Projection of the precipitation-induced landslide risk in China by 2050

Landslides can be influenced by two categories of factors: basic factors (terrain, soil conditions, land cover, and hydrology) and triggering factors (precipitation, glacial melt water, earthquakes, and human activities) (Li et al 2008). This study specifically focuses on precipitation-induced landslides owing to their predictability. Effective landslide forecasts and estimates can be achieved by integrating land surface characteristics with triggering factors. In this study, an enhanced statistical model-based framework (Wang et al 2016) is utilized for landslide prediction and forecasting, as depicted in figure 1.

Zoom In Zoom Out Reset image size

The framework comprises two aspects. One is a landslide susceptibility map, which includes slope, land cover type, elevation, surface soil texture, and river network drainage density. The other is a statistical quantity, which includes actual precipitation intensity and precipitation threshold obtained through empirical statistical methods. The landslide susceptibility map is calculated using a multiple linear regression method, whose calculation formula is as follows:

$$\begin{align}S = \mathop \sum \limits_i^k {w_i} \times {z_i},\end{align} \tag{ 1 }$$

where $z$ represents the sensitivity corresponding to the susceptibility factor, $w$ represents the weight corresponding to the susceptibility factor $z$, ${z_i}$ represents the number of $i$th factor, and $k$ represents the total number of susceptibility factors. The assignments and detailed information on the susceptibility factors are shown in tables A1 and A2, and the basic geographic data is shown in figure A1.

When assessing the triggering effect of precipitation on landslides, both the intensity and duration of precipitation are important. If the average precipitation in a specific area surpasses a certain threshold over a particular period, it indicates the potential occurrence of landslides. Numerous studies have explored the intensity–duration (I–D) threshold formula for precipitation-triggered landslides (Caine 1980, Hong et al 2006, Li et al 2017, Peruccacci et al 2017, Wang et al 2021a). He et al (2020b) developed a suitable precipitation threshold for landslides in China, thereby improving the accuracy of landslide prediction models. The threshold formula proposed is as follows:

$$\begin{equation}E = 0.49 \times {D^{0.47}}{ }\left( {1 \unicode{x2A7D} D \unicode{x2A7D} 44} \right),\end{equation} \tag{ 2 }$$

where $E$ (in millimeters) is the precipitation threshold that triggers landslides for a single precipitation event and $D$ (in hours) is the duration of precipitation event. Note that this formula is the precipitation event duration (E–D) threshold formula, which can be converted to the I–D formula (Peruccacci et al 2017). In this study, the daily precipitation-driven model is used, and the precipitation duration ($D$) in the formula (2) is set to 24. Accordingly, the threshold for precipitation intensity is 2.18 mm h⁻¹ (52.4 mm d⁻¹). Combining the susceptibility value $S$ of the landslides, the final landslide risk can be characterized by the following formula:

$$\begin{equation}{P_{\text{s}}} = \frac{{{R_{\text{d}}}}}{I} \times S,\end{equation} \tag{ 3 }$$

where ${R_{\text{d}}}$ is the actual precipitation intensity (mm h⁻¹) during a single precipitation event, $I$ is the precipitation intensity threshold, which is equal to E divided by the duration of a precipitation event, and ${P_{\text{s}}}$ is the landslide risk index. The larger the value of ${P_{\text{s}}}$, the higher the probability of landslide occurrence. In this study, ${P_{\text{s}}} > 0.5$ indicates a possible occurrence of landslides, while ${P_{\text{s}}} > 0.8$ indicates a very high probability of landslides.

The China Meteorological Forcing Dataset (CMFD) (Yang and He 2019) that demonstrates good observational representativeness is utilized in this study (He et al 2020a, Ge et al 2023). For historical simulations and future precipitation estimations, the hist-1950 and highresSST-future experiments from the HighResMIP of the sixth Coupled Model Intercomparison Project (CMIP6) are adopted. These experiments encompassed models with horizontal resolutions of less than 50 km, as detailed in table 1.

This study aims to analyze the changes in landslide risk under future land cover changes in China. The global land projection dataset for land-use and land-cover (LULC) based on plant functional types with a 1 km resolution under socio-climatic scenarios (Chen et al 2022) is used. This dataset was produced by integrating the top-down land demand constraints afforded by the CMIP6 official dataset and a bottom-up spatial simulations executed by cellular automata. Based on the climate data, the land types in the simulated products were further divided into 20 plant functional types. Moreover, this dataset also released the results of the International Geosphere-Biosphere Program land classification, which well meets the requirements of susceptibility maps in landslide statistical models for input data. The evaluation on LULC has proven that LULC has good accuracy in comparison with historical land use (ESA 2017) and future simulation data (Popp et al 2017). To align with GCMs, this study utilizes land use data from shared socioeconomic pathways (SSPs) SSP5-8.5 scenario.

This study is conducted at grid scale in China. Initially, the landslide susceptibility maps are calculated at a horizontal resolution of 1 km, with the specific data files detailed in table A1. The study consists of the following steps.

Firstly, this study validates the effectiveness of the landslide prediction framework. The period from 2001 to 2015 is taken as the historical period, and the CMFD daily precipitation data is input into the statistical model to simulate the historical distribution of landslides.

Secondly, a bias correction is applied to the HighResMIP model data to mitigate the systematic bias of the GCMs compared with observations (Piani et al 2010, Yang et al 2010). Quantile mapping (QM) has been demonstrated to effectively reduce model bias and improve mean values, interannual variability, and extreme event biases (Tong et al 2017, Han et al 2018). As landslides are often triggered by heavy precipitation, this study employs the QM to correct the biases in the HighResMIP precipitation data using the CMFD historical data. This method computes the cumulative distribution functions for the observed and simulated values separately and constructs a transfer function between them as follows:

$$\begin{align}{x_{{\text{cor}}}} = F\left( {{x_{{\text{mod}}}}} \right),\end{align} \tag{ 4 }$$

where $F$ represents the transfer function, ${x_{{\text{mod}}}}$ is the original value of the model, and ${x_{{\text{cor}}}}$ is the corrected value of the model.

To predict the spatial distribution pattern of future landslide risk in China, two time periods (2021–2035 and 2036–2050) are considered. The land cover under the SSP5-8.5 scenario in 2035 and 2050 represents the future land cover scenario, while other factors remain constant. The bias-corrected precipitation data from each HighResMIP is input into the model, and the output results are compared with the historical base period (2001–2015) to analyze the changes in future landslide risk. Nine sub-regions in China are studied based on climate and geographical features, referencing the Resource and Environmental Science and Data Center (www.resdc.cn/), as depicted in figure 2. This approach enables a better understanding of the spatial heterogeneity of future landslide risk in China.

Zoom In Zoom Out Reset image size

To assess landslide risk at the spatial scale of the susceptibility map, precipitation data was evenly interpolated to a horizontal resolution of 1 km. Additionally, to visualize the spatial patterns of landslide risk clearly, the results are plotted on a 0.25° × 0.25° grid. Each grid represents the average value of landslide risk at a 1 km × 1 km spatial scale.

The effectiveness of the landslide statistical model is validated by overlaying an inventory of 593 landslide points recorded from 2001 to 2015 onto the layer of historical landslide simulation (figure 3). The simulation of historical landslides is driven by the CMFD daily precipitation data using a statistical model. Grid points with ${P_{\text{s}}} > 0.5$ indicate the areas with potential landslide occurrence, and the frequency sum is calculated over multiple years. Analysis of the recorded landslide inventory reveals that the landslides are mainly distributed in the Sichuan Basin and surrounding regions (SBSR), as well as in the Middle-lower Yangtze Plain (MYP), Yunnan-Guizhou Plateau (YGP), and Southern China (SC). Despite potential high false-positive rates due to incomplete landslide inventory and historical monitoring, the model successfully simulates 503 landslides during the selected historical period (figure 3). Consequently, this statistical model can be deemed suitable for simulating the spatial and temporal pattern of landslide risk.

Zoom In Zoom Out Reset image size

This study utilizes the simulation data from three HighResMIP models. To assess the effectiveness of model simulation and the impact of bias correction, the daily precipitation data from 2001 to 2015 before and after bias correction are respectively input into the statistical model. The results for ${P_{\text{s}}} > 0.5$ and ${P_{\text{s}}} > 0.8$ are calculated, and the spatial correlation coefficient (SCC) is used to evaluate the consistency of model simulation with observation data. Figures 4 and 5 display the results for ${P_s} > 0.5$ and ${P_s} > 0.8$, respectively. The original model data from each model captures the concentrated distribution of landslides, with the SCC being above 0.5 for ${P_s} > 0.5$. For ${P_s} > 0.8$, except for the model of the Euro-Mediterranean Center for Climate Change (CMCC), the SCCs of the other two models with observation are around 0.6. However, compared with the observations (figure 3), the two models generally overestimate the landslide frequency. After bias correction, the model simulation results align closely with the observation, with the SCC being above 0.9. Consequently, the study will employ the bias-corrected model simulation to estimate future landslide risk.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Calculating landslide susceptibility ($S$) for future and historical periods enables the analysis of potential changes resulting from land use. As depicted in figure 6, the findings reveal high susceptibility areas in the southeastern Qinghai Tibet Plateau (QTP), the southwestern region surrounding the SBSR, and the Tianshan, Altai, and Kunlun Mountains in NASR. Some areas with high susceptibility are also observed in the MYP and SC. Analyzing susceptibility changes in the future periods reveals both increases and decreases in different regions, but the overall change is not significant. Increased susceptibility is witnessed in the southwestern, central, southern, and eastern parts of China, while decreased susceptibility is found in the northeastern and northern parts of China, the Huang-Huai-Hai plain (HHP), and the southern Himalayan Mountains of the QTP.

Zoom In Zoom Out Reset image size

The bias-corrected HighResMIP model precipitation is used in the statistical model to calculate the ${P_{\text{s}}}$ values for two future periods, 2021–2035 and 2036–2050. Subsequently, the results for ${P_s} > 0.5$ and ${P_s} > 0.8$ are derived. The results are compared with the historical period, as shown in figures 7 and 8. For the possibility of either landslides (${P_s} > 0.5$) or high-risk landslides (${P_s} > 0.8$), all three models show an obvious increasing trend in estimating the landslide risk for China in the future. Specifically, the Flexible Global Ocean-Atmosphere-Land System model (FGOALS) demonstrates the most obvious increase, exhibiting a large increase in the frequency of landslide risk in the surroundings of the SBSR, HHP and Northeast China Plain, as well as in most southeastern areas of MYP and SC (except Hainan Island). The frequency of ${P_s} > 0.5$ generally increases by more than ten times (figures 7(b) and (e)), while the frequency of ${P_s} > 0.8$ also increases by up to eight times in many regions (figures 8(b) and (e)). In contrast, the CMCC and High Resolution Atmospheric Model (HiRAM) exhibit a high degree of similarity in estimating the future landslide risk. The areas where the landslide risk increases are mainly distributed in the central, eastern, and southern parts of China.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

The geographic zoning layer is overlaid with the map of future estimates from the HighResMIP model to analyze the changes in future landslide risk across different regions. The change rate of landslide risk frequency is calculated for each region, focusing on ${P_s} > 0.5$ for the period of 2036–2050 in comparison to the historical period (2001–2015). Furthermore, the results of future landslide risk are computed based on unchanged land use scenarios using the landslide susceptibility values of 2015. These results are illustrated in table 2 and figure 7. Overall, the models indicate an increasing trend in future landslide risk estimations for China in the near future. The model of FGOALS exhibits the highest increase, with the frequency of ${P_s} > 0.5$ rising by 121% and 117% under the changing and unchanged land use scenarios, respectively. The models of CMCC and HiRAM exhibit an increase of approximately 30%. The change in landslide risk due to land cover varies across regions. Elevated landslide susceptibility contributes approximately 5%–10% to the incremental landslide risk in SC, MYP, and YGP, while the land cover changes reduce the landslide risk in HHP, Loess Plateau, and QTP.

The simulation results of different models exhibit some variations, as analyzed in the previous section. The future precipitation characteristics from the three models are analyzed to explain the differences in the model simulation results and the underlying reasons for the future changes in landslide risk. Three annual precipitation indexes calculated from daily data are selected to compare the future changes in precipitation characteristics of each model with those in the historical period. They are the total precipitation of wet days (PRCPTOT), the number of heavy precipitation days (the days with daily precipitation exceeding 20 mm, R20mm), and the frequency of the daily precipitation exceeding the threshold of 52.4 mm (Rth).

Figures 9–11 depict the spatial distribution and changes in the multi-year average values of the three indexes for the historical period (2001–2015) and the future period (2036–2050). For PRCPTOT and R20mm, the CMCC model shows a decrease in the future period in East China, while the FGOALS model exhibits a noticeable decrease in the future period in the entire Southeast China. The HiRAM model shows an increase in PRCPTOT and R20mm in China mainland except for Hainan Island and Taiwan. The HiRAM simulated precipitation patterns correspond to the simulation results of landslide risk. Specifically, the reduced future landslide risk estimated by the CMCC model in certain areas of YGP aligns with the decrease of PRCPTOT and R20mm (figures 7(a), (d), 9(c) and 10(c)). The decrease of HiRAM-estimated future landslide risk in SC especially in Hainan corresponds to a reduction in precipitation (figures 7(c), (f), 9(i), 10(i) and 11(i)). The simulations of Rth by all three models are relatively consistent. In central-eastern China, the frequency of extremely heavy precipitation is increasing, with the FGOALS model exhibiting the highest increase. This explains why the landslide risk estimated by these models is generally increasing in Southeast China.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Combined with the discussions in section 3.3, it can be inferred that much of the general increase in future landslide risk in China is attributed to the changes in future precipitation. The model simulations share commonalities in the increase of extremely heavy precipitation in southeastern China and in the increase of total precipitation in northeastern and northern China. The changes in land use impact landslide susceptibility and indirectly alter landslide risk, which varies across different geographic regions, although this 'contribution' is not as substantial as the changes in precipitation.