Satellite-Ground Data Fusion for Hourly 5-km Gridded Human-Perceived Temperature Estimation in the Yangtze River Basin, China

Highlights

What are the main fundings?

• Twelve human-perceived temperature indices were constructed for the Yangtze River Basin at 5 km hourly resolution by integrating multi-source datasets with the LightGBM model.
• The model demonstrated high accuracy across all indices and effectively captured the spatial heterogeneity and diurnal evolution of the regional thermal environment.

What is the implication of the main finding?

• Provides a reliable high-resolution HPT dataset to support heat-related health risk assessment in densely populated regions.
• Offers an important data foundation for climate adaptation and management strategies in the Yangtze River Basin and other similar regions.

Abstract

Human-perceived temperature (HPT) reflects the synergistic effects of multiple meteorological factors, and its extremes challenge human-managed and natural systems worldwide, especially in densely populated regions such as the Yangtze River Basin of China. However, detailed information on HPT at high temporal (e.g., hourly) and spatial resolution is severely lacking. In this study, we conduct a collaborative inversion for 12 HPT indices at a ~5 km spatial resolution and an hourly temporal resolution in the Yangtze River Basin from multi-source data (e.g., Himawari-8 images, meteorological stations, ERA5-Land reanalysis, and DEM data) using the LightGBM model. The model exhibited high predictive accuracy across all indices, achieving an average coefficient of determination (R²) of 0.981, root mean square error (RMSE) of 1.150 °C, and mean absolute error (MAE) of 0.860 °C. These results aligned well with observational data across spatial and temporal scales, effectively capturing the spatial heterogeneity and diurnal evolution of the region’s thermal environment. Our research provides a reliable data foundation for heat-health risk assessment and regional climate adaptation strategies.

1. Introduction

Since the 1950s, extreme heat (e.g., heat waves) over much of the world’s land mass has become more frequent and more intense under global warming [1,2,3]. Climate warming and the frequent occurrence of extreme weather events worsen human thermal discomfort and heighten exposure to unfavourable thermal environments. This, in turn, leads to adverse consequences for public health [4], socio-economics [5,6] and agricultural production [7]. The number of people affected by heat waves continues to rise, with >70,000 deaths recorded in 2003 [8] and >55,000 deaths in 2010 [9]. Heat waves can also cause other health risks, such as increasing morbidity [10,11,12]. Between 2010 and 2020, an average of 13,262 premature births per year in China were caused by heat waves [13]. In addition, heat waves can cause lower crop yields [14,15], traffic injuries [16], and lower worker productivity affecting economic activity [17].

The changes and effects of human thermal stress have received extensive attention due to their strong correlation with public health [18,19,20]. For example, the Universal Thermal Climate Index (UTCI) was used to investigate the spatial patterns of changes in human heat stress levels in South Asia [21]. The annual trends in urban thermal comfort across 183 Chinese cities from 1990 to 2016 were comprehensively assessed under rapid urbanization and climate change [22]. Eight heat stress indicators of global climate models participating in Phase 6 of the Coupled Model Comparison Program (CMIP6) were calculated and evaluated [23]. These studies are mainly based on weather station data and use low spatiotemporal resolution grid data. In high-altitude or complex terrain areas, the distribution of meteorological stations is sparse and uneven, which hinders the continuous display of temperature distribution and thermal stress trends [24]. Therefore, studying human thermal stress at higher spatial and temporal resolutions is increasingly essential, especially in regions with sparse meteorological data.

Many indicators have been proposed to evaluate and quantify the human-perceived temperature, varying widely in complexity and application [25,26,27,28]. Some indicators are based on the principles of human heat exchange, while others rely on empirical relationships derived from human responses to environmental factors such as temperature, humidity, and wind speed [25,26]. Certain indicators serve distinct purposes: the heat index (HI) is primarily used in meteorological services, wet-bulb temperature (WBT) assesses the physiological upper limit of humans [27]; while physiologically equivalent temperature (PET) and universal thermal climate index (UTCI) are widely employed to evaluate human thermal comfort [28]. Thus, collaborative research on multiple human-perceived temperature indices is crucial to increase the applicability of thermal stress assessments. For example, a study proposed constructing a daily high-resolution human-perceived temperature index dataset for the North China Plain, demonstrating high spatiotemporal consistency with ground observations [29]. To enhance understanding of the impact of heat stress on public health, travel, and work, multiple human-perceived temperature index datasets available from South and Asia (spanning 1981 to 2019) were developed. These datasets enable researchers and practitioners to study the spatiotemporal evolution of perceived temperatures and their effects on densely populated regions at a finer scale [30].

Studies have generated various human thermal stress datasets, such as ERA5-HERT [31], HDI [32], HiTiSEA [30], and HiTIC-Monthly [33]. ERA5-HERT is derived from ECMWF’s ERA5 reanalysis and includes two global hourly human thermal stress indices: Mean Radiative Temperature (MRT) and the Universal Thermal Climate Index (UTCI) [31]. HDI is a dataset generated through the quality-controlled meteorological variables from the Global Land Data Assimilation System (GLDAS) [32]. HiTiSEA is a 0.1° × 0.1° grid, incorporating multiple thermal stress indices based on the most recent ERA5-LAND and ERA5 reanalysis products for South and East Asia, covering the period from 1981 to 2019 [30]. HiTIC-Monthly is a high spatial resolution (1 km) human thermal stress index dataset covering the mainland of China at a monthly scale [33]. Previous studies have typically produced these datasets at either low temporal or spatial resolution. To conduct more detailed climate change analyses, there is an urgent need for high-resolution datasets that include multiple thermal stress indices across both spatial and temporal scales.

As the largest river in both Asia and China, the Yangtze River originates in the Tanggula Mountains of Qinghai Province and flows into the East China Sea. The main flow of the Yangtze River passes through 11 provinces, and its tributaries also spread across eight provinces, totalling 19 provincial-level administrative regions. The temperature patterns in the Yangtze River Basin are shaped by several factors, including solar radiation, East Asian atmospheric circulation, the vast topography of the Qinghai–Tibet Plateau, the North Pacific Ocean, and the varying regional topographic conditions. The Yangtze River Basin covers 18.8% of China’s total land area and supports over 40% of the nation’s population and economy. Studies have shown that the highest incidence and intensity of heat waves are concentrated in southern China [34]. Heat wave intensity is highest in the middle and lower reaches of the Yangtze River. After 2000, the frequency of heat waves significantly increased, affecting larger areas with greater intensity. In the monthly human heat index dataset of China from 2003 to 2020, it is statistically found that almost all examined human-perceived temperature indices show an increasing trend [33]. In the densely populated regions of the Yangtze River Basin, high-resolution studies on human-perceived temperature are essential to accurately analyze the spatiotemporal changes in heat waves and cold waves.

This study involves hourly collaborative inversion and monitoring of multiple human-perceived temperature indices in the Yangtze River Basin for the year 2020, conducted at a spatial resolution of 5 km × 5 km. The study utilizes 12 commonly used human-perceived temperature indices derived from the inversion process, including air temperature (TEM), indoor apparent temperature (ATin), outdoor apparent temperature (ATout), discomfort index (DI), effective temperature (ET), heat index (HI), humidity index (HMI), modified discomfort index (MDI), net effective temperature (NET), wet bulb temperature (WBT), simplified wet bulb globe temperature (sWBGT), and wind chill temperature (WCT).

2. Data and Methods

2.1. Satellite, Reanalysis and Ground Observation Data

This study utilizes data from three key sources: (1) Himawari-8 satellite data obtained from the JAXA Himawari Monitor (P-Tree System), which provides high-frequency (10 min) observations across East Asia and the Western Pacific with spatial resolutions of 2 km and 5 km, featuring 16 spectral bands from visible to infrared. Its Advanced Himawari Imager (AHI) has demonstrated utility in temperature estimation [35], solar radiation modelling [36], and land surface temperature (LST) inversion [37], with this study specifically employing thermal infrared (TIR) data from bands 11, 13, 14, and 15; (2) ERA5-Land from ECMWF, offering hourly, high-resolution (0.1° × 0.1°) global land surface data since 1950, including key variables like surface temperature, surface net solar/thermal radiation, and downward surface radiation fluxes, which have been used to derive human-perceived temperature indices; (3) Digital Elevation Model (DEM) data from the CMA CLDAS-V2.0 near real-time dataset (https://data.cma.cn/, accessed on 10 March 2025), originally sourced from NASA and NIMA’s global 30 m survey and resampled to 0.0625° resolution for Asia using an area-weighted method.

The hourly meteorological station data collected in 2020, including air temperature, relative humidity, and wind speed, are sourced from the National Meteorological Center of the China Meteorological Administration. Before utilizing the site data, rigorous quality control measures were applied, including the identification and correction of outliers as well as the treatment of missing values. The spatial arrangement of the 715 stations across the Yangtze River basin can be observed in Figure 1.

2.2. Method

This study develops an hourly human-perceived temperature estimation model by integrating multi-source datasets, including Himawari-8 satellite data, meteorological station observations, ERA5-land reanalysis, and DEM. To ensure spatial consistency, all datasets were uniformly interpolated into the 5 km resolution grid of the Himawari-8 satellite product using the bilinear interpolation method. The target values are derived from station observations using empirical formulas, while selected feature variables (

Table 1. Input feature variables in the LightGBM model.

Data Types	Abbreviation	Content	Unit
Virtual variable	Lat	Latitude	°
	Lon	Longitude	°
	Alti	Altitude	m
	Mon	Month of year	-
	Day	Day of year	-
	Hour	Hour of year	-
Himawari-8	Tbb_11	Brightness temperature (Band 11)	K
	Tbb_13	Brightness temperature (Band 13)	K
	Tbb_14	Brightness temperature (Band 14)	K
	Tbb_15	Brightness temperature (Band 15)	K
ERA5-land	PRS	Surface pressure	Pa
	V	Wind speed	m/s
	SKT	Skin temperature	K
	SSR	Surface net solar radiation	J/m2
	STR	Surface net thermal radiation	J/m2
	SSRD	Surface solar radiation downwards	J/m2
	STRD	Surface thermal radiation downwards	J/m2
	E	Total evaporation	m (water equivalent)

) were spatially matched with each of the 715 stations’ hourly human-perceived temperatures in the Yangtze River Basin. These paired datasets formed the candidate sample set for modelling. We adopted an 80:20 split for training and test sets, with the training set further subjected to 10-fold cross-validation (10-CV) to optimize parameters [38]. In 10-CV, the training data was divided into 10 equal parts, iteratively using each part for validation while the remaining nine trained the model. The average performance metrics (R², RMSE, MAE) across folds guided the selection of the optimal LightGBM algorithm (implemented via Scikit-Learn). The final model, trained on the full training set, was evaluated on the independent test set to assess its accuracy in generating high-resolution estimates for thermal environment assessment (Figure 2).

In this study, the hourly human-perceived temperature was calculated using meteorological station data, including air temperature (TEM), relative humidity (RH), wind speed (WS), and derived actual vapor pressure ($E_a$), which was computed from TEM and RH rather than direct observations. The human-perceived temperature indices were determined hour by hour based on these variables (

Table A1. Computation of 12 human-perceived temperature indices.

Name	Human-Perceived Temperature	Computation Equation	Reference
ATin	Apparent Temperature (indoors)	$ATin=-1.3+0.92\times TEM+2.2\times E_a$	Steadman [49]
ATout	Apparent Temperature (outdoors, in the shade)	$ATout=-2.7+1.04\times TEM+2\times E_a-0.65\times WS$	Steadman [50]
DI	Discomfort index	$DI=0.5\times WBT+0.5\times TEM$	Epstein et al. [51]
ET	Effective Temperature	$ET=TEM-0.4\times \left(TEM-10\right)\times \left(1-0.001\times RH\right)$	Gagge et al. [52]
HI	Heat Index	$HI=-8.784695+1.61139411\times TEM$ $-2.338549\times RH-0.14611605\times TEM\times RH$ $+2.211732\times {10}^{-3}\times {TEM}^2\times RH+7.2546\times {10}^{-4}$ $\times TEM\times {RH}^2-1.2308094\times {10}^{-2}\times {TEM}^2$ $-1.6424828\times {10}^{-2}\times {10}^{-2}\times {RH}^2$ $+3.582\times {10}^{-6}\times {TEM}^2\times {RH}^2$	Anderson et al. [53]
HMI	Humidex	$HMI=TEM+0.5555\times \left(0.1\times E_a-10\right)$	Masterton and Richardson [54]
MDI	Modified DiscomfortIndex	$MDI=0.75\times WBT+0.38\times TEM$	Moran et al. [55]
NET	NetEffective Temperature	$NET=37-{\displaystyle \frac{37-TEM}{0.68-0.0014\times RH+\frac{1}{\left(1.76+1.4\times {WS}^{0.75}\right)}}}$ $-0.29\times TEM\times \left(1-0.01\times RH\right)$	Houghton [56]
sWBGT	Simplified Wet-bulb Temperature	$sWBGT=0.567\times TEM+0.0393\times E_a+3.94$	Willett and Sherwood [57]
TEM	Surface Air Temperature	-	-
WBT	Wet-bulb Temperaturfe	$WBT=TEM\times atan\left(0.151977\times {\left(RH+8.313659\right)}^2\right)$ $+atan\left(TEM+RH\right)-atan\left(RH-1.676331\right)$ $+0.00391838\times {RH}^{1.5}\times atan\left(0.02301\times RH\right)$ $-4.686035$	Stull [58]
WCT	Wind Chill Temperature	$WCT=13.12+0.6215\times TEM-11.37$ $\times {\left(WS\times 3.6\right)}^{0.16}+0.3965\times TEM\times {\left(WS\times 3.6\right)}^{0.16}$	Osczevski and Bluestein [59]

TEM is the air temperature (°C), RH is the relative humidity (%), WS is the wind speed (m/s), and $E_a$ is the actual water vapor pressure (kPa). All human-perceived temperature indices are measured in degree Celsius (°C).

). To evaluate the accuracy of the LightGBM model, three statistical metrics were used: the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE). R² (ranging from 0 to 1) measures the proportion of variance in the dependent variable explained by the model, with higher values indicating better fit. RMSE reflects the overall prediction error and is sensitive to outliers, while MAE provides the average absolute error. Lower RMSE and MAE values indicate greater model accuracy.

$$E_s=6.112\times \mathit{exp}\left({\displaystyle \frac{17.67\times TEM}{TEM+243.5}}\right)$$

(1)

$$E_a={\displaystyle \frac{RH}{100}}\times E_s$$

(2)

$$MAE={\displaystyle \frac{1}{N}}\times {\sum }_{i=1}^N|y_i-\widehat{y}|$$

(3)

$$RMSE=\sqrt{{\displaystyle \frac{1}{N}}\times {\sum }_{i=1}^N{\left(y_i-\widehat{y}\right)}^2}$$

(4)

$$R^2=1-{\displaystyle \frac{{\sum }_{i=1}^N{\left(y_i-\widehat{y}\right)}^2}{{\sum }_{i=1}^N{\left(y_i-\overline{y}\right)}^2}}$$

(5)

where $\widehat{y}$ is the predicted value of human-perceived temperature indices, $\overline{y}$ is the mean of the observed human-perceived temperature indices calculated from meteorological stations, and $N$ is the number of samples.

We used LightGBM to estimate perceived temperature in the Yangtze River Basin in 2020. Developed by Microsoft Research, LightGBM is an efficient gradient-boosted decision tree (GBDT) algorithm with strong performance in machine learning tasks [39]. It has been applied in fuel consumption prediction [40], landslide analysis [41], financial risk assessment [42], evapotranspiration estimation [43], air quality prediction [44], and PM_2.5 reconstruction [45]. LightGBM accelerates training via Gradient-based One-Side Sampling (GOSS) and Exclusive Feature Bundling (EFB) [46], which improve efficiency by excluding low-gradient data and reducing feature dimensionality, respectively. Compared to traditional GBDT methods (e.g., XGBoost), it maintains accuracy while significantly cutting training time [39,47].

3. Results

3.1. Accuracy Assessment of Spatiotemporal LightGBM Model

The model fitting and 10-CV accuracy of the LightGBM model are summarized in

Table 2. Performance evaluation of time-, site-, and sample-based LightGBM models using 10-fold cross-validation and test datasets.

Indices	Site-Based				Time-Based				Sample-Based
	Model CV		Model Tested		Model CV		Model Tested		Model CV		Model Tested
	R2	RMSE	R2	RMSE	R2	RMSE	R2	RMSE	R2	RMSE	R2	RMSE
ATin	0.976	1.282	0.976	1.504	0.987	1.168	0.984	1.272	0.987	1.176	0.986	1.179
ATout	0.969	1.556	0.969	1.846	0.986	1.317	0.983	1.428	0.986	1.328	0.985	1.325
DI	0.978	1.056	0.978	1.221	0.988	0.957	0.985	1.030	0.988	0.964	0.987	0.963
ET	0.950	1.161	0.950	1.228	0.968	1.019	0.964	1.070	0.967	1.026	0.967	1.025
HI	0.961	1.368	0.961	1.556	0.978	1.252	0.973	1.343	0.978	1.261	0.976	1.257
HMI	0.968	1.331	0.968	1.573	0.985	1.177	0.980	1.299	0.985	1.187	0.982	1.187
MDI	0.978	1.195	0.978	1.350	0.987	1.099	0.985	1.165	0.987	1.105	0.987	1.102
NET	0.920	2.302	0.920	2.715	0.984	1.334	0.979	1.435	0.984	1.343	0.982	1.342
sWBGT	0.967	0.778	0.967	0.901	0.983	0.702	0.978	0.768	0.983	0.708	0.981	0.707
TEM	0.967	1.350	0.976	1.137	0.983	1.215	0.978	1.331	0.983	1.227	0.981	1.226
WBT	0.980	1.998	0.979	1.146	0.988	0.922	0.986	0.971	0.988	0.927	0.987	0.925
WCT	0.939	2.211	0.939	2.442	0.978	1.556	0.973	1.679	0.978	1.567	0.976	1.564

, based on site-based, time-based, and sample-based 12 human-perceived temperature indices. The fitting accuracy of all human-perceived temperature indices using the LightGBM model across different cross-validation methods exceeds 0.920, indicating that the model achieves a good fit to the data. The accuracy of the time-based method is slightly lower than that of the sample-based method, while the site-based cross-validation shows the lowest accuracy among the three methods. These results indicate that while the LightGBM model performs well overall, its accuracy varies depending on the cross-validation method used.

Table 3. LightGBM model fit, model 10-fold cross-validation, and model test results for different temperature indices.

Indices	Model Fitted			Model CV			Model Tested
	R2	RMSE	MAE	R2	RMSE	MAE	R2	RMSE	MAE
ATin	0.987	1.126	0.846	0.987	1.176	0.879	0.986	1.176	0.882
ATout	0.986	1.268	0.953	0.986	1.328	0.993	0.985	1.325	0.992
DI	0.988	0.923	0.694	0.988	0.964	0.721	0.987	0.963	0.721
ET	0.967	1.018	0.763	0.967	1.026	0.768	0.967	1.025	0.768
HI	0.978	1.207	0.916	0.978	1.261	0.953	0.976	1.257	0.951
HMI	0.985	1.111	0.829	0.985	1.187	0.879	0.983	1.187	0.879
MDI	0.987	1.078	0.816	0.987	1.105	0.834	0.987	1.102	0.832
NET	0.984	1.273	0.951	0.984	1.343	0.995	0.982	1.342	0.994
sWBGT	0.983	0.675	0.504	0.983	0.708	0.526	0.981	0.707	0.525
TEM	0.983	1.169	0.875	0.983	1.227	0.912	0.981	1.226	0.912
WBT	0.988	0.906	0.690	0.988	0.927	0.704	0.987	0.925	0.704
WCT	0.978	1.525	1.136	0.978	1.567	1.164	0.976	1.564	1.161

shows the model fitting results, 10-fold cross-validation results and testing results of 12 human-perceived temperature indices in 2020. A scatter plot comparing observed versus estimated human-perceived temperature values is shown in Figure 3. The points in the scatter plot cluster closely around the 1:1 line, indicating strong agreement between the observed and estimated values. As shown in the table, the model fitting accuracy for the 12 human-perceived temperature indices ranges from 0.967 to 0.988 in R², 0.906 to 1.525 °C in RMSE, and 0.504 to 1.136 °C in MAE. The 10-CV accuracy R² values range from 0.967 to 0.988, with RMSE values between 0.708 and 1.567 °C, and MAE values between 0.526 and 1.164 °C. For testing accuracy, R² values range from 0.976 to 0.987, RMSE ranges from 0.707 to 1.564 °C, and MAE ranges from 0.525 to 1.161 °C. The top-performing indices include sWBGT, WBT, DI, ET and MDI (RMSE ≈ 0.7–1.1 °C). These indices are largely smooth and monotonic functions of air temperature and humidity, and our features capture these drivers well (SKT as a temperature indicator, E as a near-surface moisture indicator, Tbb as cloud/radiation indicators, and temporal-spatial priors such as Hour, Mon, and Lat), leading to stable mappings. The intermediate-performance group comprises ATin, HMI, TEM, HI, ATout, and NET (RMSE ≈ 1.18–1.34 °C). These indices are more sensitive to humidity and wind and often contain polynomial terms or interaction effects. However, the current feature set lacks direct humidity variables (e.g., RH, $E_a$), and reliance on indirect proxies (E, Tbb, radiation fluxes) introduces structural noise. Additionally, ATout is dependent on wind speed, and the representativeness of grid-based data for station observations further compromises its prediction accuracy. The relatively weakest index is WCT (RMSE ≈ 1.56 °C), which is dominated by infrequent cold and windy regimes (class imbalance) and includes nonlinear wind exponents, making it highly sensitive to wind representativeness.

Annual and seasonal estimation accuracy, quantified through R², RMSE, and MAE metrics, is displayed in Figure 4. The annual estimation accuracy is higher than the seasonal estimates, with the highest R² values overall. Among the seasons, R² values are consistently lowest in summer and highest in autumn for each index. However, except for the HI, the remaining human-perceived temperature index does not show the highest RMSE and MAE values in the four seasons in summer, but the R² value is not the highest in spring. All human-perceived temperature indices have R² greater than 0.88 in all seasons and RMSE and MAE values less than 1.8 °C. Overall, these accuracy indicators demonstrate that the estimation quality of the 12 individual perceived temperature indices is satisfactory.

3.2. Spatiotemporal Accuracy of LightGBM in Human-Perceived Temperature Estimation

The spatial distribution of R², RMSE and MAE at each station in the Yangtze River Basin is shown in Figure 5. All 12 indices exhibit high R² values (>0.96), demonstrating the strong performance of our LightGBM model. Stations with higher estimation accuracy are primarily in the eastern Yangtze River Basin, particularly in the Yangtze River Delta. In contrast, stations in the western part of the basin, such as those on the Qinghai–Tibet Plateau and Yunnan-Guizhou Plateau, show relatively lower R² values (<0.975). The R² values of HI, ET and WCT were lower than 0.985 in almost all sites, which was relatively low in accuracy compared with other indices. In terms of RMSE and MAE values, all indices show relatively small values (<2.0 °C) at most stations across the Yangtze River Basin. WCT has the highest RMSE and MAE values, followed by ATin and ATout, while sWBGT shows the smallest RMSE and MAE values (<0.8 °C). From the west to the east of the Yangtze River Basin, the RMSE and MAE values of all indices have decreased. In general, stations with lower RMSE and MAE values are located in plain regions such as the Yangtze River Delta and the Sichuan Basin. Conversely, stations with higher RMSE and MAE values are found in areas with complex terrain, such as the higher-altitude Qinghai–Tibet Plateau and the mountainous Yunnan-Guizhou region. This difference may be related to the uneven distribution of meteorological stations, that is, there are more meteorological stations in plain areas and fewer in areas with complex terrain. In summary, RMSE and MAE values decrease from high-altitude to low-altitude areas. Lower-altitude regions exhibit smaller RMSE and MAE values, whereas higher-altitude mountainous areas and plateaus tend to have relatively higher values.

Figure 6 shows hour-by-hour accuracy of 12 human-perceived temperature indices for the whole year of 2020, including R², RMSE, and MAE values. RMSE and MAE values were less than 1.6 °C and R² values were greater than 0.97 at almost all times throughout the year. WCT exhibited higher RMSE and MAE values throughout the year. The RMSE and MAE values for DI, ET, sWBGT, and WBT remained relatively stable across the 24 h period, while other indices showed a significant increase in RMSE and MAE values between 2:00 and 14:00 UTC. This increase in RMSE and MAE values may be due to the significant temperature fluctuations during this period, corresponding to sunrise and the transition from night to day. DI, MDI, and WBT all showed better estimation accuracy at 24 times, i.e., R² was over 0.98, and ET and sWBGT have RMSE and MAE less than 1.2 °C at all times. This shows that the estimation of the 12 human-perceived temperature indices is highly reliable across all hours.

3.3. Spatial Variation in Human-Perceived Temperatures

The evaluation results demonstrate that the LightGBM model is capable of high-precision estimation of human-perceived temperatures at a regional scale. Based on this finding, the model was employed to conduct hourly seamless estimations of 12 human-perceived temperature indices across the Yangtze River Basin for the year 2020. To illustrate the model’s performance, we analyzed the spatiotemporal evolution patterns of these indices at diurnal and hourly timescales. The LightGBM model can perform highly accurate estimation of human-perceived temperature across the Yangtze River Basin, where 12 human-perceived temperature index distributions are presented on 1 January (winter), 1 April (spring), 1 July (summer), and 1 October (autumn), 2020. Overall, the 12 indices show a similar distribution pattern, with temperatures generally decreasing from west to east, with the lowest and highest temperatures in the northwest and southeast, respectively. For all indices, temperatures decrease with increasing altitude. The Tibetan Plateau consistently has the lowest temperatures, while the Sichuan Basin and the eastern part of the Yangtze River Basin record the highest temperatures. In different seasons, different indices also show different rules. In winter (refer to Figure 7), the overall temperature of WCT index is the highest, and the temperature in most areas is above 15°, while the overall temperature of HMI and WBT index is lower, and the temperature in most areas is below 15°. In addition, Figures S1–S3 in the Supplementary Information show the distribution of human-perceived temperature indices in spring, summer and autumn. The distribution patterns of spring and winter were similar, while the ATin and ATout indices were generally higher in summer and autumn, while the HMI and sWBGT indices were generally lower.

Figure 8 presents the spatial distribution of the hourly evolution of annual mean TEM across the entire Yangtze River Basin in 2020. The spatial distribution of perceived temperature at different times reveals similar regional patterns. Overall, temperatures were lower in the western region and higher in the eastern region, primarily due to the decreasing altitude from west to east, as temperatures generally increase with decreasing elevation. Across the 24 time periods, significant temperature differences are observed between 0:00 and 11:00 UTC and 12:00–23:00 UTC. Temperature begins to rise from 0:00 UTC, peaking at 8:00 UTC, primarily due to solar radiation reaching the surface. Notably, the Tibetan Plateau experiences a temperature peak throughout this period. This peak lasts until 11:00 UTC, after which the sun’s altitude decreases, solar radiation weakens, and surface heat is gradually lost, leading to a decline in temperature that stabilizes by 23:00 UTC. During the night, temperature changes are minimal, likely due to the absence of solar radiation, as the surface gradually releases stored heat, resulting in relatively stable conditions. The supplementary shows the spatial distribution of the other 11 indices during the hourly evolution of average temperature across the whole Yangtze River Basin in 2020 (Figures S4–S14). Overall, these indices exhibit similar spatial patterns and temporal variation trends. All indices can effectively capture temperature changes across 24-h intervals, indicating that the trained LightGBM model is reliable for estimating human-perceived temperature in the Yangtze River Basin. These data not only provide high-resolution spatial information on perceived temperature dynamics, but also offer a valuable basis for further analysis of heat wave and cold wave patterns at finer spatiotemporal scales.

3.4. SHAP Analysis for LightGBM Model Interpretability

Understanding how a model’s predictions are affected by features can help professionals identify potential problems and improve model performance. To address this, an effective approach is to use the SHAP (SHapley Additive exPlanations) theory proposed by Lundberg and Lee [48], which demystifies the behaviour of a machine learning model by calculating the contribution of each feature to the model estimate. The main goal of SHAP is to provide a transparent, easy-to-understand explanation for each estimate of the model, which is particularly suitable for tree-based models.

Figure 9 illustrates the SHAP feature interpretation for LightGBM model predicting TEM. This graph visually represents how different input features influence the model’s predictions. Each dot in the plot corresponds to the SHAP value of a specific feature for an individual sample, indicating the extent to which that feature influences the model’s prediction. Features are listed along the y-axis in descending order of overall importance, while the x-axis displays the range of SHAP values. The color gradient from red to blue reflects the original feature values from high to low. To better visualize data density, overlapping points are slightly dispersed along the vertical axis, illustrating the distribution of SHAP values for each feature. SKT is identified as the most critical feature influencing TEM prediction, with higher values corresponding to larger positive SHAP values. This indicates that an increase in skin temperature substantially raises the predicted TEM. Alti exerts a significant negative influence, as higher elevations are associated with lower predicted temperatures, aligning well with the precious analysis in this study. Hour and Mon also exhibit strong effects, highlighting the essential roles of diurnal and seasonal variations in temperature prediction. PRS and STRD contribute notably to the model, likely due to their close association with near-surface thermodynamic processes. Spatial features such as Lon and Lat also show considerable importance, suggesting the model effectively captures regional geographic variations in temperature. Other radiation-related variables, including SSRD, STR, and SSR, demonstrate moderate importance, indicating that both shortwave and longwave radiation variability influence TEM predictions to some extent. In contrast, satellite-derived brightness temperature variables (e.g., Tbb_11, Tbb_13, Tbb_14, and Tbb_15) show relatively low importance, possibly due to their indirect effects or redundancy with other features.

4. Conclusions

In this study, we integrate Himawari-8 satellite data, meteorological station observations, ERA5-Land reanalysis, and DEM data to estimate and invert 12 hourly human-perceived temperature indices (ATin, ATout, DI, ET, HI, HMI, MDI, NET, sWBGT, TEM, WBT, and WCT) for the year 2020 across the Yangtze River Basin at ~5 km spatial resolution, using the LightGBM model. The estimated results show that all indices achieved high accuracy (average R² = 0.9815, RMSE = 1.1499 °C, and MAE = 0.8599 °C) and demonstrated strong agreement with the observations across both spatial and temporal scales.

In the Yangtze River Basin, all 12 human-perceived temperature indices exhibit an overall increasing trend from west to east, with higher values observed at lower elevations. Additionally, the indices tend to decrease with increasing latitude. High-altitude regions, such as the Qinghai–Tibet Plateau and the Yunnan-Guizhou Plateau, consistently record significantly lower temperatures. SHAP-based interpretability analysis further indicates that TEM predictions are primarily influenced by SKT, Alti, Hour, Mon, PRS, and STRD. This study is of great significance for monitoring the spatiotemporal variations in human heat stress and for the refined assessment of health risks under extreme heat conditions.