Contrasting Population Projections to Induce Divergent Estimates of Landslides Exposure Under Climate Change

Abstract

At first glance, assessing future landslide-exposed population appears to be a straightforward task if landslide hazard estimates, climate change, and population projections are available. However, the intersection of landslide hazard with socioeconomic elements may result in significant variation of estimated landslide exposure due to considerable variations in population projections. This study aims to investigate the effects of different sources of population data on the evaluation of landslide-exposed population in China under four Shared Socioeconomic Pathways (SSPs) scenarios. We utilize multiple global climate models (GCMs) from Coupled Model Intercomparison Project Phase 6 and six high-resolution spatially explicit static and dynamic population data sets to drive available landslide models. The results indicate an overall rise in landslide hazard projections, with an increase in the potential impact area of 0.4%–2.7% and an increase in the landslide frequency of 4.7%–20.1%, depending on the SSPs scenarios and future periods. However, the likely changes in future landslide exposed population, as modeled by incorporating population data from different sources with landslide hazard, yield divergent outcomes depending on the population data source. Thus, some of the projections depict an increase in future landslide exposure, while others show a clear decrease. The nationwide divergence ranged from −64% to +48%. These divergent findings were mainly attributed to differences in population data and a lesser extent to variations in GCMs. The present findings highlight the need to pay closer attention to the dynamic evolution of the elements at risk and the associated data uncertainties.

Introduction

Landslides are widespread geomorphic phenomena around the globe and lead to severe human fatalities every year. It has been reported that between 1998 and 2017, approximately 4.8 million individuals were impacted by landslides globally, resulting in over 18,000 deaths (Mizutori & Guha-Sapir, 2017). Statistics from the United Nations Office for Disaster Risk Reduction (UNDRR) indicate that approximately five million persons worldwide were exposed to rainfall-induced landslides between 2000 and 2019 (UNDRR, 2020). The risks of landslide disasters due to extreme weather conditions persist even in recent years. For instance, the landslides and flash floods triggered by exceptionally heavy rains in Brazil between late May and early June 2022 caused 130 deaths in the states of Pernambuco, Alagoas, and Paraíba (Marengo et al., 2023). China represents one of the countries with the most human casualties associated with slope instabilities worldwide, with 725 reported fatal landslides resulting in 10,340 deaths, approximately 8 fatalities per million population, from 2004 to 2017 (Froude & Petley, 2018). Considering the estimated likely intensification and increased frequency of extreme rainfall events within several Chinese regions under climate change, significant impacts and risks due to landslides are expected in the future (He et al., 2019; Li et al., 2018; Lin et al., 2020, 2022). Hence, more tailored risk mitigation and climate change impact adaptation measures are needed to alleviate the probable adverse effects of slope instabilities.

Landslide risk assessment is based on three main components, namely hazard, exposure, and vulnerability (Corominas et al., 2014; Pereira et al., 2020). More specifically, landslide risk, as defined by United Nations International Strategy for Disaster Reduction (UNISDR), is a function of the severity and frequency of landslide hazard, the amount and value of the element at risk exposed to the hazard, as well as their susceptibility to damage and loss (UNISDR, 2015). A wealth of recently published studies has focused on investigating various aspects of landslide hazard, including landslide susceptibility (Goetz et al., 2015; B. Wang et al., 2022), critical triggering conditions, intensity and magnitude, and slope instability (Segoni et al., 2018; Steger et al., 2023; N. Wang et al., 2021). Moreover, by coupling spatial susceptibility and landslide triggering rainfall conditions, spatially explicit dynamic landslide prediction systems for rainfall-induced landslides have recently been realized (Huang et al., 2022; Monsieurs et al., 2019; Segoni et al., 2015; Stanley et al., 2021). Efforts have been made to integrate satellite precipitation or reanalyze precipitation data with the aforementioned spatial and temporal landslide models for early warning. These results showed that there was a lot of room for improvement of spatiotemporal prediction of landslides (Kirschbaum & Stanley, 2018; Ozturk et al., 2021). A considerable amount of studies used precipitation data extracted from global climate models (GCMs) or regional climate models to feed well-established landslide analysis procedures to explore potential changes in landsliding in the context of climate change (Alvioli et al., 2018; Gariano et al., 2017; Lin et al., 2020).

However, besides the landslide hazard situation, landslide risk also strongly depends on the exposure and vulnerability of socioeconomic systems. Among these, landslide vulnerability is comparatively difficult to qualify in terms of a uniform metric or approach due to its multiple facets (including physical, social, and ecological vulnerability, etc.). As a result, vulnerability has been rarely addressed in landslide risk assessment (H. Y. Luo et al., 2023; Mirdda et al., 2022). The present study will mainly focus on the hazard and exposure components of the landslide risk and will not involve vulnerability. In fact, potential adverse consequences may largely depend on the distribution of population, buildings, property, and transportation infrastructure that may be affected by landslides. Landslide exposure is described by the spatial intersection of the landslide hazard footprint and its temporal occurrence and the location of main elements at risk, that is, the number or value of the elements at risk that may experience potential losses due to slope instability (IPCC, 2021; UNISDR, 2009). Although landslide exposure is conceptually different from risk (since it does not explicitly address consequences, and it does not consider vulnerability), it can still be considered a better proxy to landslide risk than susceptibility or even hazard alone (e.g., Emberson et al., 2020). Its quantification, however, remains a challenge, due to the localized impacts caused by landslide disasters and the lack of high-resolution dynamic socioeconomic data (Gariano & Guzzetti, 2016, 2022). Existing literature proposed a multi-indicator analysis based on the demographic data in administrative units, such as provinces or municipalities, to estimate the exposure of elements at risk to landslide by incorporating the landslide susceptibility or frequency of landslides occurrence within the same areas (Gariano et al., 2017; Martha et al., 2021). Several studies used dasymetric mapping procedures to refine the estimation of the spatial distribution of elements at risk to provide comparatively fine-scaled estimations of landslide exposure (Emberson et al., 2021; Garcia et al., 2016; Pellicani et al., 2014; Santangelo et al., 2021). Efforts have attempted to estimate landslide exposure at continental or global scales (Emberson et al., 2020; Schlögl & Matulla, 2018). In addition, the interaction between risk components has been investigated in some studies (Depicker et al., 2021; Ozturk et al., 2022). Findings indicate that the distribution of population and transport infrastructure, apart from influencing landslide exposure as elements at risks, could interact with changes in landslide hazard (Bozzolan et al., 2020; Brenning et al., 2015). For instance, the unplanned sprawl of urbanization in mountainous areas would not only lead to an increased exposed population and buildings, but also to changes in the morphology, hydrological or material properties of the landslides, and thus changes in the landslide hazard for individual or regional landslides (Dille et al., 2022; Ozturk et al., 2022), due to the associated changes in land use, especially the destruction of vegetations, the cutting of slopes for road construction or housing, and the excavation at the foot of the slopes (Bozzolan et al., 2020; Depicker et al., 2021; Dille et al., 2022; Ozturk et al., 2022). Nevertheless, previous studies on the exposure to landslides relied on a static representation of the elements at risk, dismissing the dynamic evolution of the underlying socioeconomic system. While this simplification can be acceptable when considering stationary processes within short time frames, disregarding changes in the socioeconomic systems may significantly hamper the assessment of potential impacts associated to landslides in the future under climate change.

In recent years, the CMIP6 GCMs have provided several scenarios that combine the Shared Socioeconomic Pathways (SSPs) of the Global SocioEconomic Development Framework with the Representative Concentration Pathways (RCPs) of future climate change (KC & Lutz, 2017; O'Neill et al., 2016). This allows for the dynamics of socioeconomic components to be considered along with the changing hazards. A considerable number of global, high-resolution dynamic population projections based on SSPs have been released (Gao, 2020; Li et al., 2022; Olén & Lehsten, 2022; X. Wang et al., 2022a). Combining this data with future climate projections (e.g., CMIP6) and spatially explicit landslide modeling results for very large areas (e.g., Lin et al., 2022) provides new possibilities for the estimation of future landslide exposure. However, each of these components is known to be affected by uncertainties which renders their meaningful combination anything but trivial. Particularly the population projections available for China (Gao, 2020; Li et al., 2022; Olén & Lehsten, 2022; X. Wang et al., 2022a) depict a partly contrasting development for the future. The potential differences in landslide exposure estimated in this context stand out to be investigated.

Thus, the objectives of this study are to assess the evolution of the landslide-exposed population in China under climate change scenarios using multiple GCM data from the Coupled Model Intercomparison Project Phase 6 (CMIP6) and static and dynamic projection population data from various institutions. A particular focus is set on investigating the impact of different sources of population data on the assessment results and determining the consistency of signals regarding the impact of climate change on the population exposed to landslides in China.

Data and Methods Study Area

China is one of the countries with the highest human casualties due to landslides globally (Figure 1) (Lin & Wang, 2018). During 2004–2017, the global catastrophic landslide database reported 725 fatal landslides that occurred in China, resulting in 10,340 fatalities, which is second highest only to India in terms of frequency and number of fatalities (Froude & Petley, 2018). China covers a wide geographical territory, with significant topographic relief and altitudes generally decreasing from west to east. Mountainous terrain accounts for two thirds of the national territory and hosts approximately one third of the total population (Cui, 2022). Being situated in the eastern Asia-Europe continent, on the western coast of the Pacific Ocean and close to the Indian Ocean to the south-west, the monsoon-driven climate of China is very distinctive (Ding et al., 2019). This climate is dominated by the rainy season in summer, with southerly winds blowing from the ocean to land, and warm and humid air currents from the Pacific Ocean to the southeast and the Indian Ocean to the southwest, resulting in widespread rainfall throughout the country in summer. The volatility of the monsoon climate leads to more frequent catastrophic events such as rainstorms and typhoons, and further triggers landslides and debris flows in vast mountainous areas (Wu et al., 2019). Such extreme weather events and the associated landslide hazards are likely to increase as a result of global climate change (Lin et al., 2022; Marengo et al., 2021). Since China has a large territory, the geomorphic and climatic environments of different areas have prominent spatial heterogeneity. Based on the geomorphic zoning proposed by N. Wang et al. (2020), the national landscape was divided into six main geomorphic regions (Figure 1), including eastern plains (EP), southeastern mountains (SEM), north-central plateaus (NCP), southwestern mountains (SWM), northwestern basins (NWB), and Tibetan Plateau (TP). The geomorphic region represents the combination of different physiographic and climatic environments and anthropogenic characteristics.

Figure 1. Six geomorphic regions and spatial distribution of reported historical fatal landslides in China from 1950 to 2020 (Lin & Wang, 2018) (EP: eastern plains; SEM: southeastern mountains; NCP: north-central plateaus; SWM: southwestern mountains; NWB: northwestern basins; TP: Tibetan Plateau).

Observed and GCM Precipitation Data

Data on observed precipitation was obtained from the CN05.1 gridded observation data set. This daily data set is based on an interpolation of more than 2,400 meteorological stations in China by using an anomaly approach (Wu & Gao,  2013). The data set covers the period between 1961 and 2018 and is represented via a spatial resolution of 0.25° × 0.25°. It has been widely used for precipitation analysis at national and regional scales due to its high-quality representation of the spatial and temporal distribution of climatological mean and extreme precipitation (N. Luo et al., 2022; H. Xu et al., 2022). Within this study, the CN05.1 data set is used to evaluate the simulation performance of the CMIP6 GCMs.

GCMs are essential tools for analyzing the mechanisms of climate system variability and for projecting future changes in the climate system. GCMs have been commonly used in studies of global and regional climate change and associated impact modeling attempts (Eyring et al., 2016). The GCMs data for this study were obtained from the latest CMIP6. As compared to the previous program CMIP5, the CMIP6 includes more members of the GCMs and provides improvements in both, model resolution and physical parameterization scheme. More importantly, the latest CMIP6 considers both socioeconomic scenarios and Greenhouse Gas (GHG) emission scenarios when using numerical models to project global and regional climate change. That would be a development from the RCPs of CMIP5, which only approached GHG emissions, to SSPs scenarios, which combine future climate change and global socioeconomic development frameworks (KC & Lutz, 2017; O'Neill et al., 2016). The four scenarios of SSPs-RCPs (SSP1-2.6, SSP2-4.5, SSP3-7.0, and SSP5-8.5) were selected given that more climate model ensembles and future scenarios can reduce uncertainty in future projections as well as depict multiple possible future pathways. SSP1-2.6 represents the combination of the sustainable development pathway (SSP1) and low radiative emissions scenarios; SSP2-4.5 depicts a combination of medium social vulnerability and medium radiative forcing scenarios; SSP3-7.0 represents a combination of relatively high social vulnerability and relatively high radiative forcing scenarios; and SSP5-8.5 relates to a combination of traditional fossil fuel-dominated pathways and high radiative forcing scenarios. These four combined SSPs-RCPs scenarios presently contain a total of 24 available CMIP6 GCMs. Details of the 24 models are provided in Table S1 in Supporting Information S1. Each of the climate models consists of daily precipitation data for the historical period (1850–2014) and the future scenario (2015–2100). For the SSP scenarios, the historical baseline period was 1995–2014, and the scenarios for the near-future, mid-future, and far-future were 2021–2040, 2041–2060, and 2080–2099. The choice of 2021–2040, 2041–2060, and 2080–2099 to represent the near-term, mid-term, and long-term future periods is based on the recommendations of the IPCC Sixth Assessment Report (IPCC, 2021). These three different future periods could facilitate comparison with the latest 20-year baseline reference period (1995–2014) of IPCC to explore the possible climate change impacts. Such a choice has been adopted in many existing studies of future climate change impact projections as well (Kwiatkowski et al., 2020).

Historical and Future Projections of Grid Population Data

The static population data for the historical periods were obtained from the spatially distributed gridded data set of population in China from the Resource and Environmental Sciences Data Center of the Chinese Academy of Sciences (CAS population). This data set is based on the demographic data of the counties (i.e., county-level region), and the multi-annual population distribution gridded data were produced by spatialization considering the land use types (X. Xu et al., 2018), the night light brightness, the residential density, transportation and topography conditions. The data set has six snapshots (six periods of data in 1995, 2000, 2005, 2010, 2015, and 2019) with a spatial resolution of 1 km. In this study, the latest available snapshot of the year 2019 acted as a reference data set that represents the static population spatial distribution in China (Xu, 2017).

The future projected spatial distribution of the gridded population is obtained from five sets of population projections under the SSP developed by different research groups. In particular, the study builds upon projects from the Centre for Environmental and Climate Research at Lund University in Sweden (Lund population), The NASA Socioeconomic Data and Applications Center (NASA population), the East China Normal University in China (ECNU), the School of Environment at Tsinghua University (Thua_Cai population) and the School of Architecture at Tsinghua University in China (Thua_Long population). These five data sets of future population projections exhibit a spatial resolution of 30″ (i.e., approximately 1 km), with the spatial coverage of Thua_Cai for mainland China and global coverage for the remaining four products (further details are shown in Table 1). Although all five data sets produce population projections within the framework of the SSPs, differences in the underlying population projections, downscaling procedures, and environmental parameters resulted in diverse final outcomes.

Table 1 List of the Population Data Sources Used in This Study

Categories	Names	Data	Main differences
Static population data	CAS grid population data (Xu, 2017)	Historical grid population in 1995, 2000, 2005, 2010, 2015, and 2019. The year 2019 acts as a reference within this study.	Distributing counties census population as spatial grid data considering land use, nighttime light, density of settlements, and so on, as auxiliary information.
Dynamic population data	ECNU grid population data (Li et al., 2022)	Data set of 1-km resolution global population projections under SSPs (SSP1–SSP5) from 2020 to 2100 for every 10 years.	The global-scale random forest population projection model was established by considering environmental auxiliary data, including distance from urban centers, distance from roads, distance from water bodies and slope and future global built-up land (G. Chen et al., 2020) to produce 1-km grid population data.
	Lund grid population data (Olén & Lehsten, 2022)	Annual population data at 1 km resolution consistent with both SSPs and RCPs within the framework of the IPCC from 2010 to 2100.	The WorldPop of 2010 (Tatem, 2017) as the base year determines the ranking of the population weights for each grid cell, and the RCP urban fraction data (Hurtt et al., 2011) determines the extent of the urban and rural masks, whereby the SSP projected population at the national level (Riahi et al., 2017) is allocated to 1-km grid cells.
	NASA grid population data (Gao, 2020)	Spatial distribution of the projected population at 10-year intervals of 1 km from 2010 to 2100 under SSPs (SSP1–SSP5).	Using the gravity-parameterized models (Jones & O’Neill, 2016), the projected population of each SSP country was downscaled to 1/8-degree and then distributed to 1-km grid cells by incorporating human settlement point records and nighttime light data.
	Tsinghua_Long grid population data (X. Wang et al., 2022a)	Global gridded population data set covering 248 countries or areas at 30″ (approximately 1 km) spatial resolution with 5-year intervals from 2020 to 2100.	Random sampling was carried out in each of the eight regions of the world stratified by population density, that is, Europe (EU), Latin America (LA), Middle East & North Africa (MENA), Russia & the Near Abroad (RNA), Sub-Sahara Africa (SSA), United States & Canada (USC), Oceania (OC) and South & East Asia (SEA), and the resulting samples were combined with environmental auxiliary data to generate 1-km grid cells in a random forest population projection model for each of the eight regions.
	Tsinghua_Cai grid population data (Y. Chen et al., 2020a)	Spatially explicit population grids for each year at 30″ resolution under shared socioeconomic pathways from 2010 to 2100.	The urbanization rate and population projection models for each province were built based on the global narratives of SSPs and the census data of each Chinese province, respectively. The urban and rural mask extent was determined based on RCP urban fraction data, whereby the projected population of each province was distributed to 1-km grid cells.

Methods

The methodological workflow of this study is shown in Figure 2 and consists of three main steps: (a) using GCMs to estimate landslide susceptibility and landslide occurrence frequency under climate change; (b) leveraging these landslide hazard indicators and different population information to model population exposure to landslides; and (c) comparing the directional differences in the effects of climate change on estimated exposed population to landslides obtained from different, partly contrasting, population simulations. A detailed description of each part is provided in the following.

Figure 2. Overview of the methodological framework adopted in this work.

Landslide Susceptibility and Landslide Occurrence Frequency Modeling

The landslide hazard indicator assessment used historical simulations of GCMs and future projections data to estimate possible changes in future landslide hazards under climate change scenarios. The landslide hazard indicator in this study includes both the spatial component (i.e., susceptibility) and a temporal component (i.e., occurrence frequency). Before this assessment, the simulation performance of the CMIP6 climate models was evaluated based on observed data (i.e., CN05.1). The applicability of the CMIP6 multi-model ensemble for landslide hazard estimations in China was evaluated by calculating the mean annual precipitation, frequency of exceeding empirical rainfall thresholds for each geomorphological region, for the CN05.1 gridded observation data and the CMIP6 multi-model ensemble, respectively.

The assessment of the spatial susceptibility of the terrain to landslides in China is based on a previous study (Lin et al., 2022). Within this study, the results of landslide susceptibility were directly used and the key processes of analysis are outlined briefly in the following. The model accounts for the factors of mean annual precipitation, lithology, slope, and soil moisture index as influencing factors of landslide spatial predisposition, and additionally introduces geological environment zones and land use as random effect factors to reduce the potentially incompleteness effects of landslide inventory, based on the generalized additive mixed effects method (GAMM) and the historical rainfall-induced landslides database. More details on modeling landslide susceptibility using GAMMs are available in Lin et al. (2021) and Steger et al. (2021). In this modeling context, changes in landslide susceptibility are exclusively stemming from variations in mean annual precipitation patterns, while the remaining variables are considered static in time. For each GCM, the multi-year mean precipitation for the historical baseline period (1995–2014) and different SSPs scenarios for the near-future, mid-future, and far-future were calculated. The outcomes were used to drive the GAMM-based national landslide susceptibility assessment model to estimate the spatial susceptibility of landslides in China for the historical and future periods.

To estimate the temporal landslide component, the frequency of rainfall events exceeding a critical landslide rainfall threshold was used. In this study, the results of landslide occurrence frequency from the previous study (Lin et al., 2022) were employed and the key processes of analysis are outlined briefly in the following. In analogy to previous studies, these calculations were used as proxy for representing historical and future landslide frequency across different geomorphological subdivisions in China (Bezak & Matjaž, 2021; Lin et al., 2020). The rainfall thresholds for triggering landslides in different geomorphological subdivisions in China were constructed by N. Wang et al. (2021) as a power law relationship between cumulative event rainfall and rainfall duration for each subdivision based on a data set of historical hydrogeomorphic processes in China (containing all hydrogeomorphic processes between the defined range of flash floods to debris flows). In contrast to the single rainfall threshold curve used at the large-scale scale in China, the thresholds used within this study account for the variability of geomorphological features and climatic conditions in different regions, which may enable better exploration of the spatial heterogeneity of landslides induced by different geomorphological and rainfall conditions. For each GCM, the algorithm CTRL-T (Calculation of Thresholds for Rainfall-Induced Landslides-Tool), proposed by Melillo et al. (2018) for reconstructing rainfall events, was used to extract and reconstruct rainfall events for historical and future periods for each GCM, and the reconstructed rainfall events were compared with empirical rainfall thresholds in different geomorphological subdivisions to count the number of rainfall events exceeding empirical thresholds for historical and future periods, that is, the annual average landslide frequency.

Landslide Exposed Population Modelling

Landslide exposed population in this study represents the number of people in areas exposed to the hazard of landslide damage and serves as an estimate of how seriously an area might potentially be affected by landslides. The exposed population to landslides can be calculated as follows in Equation 1. It includes three components: landslide susceptibility (LS), landslide frequency (Freq), and population count (P). [Image Omitted. See PDF]where Exposure is the landslide-exposed population of region j; LS_i is the landslide susceptibility of grid cell i. If the grid relates to modeled moderate to very high landslide susceptibility, the areas are considered to contribute to landslide exposure (value 1), otherwise do not (value 0); Freq_i is the average annual landslide frequency of grid cell i in the corresponding period and scenario; P_i is the population of grid cell i in the corresponding period and scenario; n is the total number of grid cells in region j.

In this study, landslide susceptibility was considered a dichotomous variable (i.e., susceptible vs. no-susceptible terrain) used to characterize the spatial extent of the potential impact on the population. To determine potential landslide impact terrain, the outcomes of the national-scale susceptibility model in Lin et al. (2021) were reclassified, and areas classified as having “moderate” to “very high” landslide susceptibility were considered as potential impact areas. Within this study, landslide frequency in a given grid cell refers to the expected average annual frequency of exceedance of the empirical rainfall event-duration threshold defined in the geomorphological area the grid cell belongs to, and is expressed in times/year. The population component is represented by the number of persons in the corresponding period and scenario. In the case of the static population data from CAS, the same population distribution is used for all future periods. For the five sets of dynamic population data, the historical reference period is modeled uniformly using the population of 2020, the year closest to the static population data. The reason for choosing the year 2020 is that the beginning of the population projections is not consistent across the dynamic population data, and therefore it is straightforward to use the 2020 population consistently for the modeling of the historical baseline period for the comparison between the different population data sets. For the near-term (2021–2040), mid-term (2041–2060), and far-term (2080–2099) periods under different SSPs, a 20-year average of the corresponding periods under each SSP was used for future exposure simulations for the year-by-year Thua_Cai data set and Lund data set; the intermediate years corresponding to the near-future, mid-future, and far-future periods under each SSP were used for simulations for the 5-year and 10-year interval Thua_Long, NASA and ECNU data sets, that is, 2030, 2050 and 2090, respectively. That is, the exposed population to landslides results from modeling using different sets of demographic data from different sources are the same in terms of landslide susceptibility and average annual landslide frequency for the historical reference period and for the near-future, mid-future, and far-future under the four SSPs. This means that the final differences in exposed population to landslides obtained from different sources of demographic information are all accounted for by differences in the respective population data. It should be noted that the spatial resolution of the different population data products requires a uniform spatial resolution of 1 km to be consistent with the results of the landslide hazard indicators prior to the calculation of the population exposure.

Comparing the Expected Change in Landslide-Exposed Population Under Climate Change Considering Different Population Data Sources

The expected change in landslide-exposed population (ECLEP) under climate change was estimated using Equation 2. For different sources of population data, the signal of the ECLEP was quantified by estimating the change in their projected landslide exposed population for different SSPs relative to their modeled exposed population for the historical reference period. [Image Omitted. See PDF]where ECLEP refers to the expected change in the landslide-exposed population under climate change and considering the population dynamics; Exposure__fut refers to the projected landslide-exposed population in the future period; and Exposure__his refers to the simulated landslide-exposed population in the historical baseline period.

Results Increasing Projected Change of Landslide Hazards

The simulation performance of the CMIP6 multi-model ensemble for climatic mean and extreme precipitation in China was evaluated using the CN05.1 gridded observations (Table 2). The CMIP6 multi-model ensemble adequately reproduced the mean annual precipitation in China and across the different geomorphological subdivisions. The absolute values of the simulated mean annual precipitation (i.e., 651.4 mm) from the CMIP6 multi-model ensemble are closely aligned with the results from the observations (i.e., 655.9 mm). At the same time, the spatial correlation coefficients R (0.99) are very high and the RMSE (28.64) are relatively low.

Table 2 Evaluation of CMIP6 Global Climate Model for Simulating Mean and Extreme Precipitation During the Historical Baseline Period (1995–2014)

Region	Mean annual precipitation (mm)				Frequency of exceeding the rainfall thresholds (times/a)
	OBS	MME	RMSE	R	OBS	MME	RMSE	R
China	655.9	651.4	28.64	0.99	9.96	9.19	1.62	0.91
EP	652.1	669.3	36.16	0.99	11.12	11.26	1.03	0.95
SEM	1,489.3	1,452.1	58.34	0.98	15.71	12.38	4.75	0.17
NCP	469.2	480.8	29.87	0.98	11.23	10.96	1.10	0.92
NWB	161.2	161.4	9.07	0.99	10.78	10.50	1.44	0.98
SWM	1,134.7	1,112.2	44.69	0.99	11.77	9.49	5.02	0.55
TP	374.3	372.7	18.82	0.99	4.31	4.68	1.29	0.95

Note. OBS indicates the observation results; MME indicates the multi-model ensemble results; RMSE denotes root mean square error; R denotes the spatial correlation coefficient; the acronyms of the six geomorphic regions in the first column including EP: eastern plains; SEM: southeastern mountains; NCP: north-central plateaus; SWM: southwestern mountains; NWB: northwestern basins; TP: Tibetan Plateau, the spatial distribution is shown in Figure 1.

Extreme precipitations that were used to calculate landslide frequencies, were reproduced fairly well by the CMIP6 multi-model ensemble at both, national-scale levels and across the various geomorphological sub-regions. The frequency of precipitation exceeding the landslide thresholds was comparable between the CMIP6 multi-model ensemble and the observations with a relatively small RMSE of 1.62. The spatial correlation coefficients between the multi-model ensemble and the observed data are relatively high for the whole of China (0.91) and most of the geomorphological regions, with the exception of the SEM for the Southeast Region and the SWM for the Southwest Mountains, both of which are higher than 0.9. It should be noted that the CMIP6 multi-model ensemble remained somewhat underestimated for the simulation of extreme precipitation in regions with higher precipitation in southern China, including the southeastern and the southwestern region, and attention needs to be drawn in the interpretation of the results for future projections. In general, the results of the 24 GCMs ensembles were overall in good agreement for the simulation of mean climate regimes and extreme precipitation in China.

The CMIP6 multi-model ensemble simulations and projected precipitation estimates were used to model the likely changes in spatial susceptibility and frequency of future landslides under climate change scenarios with respect to the historical baseline period, as shown in Figure 3. In general, the CMIP6 multi-model ensemble projected an increase in both, the extent of areas potentially affected by landsliding (i.e., moderate to very high landslide susceptibility) and landslide frequency in the future compared to the historical reference period. The estimated increase in landslide impact area and frequency is generally higher in the long-term future and under high emission scenarios. From a national-scale viewpoint, the multi-model ensemble projects an increase of approximately 0.4%–2.7% in the areal extent of potentially impacted terrain and an increase of 4.7%–20.1% in the frequency of landslide occurrence in China under different SSPs scenarios in the future compared to the historical reference period. Landslide susceptibility and occurrence frequency are projected to increase for each geomorphological region (except for the landslide susceptibility in the SEM region) in comparison to the historical period, though there exist spatial differences. For the areas potentially impacted by landslides, the Tibetan Plateau region is projected to have the highest increase relative to the historical baseline period, with the multi-model ensemble projections under the SSP5-8.5 scenario increasing by approximately 4% relative to the historical baseline period, while the remaining geomorphological regions are projected to increase by 1.8%–2.5%. Minor changes to the baseline period are projected for the SEM region under different scenarios. Regarding the frequency of landslides, the multi-model ensemble projections for all geomorphological regions are projected to increase, with the highest increases in the Tibetan Plateau and Northwest Basin regions, representing 25.8% and 27.4% increases under the SSP5-8.5 scenario, respectively, while the increases in the other geomorphological regions are in the range of 10%–20%.

Figure 3. Variations of areas potentially affected by landslides (upper panels) and frequency (lower panels) of exceeding triggered landslide rainfall thresholds for different future periods under SSPs scenarios in China and each geomorphic region with respect to the historical baseline period. Each point shows the median ensemble of 24 global climate models; the acronyms of the six geomorphic regions in the subplot of each column including EP: eastern plains; SEM: southeastern mountains; NCP: north-central plateaus; SWM: southwestern mountains; NWB: northwestern basins; TP: Tibetan Plateau, the spatial distribution is shown in Figure 1.

Landslide-Exposed Population During Historical Baseline Period

Figure 4 presents the spatial distribution of the population of China at the beginning of the reference period (2019 for the CAS data set, 2020 for the other) for the different population data sets, while a descriptive statistic is provided in Table 3. It can be noted how the total population is comparable across population data products in the baseline period, ranging from 1.37 to 1.42 billion. The different products consistently show comparatively denser population in the east and sparser population in the west of the country, although significant differences in the spatial distribution of the population data can be observed across the different sources. In particular, the population data from CAS and NASA were generally presented in more contiguous clusters, that is, more homogeneous in the high-density areas such as contiguous grids of red-range population density values in the southern part of the Eastern Plains region, and a more contiguous grids of yellow-range population density values in the Northwest Basin region. The apparently lowers coefficient of variation for CAS and NASA reflects their more concentrated distribution, that is, lower spatial variability. In contrast, the other four sets of demographic products showed a more sporadic and discrete distribution, that is, higher spatial variability, particularly the noticeably higher coefficient of variation for the Lund and Tsinghua_Cai data sets (Table 3).

Figure 4. Spatial distribution of grid population (reference period) obtained from different research sources (a) CAS population in 2019; (b) ECNU population in 2020 (mean of SSP1, SSP2, SSP3, and SSP5 in 2020, the same as the following); (c) Lund population; (d) NASA population; (e) Tsinghua_Cai population; and (f) Tsinghua_Long population.

Table 3 Statistical Metrics for Each Grid Population Data

[Table omitted. See PDF]

The aforementioned population data from different sources was used to estimate landslide exposed population for the reference period (Figure 5). The observed differences in estimated exposed population for the different regions mainly result from different sources of population data (bar heights in Figure 5). The size of the error bars shown for each region reflects the (generally smaller) differences in population exposure estimates resulting from the use of different GCMs. Figure 5 shows that there were significant differences in exposed population to landslides based on simulations of different sources of population products for the historical baseline period. Furthermore, the differences caused by the different sources of population data are considerably higher than those associated with the different GCMs. At the national scale, estimated annual average landslide exposure levels range from 2,314 to 4,078 million per year, depending on the respective population product sources. The largest exposure is shown by the more clustered distribution of CAS and NASA population data, both exceeding 4,000 million persons per year, while the smallest exposure is modeled by the Lund and Tsinghua_Cai population data with a range of 2,314 million to 2,515 million persons per year. It should be noted that the estimated landslide-exposed population in this study is defined by a multiplication of the hazard (in terms of how often a hazard threshold is exceeded) and the number of elements at risk (i.e., number of people). As the population located in the area with a non-zero hazard may be exposed to multiple events (estimated frequency of events more than one), this would result in the estimated annual average landslide exposed population exceeding the actual total population. Comparable approaches have been used to estimate the exposed population to landslides on a global scale (Emberson et al., 2020).

Figure 5. Estimated population exposed to landslide for China (a) and single geomorphic regions (b)–(g) during the historical reference period. The bar chart shows the median ensemble of 24 global climate models with error bars at the 25th and 75th percentiles, respectively.

In general, variations in estimated landslide exposure across geomorphological regions are similar when comparing different sources of population products. Higher landslide exposure was estimated by building upon population data from contiguous clusters, while lower estimates were obtained from population data with dispersed distribution patterns. In addition, in the less populated NCP region (range from 21.7 to 110.3 million persons per year in Figure 5, NCP panel) and NWB region (range from 3.4 to 37.2 million persons per year in Figure 5, NWB panel), the maximum difference between the modeled results of different population products could be of fivefold and tenfold, respectively. Therefore, the signals of climate change impacts on landslide-exposed populations obtained in the subsequent sections based on modeling-driven simulations from different sources of population data are represented by calculating the changes in their projected landslide-exposed populations for different SSPs scenarios and periods relative to the modeled exposed populations for their own historical reference periods.

Divergent Signals of the ECLEP Based on Different Population Data

Figure 6 presents projected population trends from the year 2020–2100 under the various SSPs in China for five different data sources. In general, the highest population numbers were constantly projected for the regional competitive path of SSP3, followed by the intermediate path of SSP2. The lowest population numbers relate to the sustainable development path of SSP1 or to the traditional fossil fuel-based path of SSP5. With the exception of the SSP3 for the ECNU population data, the overall population trend was associated with an increasing trend to a certain year, reaching a peak and then proceeding to decrease. The Lund, NASA, and Tsinghua_Long products are expected to reach their peaks around the year 2030, Tsinghua_Cai at around 2040 and the SSP1, SSP2, and SSP5 of the ECNU product reach their peaks around 2060. These differences between different socioeconomic projections may further contribute to discrepancies in the signals of climate change impact assessment results.

Figure 6. Projected population of China in 2020–2100 under each SSP scenario by different research institutions.

Figure 7 depicts the signals of the expected change in landslide-exposed population (ECLEP) separately for different SSPs scenarios, different sources of population data, and different time periods (i.e., near-future, mid-future, and far-future). The shown variations provide a comparison to the historical reference period. It should be mentioned that the static CAS-based population data in the first row of each panel represent changes in future landslide risk proxy due to solely future landslide hazard variations. This allows for a more convenient comparison with the results based on dynamic population data which are driven by joint changes in landslide hazard and population (shown in the subsequent rows of the panel). In general, there is a noticeable divergence in the ECLEP signals modeled using different population data. The simulations based on static CAS and dynamically changing ECNU population data indicate that under climate change, an increase in exposed population to landslides could be expected at national scale and within the geomorphological regions (cf., positive variations with yellow to red colors in the first two rows of the subplots in Figure 7). The simulations based on the remaining four population data sets indicate that the ECLEP might tend to decrease compared to the historical baseline period, that is, the last four rows of the subplots in Figure 7 show mainly negative variations (blue color). For instance, the multi-model ensemble median estimates of the future landslide-exposed populations across the country under the medium emission scenario (SSP2-4.5) are projected to increase by about 0%–19.5% compared to the baseline period based on the CAS and the ECNU population data. In contrast, the results of the exposed population for the same scenario based on the other four sets of population data are projected to decrease by 0.4%–49.1% relative to the baseline period. In terms of the different scenarios and periods in the future, the SSP3-7.0 is most often associated with an increase in the future exposed population to landslides in China. As for the three scenarios SSP1-2.6, SSP2-4.5, and SSP5-8.5, in the far future, the decrease of exposed population driven by the population data of Lund, NASA, Tsinghua_Long and Tsinghua_Cai is expectedly more relevant. These variations across the different time periods and scenarios indicate an intricate evolution of the ECLEP signal under climate change, although in specific regions the ECLEP signals resulting from different demographic data show a remarkable agreement. For instance, results from multiple sets of population data for the NCP region consistently indicate that the climate change impact signal is projected to increase in future.

Figure 7. Expected changes in future landslide exposed population with respect to the historical baseline period. The different panels represent China and the different geomorphological regions; the rows in each panel represent modelled signals based on different sources of population data; the columns depict different combinations of SSPs for separate time periods.

Figure 8 exemplifies estimated landslide-exposed population trends for different demographic data sources under the moderate social vulnerability and moderate radiative forcing SSP2-4.5 scenarios. The analogous graphs for the remaining three SSP scenarios are included in Figures S1–S3 in Supporting Information S1. Figure 8 indicates that at the national scale, the projected impacts of climate change on exposed population differ substantially for the near-year, mid-future, and far-future. The simulations based on static CAS and dynamic ECNU demographic data indicate an increasing trend of future climate change impacts on exposed population, while simulations based on the other four demographic data sets point to a decreasing trend. More specifically, for the far future under the SSP2-4.5 scenario, the projections based on the ECNU product show a 16% increase in the projected future landslide-exposed population compared to the baseline period, while the corresponding projections based on the Lund product result in a 49.1% decrease. Substantial variations in the projected future landslide exposure can also be observed when comparing different geomorphological regions and different population data sources. For instance, for the projected far-future, CAS and ECNU anticipate an increase in population exposure, while the remaining products point to a decreasing number of people exposed to landsliding. There are consistent change signals across different population data simulations for specific time periods in some regions, such as in the near-future and mid-future for the North Central Plateau region under the SSP2-4.5 scenario; and in the near-future for the Eastern Plains and Northwest Basin regions. Furthermore, the range of error bars also reveals that the overall direction of change in landslide-exposed populations simulated by different GCMs is relatively consistent with respect to the same population data.

Figure 8. Changes in future landslide exposed population with respect to the historical baseline period under SSP2 scenario. The bar chart shows the median ensemble of 24 global climate models with error bars at the 25th and 75th percentiles, respectively.

Discussion

In this study, the possible impacts of climate change on landslide-exposed population in China under multiple SSPs scenarios are simulated and projected using multiple GCMs from CMIP6 and multiple sets of population spatial distribution projections. Contrasting with existing studies that have focused only on the changes in the landslide hazard, the present study, based on a combination of high-resolution static and dynamic population gridded data sets, aims at exploring the possible impacts of landslide hazard under future climate change scenarios, considering the changes in the population exposed to landslides on a large spatial scale. Furthermore, through the use of high-resolution dynamic population projections, this work could provide a more comprehensive picture of the potential impacts of landslide hazards than existing studies that only use static population distribution data or administrative region-based demographic data (Gariano et al., 2017; Martha et al., 2021), hence providing more meaningful guidance for the development of landslide risk mitigation and climate change adaptation measured (Galasso et al., 2021).

The results of the study revealed remarkable heterogeneity in the spatial patterns and the temporal evolution of different sources population data. These disparities between the various socioeconomic estimates would further propagate into the signals of climate change impact assessment results. The derived findings of the present work indicate a divergence in the change in the population exposed to future landslides in China based on different sources of demographic data (different rows of each panel in Figure 7). Since the CMIP6 multi-model ensemble is projected to increase in the extent of susceptibility and frequency of landslide occurrence in China, compared with the historical reference period (Figure 3). The first row of each panel in Figure 7 based on the CAS static population showed that considering only the effects of climate change overall would result in an increase in the future landslide-exposed population. Whereas accounting for the joint effects of both climate change and population change on the future landslide-exposed population could result in an increase or decrease (the last five rows of each panel in Figure 7), depending on the considered dynamical population projections. These findings underline that variations in projected exposed population to landslide strongly depend on the sources of the underlying population data (i.e., population changes), and to a lesser extent, on the corresponding GCM (i.e., landslide hazard changes).

The possible reasons for this divergent result are the varying assumptions, targeted users, and population downscaling methods used by the different research institutions in performing the population projections, as well as the ancillary data on environmental factors, even though the future population projections were carried out by each research institution within the framework of the SSPs scenarios. For instance, the population projections developed by ECNU are based on modeling the results of their future urban land change projections and use the same parameters at the global scale, which may be more suitable for urban areas rather than rural areas (Y. Chen et al., 2020a; Li et al., 2022). The ancillary data of nighttime light data considered in the population grid modeling may also overestimate areas of abundant electricity supply (i.e., developed urban agglomerations) and underestimate areas of electricity relative shortage (i.e., rural or economically underdeveloped poor areas) (Andersson et al., 2019; Bennett & Smith, 2017; Kuffer et al., 2022). As a consequence, this may introduce uncertainty in their use for landslide-exposed population modeling.

These findings highlight the critical importance of exposure factors, such as socioeconomic aspects, in the assessment of climate change impacts on landslides, and the need to give more attention and consideration to the element at risk in disaster impact assessments. However, it does not necessarily imply that impact and risk assessments can be carried out directly by leveraging the increasing amount of publicly available exposure data simply overlaid with the hazards, which may lead to misleading conclusions. For instance, an assessment of the future exposed population to landslide hazards in China based merely on ECNU population data compared to a historical baseline period would result in an increase in the likely landslide exposure in the future, whereas the opposite findings would be reached if only the Tsinghua_Long population data were used. The same inconsistence has been noted in studies on the assessment of flood-exposed populations. Smith et al. (2019) have compared the differences in the use of different resolutions of historical population-gridded spatial distribution data for estimating population exposure to flood hazards and showed that simulations based on lower-resolution population data could noticeably overestimate the population potentially exposed to flood hazards. This study shows that various sources for population projections under climate change scenarios might produce ambiguity in estimates of the impact of disasters. As a result, it is necessary to assess and understand the underlying hypotheses and data quality of the input data as well as to take into account the dynamics of the data on the elements at risk. This will make it easier to choose suitable inputs and produce more trustworthy results for assessments of the impact and risk of disasters.

Results of the study also show consistent trends in some of China's morphological regions for given periods in the future. For instance, for the NCP region, in both near-future and mid-term future, all models agree on a potential increase of landslide exposure, albeit with different magnitudes. Similar consistency but with weaker trends is also observed in the regions SWM and NWB. These areas may require more effective climate change adaptation strategies and tailored landslide risk mitigation interventions.

It should be noted that the categorized landslide susceptibility used in this study to determine the potential impact area of landslides represents a simplified approach. Such categorical susceptibility has been used in existing studies for situational awareness nowcasts (i.e., as a proxy for landslide hazard) and exposure estimates (Emberson et al., 2020, 2021; Kirschbaum & Stanley, 2018). Future research could incorporate continuous spatial landslide susceptibility probabilities (Ozturk et al., 2020), dynamic temporal probabilities of rainfall-induced landslides (Steger et al., 2023), and underlying elements at risk to estimate potential landslide consequences for dynamic impact-based and risk-based early warning. It is expected to be a considerably effective practice for mitigating landslide risk (Galasso et al., 2021). In the light of the insights gained from this study, further research is needed to better analyze and understand the reasons for the apparent divergency in the results, but the preliminary findings might also suggest that an ensemble approach could also be applied to population projections, to better represent the effect of epistemic uncertainties of the underlying SSP projections into the risk assessment. Besides, the interaction of risk components, that is, elements at risk may simultaneously have an effect on hazard and exposure, could be investigated closer in future studies to more accurately assess and quantify disaster risk (Depicker et al., 2021; Ozturk et al., 2022).

Conclusions

This study uses CMIP6 multiple GCM data with multiple sets of population grids under different SSPs scenarios to drive the landslide model for simulating possible changes in future landslide exposed population under four SSPs scenarios. The results show that the CMIP6 multi-model ensemble describes a likely overall increase of future landslide hazard in China, considering the extent of susceptibility and frequency of landslide occurrence, compared to the historical reference period. However, assessing the possible changes in the future exposed population to landslides in China under climate change scenarios considering dynamic population data from different sources leads to divergent results, depending on the considered population projection.

These results highlight the uncertainty that may result from different sources of future population projections and stress the need to consider not only the dynamics of the element at risk data in disaster impact and risk assessment studies, but, more importantly, the need to understand the quality and underlying assumptions of the data used in the assessment. This will allow for a more appropriate selection of suitable input data in the practical development of potential disaster impact and risk assessments.

The findings highlight the importance of employing a holistic perspective when assessing the impact of climate change on landslides. It is imperative to expand the focus beyond the evolution of the hazards alone and account for the potential changes in the exposed elements at risk. Both two components have a considerable influence on the ultimate outcome of the disaster risk assessment. Additionally, it is essential to carefully examine the quality and applicability of the input data before utilizing it in the disaster impact and risk assessment. On this basis, appropriate data-driven models can be used to derive meaningful findings and avoid misleading outcomes due to problems with the applicability of the data inherently.

Acknowledgments

This research was supported by the Natural Science Foundation of Jiangsu Province (BK20220456), the China Postdoctoral Science Foundation (2023M731751), and the Guangxi Key Research and Development Program (Guike AB22080060).

Conflict of Interest

The authors declare no conflicts of interest relevant to this study.

Data Availability Statement

The data used to derive all the related results can be available in Lin et al. (2023). The data of 24 global climate models CMIP6 can be accessed at https://esgf-node.llnl.gov/projects/cmip6/. The CN05.1 gridded daily observation data set over China can be accessed at https://ccrc.iap.ac.cn/resource/detail?id=228. The CAS static grid population data are available from the Resource and Environmental Sciences Data Center in Xu (2017). The projected gridded population of Lund can be accessed in Olén and Lehsten (2022). The projected gridded population of NASA can be accessed in Gao (2020). The projected gridded population of ECNU can be accessed at https://www.geosimulation.cn/FPOP.html. The projected gridded population of Thua_Cai can be accessed in Y. Chen et al. (2020b). The projected gridded population of Tsinghua_Long can be accessed in X. Wang et al. (2022b).