A daily high-resolution (1 km) human thermal index collection over the North China Plain from 2003 to 2020

Abstract

Human-perceived temperature (HPT) describes the joint effects of multiple climatic factors such as temperature and humidity. Extreme HPT events may reduce labor capacity and cause thermal discomfort and even mortality. These events are becoming more frequent and more intense under global warming, posing severe threats to human and natural systems worldwide, particularly in populated areas with intensive human activities, e.g., the North China Plain (NCP). Therefore, a fine-scale HPT dataset in both spatial and temporal dimensions is urgently needed. Here we construct a daily high-resolution (~1 km) human thermal index collection over NCP from 2003 to 2020 (HiTIC-NCP). This dataset contains 12 HPT indices and has high accuracy with averaged determination coefficient, mean absolute error, and root mean squared error of 0.987, 0.970 °C, and 1.292 °C, respectively. Moreover, it exhibits high spatiotemporal consistency with ground-level observations. The dataset provides a reference for human thermal environment and could facilitate studies such as natural hazards, regional climate change, and urban planning.

Background & Summary

Under global warming, the frequency and intensity of extreme temperature events (e.g., heatwaves and cold spells) are increasing in most parts of the world^1,2,3,4,5. They pose severe impacts on human society and the ecological systems^6,7,8,9,10, putting more people at risk of extremely hot or cold^11,12. For example, heatwave led to more than 70,000 deaths in Europe in 2003 and 55,000 deaths in Russia in 2010^13,14. Extreme cold events also pose a great threat to human health and often cause a higher proportion of deaths than extreme heat^15,16.

Human-perceived temperature (HPT) has attracted much attention in recent years due to its close connection with human health^17,18. HPT is not only affected by air temperature but also by other meteorological variables (e.g., humidity and wind speed). For instance, HPT would become higher on hot days with high humidity while becoming lower on cold days under the additional influence of wind^19,20,21,22. Relevant studies revealed that discomfort HPT could impact human health, such as respiratory distress and reduced skin evaporation, and significantly increase the health risks and mortality in heat- and cold-exposed populations under the joint effects of multiple meteorological factors^23,24,25. Li et al.²⁶ showed that global HPT increases faster than the actual air temperature. Faster increases in HPT were also confirmed by regional studies^27,28. However, previous studies mainly quantitated the human thermal environment and human heat exposure assessment based on meteorological station data and homogeneous climatic datasets^29,30,31,32. These station-based or raster-based datasets have a coarse spatial resolution (e.g., 0.5° × 0.5° or 2.5° × 2.5°) and insufficient temporal resolution (e.g., monthly), which cannot meet the requirements of long time series, full regional coverage, and fine spatial scale studies²⁸. A seamless dataset of HPT with high temporal frequency and spatial resolution is still lacking and sorely needed.

A better understanding of the thermal environment helps prevent or mitigate heat/cold stress effectively¹⁸. Various HPT indices have been proposed to investigate human thermal stress previously. For example, the wet-bulb temperature (WBT), effective temperature (ET), and heat index (HI) have been widely used for quantitative assessment of human thermal environment research^{17,28,33,34,35}. However, there is still a lack of unified evaluation standards or comprehensive datasets for assessing the human thermal environment and stress³⁶. Therefore, it is imperative to develop a new collection with multiple indices for evaluating the thermal environment at a fine scale (e.g., at daily frequencies, and ~0.00833° spatial resolution in a geographical coordinate system, which is equal to around 1 km near the equator).

Previous studies have projected that the morbidity and mortality induced by extreme temperatures may continue to rise, especially in populated Asia^37,38,39. In particular, the North China Plain (NCP), as one of the most populated and urbanized areas of the world, is suffering from severe thermal stress^40,41. It is located in a typical East Asian monsoon region (113°E–121°E and 34°N–41°N)^40,42, and spans eleven provincial units of China, including one of the regions with the most profound human activities in China, i.e., the Beijing-Tianjin-Hebei region⁴³. NCP has a total population of 0.3 billion, which accounts for around one-fifth of the population of the country⁴⁴. The population exposed to heatwaves in this area is up to 0.66 million persons/day during 1961–2015⁴⁵. By examining the long-term changes of major urbanization agglomerations in China, Wang et al.⁴⁶ indicated that HPT in northern China grew faster than that in the south, particularly in NCP. Zhang et al.⁴⁷ have generated a 1-km-resolution dataset of monthly human thermal index collection (HiTIC-Monthly) over China during 2003–2020. However, this monthly dataset is not capable of detecting extreme weather events such as heatwaves and cold spells, which are usually defined based on daily temperatures. The lack of a daily HPT dataset limits accurate heat stress-related studies in this region, and a multi-index dataset of human thermal stress at a daily scale over NCP is highly essential.

To fill the aforementioned research gap, this study aims to construct a high spatial resolution (1 km × 1 km) human thermal index collection at a daily scale over NCP from 2003 to 2020 (i.e., HiTIC-NCP). The 12 widely-used indices include Surface Air Temperature (SAT), indoor Apparent Temperature (AT_in), outdoor shaded Apparent Temperature (AT_out), Discomfort Index (DI), Effective Temperature (ET), Heat Index (HI), Humidex (HMI), Modified Discomfort Index (MDI), Net Effective Temperature (NET), Wet-Bulb Temperature (WBT), simplified Wet-Bulb Globe Temperature (sWBGT), and Wind Chill Temperature (WCT). Compared with our previous HiTIC-Monthly dataset, the current HiTIC-NCP dataset has improved the temporal resolution from a monthly to a daily scale. It also incorporates different predictors by including new variables such as daily precipitable water vapor, which reflect the high-frequency (day-to-day) fluctuations of atmospheric humidity that induces excessive heat stress as perceived by human body. This study is thus of great significance for regional climate change, urban planning, urban thermal environment, and public health research.

Methods

Overview

The data used in this study mainly include meteorological station observations and four gridded datasets. The Light Gradient Boosting Machine (LightGBM) algorithm proposed by Ke et al.⁴⁸ is used to produce 12 human thermal indices. The workflow of this study is shown in Fig. 1. First, we derived daily HPT indices from daily values of temperature, humidity, and wind speed based on station observations, and preprocessed the daily gridded datasets. We extracted the covariate values at meteorological stations from gridded datasets to form candidate samples (which include station-level HPTs and their co-located covariate values), then randomly selected 80% of the samples as a training set and 20% of that as a validation set. Hyperparameter optimization based on a 5-fold cross-validation was used to determine the optimal parameters of the model by the training set, and a systematic evaluation of model performance was conducted based on the validation set. Finally, a collection of 12 HPTs was generated by the trained LightGBM model. The detailed data and methods are described as follows.

Fig. 1

Framework of generating the daily human thermal index collection over the North China Plain (HiTIC-NCP). HPT: Human-perceived Temperature; LightGBM: Light Gradient Boosting Machine.

Meteorological observations

The station-observed daily data including surface air temperature, relative humidity, and wind speed from 2003 to 2020 were collected from the China Meteorological Data Service Center (CMDSC, http://data.cma.cn). All station data have been strictly quality-controlled based on the method provided by Xu et al.⁴⁹, including homogenizing climate data and processing outliers in the data series. The observational network includes 374 stations across NCP (Fig. 2). In this study, we calculated daily human-perceived temperature from meteorological observations and extracted the covariate values from gridded datasets at the same locations. The data were stored in the form of one per station per day, then processed and divided into training (80%) and validation (20%) sets for model calibration. The calculation of the 12 human thermal indices is described in Table 1.

Fig. 2

Study area and the spatial distribution of meteorological stations in the North China Plain (NCP). The shaded color indicates elevation, and the black dots denote the stations.

Table 1 Equations of the 12 thermal indices.

Gridded datasets

Four gridded datasets with a total of seven independent variables (Table 2) are used as input in the LightGBM model. These variables include land surface temperature (LST), population density, elevation, slope, aspect, precipitable water vapor, and day of the year.

Table 2 Gridded datasets and covariates used in this study.

Table 3 Overall R², MAE, and RMSE values of the 12 human thermal indices over NCP from 2003 to 2020.

LST is closely related to near-surface air temperature⁵⁰. In this study, we adopted an LST dataset named MODIS-like LST⁵¹. This dataset was developed based on the Moderate Resolution Imaging Spectroradiometer (MODIS) LST dataset and the gap-filling algorithm⁵¹ to make it seamless at a global scale in 1 km grids from 2003 to 2020. The average root mean squared error (RMSE) of this LST dataset is 1.80–2.03 °C and 1.23–1.45 °C for daytime and nighttime temperatures, respectively. The daily mean LST was calculated by averaging the corresponding daytime and nighttime LSTs. In addition, population distribution and regional population density are demonstrated to affect regional temperature^52,53,54,55, and we used a high-accuracy population density dataset provided by the WorldPop project team as a predictor⁵⁶.

The Multi-Error-Removed Improved-Terrain (MERIT) Digital Elevation Model (DEM)⁵⁷ with a spatial resolution of 90 m was used in this study to extract slope and aspect. Compared with previous DEM data (e.g., SRTM and AW3D-30m), the errors of MERIT DEM are significantly reduced, and the accuracy in the flat area has been effectively improved^57,58,59. In addition, precipitation has been proven to affect air temperature and humidity^60,61. Therefore, a global daily precipitable water vapor dataset with a spatial resolution of 1 km was used.

The LightGBM machine learning algorithm

LightGBM is a gradient boosting decision tree (GBDT) based machine learning algorithm. Compared with other algorithms based on GBDT such as XGBoost and pGBRT, LightGBM has improved and optimized by the Gradient-based One-Side Sampling (GOSS) and Exclusive Feature Bundling (EFB) technologies, which allow LightGBM to better retain samples with large gradients without significantly altering the original data shape and focus more on undertrained samples⁴⁸. LightGBM consumes much less memory and improves training efficiency by more than 20 times while maintaining nearly the same accuracy⁴⁸. Due to its superior degree of generalizability, extremely fast processing speed, and high accuracy, LightGBM has been widely applied in different fields (e.g., climatology, air pollution, social psychology)^{47,62,63,64,65,66,67}. As shown in Fig. 1, the daily gridded datasets were fed to the trained LightGBM model for HPT prediction.

Hyperparameters tunning

The accuracy of machine learning relies heavily on parameterization⁶⁸. Hyperparameter optimization configures the model by searching for various model hyperparameters and selecting a subset that achieves the best performance on the given dataset. The grid search-based hyperparameters tunning method was adopted in our study⁶⁹.

Accuracy evaluation

Three statistical metrics were used to assess the performance of the training model, including the determination coefficient (R²), root mean absolute error (MAE), and RMSE. The R², RMSE, and MAE values are computed as follows:

$$\begin{array}{c}{R}^{2}=1-\frac{{\sum }_{i=1}^{N}{\left({x}_{oi}-{x}_{pi}\right)}^{2}}{{\sum }_{i=1}^{N}{\left({x}_{pi}-\overline{{x}_{o}}\right)}^{2}}\end{array}$$

(1)

$$\begin{array}{c}RMSE=\sqrt{\frac{1}{N}{\sum }_{i=1}^{N}{\left({x}_{oi}-{x}_{pi}\right)}^{2}}\end{array}$$

(2)

$$\begin{array}{c}MAE=\frac{1}{N}{\sum }_{i=1}^{N}\left|{x}_{oi}-{x}_{pi}\right|\end{array}$$

(3)

where N represents the number of data points, x_p and x_o are the i_th prediction and observation values, respectively.

Data Records

The high spatial resolution (1 km) human thermal index collection at a daily scale from 2003 to 2020 over NCP (HiTIC-NCP) shows a high consistency between observations and predictions, and the product has respectable accuracies. HiTIC-NCP contains 216.rar files, with individual files decompressed into the NetCDF format. The dataset is available on the general-purpose repository Zenodo, at https://doi.org/10.5281/zenodo.7528001⁷⁰. This dataset is stacked by year and each stack consists of daily images by day of the year in the NetCDF format. The unit of the dataset is 0.01 degree Celsius (°C), and the values are stored in an integer data type (Int16) to save storage space, and thus need to be divided by 100 to get the values in degree Celsius when in use. The geographic coordinate system of the dataset is World Geodetic System 1984 (WGS84) Coordinate System. The spatial distribution of the 12 human thermal indices on August 13, 2013, one of the severest heat stress days in the history of NCP, and January 3, 2008, when the sharpest snowstorm struck NCP, are shown as examples in Figs. 3, 4.

Fig. 3

Spatial distribution of the 12 human thermal indices over NCP on August 13, 2013: (a) AT_in, (b) AT_out, (c) DI, (d) ET, (e) HI, (f) HMI, (g) MDI, (h) SAT, (i) sWBGT, (j) WBT, (k) NET, and (l) WCT.

Fig. 4

As Fig. 3 but on January 3, 2008.

Technical Validation

Overall accuracy of 12 predicted human thermal indices

We assessed the overall accuracy of 12 human thermal indices over NCP during 2003–2020 in terms of R², MAE, and RMSE. All indices exhibit desirable performance, with the averaged R², MAE, and RMSE being 0.987, 1.292 °C, and 0.970 °C, respectively. It demonstrates that the LightGBM model has good effectiveness in the prediction of human thermal indices at a daily scale. Figure 5 displays the scatter plots of observations versus predictions of 12 human thermal indices. The closer the scattered points are to the 1:1 line, the better agreement between the predicted and observed values.

Fig. 5

Scatter plots of predicted versus observed daily human thermal indices over NCP from 2003 to 2020: (a) AT_in, (b) AT_out, (c) DI, (d) ET, (e) HI, (f) HMI, (g) MDI, (h) SAT, (i) sWBGT, (j) WBT, (k) NET, and (l) WCT. Colors mean the density of data points, and the black dashed line is the 1:1 line.

The R², MAE, and RMSE values of 12 predicted thermal indices during 2003–2020 are shown in Fig. 6. The R² values range from 0.975 to 0.990, the MAE values are all smaller than 1.4 °C, and the RMSE values are smaller than 1.9 °C (Table 3). Particularly, sWBGT has the highest R² values of 0.990, followed by HMI, WBT, and DI, and its MAE and RMSE values are 0.698 °C and 0.926 °C, respectively. ET is with an R² of 0.987 and has the lowest MAE and RMSE values (i.e., 0.598 °C and 0.785 °C, respectively). WCT and NET exhibit the highest MAEs (i.e., 1.17 and 1.37 °C, respectively) and RMSEs (i.e., 1.61 and 1.84 °C, respectively). It is likely caused by the involvement of wind speed in the calculation of WCT and NET (recall Table 1).

Fig. 6

R², MAE, and RMSE values of 12 predicted human thermal indices over NCP from 2003 to 2020.

Spatial variations of prediction accuracy

Figures 7–9 depict the distributions of R², MAE, and RMSE values in 12 human thermal indices at individual stations of NCP during 2003–2020. As shown in Fig. 7, the high R² values (i.e., ranging from 0.970 to 0.995) are seen at most stations across NCP, especially in the central, southwestern, and northern parts. The R² values of NET at most stations have relatively lower accuracies, likely caused by the inclusion of wind speed. Figures 8, 9 show that ET has the best results with the lowest MAEs and RMSEs (i.e., < 0.911 °C and < 1.161 °C, respectively, see Figs. 8d, 9d). Compared with other 11 human thermal indices, NET has higher MAE and RMSE values (i.e., > 1.058 °C and 1.415 °C, respectively, see Figs. 8k, 9k). Spatially, MAEs and RMSEs exhibit an increasing tendency from low to high elevation, with smaller MAE and RMSE values in inland areas with lower elevation and relatively larger values in mountainous areas with higher elevation.

Fig. 7

Spatial distribution of R² of the 12 predicted human thermal indices at individual meteorological stations of NCP from 2003 to 2020: (a) AT_in, (b) AT_out, (c) DI, (d) ET, (e) HI, (f) HMI, (g) MDI, (h) SAT, (i) sWBGT, (j) WBT, (k) NET, and (l) WCT.

Fig. 8

As Fig. 7 but for MAE.

Fig. 9

As Fig. 7 but for RMSE.

Prediction accuracy in different periods

Figure 10 illustrates the averaged R², MAE, and RMSE values of 12 indices in individual years of 2003–2020. All R² values of 12 indices in all years exhibit well performance (i.e., > 0.955 °C, Fig. 10a). Better annual predictions appear in WBT over 2008–2013 and sWBGT over 2008–2014, with R² exceeding 0.990. The lowest R² is observed in NET in 2003 (i.e., 0.956). Annual averaged MAEs and RMSEs of all indices in nearly all years are less than 1.5 °C and 2.0 °C, respectively (Fig. 10b,c), demonstrating that our product is reliable in all years.

Fig. 10

Prediction accuracies of 12 human thermal indices over NCP in individual years from 2003 to 2020.