from the Chinese Academy of Sciences’ Resource and Environmental Science Data Center (RESDC; https://www.resdc.cn/). Also, 2020 land use layers at 30 m resolution were likewise sourced from RESDC. Topography is represented by a 12.5 m NASA DEM from the Earthdata Open Sharing Platform (https://search.asf.alaska.edu/), while soil classes originate from the 1:4 million scale Chinese Soil Map (2000) provided by the Nanjing Institute of Soil Science. Historical and future (SSP245/SSP585) CMIP6 simulations were retrieved via ESGF MetaGrid (https://aims2.llnl.gov/search) and bilinearly interpolated to the ten stations [43]; specifics are given in

Table 1. Global model information.

No.	Model	Institution	Spatial Resolution
1	ACCESS-ESM1-5	CSIRO-ARCCSS	1.875° × 1.25°
2	BCC-CSM2-MR	Beijing Climate Center	1.12° × 1.12°
3	CanESM5	The Canadian Centre for Climate Modelling and Analysis	2.81° × 2.77°
4	CMCC-ESM2	CMCC	1.12° × 1.12°
5	CNRM-CM6-1	CNRM	1.406° × 1.389°
6	CNRM-ESM2-1	CNRM	1.406° × 1.389°
7	INM-CM4-8	Russian Institute for Numerical Mathematics Climate Model	2° × 1.5°
8	INM-CM5-0	Russian Institute for Numerical Mathematics Climate Model	2° × 1.5°
9	IPSL-CM6A-LR	IPSL (Institute Pierre-Simon Laplace)	2.5° × 1.27°
10	MIROC6	MRI (Meteorological Research Institute)	1.4° × 1.4°
11	MRI-ESM2-0	MRI (Meteorological Research Institute)	1.4° × 1.4°
12	CESM2-WACCM	NCAR (National Center for Atmospheric Research)	1.25° × 0.9°

.

2.3. Method

2.3.1. Statistical Downscaling Methods

Global model data have a relatively coarse spatial resolution, which necessitates the application of downscaling methods for regional climate prediction. Downscaling techniques are generally grouped into statistical, dynamical, and hybrid (statistical–dynamical) approaches. Because of its modest computational burden, statistical downscaling is the most frequently adopted, with probability bias correction, quantile mapping, and the Delta technique dominating the literature. Among these options, the Delta technique excels at eliminating systematic discrepancies between global model outputs and regional climates while retaining the large-scale variability embedded in the original simulations through land-surface and circulation parameterizations [44,45]. In this study, we employ the Delta method, which calibrates global model projections by differencing historical modelled and observed values; the corresponding formulations are given in Equations (1) and (2).

$$\begin{array}{c}{delta}_{ymon{i}_p}={\displaystyle \frac{{obs}_{ymon{i}_p}}{{gcm}_{ymon{i}_p}}}\end{array}$$

(1)

$$\begin{array}{c}{gcm}_{daily{i}_p}={gcm}_{daily{i}_p}\times {delta}_{ymon{i}_p}\end{array}$$

(2)

where ${obs}_{ymoni\_p}$ represents the multi-year monthly average of the observed data rate for the meteorological station, ${gcm}_{ymoni\_p}$ denotes the multi-year monthly average of the data rate from the global model’s historical data, ${delta}_{ymoni\_p}$ is the proportionality factor for the rate period, ${gcm}_{dailyi\_p}$ is the future daily precipitation predicted by the global model, and ${gcm}_{dailyi\_p}$ is the future daily precipitation after downscaling by the global model.

2.3.2. Downscaling Effectiveness Evaluation Methods

The Delta downscaling method can effectively reduce the systematic bias between global models and regional climate, while preserving the variability characteristics of global models under multiple scenarios. Based on subsequent research methodologies and station data, this study selects the downscaling period from 1965 to 2000 and the evaluation period for the downscaling effectiveness from 2000 to 2015. This study employs the correlation coefficient and Root Mean Square Error (RMSE) as the evaluation metric for the downscaling effectiveness, with the calculation method provided in Equations (3) and (4).

$$\begin{array}{c}{\rho }_{xy}={\displaystyle \frac{N\sum XY-\sum X\sum Y}{\sqrt{N\sum X^2-{\left(\sum X\right)}^2\sqrt{N\sum Y^2-{\left(\sum Y\right)}^2}}}}\end{array}$$

(3)

$$\begin{array}{c}RMSE=\sqrt{{\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N}{\left(Y_i-f\left(x_i\right)\right)}^2}\end{array}$$

(4)

where ${\rho }_{xy}$ represents the correlation coefficient, $X$ and $Y$ denote the value of $X$-variable and $Y$-variable, respectively, $RMSE$ stands for Root Mean Square Error, $N$ represents the sample size, $Y_i$ is the observed values, and $f\left(x_i\right)$ is the predicted values.

2.3.3. Determination of Model Weights

Projecting future climate with GCMs remains the dominant approach in climate change research; however, the associated uncertainty originates from the GCMs themselves. In this study, we used Bayesian Model Averaging (BMA), a statistical method for dealing with model uncertainty, as a principled framework for multi-model integration. BMA quantifies predictive uncertainty by weighting all GCM projections according to their posterior model probabilities [46]. BMA derives each model’s posterior probability from the data itself, then aggregates the individual projections through a posterior-weighted mean. By avoiding exclusive reliance on any single model, this strategy mitigates selection bias and improves predictive fidelity. To streamline implementation, model weights are approximated by the inverse of their respective Mean Square Error (MSE). The resulting BMA formulation is expressed in Equation (5).

$$\begin{array}{c}BMA={\sum }_i^n{w_{i}Model}_i\end{array}$$

(5)

where ${Model}_i$ represents the predicted value of the i-th model, BMA is the predicted value after weighted averaging, and $w_i$ is the weight of the i-th model, calculated as $w_i={\displaystyle \frac{1/{MSE}_i}{{\sum }_j^n{MSE}_j}}$.

The model weights for daily maximum temperature and precipitation are shown in

Table 2. Model weights for daily maximum temperature and precipitation.

Model	Weight for Daily Maximum Temperature	Weight for Precipitation
ACCESS-ESM1-5	0.082	0.080
BCC-CSM2-MR	0.078	0.074
CanESM5	0.080	0.067
CMCC-ESM2	0.077	0.077
CNRM-CM6-1	0.074	0.076
CNRM-ESM2-1	0.075	0.075
INM-CM4-8	0.085	0.080
INM-CM5-0	0.086	0.083
IPSL-CM6A-LR	0.068	0.073
MIROC6	0.068	0.082
MRI-ESM2-0	0.074	0.074
NorESM2-LM	0.079	0.079
NorESM2-MM	0.074	0.079

.

2.3.4. Determining the Return Period of Extreme Precipitation

The return period of precipitation is determined using the Peaks Over Threshold (POT) model. Existing research has indicated that precipitation in the East China region, once exceeding a certain threshold, conforms to the Generalized Pareto Distribution (GPD), with the threshold often selected as the 90–94th percentile. Therefore, this study adopts the 92nd percentile of precipitation from June to August for each station in Shanghai as the threshold. The Generalized Pareto Distribution is fitted to the excess precipitation over the threshold to obtain the precipitation amounts with return periods of 20, 50, and 100 years for each meteorological station. The probability density function of the Generalized Pareto Distribution [47] is presented in Equation (6).

$$y\hspace{0.17em}\hspace{0.17em}=\hspace{0.17em}f\left(x∣k,\sigma ,\theta \right)=\left({\displaystyle \frac{1}{\sigma }}\right){(1+k{\displaystyle \frac{(x-\theta )}{\sigma }})}^{-1-{\displaystyle \frac{1}{k}}}$$

(6)

where $k$ denotes the shape parameter, $\sigma$ represents the scale parameter, and $\theta$ signifies the threshold parameter.

2.3.5. Hydrological Model

Current research indicates that the Soil Conservation Service Curve Number (SCS-CN) model [48] has demonstrated effective simulation of precipitation runoff in Shanghai. The runoff volume is determined based on the Curve Number (CN) values corresponding to different land use types in Shanghai. The computational method is presented in Equation (7).

$$\begin{array}{c}\left\{\begin{array}{c}Q={\displaystyle \frac{{\left(P-0.2S\right)}^2}{P+0.8S}}\left(P\ge 0.2S\right)\\ Q=0\left(P<0.2S\right)\end{array}\right.\end{array}$$

(7)

where $S={\displaystyle \frac{\mathrm{25,400}}{CN}}-254$, P (mm) represents the precipitation amount; Q (mm) denotes the runoff depth; and CN is a dimensionless parameter used to describe the relationship between precipitation and runoff.

Curve Number (CN) magnitude reflects the interplay among soil texture, antecedent moisture, and land-cover characteristics. Previous assessments indicate that Shanghai’s soil can be treated as moderately moist when the SCS framework is applied. Integrating local soil taxonomies, land use maps, and prior studies, we assigned CN values accordingly.

As a hydrologic index, CN scales inversely with infiltration potential: larger values denote diminished retention and heightened runoff generation. First, the hydrologic soil group is identified from minimum infiltration rates via Equations (8) and (9) (Appendix

Table A1. Hydrologic group rating criteria.

Criteria	A	B	C	D
Final constant infiltration rate (mm/h)	7.6–11.4	3.8–7.6	1.3–3.8	0–1.3

). The CN is then retrieved from the intersection of this group and the land use class (Appendix

Table A2. CN value for different land use types in Shanghai.

Land Use	A	B	C	D
Impervious Surface	85	90	94	98
Cultivated Land	65	74	83	86
Forest	65	74	83	86
Grassland	65	74	83	86
Water	98	98	98	98
Other	78	87	82	95

).

$$\begin{array}{c}X={\left(20Y\right)}^{1.8}\end{array}$$

(8)

$$\begin{array}{c}Y={\displaystyle \frac{Z}{10}}\ast 0.03+0.002\end{array}$$

(9)

$X$ denotes the soil’s minimum infiltration rate (mm h⁻¹), $Y$ signifies the mean particle diameter (mm), and $Z$ represents the sand fraction (%). If $Z$ = 0%, $Y$ is set to 0.01 mm; if $Z$ = 100%, $Y$ equals 0.3 mm; and in the case of 100% clay, $Y$ is assigned 0.002 mm.

2.3.6. Watershed Delineation

Urban flood risk is closely associated with urban topography. This study assumes that precipitation is solely influenced by gravity. Based on the 12.5 m DEM data of Shanghai, the runoff direction in Shanghai is determined. The number of flow accumulation units is derived from the analysis of runoff direction. Considering the current conditions in Shanghai and existing research, a threshold of 30,000 is selected to extract the water-collecting basins in Shanghai. Ultimately, 849 water-collecting basins are identified.

2.3.7. Flood Risk Assessment Indicators

Flood risk assessment indicators are utilized to describe the capacity of an area to withstand precipitation under current conditions. The flood risk in a region is positively correlated with the amount of precipitation in that area and inversely correlated with its drainage capacity. The construction of flood risk assessment indicators is formulated as shown in Equation (10).

$$\begin{array}{c}R={\displaystyle \frac{Q_{extreme}}{D}}\end{array}$$

(10)

where $R$ denotes the flood risk assessment indicator, $Q_{extreme}$ represents the runoff depth under precipitation conditions, and $D$ signifies the drainage capacity of the area.

In this study, data processing is conducted using Python 3, spatial analysis and mapping are based on the ARCGIS 10 platform, and the spatial interpolation is conducted using the inverse distance weighting method [49].

3. Results

3.1. Evaluation of Downscaling Effectiveness

Optimization of the physical parameterization equations for land-surface processes models can enhance the accuracy of global model simulations. Various downscaling methods can reduce, to a certain extent, the systematic bias between models and regional climate [50]. The Delta downscaling calculates the deviation between historical observed data and model historical data, assuming this deviation to be an inherent bias of the model that remains applicable in future prediction periods [51,52]. A higher correlation coefficient and a lower Root Mean Square Error (RMSE) indicate a closer alignment between model and observed data, signifying better downscaling effectiveness.

In the context of climate change, uncertainties in global climate increase. Employing a multi-model ensemble approach can effectively reduce model uncertainties. This study utilizes Bayesian Model Averaging (BMA).

Table 3. Correlation coefficients, RMSE, and BMA for each model.

Models	Daily Precipitation
	Correlation Coefficient	RMSE
ACCESS-ESM1-5	−0.009	10.418
BCC-CSM2-MR	0.005	11.325
CanESM5	0.007	12.512
CMCC-ESM2	0.029	10.873
CNRM-CM6-1	0.014	11.039
CNRM-ESM2-1	0.017	11.135
INM-CM4-8	0.013	10.468
INM-CM5-0	0.029	10.134
IPSL-CM6A-LR	0.017	11.391
MIROC6	0.024	10.158
MRI-ESM2-0	0.019	11.316
NorESM2-LM	0.047	10.564
NorESM2-MM	0.028	10.550
BMA	0.056	8.105

presents the correlation coefficients, RMSE and BMA for each model. After downscaling, the RMSE values of daily precipitation for individual models range from 10.158 to 12.512, with correlation coefficients between −0.009 and 0.0047. The BMA exhibits an RMSE of 8.105 and a correlation coefficient of 0.056, demonstrating lower uncertainty compared to individual models. This study selects the BMA-weighted values as the basis for calculating precipitation values.

3.2. Analysis of the Spatiotemporal Characteristics of Precipitation in Shanghai Under Multiple Scenarios

Currently, there are numerous approaches to climate prediction scenario simulation, with the primary baseline scenarios of CMIP6 being SSP119, SSP245, SSP370, and SSP585. SSP585 represents a socioeconomic development pathway with high anthropogenic radiative forcing, reaching 8.5 W/m² by 2100, while SSP245 represents a socioeconomic development pathway with moderate radiative forcing and vulnerability, characterized by less extreme land use and aerosol pathways. Compared to the other two socioeconomic development pathways, SSP2-45 and SSP5-85 may hold greater representation.

Under the SSP245 medium emission scenario, extreme precipitation in Shanghai is projected to show limited temporal change but increasing spatial disparity throughout the 21st century. By the mid-century, the estimated precipitation amounts for the 20-, 50-, and 100-year return periods range from 30.88 to 68.54 mm, 36.54 to 97.02 mm, and 40.91 to 134.08 mm, respectively. By the end of the century, these ranges increase slightly to 32.81–68.60 mm, 39.75–96.18 mm, and 45.56–133.34 mm. Although there is a minor rise in the upper bounds, the overall change is not statistically significant, indicating that extreme precipitation intensity in Shanghai is likely to remain stable under the SSP245 pathway. Nevertheless, the broader spread of high-end values suggests a potential increase in the frequency of localized intense rainfall events (Figure 2). Spatially, extreme precipitation displays clear regional differentiation. Chongming Island consistently experiences higher precipitation values, while the central urban areas—such as Yangpu, Hongkou, Zhabei, Huangpu, Putuo, and the northwestern part of Pudong New District—tend to have lower values. In contrast, the outer districts including Qingpu, Songjiang, Jinshan, Minhang, and Fengxian show comparatively higher precipitation intensities. By the mid-21st century, as the return period increases, the western parts of Songjiang, Qingpu, and Jinshan are projected to experience more substantial precipitation, exhibiting a distinct west-to-east gradient. Toward the end of the century, two precipitation hotspots emerge: the first in western Songjiang, Qingpu, and Jinshan; the second in the southeastern part of Pudong New District. Across both areas, projected rainfall magnitudes for 20-, 50-, and 100-year events markedly exceed those in the central districts, underscoring a pronounced shift of extreme precipitation risk toward the urban periphery.

Under SSP585 (Figure 3), Shanghai’s projected precipitation ranges from 33.39 to 69.18 mm (20-year return period), 47.97 to 124.15 mm (50-year return period) and 68.69 to 246.85 mm (100-year return period) for the mid-century period. By 2100, the corresponding intervals are 36.14–83.96 mm, 46.02–140.59 mm and 55.08–222.96 mm. Comparing the two horizons, the 20- and 50-year maxima rise by 25% and 17.7%, respectively, whereas the 100-year peak declines by 9.7%. Consequently, the city faces an intensified precipitation threat by the end of the century under high emissions. Spatially, mid-century extremes are highest on Chongming Island, while the southwest mainland exceeds the east. The downtown core records the lowest intensities. By 2100, extremes exhibit a “south-rich, north-poor” pattern, with maxima anchored in southern Fengxian and the southeastern corner of Pudong.

Compared to the SSP585 scenario, under the SSP245 scenario, the terrestrial precipitation in Shanghai exhibits a stepped distribution from east to west, while under the SSP585 scenario, the terrestrial precipitation in Shanghai primarily shows a north–south difference. In terms of precipitation amounts, the differences in precipitation for the 20-year return periods during the mid-20th century are not significant between the two scenarios. However, for the 50-year and 100-year return periods, the precipitation amounts in the SSP585 scenario are 21.5% and 71.5% higher than those in the SSP245 scenario, respectively. At the end of the 21th century, for the difference in precipitation for the 20-year, 50-year, and 100-year return periods, the precipitation amounts in the SSP585 scenario are 14.03%, 23.28%, 40.38%, and 63.31% higher than those in the SSP245 scenario, respectively.

3.3. Multi-Scenario Urban Flood Risk Assessment in Shanghai

Due to the varying catchment areas and differences in surface cover and permeability, the runoff generated from the same amount of precipitation varies across regions. Therefore, precipitation levels alone cannot reflect the flood risk of an area. Runoff, which occurs when rainfall exceeds the area’s capacity to absorb water, is an important parameter for measuring flood risk. The SCS-CN model, one of the most widely used models, requires the determination of parameters such as the underlying surface’s Curve Number (CN) value and precipitation levels. By running the SCS-CN model on the generated topographic model of Shanghai, the runoff depth for each water-collecting basin can be obtained. Figure 4 and Figure 5 show the runoff depth distribution under different return periods in shanghai under the SSP245 scenario and SSP585 scenario.

For the mid-21st century simulation, under the 20-year return period of precipitation, the runoff depths in Shanghai under the SSP245 and SSP585 scenarios are 3.11–59.94 mm and 3.68–59.78 mm, respectively. The spatial distribution of runoff exhibits similar characteristics, with high values occurring in Songjiang District, and the Jing’an, Huangpu, and Changning districts. Under the 50-year return period of precipitation, the runoff depths in Shanghai under the SSP245 and SSP585 scenarios are 5.74–84.70 mm and 10.10–100.67 mm, respectively. The spatial distribution of runoff under the 50-year return period remains similar across both scenarios, with high values in Songjiang District and the Jing’an, Huangpu, and Changning districts. This is primarily due to the relationship between runoff depth and impervious area [53]. In this study, it is assumed that the impervious area in Shanghai remains constant across all scenarios; thus, precipitation is the main factor affecting runoff depth. Under the 100-year return period of precipitation, the runoff depths in Shanghai under the SSP245 and SSP585 scenarios are 8.49–117.79 mm and 18.78–200.63 mm, respectively. In terms of spatial distribution of runoff, under the SSP245 scenario, high values are observed in Songjiang and Qingpu districts, and the Jing’an, Huangpu, and Changning districts, while under the SSP585 scenario, high values are noted in Songjiang District, and the Jing’an, Huangpu, and Changning districts.

By the end of the 21st century, under the 20-year return period of precipitation, the runoff depths in Shanghai under the SSP245 and SSP585 scenarios are 3.67–60.51 mm and 5.04–72.23 mm, respectively. Spatially, high values of runoff are observed in Songjiang District, and the Jing’an, Huangpu, and Changning districts. Under the 50-year return period of precipitation, the runoff depths in Shanghai under the SSP245 and SSP585 scenarios are 6.91–85.16 mm and 10.25–113.43 mm, respectively. The high-value areas of runoff are similarly located in Songjiang, Jing’an, Huangpu, and Changning districts. Under the 100-year return period of precipitation, the runoff depths in Shanghai under the SSP245 and SSP585 scenarios are 8.49–117.79 mm and 16.40–174.92 mm, respectively. The spatial distribution characteristics of runoff are consistent with those observed in the mid-21st century under the 100-year precipitation return period, with high values in the SSP245 scenario occurring in Songjiang, Qingpu, Jing’an, Huangpu, and Changning districts. In the SSP585 scenario, high-value areas are found in Songjiang District, Jinshan District, Jing’an, Huangpu, and Changning districts.

Under the SSP585 scenario, no significant change was observed for the 20-year return periods of precipitation compared to the SSP245 scenario. However, for the 50-year and 100-year return periods, the precipitation levels under the SSP585 scenario are higher than those under SSP245. By the mid-21st century, the maximum values for the 50-year and 100-year return periods increase by 15.86% and 41.29%, respectively, and by the end of the 21st century, they increase by 24.92% and 32.66%, respectively. Spatially, the central urban areas of Shanghai exhibit higher runoff depths under both scenarios. Notably, under the 100-year return period of precipitation, the Chongming Island area shows high runoff depths, indicating a challenge for flood risk in this region.

Flood risk is not only related to the runoff generated by precipitation but also to the area’s drainage capacity [54]. When drainage capacity exceeds runoff, the flood risk in an area is relatively low. Conversely, when drainage capacity is less than runoff, the area faces a higher flood risk. Therefore, in flood risk assessment, the area’s drainage capacity is an essential factor affecting the flood risk [55,56]. According to the Shanghai Flood Control and Drainage Plan (2025–2035), flood control and drainage are divided into 15 districts. The drainage capacity of Shanghai’s districts is derived from relevant planning documents (Shanghai Urban Rainwater Drainage Plan (2020–2035). https://swj.sh.gov.cn/ghjhua/20211009/ae9ce5cd33384864b345c75a68e655d4.html, accessed on 20 March 2025.). The plan sets the flood control standard for the entire city at a 20-year return period, with 50-year and 100-year return periods being manageable. Thus, this study selects the 50-year and 100-year precipitation standards and sets the regional drainage capacity based on planning and literature values. By applying the regional runoff depth and district drainage capacity values to Equation (6), the flood risk assessment indicator R is obtained. Using the natural breaks method, R values are divided into five segments, corresponding to low risk, lower risk, moderate risk, higher risk, and high risk.

Under the SSP245 scenario, facing the 50-year return period of extreme precipitation, the flood risk distribution in Shanghai at the mid-century and end-century is similar, with most areas in the mainland part of Shanghai at low to moderate risk, while higher- and high-risk areas are mostly found in the Chongming District and the outskirts of the mainland (Figure 6). The central and western parts of the Chongming District are at moderate to high risk, and the far western part of the Qingpu District presents a high flood risk area. Notably, there is a higher-risk area located in the Pudong area at the junction with Puxi. When facing the 100-year return period of extreme precipitation, the moderate to high-risk areas in Shanghai increase significantly by the mid-20th century, with most of the Chongming District being at moderate to high risk and large areas of moderate to high risk appearing in Pudong New Area, Fengxian District, Jinshan District, and Qingpu District. By the end of the 21st century, the moderate to high-risk areas are primarily located in the Chongming District, Pudong New Area, and the western part of Qingpu District.

Under the SSP585 scenario, facing the 50-year return period of extreme precipitation, the mid-21st century sees moderate to high-risk areas mainly distributed in the Chongming District, the western part of Qingpu District, most of Songjiang District, and the western part of Jinshan District, as well as the eastern part of Pudong New Area (Figure 7). By the end of the 21st century, the moderate to high-risk areas are primarily located in Pudong New Area, Fengxian District, Jinshan District, and the western part of Qingpu District. When facing the 100-year return period of extreme precipitation, the mid-21st century shows higher- and high-risk areas in the Chongming District, Jinshan District, and Songjiang District, with moderate risk mainly in Pudong New Area and Fengxian District. By the end of the 21st century, there is little change in the moderate to high-risk areas, with higher- and high-risk areas appearing in the southeastern part of Pudong New Area.

Overall, under the condition that all districts can achieve the theoretical drainage capacity, the central urban area of Shanghai is in a low-risk state. Compared to the central urban area, other districts have a higher risk level. To meet the planning standards for controllable urban flooding at 50-year and 100-year return periods, the focus in the central urban area should be on ensuring smooth drainage, while other surrounding areas need to further increase their drainage or water retention capacity.

4. Discussion

4.1. Analysis of Downscaling Effectiveness Based on Bayesian Approach

In the field of climate research, enhancing the accuracy of Global Climate Models in precipitation prediction is of crucial importance for accurately assessing the impacts of climate change. This study focuses on optimizing precipitation prediction methods by introducing the Bayesian Model Averaging (BMA) technique, which has significantly improved the accuracy of precipitation prediction in the CMIP6 models. Traditional single models often exhibit considerable uncertainty in precipitation prediction. In this study, the Root Mean Square Error (RMSE) values of daily precipitation after downscaling for each individual model range from 10.158 to 12.512, with correlation coefficients between −0.009 and 0.0047. This indicates that there is a certain deviation between the simulation results of single models and the observed data, and the degree of agreement needs to be improved. In contrast, after adopting the BMA method, the RMSE decreases to 8.105, and the correlation coefficient increases to 0.056. This result clearly demonstrates the advantage of BMA in reducing model uncertainty and improving prediction accuracy, and this finding is in keeping with the work of Liu et al. [57]. They highlighted that ensemble approaches, such as Bayesian Model Averaging (BMA), exhibit significant advantages in mitigating uncertainty and bolstering the dependability of hydrological forecasts. The reduction of BMA’s RMSE to 8.105 and the increase in its correlation coefficient to 0.056 suggest that BMA effectively integrates the strengths of different models by combining information from multiple models, thereby reducing model uncertainty and bringing the prediction results closer to the observed data, with a higher degree of agreement.

4.2. Uncertainty Analysis of the Flood Risk Model

The flood risk model integrates the CMIP6 multi-model ensemble with the SCS-CN model, effectively analyzing future flood risks under climate change scenarios. The precipitation forecast based on CMIP6 can effectively reflect urban extreme weather conditions under specific emission scenarios and is widely applied globally. The flood risk model employs Bayesian weighted averaging, synthesizing predictive data from multiple models, which can effectively avoid uncertainties arising from biases in single model predictions. This is currently one of the important methods to reduce uncertainties in future climate predictions [58].

Another source of uncertainty in this model is the determination of the Curve Number (CN) in the SCS-CN model. The CN is one of the most important parameters when applying the SCS-CN model and is widely used for runoff estimation in ungauged areas. In previous studies, some scholars have established a connection between land use and CN values using empirical and literature search methods [59,60,61], but this approach has significant uncertainties. CN values often vary for the same land use type [62]. For example, a city park with sandy soil will have a different CN value than one with clay soil. Therefore, in this study, CN is determined based on the hydrologic soil group and land use type, which has been shown in later studies to be an important measure to reduce uncertainties in runoff estimation in ungauged areas.

Lastly, the uncertainty in the flood risk model also stems from the prediction of urban drainage capacity. This study uses the drainage capacity from Shanghai’s drainage planning. The model assumes that drainage areas have the same drainage capacity, which does not align with the actual situation. There are areas within the drainage planning region with better drainage conditions and areas where the drainage capacity does not meet planning requirements. Nonetheless, this model is used for future risk prediction, serving urban planning and design. Using planned drainage capacity as the average level of regional drainage for analyzing urban flood risks still has significant reference value and importance.

4.3. Spatial Pattern of Flood Risk and Planning Strategies in Shanghai

The results indicate that, temporally, under the SSP245 scenario, there is no significant change in Shanghai’s flood risk from the mid to the end of the 21st century. Under the SSP585 scenario, the high-risk areas increase significantly by the end of the 21st century. Spatially, under all scenarios, there is a pattern of lower flood risk in the central urban areas and higher risks in the peripheral areas of Shanghai. The distribution of precipitation also shows that high precipitation values do not occur in the central urban areas.

From a watershed perspective, this study proposes planning recommendations for different risk areas, focusing on source control, midstream transportation, and end disposal. The purpose of source control is to reduce surface runoff depth by utilizing green infrastructure to maximize the absorption, retention, and slow release of rainwater by roads, water systems, and green spaces, mainly for risk areas with deeper surface runoff. The purpose of midstream transportation is to accelerate the drainage of rainwater, including integrating the drainage system from a watershed perspective, connecting water networks, and maintaining pumping systems. The purpose of end disposal is to increase the city’s water retention capacity, with main measures including controlling construction areas, clearing rivers and lakes, increasing the capacity of natural water bodies, and setting up floodable areas. Based on the runoff depth and land use conditions of the risk areas in this study, combined with the actual situation and flood characteristics of each district in Shanghai, the planning responses for flood risk in each district are shown in

Table 4. Flood characteristics and planning responses for each district.

District	Flood Characteristics	Planning Responses
Central Urban Area	High runoff depth, strong theoretical drainage capacity, low risk	Increase permeable pavements, ecological roads, etc., to reduce runoff depth, focus on the maintenance and transformation of pumping systems, and strengthen drainage capacity in low-lying areas.
Chongming District	High runoff depth, average drainage capacity, moderate to high risk	Flat terrain and high groundwater levels in Chongming District lead to poor water retention capacity. It is necessary to connect surface water systems and improve the pumping capacity of urban spaces. Control the area of construction land and optimize ecological space, implementing the principle of “adapting to water if suitable” in spaces prone to water disasters.
Qingpu District	Moderate runoff depth, average drainage capacity, low to moderate risk	Plan regional drainage from a watershed perspective, integrate urban blue–green spaces in new city planning, and coordinate planning. The risk of flood and water in the Dianshan Lake area is high, so it is necessary to strengthen the pumping capacity of pumping stations and further complete ecological restoration to increase the capacity for rainwater retention and purification.
Pudong New Area	Moderate runoff depth, average drainage capacity, moderate risk	Ensure pumping capacity in the main urban area, increase the area of permeable pavements to reduce runoff. When constructing new cities and sub-centers, plan holistically from a watershed perspective, set high runoff reduction targets, and restore the ecology of tidal flats.
Baoshan District	Moderate runoff depth, strong drainage capacity, low risk	Strengthen the maintenance and construction of pumping systems, further enhance the drainage capacity of the central urban area and Baoshan main urban area. Add green infrastructure in suitable areas to reduce runoff volume and control runoff pollution.
Jiading District	Moderate runoff depth, strong drainage capacity, low risk	In the central urban area and Jiading main urban area, focus on pumping with supplementary self-draining, strengthen the maintenance and construction of pumping systems, and in other areas, strengthen the construction of green infrastructure to reduce total runoff and control runoff pollution.
Minhang District	Moderate runoff depth, strong drainage capacity, low risk	In the central urban area and Minhang main urban area, focus on pumping with supplementary self-draining, strengthen the maintenance and construction of pumping systems, and in other areas, strengthen the construction of green infrastructure to reduce total runoff and control runoff pollution.
Fengxian District	Average runoff depth, average drainage capacity, low to moderate risk	Coordinate regional drainage planning from a watershed perspective, control sources and intercept pollution, connect water systems within the region, and strengthen ecological restoration. Maintain and further increase the water retention capacity of forest land within the region. Restore rivers ecologically to reduce runoff pollution.
Jinshan District	Moderate runoff, average drainage capacity, moderate to high risk	Plan regional drainage from a watershed perspective, control the scale of construction land, connect water systems within the region to increase drainage capacity. Treat sewage from docks and ports before discharge, and restore the ecology of coastal areas.
Songjiang District	Moderate runoff, average drainage capacity, moderate to high risk	Plan regional drainage from a watershed perspective, control the scale of construction land, implement the principle of “afforesting if suitable,” increase forest coverage. Strengthen the construction of emergency drainage facilities to improve the capacity.

.

The aforementioned planning responses can overall increase Shanghai’s ability to resist flood risks. However, due to the limitations of the water retention capacity of green infrastructure and the speed of drainage in pipes, these planning measures are not effective against short-term heavy rainfall. Planning based on short-term precipitation intensity would lead to increased costs and low efficiency in the use of flood facilities. For such flood risks, a management approach can be taken to establish a sound early warning system for flood risks, and set up meteorological stations, runoff monitoring stations, and manual inspection points in high-risk, high-value, and high-sensitivity areas. At the same time, equip with emergency flood equipment and establish emergency plans for flood risks. In case of a surge in precipitation or runoff depth, verify manually in a timely manner and activate emergency plans promptly if necessary to prevent the expansion of flood risks and reduce flood risk losses.

4.4. Limitations

This study established a coupled CMIP6-SCS-CN model to assess the flood risk in Shanghai, providing a scientific basis for the large cities to address climate change. However, there are several areas that require improvement in future research.

Firstly, in terms of extreme climate prediction, CMIP6 is a widely used climate forecasting tool, but its ability to predict extreme weather is relatively weak. Predicting extreme weather conditions remains a challenge and a focal point for future climate forecasting. Existing studies suggest that methods such as probability correction or quantile fitting can improve the accuracy of extreme weather prediction to some extent compared to the Delta method, and these will be incorporated into the next phase of research.

Secondly, the determination of the CN value depends on soil type and land use, but it does not consider the dynamic changes of urbanization. Future research should introduce scenario-dependent CN value corrections (such as based on the historical impervious surface change rate) to account for urban development impacts, or at least discuss their potential overestimation/underestimation of the results.

Thirdly, the drainage capacity data of the 849 sub-basins and 15 administrative districts are directly matched, ignoring the heterogeneity of the drainage systems within the basins. A more detailed spatial analysis of drainage infrastructure is needed to capture intra-basin variability accurately.

Fourthly, urban stormwater pipelines are a crucial component of urban drainage, which were not utilized in this model, potentially limiting its application at the community scale. Therefore, when assessing flood risks at the community scale, local drainage pipelines should be considered for a more accurate evaluation of local runoff.

Additionally, if observed runoff data are available, calibrating the CN values based on these observations could enhance model accuracy. Given that this method is used for a local study, employing a regional climate model tailored to the specific climatic conditions of Shanghai could provide more nuanced results. Lastly, the study assumes constant land use, which may not reflect real-world urban dynamics. Incorporating projections of land use changes could improve the model’s applicability over time.

5. Conclusions

This study utilized climate data from 12 global models from CMIP6, coupled with historical data from 10 meteorological stations in Shanghai. The Delta downscaling method was selected to predict precipitation for the years 2025–2055 and 2070–2100 under the SSP245 and SSP585 scenarios, with return periods of 5, 10, 20, 50, and 100 years. The SCS-CN model was employed to simulate runoff depths under multiple scenarios, and further combined with the drainage capacity of each region to categorize the flood risk levels in Shanghai under various scenarios. Results indicate that the central urban area of Shanghai experiences less precipitation but has greater runoff depths and strong drainage capacity, resulting in lower flood risks. In contrast, the surrounding urban areas, with average drainage capacity, face higher flood risks. The flood risk at the end of the 21st century is higher than in the middle, especially under high-emission scenarios with return periods of 50 and 100 years, where the area of high-risk regions significantly increases. It should be noted that while the absolute values of precipitation changes (such as the 32.6% increase in 100-year return period precipitation) provide a quantitative indication, the reliability of these absolute values is subject to the limitations of CMIP6 in simulating extreme precipitation events. Therefore, the relative changes in precipitation and flood risk are considered more reliable indicators for planning and decision-making. Based on the risk level categorization and the actual conditions of each district, planning responses for urban flood facilities were proposed from a watershed perspective, with a focus on coordinating flood control, drainage, and stormwater management for different risk level areas. These strategies aim to provide a rational layout and improve the resilience of the city’s infrastructure against flood risks.