Uncertainty analysis of potential population exposure within the coastal lowlands of mainland China

With accelerating global sea level rise driven by climate change, accurate estimates of potential population exposure (PPE) within the low-elevation coastal zones (LECZ) are critical for coastal planning and assessing the benefits of climate mitigation. Multiple digital elevation models (DEM) and population grid datasets have been used for the PPE assessment of coastal lowlands. However, the uncertainty arising from differences in data sources and production methods results in poorly guided estimates. In this study, four global DEM and five population datasets were used to estimate the PPE in the LECZ of China and to assess the uncertainty of PPE estimation. Based on the DEM and population grid with the best accuracy, we found that more than 13.82% of China’s residents lived in the LECZ in 2010. Different DEM-population combinations yielded significantly different PPE estimates, ranging between 3.59–24.61 million and 31.56–112.24 million people in the LECZ below 1 m and 4 m elevation, respectively. The satellite Lidar-based DEM improves the estimates of the LECZ and obtains the PPE within LECZ below 4 m elevation that far exceeds those of other DEM datasets. The usage of WorldPop and LandScan population datasets leads to an underestimation of PPE within the LECZ of China. In contrast, integrating more geospatial big data helps generate better population grids, thus reducing the uncertainty of coastal PPE estimates. There is still a need to improve the availability and accuracy of coastal geospatial data and to deepen the understanding of coastal vulnerability.


Introduction
Considering the carbon emission scenarios and the land subsidence caused by drainage and groundwater abstraction, relative sea-level rise (hereafter: SLR) will reach 1 m before 2150 and 4 m in centuries to come [1][2][3] in the most densely populated coastal areas.This immediately implies the increased coastal exposure to flooding and the cost of climate change adaptation measures [4].Low-elevation coastal zones (LECZ), areas below 10 m elevation and hydrological connected to the ocean [5], may be the most meaningful areas for exposure projection as this suggests the greatest potential for this site to be permanently submerged in the ocean [6,7].Many studies have been devoted to assessing the risks of LECZ, especially for the population within the LECZ in recent decades [8][9][10].McGranahan et al [5] first undertook the global review of the LECZ and its inner potential population exposure (PPE), reporting that LECZ covers 2% of the global land area while containing 10% of the world's population.Then country and regionalwide fine-grained assessments facilitated by enhanced data availability further confirmed the prevalence of dense population distributions within LECZ [4,11].However, the PPE with uncertainty in the LECZ was thus widely generated, i.e. differential estimates within the same study area, including the area of the LECZ, the number of exposed populations, together with the spatial distribution of the LECZ and exposed populations [9,12,13].This ultimately hinders the development of policies about resisting the flooding and inundation risks posed by near-future SLR to the LECZ.

F Li et al
Digital elevation model (DEM) and population grid datasets are the major reasons for this discrepancy in PPE estimates within the LECZ.Satellitebased (near) global-DEM are the most commonly used DEM datasets in global coastal SLR, including DEM acquired from Shuttle Rader Topography Mission (SRTM), Advanced Land Observing Satellite (ALOS), etc [14][15][16].However, multiple global DEM were investigated to show limitations in SLR-related assessments due to the dense vegetation and built-up areas in the non-bare earth terrain involved in measured data [16][17][18][19][20].The SRTM was the first global DEM but was found to have a global mean error of 1.9 m when using the DEM from NASA's ICESat satellite as ground truth [21].Despite the improvement in accuracy compared to SRTM, ALOS and the improved SRTM-based NASADEM still lack the necessary precision and minimum uncertainty for SLR assessment [3,22].And their vertical errors which approach or even exceed the projected SLR heights for this century incline to trigger serious misestimation of PPE [2,21].
As for population datasets, decomposing census data through machine learning (ML) algorithms and multi-source covariates is currently the dominant method for generating population grids [23,24].ML algorithms such as random forest (RF), and artificial neural network (ANN) have been shown to have higher fitting accuracy than area interpolation and multiple regression methods for population spatialization [25,26].The resulting global population datasets including WorldPop and LandScan are usually used for PPE estimation worldwide [27,28], but the estimated population below 1 m elevation varies from 1% to 2.3% of the total global population [29].Differences in covariates and algorithms as well as the resolution of the population grid could be responsible for the differences in estimates [12,24].Remote sensing data represented by nighttime light (NTL) and normalized vegetation index, which are typical covariates for population spatialization [30], are mostly limited in resolution and can only reflect the characteristics of population distribution from the perspective of the residential environment.Geospatial big data complements this gap and shows great a correlation with population activity characteristics [31,32].Currently, population datasets generated by fusing remote sensing data and socio-aware big data based on ML algorithms have been found to yield better-fitting population models [24,33].However, these population datasets produced by new techniques are still not widely used for coastal risk assessment and comparison with previous population datasets.
China has a huge LECZ and a dense coastal population, with more than 10% of the country's population living in the LECZ [5].Three major economic zones in eastern China, the Yangtze River Delta, the Pearl River Delta, and the Bohai Rim, are all located in the low-lying estuary alluvial plains and are at risk of inundation from SLR.However, although the extremely high PPE within the LECZ of China is well established, previous studies have ignored the uncertainty introduced by different datasets in the estimates.Recently, new satellite Lidarbased DEM data were produced for global SLR assessment and have been updated to a second version (GLL_DTM_v2, hereafter: GLL) at 0.01-degree horizontal resolution (∼1 km near the equator) [17,34].Validation with existing well-described and accurate local DTM revealed that GLL exhibits higher vertical accuracy than other global DEM and is more suitable for global SLR assessment [3].Therefore, it is now necessary to combine new elevation data and population grids to quantify the PPE of the LECZ and compare it with previous data sources for uncertainty analysis.
In this study, four global DEM (SRTM, NASADEM, ALOS, and GLL) and five population datasets (WorldPop, LandScan, and three of our own) were used to estimate the PPE within the LECZ in mainland China.The manuscript is organized as follows: section 2 describes the data sources and the data pre-processing process.Section 3 demonstrates the methodology to perform the PPE estimation.The results are presented in section 4, including the distribution of population and LECZ along the coast of China, and the uncertainty in PPE estimates.Finally, a discussion and summary are presented in sections 5 and 6, respectively.

Global population datasets
WorldPop: Covering 166 countries and 82% of the world's population, WorldPop is one of the best population datasets in the world today [28].Based on the RF algorithm, WorldPop decomposes administrative-level population data using remote sensing and geospatial datasets to generate population grid data with a resolution of 100 m.In this study, WorldPop for China in 2010 was downloaded from www.worldpop.org/.
LandScan: LandScan is a global population dataset with a resolution of 1 km developed by the Department of Energy's Oak Ridge National Laboratory using spatial data, high-resolution image exploitation, and multivariate dasymetric modeling methods [27,35].The LandScan Global 2010 was downloaded from https://landscan.ornl.gov/.

DEM datasets ALOS:
This DEM dataset is a global digital surface model acquired by the Panchromatic remote-sensing Instrument for Stereo Mapping aboard the ALOS with a horizontal resolution of 30 m (approximately 1 arc second) [36].Download from: www.eorc.jaxa.jp/ALOS/jp/dataset/aw3d30/aw3d30_j.htm.

SRTM:
This dataset was obtained by the SRTM [37], a joint survey conducted by NASA and the National Imagery and Mapping Agency with a resolution of 30 m. SRTM reflects the highest elevation of an object on the ground and provides free data for all land areas from 60 GLL_DTM_v2: GLL is a global DEM dataset generated by Vernimmen and Hooijer [3] based on the most accurate new satellite LIDAR ICESat-2 data, with a 0.01 • horizontal resolution (∼1 km near the equator).This DEM data has been shown to have higher vertical accuracy compared to other global DEM [3,34].The GLL data for this paper are from the open study by Vernimmen and Hooijer [3], downloaded from https://doi.org/10.5281/zenodo.6534526.

Auxiliary data in population spatialization
Census data: Census population data of counties (978 units) and townships (13 065 units) in coastal China were obtained from the Sixth National Population Census of mainland China, 2010.Hongkong, Macao, and Taiwan were excluded due to their distinct political and economic status.In this study, county-level data were used to fit the ML models, while township-level data were employed to evaluate the accuracy of the population grids.
Remote sensing data: NTL data were obtained from the DMSP/OLS sensor of the National Oceanic and Atmospheric Administration, which provides the 2010 NTL product with a resolution of 500 m [39] (https://ngdc.noaa.gov/eog/dmsp/download_radcal.html).Land use data were downloaded from a multi-temporal global surface product produced by Liu et al [40], with a spatial resolution of 30 m (www.geosimulation.cn/GlobalUrbanLand.html).This product is a binary dataset that distinguishes urban land from non-urban land and contains not only details of large cities but also information on smaller settlements.In this study, the proportion of urban land per 100 m grid (i.e.land use layer) was calculated as one of the model covariates.The normalized difference vegetation index (NDVI) data were obtained from the Vlaamse Instelling Voor Technologisch Onderzoek with a resolution of 250 m (www.vito-eodata.be/PDF/portal/Application.html).All ten-day NDVI data products of SPOT-VEGETA-TION in 2010 were downloaded, and the annual maximum NDVI images were calculated (equation ( 1)) using the maximum value composite method [30].In addition, slope data with a resolution of 100 m were created based on DEM data where NDVI 1 , NDVI 2 , …, and NDVI 36 are NDVI values of the same pixel in the 36 ten-day SPOT images.
Geospatial Big Data: Two types of geospatial big data were used for modeling in this study, namely road network data and point of interests (POI).The road network data were obtained from the Data Center for Resources and Environment Sciences, Chinese Academy of Sciences.Based on the definition of road density (road length per unit area), all road types were fused to generate a road density raster layer with a resolution of 100 m.POI are the point data abstracted from geographic entities, recording four dimensions of latitude, longitude, name, category, and attribute [41,42].The POI data used in this study were collected from the Baidu map, with a total of 3605 250 entries divided into 20 categories, including government agencies, commercial buildings, hotels, residential communities, etc [43,44].All types of POI were processed to generate POI density layers according to the kernel density method, and then the principal component analysis was used to fuse the various POI densities layers into a single density layer [45].See supplementary 2 for detailed processing steps.

Evaluation methodology
The flowchart of this study is shown in figure 1.While using two global population datasets, WorldPop and LandScan, we additionally produced coastal population grids for China based on recent studies using three ML algorithms and multi-sources data [24,33].
Then, the LECZ extracted from the four DEM were superimposed on the population datasets to assess the PPE of mainland China.

Population spatialization
Referring to previous studies [24,33,46,47], this study selected the logarithm of the county average population density as the dependent variable and the mean of each covariate at the county-level as the independent variable to build ML models.Three algorithms, namely RF, Cubist, and ANN, were used in this study.All algorithms have been widely used in population spatialization, and the details of these algorithms can be referred to Yang et al [13] and Krogh [48].
The spatial distribution of the population is influenced not only by the natural environment but also where Pop d , Road d , and POI d represent the population density, road density, and POI density, respectively.The function F is a nonlinear function mapping between dependent variables and the covariates computed by ML algorithms.Then the corresponding independent variable raster layers were input to the trained models, and the three weight layers of RF, Cubist, and ANN about population distribution were successfully obtained.Finally, referring to the dasymetric method [24,25], the county-level population census data were allocated according to the weight layers (equation ( 3)), and the population distribution maps of coastal China with a resolution of 100 m were generated.Based on the fact that humans cannot live in the water, this study changed the pixel value in the water body to 0 while the non-water body remained unchanged where W grid is the Population-distribution weight for a 100 m × 100 m gridded area, W county is the summed Population-distribution weight of a county, Pop county represents the county census Population, and Pop grid is the population density for the gridded area.
All population datasets will be validated at the township-level census, and evaluated on the mean absolute error (MAE), mean relative error (MRE), root mean square error (RMSE), percentage root mean square error (%RMSE), and correlation coefficient (R).The MAE and MRE are used to assess the numerical deviation of the predictions from the census data.RMSE and %RMSE reveal the overall distribution of errors between predictions and census data.R reflects the linear relationship between the two datasets.(See supplementary 1, equations (A1)-(A5) for the corresponding equations.)

Extraction of LECZ
The study area covers 11 coastal provinces of mainland China from north to south: including Liaoning, Hebei, Tianjin, Shandong, Jiangsu, Shanghai, Zhejiang, Fujian, Guangdong, Guangxi, and Hainan.According to the definition of McGranahan et al [5], the LECZ refers to an area below 10 m elevation and hydrological connected to the ocean.In this study, four DEM were used to extract LECZ respectively based on the following summary rules [29]: (1) Extract areas below 10 m in elevation in the study area.(2) Eliminate scattered patches within the coastline to ensure the continuity of LECZ.(3) Further breakdown of LECZ (hereinafter referred to all coastal lowlands) based on altitude, including areas below 1 m elevation (hereinafter: LECZ 1 , the same below), areas below 2 m elevation (LECZ 2 ), areas below 4 m elevation (LECZ 4 ) and areas below 10 m elevation (LECZ 10 ), respectively.

Variability in the LECZ
The results of different DEM datasets show that LECZ in China is mainly located in the Bohai Rim, Yangtze River Delta, and the Pearl River Delta, with a few scattered along the coastline of other cities (figure 2).For the defined 10 m altitude, the SRTM dataset obtained the largest LECZ 10 territory (221.60 × 10 3 km 2 ) but the GLL obtained the smallest LECZ 10 territory (173.41 × 10 3 km 2 ), and the ALOS and GLL extracted the most similar LECZ 10 extent in mainland China (table 1).The differences between the DEM datasets are also evident in the details within the LECZ.The GLL presents more detailed vertical height information at lower elevations in coastal areas, with more areas considered to be between 1 m and 2 m above sea level.In contrast, ALOS and NASADEM obtained a larger area of LECZ 1 , while SRTM extracted the largest area between 4 m and 10 m attitude.
Specifically, when analyzed from the provincial perspective (first administrative unit in China), Jiangsu, Guangdong, and Shandong are the three provinces with the largest LECZ area according to ALOS and NASADEM (figure 3).While in SRTM and GLL, Hebei overtakes Shandong as the third coastal province in terms of the LECZ 10 area.In the other coastal provinces, the differences in the LECZ 10 area obtained from the four DEM are not significant.However, this difference was found to increase with decreasing elevation thresholds, manifested as an increased relative standard deviation (RSD, equation ( 4)), especially in the provinces of Tianjin, Jiangsu, Shanghai, Zhejiang, and Guangdong.According to SRTM, Shanghai only has LECZ 1 for an area of 30 km 2 and LECZ 2 for an area of 120 km 2 .While in other DEM, the area of LECZ 1 and LECZ 2 in Shanghai are maximally estimated to be 1150 km 2 and 2410 km 2 , respectively.Assuming that the GLL is a relatively accurate ground truth elevation, all other DEM datasets result in relatively high misestimations.Among them, NASADEM and ALOS show mostly overestimation of coastal lowlands, and SRTM shows mostly underestimation of coastal lowlands below 4 m elevation and overestimation of coastal lowlands from 4 m to 10 m elevation.And in individual cities, the situation may still become different.
where n is the number of estimates, x i donates the ith estimate, and x represents the mean of all estimates.

Variability of population datasets
The population datasets generated in this study (hereafter referred to as RFPop, CubistPop, and ANNPop, respectively) and WorldPop exhibit consistent spatial distribution patterns along the coast of mainland China (figure 4), with the following characteristics: (1) The Yangtze River Delta, Pearl River Delta, and Bohai Rim are the main centers of population concentrations in eastern China.(2) The population is distributed in a dotted pattern, with high population density in the urban centers and decreasing toward rural areas.Especially in Shanghai and the Pearl River Delta, the population concentration is highly spatially consistent with the cities. (3) In the southeastern coastal provinces of China, specifically Zhejiang, Fujian, Guangdong, and Guangxi, the population is concentrated near the coastline, while the density is generally low in the inland areas.Verifying the above population datasets at a finer scale (township level), and the results are presented in table 2.
All four datasets present good accuracy, with high R-value and low validation error.However, the consistency of the RFPop, CubistPop, and ANNPop with the township statistics is significantly improved compared to WorldPop.Overall, RFPop achieves the best results among all population datasets.However, the differences between population datasets were still pronounced in local areas.Even within the same location, the population densities obtained from different population grids tend to differ.For example, the most densely populated areas in all datasets were predicted to be located in the center of Shanghai, but CubistPop and ANNPop projected higher population concentrations in downtown areas compared to RFPop and WorldPop (figure 5).In addition, the principles of dasymetric mapping dictate that the total population is the same for all datasets within county administrative areas.Therefore, higher population density in downtown areas also implies a more uneven population distribution, i.e., a larger gap in population density between urban and rural areas.On the other hand, there are differences in the details of spatial distribution characteristics within urban population concentration hotspots captured by different population datasets.Compared with Google Earth satellite images, the population distribution of RFPop and WorldPop are more closely matched to urban land use types, highlighting the  central and peripheral areas of the city.CubistPop presents a more significant road trace feature with high population density in downtown distributed in strips.In contrast to other datasets, the urban population densities in ANNPop show a circular distribution pattern, with higher and more concentration in urban centers, but unfortunately, it seems to be inferior to the other datasets in terms of identifying urban details.

Uncertainty estimation by combining DEM and population datasets
The PPE uncertainty for the LECZ was further estimated by overlying the DEM and population datasets.
The results show that the PPE within the LECZ 10 of mainland China generated from different data sources ranges from 121.55 million to 185.15 million, accounting for 9.07%-13.82% of the national population (table 3).The Bohai Rim, the Yangtze River Delta, and the Pearl River Delta account for the major portion of the PPE in mainland China (figure 6).The Yangtze River Delta, in particular, not only has the most widely distributed LECZ, but also many densely populated metropolitan cities such as Shanghai, Suzhou, Hangzhou, and Wuxi.Therefore, the accuracy of both DEM and population datasets will have a significant impact on PPE in this region.
A similar situation occurs in the cities of Guangzhou, Zhuhai, and Shenzhen in the Pearl River Delta, where the variability in DEM and population datasets is particularly significant in the urban centers (figure 2).In contrast, despite the large differences among DEM in the southwest Bohai Rim, the uncertainty in PPE is significantly smaller than in the Yangtze River Delta region due to the lower population density in the coastal lowlands, excluding Tianjin (figure 4).
Undoubtedly the superposition of these variability further amplifies the uncertainty in the estimates.Differences in estimates due to different DEMpopulation combinations amount to 21.02 million in LECZ 1 , 33.14 million in LECZ 2 , 80.68 million in LECZ 4 , and 67.11 million in LECZ 10 (table 3).The uncertainty in the estimates increases rapidly with decreasing altitude, a trend similar to the uncertainty present in the LECZ distribution in the previous section, i.e. the lower the elevation in the coastal zone, the higher the inaccuracy of the DEM measurements.A significant point of divergence in PPE estimates occurs at 4 m above ground level, with SRTM always underestimating PPE at elevations below 4 m, and ALOS underestimates PPE at elevations above 4 m in all DEM datasets.As one of the most vertically accurate DEM in coastal zones globally, GLL achieves much higher PPE estimates within LECZ 4 than any other DEM datasets, implying that early SLR impacts will be more severe than predicted by previous studies.In addition, despite the differences in the spatial distribution of the population, similar PPE estimates are obtained for the three population grids we produced on the same DEM benchmarks (table 3).In contrast, WorldPop produces similar results to our population grids in LECZ 4 and lower altitudes but underestimates across the LECZ 10 .For LandScan, it caused an underestimation of PPE at all altitudes compared to the other population grids.Combined with the validation of the different datasets at the township-level (table 2), datasets with better validation accuracy tend to give similar estimates, and unfortunately, it appears that the results of this more accurate estimation are currently higher.
Estimation results at the provincial level more clearly demonstrate the uncertainty in PPE estimates due to the superimposition (figure 7).Provinces with larger territories located in the LECZ tend to have high PPE estimates as well, with typical cities such as Jiangsu and Guangdong.However, the variability of different estimates for these regions is correspondingly high (higher variance in the box plots) and shows a similar trend to the uncertainty of LECZ distribution (figures 7 and 3).In contrast, provinces with smaller LECZ areas may also have high PPE estimates, such as Shanghai, Tianjin, and Zhejiang, which   is attributed to the dense urban population within the LECZ.In these areas, the superposition of uncertainties in the downtown population datasets and the DEM often leads to higher uncertainty in the PPE estimate, i.e., a higher RSD of the estimate.Only in provinces with lower population densities and smaller LECZ areas are the PPE estimates and estimation uncertainties likely to be smaller, such as Guangxi and Hainan.However, the uncertainty of PPE estimates within the Chinese LECZ remains high because the LECZ are mostly located in estuarine deltas with well-developed economies and dense populations.It is also clear that the uncertainty in the PPE estimates increases rapidly as the elevation of the coastal provinces decreases.

Discussion
Despite all previous studies have indicated that a large global population resides in the LECZ, our study found that there are significant data-related uncertainties in the estimation of PPE [5,13].The difference in PPE estimates for China' LECZ generated by different DEM and population grids reached a maximum of 67.11 million people.Based on the most accurate population grid (RFPop) and the DEM (GLL) datasets [3], we obtained an estimation of PPE within the LECZ of 185.15 million people in mainland China, which accounted for 13.82% of the country's total population in 2010 (table 4).This result is much higher than those obtained by Yang et al [13], F Li et al   Liu et al [12], and McGranahan et al [5] for mainland China.Compared with previous studies, the results of PPE estimation in this study are more accurate and reliable due to the use of higher quality population and DEM data.Previous studies have underestimated the risk of inundation of China's coastal populations, and have led to differences in the spatial distribution of PPE, especially in the three metropolitan areas in eastern China, which hinders the formulation of coastal population management policies.The increasing availability of data worldwide in recent years has facilitated coastal risk assessments on the one hand, and raised the potential for increased uncertainty in the results on the other.This implies the need for more careful selection of source data in assessments, as well as efforts to produce    [16,49], and are validated to have low vertical accuracy by any measure [50].Local validation results show that the average deviation of these global DEM is generally above 1 m, which is higher than the projected SLR at the end of this century, which undoubtedly leads to a lack of credibility in their PPE estimation [21,49,51].Moreover, such biases tend to vary across geographic regions and between DEM, making it difficult to reasonably correct them.Therefore, there is a need to develop low-cost, high-precision elevation sounding methods, especially for densely populated and economically developed coastal areas [52].This may change the original perception of SLR risks.For example, Kulp and Strauss [16] found that the global vulnerability to coastal flooding based on ICESat-2 Lidar-corrected CoastalDEM is three times greater than that of SRTM.In this study, the new satellite Lidar-based DEM reveals more accurate coastal vertical heights and yields a different LECZ distribution compared to previous studies [3,34,50].Data accuracy issues also arise in population distribution mapping, where different population datasets tend to show different spatial distributions of the population [27,53].In general, more covariates improve the ability of ML to capture human activities [33,54], which promotes the development of fusing data such as DEM, NDVI and, NTL for population spatialization [30,55].However, the spatial and temporal resolution of satellite data limits most population datasets to focus on group aggregation characteristics of populations, and detailed information from an individual perspective is usually missing [24,55].Meanwhile, the large number of covariates substantially increases the computational effort, leading to the difficulty of black-box structured ML to capture the complete features of human activities, and ultimately generate differentiated population datasets.The strong relationship between geospatial big data and population distribution is now well established [33] and complements this gap from a sociological perspective.POI and road network data have amply demonstrated the advantages of this feature in population spatializing and produce better coastal population datasets in this study.Nowadays, there is a need for more access to widely available geospatial big data globally, especially those related to individuals (e.g., social media location data), which will further improve population spatialization and thus reduce the estimation error of coastal PPE [28,46].
In addition, imperfections in definitions and assessment schemes hinder more accurate PPE estimates.The previous definition of 10 m altitude in the LECZ somewhat obscures the uncertainty introduced by different data in the assessment of PPE [9,29].
As in this study for LECZ 1 , LECZ 2 , and LECZ 4 , the variability between DEM was found to increase rapidly with decreasing altitude.All global DEM data are observed to lack detailed information below 4 m elevation compared to GLL.The SRTM data [15,20,56,57], which are widely used in global SLR risk assessment, underestimates land area below 4 m and overestimates land area above 4 m in LECZ of China.However, this phenomenon is often overlooked in previous PPE estimations due to the lack of corresponding definitions.On the other hand, provincial authorities are the major formulators of local policies in China.However, few studies have assessed the SLR risks faced at provincial level, naturally overlooking the great variability of PPE among provinces [9,58,59].In this study, this variability has been confirmed and found to depend not only on the area of the provinces' territories falling within the LECZ but also on the population or economic density of the location where the LECZ is situated.Developed regions such as Guangdong and Shanghai have the tendency to show an extremely high ratio of PPE, due to the lower coastal elevations and highly concentrated populations [60,61].Unfortunately, numerous metropolises around the world are located in these estuarine regions [62], which implies great PPE as well as estimation uncertainties that are not easy to neglect.Considering the recently published SLR projection [63][64][65][66] and the latest research showing that earlystage SLR leads to the greatest increase in global coastal exposure to flooding [3], there is an urgent need for detailed study of LECZ.This includes a more detailed precise definition and a focus on smaller regions to improve the accuracy of PPE estimation.
Although this study obtains the most accurate PPE estimates currently available in China, there are still some other limitations.Firstly, shoreline change, especially reclamation, and levees in coastal cities, can seriously affect the delineation of the LECZ boundaries and the implementation of policies.Moreover, a more detailed and exact definition of the LECZ needs to be proposed given the rate of SLR in this century.Secondly, differences in the acquisition time and resolution between DEM data and spatialized population covariates potentially result in the uncertainty of the PPE estimation.The spatial-temporal consistency of geospatial datasets from multiple sources needs to be considered in future estimations and needs to be integrated with government policy planning to provide more accurate results.Finally, both ocean circulation and climate change potentially affect the risk of inundation of coastal populations, which deserves to be discussed in subsequent studies.

Conclusion
In this study, we estimated the PPE within the LECZ of China based on four DEM datasets and five population grids.Different DEM-Population combinations produced large estimation uncertainties that increased rapidly with decreasing altitude.In LECZ 1 and LECZ 4, the difference between the highest and lowest estimates reaches 21.02 million and 80.68 million people, respectively.The satellite Lidar-based DEM improves the estimates of the LECZ and obtains estimates of PPE within LECZ 4 that far exceed those of other DEM datasets.WorldPop and LandScan, the most widely used population grids, lead to underestimation in the assessment of PPE within the LECZ of China.In contrast, new gridded population datasets that incorporate geospatial big data such as POI achieve higher accuracy in coastal China.Based on the DEM and population grids with the best accuracy, we found that in 2010, 13.82%, 8.29%, and 0.56% of the total population were located in LECZ 10 , LECZ 4 , and LECZ 1 in China, respectively.Considering the significant uncertainties in current PPE estimates, there is a need to refine the framework for LECZ assessment and to improve our understanding of population spatio-temporal dynamics in coastal areas by integrating more geospatial big data.

Figure 1 .
Figure 1.Flowchart for potential population exposure estimates within the LECZ of mainland China.The gray boxes represent input data for PPE estimation.The blue boxes represent the processing methods in estimation.The green boxes are the datasets and results generated during the estimation process.

F
Li et al

Figure 5 .
Figure 5.Comparison of population distribution in central Shanghai using (a) Random forest, (b) Cubist, (c) Artificial neural network, (d) WorldPop (unit: people/ha), and (e) Google Earth satellite images.The black circles show the main urban areas of Shanghai.

Figure 6 .
Figure 6.Population distribution of RFPop in LECZ of Yangtze River Delta, Pearl River Delta and Bohai Rim.Figure (a) represents the PPE estimates based on RFPop and GLL.Figures (b)-(e) are PPE estimates based on the four DEM and RFPop in the Pearl River Delta; figures (f)-(i) are estimates in the Yangtze River Delta, and figures (j)-(m) are estimates in the Bohai Rim.
Figure 6.Population distribution of RFPop in LECZ of Yangtze River Delta, Pearl River Delta and Bohai Rim.Figure (a) represents the PPE estimates based on RFPop and GLL.Figures (b)-(e) are PPE estimates based on the four DEM and RFPop in the Pearl River Delta; figures (f)-(i) are estimates in the Yangtze River Delta, and figures (j)-(m) are estimates in the Bohai Rim.

Figure 7 .
Figure 7.Estimated PPE by province in LECZ.The box plot shows the PPE obtained from the estimation of different DEM-population combinations, with the horizontal line representing the median of the estimates, and the dashed line representing the relative standard deviation (RSD) between estimates.(a) PPE estimates in the LECZ1; (b) PPE estimates in the LECZ2; (c) PPE estimates in the LECZ4; (d) PPE estimates in the LECZ10.

Table 1 .
Estimation of China's coastal land area with different increments based on four kinds of elevation data.

Table 2 .
Validation results of each population dataset at the township level.

Table 3 .
The potential population exposure of China's low elevation coastal zone under different elevation increments according to RFPop, CubistPop, ANNPop, WorldPop, and LandScan.

Table 4 .
The percentage of PPE in the national population obtained based on GLL and different population datasets.