Table 1 lists the information of the required 15 datasets for the three main steps in the generation of ANCFDS-LST (namely, LST angular normalization, all-weather LST estimation, and spatial downscaling). The LST dataset produced in this study spans the period from 2018 to 2023, covering the full operational phase of the FY-4A LST product (note that the official FY-4A LST product has not been generated since 4 March 2024, as it has been superseded by a subsequent operational satellite). First, FY-4A and MODIS directional LSTs (i.e., 2 datasets) were used to generate angular-normalized nadir LST (Tnadir) and hemispherical LST (Themi). Next, 10 remote sensing and reanalysis products were employed to drive the generation of hypothetical clear-sky LST and the application of CRF correction, for estimating all-weather Tdir, Tnadir, and Themi at a 0.05° spatial resolution. Finally, these three 0.05° all-weather LST products were downscaled using 0.01° regression kernels obtained from 3 datasets, including MYD11A1, GTOPO30 DEM, and ERA5 Land.

For the first step (i.e., LST angular normalization), the FY-4A directional LST (LST_FY) and MODIS LST (LST_MxD) products (i.e., datasets 1–2 in Table 1) were jointly utilized to calibrate the TEKDM model. Based on the calibrated model, clear-sky Tnadir and Themi were derived using the corresponding kernels and kernel coefficients. The official FY-4A LST is retrieved using a SW algorithm, with land surface emissivity (LSE) estimated via NDVI-based threshold method (Dong et al., 2013, 2023a). The MODIS MxD11A1 LST was retrieved using a generalized split-window (GSW) algorithm, incorporating land-cover-based LSE as input (Wan and Dozier, 1996). Both FY-4A and MODIS LST products were resampled to the same spatial resolution (i.e., 0.05°) using simple averaging before the joint estimation. Pixels with a view zenith angle (VZA) greater than 70° were masked (Freitas et al., 2013). To reduce systematic discrepancies between these two datasets, the 0.05° FY-4A LST was linearly adjusted to match the 0.05° MODIS MxD11A1 LST using a linear transformation (i.e., LSTMxD=a×LSTFY+b). The parameters a and b were determined based on nighttime matchups under the condition of VZA < 50°, VZA difference < 5°, and LST difference < 5 K, referenced to Ermida et al. (2017, 2018a). A total of 132 972 698 matchup pairs were collected from 1 January 2018 to 31 December 2023 in the disk of FY-4A, resulting in a slope of a=1.0240 and an intercept of b=-5.7378. It should be noted that all subsequent cloud-filling and downscaling procedures were performed based on the bias-corrected FY-4A LST, and thus the final released products are provided on the MODIS-referenced LST scale.

The second step (i.e., all-weather LST estimation) includes the estimation of hypothetical clear-sky LST and the CRF correction. In the hypothetical clear-sky LST estimation, datasets 3–7 in Table 1 – including the MCD12Q1 IGBP land cover type (LCT) product (dataset 3), the Köppen-Geiger climate type (CT) product (dataset 4), the Global LAnd Surface Satellite (GLASS) fractional vegetation cover (FVC) product (dataset 5), the GTOPO30 DEM data (dataset 6), the ERA5-Land 2 m air temperature (T2m), and dew-point temperature (D2m) data (dataset 7) – were employed to characterize surface properties (Wei et al., 2019). Moreover, datasets 8–12 in Table 1 were further used to calculate three additional variables to depict surface thermal conditions – including the ATC-simulated LST, the surface incoming radiation (Rin), and hypothetical clear-sky skin temperature (TERA5-skin).

The ATC model has three parameters as shown in Eq. (1), which was calibrated using the MxD11A1 LST product as input (dataset 8). In total, four ATC models were obtained since TERRA and AQUA MODIS sensors provide observations during both day and night. A 3×3 spatial median filter was applied to fill invalid ATC parameters caused by persistent cloud coverage. Then, the LST values simulated using the calibrated ATC model were used in the estimation of hypothetical clear-sky LST. These ATC-simulated LSTs are denoted as TATC,d1, TATC,d2, TATC,n1, and TATC,n2, where the subscripts “d” and “n” refer to daytime and nighttime data, respectively, and “1” and “2” indicate TERRA and AQUA platforms. 1TATC(DOY)=MAST+YAST×cos⁡2π365(DOY-dx) where TATC(DOY) is the ATC-simulated LST as a function of day-of-year (DOY). MAST is the annual mean surface temperature, YAST is the yearly temperature amplitude, and dx is the phase shift. The QC band was filtered to ensure high-quality inputs for the ATC model calibration.

Rin is the sum of absorbed shortwave and longwave radiation as calculated in Eq. (2), with the input of surface downward shortwave radiation (SDSR_clr), land surface albedo (Albedo), surface downward longwave radiation (SDLR_clr) and broad band emissivity (BBE, εbb). Both SDSR_clr and SDLR_clr under hypothetical clear-sky condition are extracted from ERA5 products (dataset 9). The Albedo is from the GLASS product for 2018–2022 (dataset 10) and supplemented by MCD43C3 for 2023 (dataset 11). A Harmonic ANalysis of Time Series (HANTS) method was employed to smooth and fill the gaps in MCD43C3 product in this study (Zhou et al., 2022, 2023). The BBE is generated from MUlti-source data SYnergized Quantitative (MuSyQ) remote sensing system (dataset 12) (Li et al., 2019). Datasets 9–12 were resampled to a 0.05° resolution to match LST_FY. 2Rin=(1-Albedo)⋅SDSRclr+εbb⋅SDLRclr Based on the Stefan–Boltzmann law, the ERA5 hypothetical clear-sky surface thermal radiation – e.g., surface downward longwave radiation (SDLR_clr) and surface net thermal radiation STR_clr – could be used to calculate the hypothetical clear-sky skin temperature (TERA5-skin) at a spatial resolution of 0.25° to depict temporal dynamics of LST. It was then physically downscaled to a 0.05° resolution using DEM data, assuming a temperature lapse rate (K) of 6.5 K km⁻¹ (Minder et al., 2010). Therefore, the TERA5-skin was calculated by the following equation: 3TERA5-skin=SDLRclr-STRclr-1-εbbSDLRclrεbbσs4+K⋅DEMi-DEMm where σs is the Stefan–Boltzmann constant 5.67×10-8. The DEM_i and DEM_m refer to the elevation of the 0.05° pixel and the mean elevation of the corresponding 0.25° pixel, respectively.

After generating the hypothetical clear-sky LST, a CRF correction is required to convert it into cloudy-sky LST. This correction is physically based on the surface energy balance differences between clear-sky and cloudy conditions. Specifically, cloud-free and cloudy SDSR from ERA5 reanalysis data (dataset 9), together with albedo data (dataset 10 and 11), are used to describe the shortwave radiation budget. For the longwave radiation budget, ERA5 cloud-free and cloudy-sky SDLR data (dataset 9), BBE data (dataset 12), the hypothetical clear-sky LST and cloudy-sky LST are required. Then, the only unknown variable, cloudy-sky LST (expressed as the sum of hypothetical clear-sky LST and the CRF correction value) can be retrieved by solving the SEB equation. More details will be given in Sect. 3.2.

In the third step (i.e., spatial downscaling), a new algorithm combining existing kernel- and fusion-based methods was proposed to downscale LST to the target time, consisting of two main processes. First, a 0.05° gap-free high-resolution LST at the initial time is required. In this study, TATC,d2 simulated by the ATC model (dataset 13) is resampled as the initial LST. Second, the 0.05° LST difference between the initial and target time is predicted using an ML method. The input features include TATC,d2; AQUA MODIS ATC parameters for both day (MAST_d2, YAST_d2, and dxd2) and night (MAST_n2, YAST_n2, and dxn2), derived from dataset 13; DEM data (dataset 14); and ERA5-Land T2m (dataset 15). To apply this spatial downscaling model at the 0.01° scale, the T2m data were further downscaled to 0.01° resolution using high-resolution DEM and the temperature lapse rate (K) of 6.5 K km⁻¹. Finally, the gap-free high-resolution LST at the target time is obtained by adding the estimated 0.01° LST difference to 0.01° TATC,d2.

For cross-validation of the generated Tnadir, the VIIRS VNP21A1 LST data (https://lpdaac.usgs.gov/product_search/, last access: 16 June 2026) in 2020 was adopted as reference. The VIIRS instrument onboard the Suomi National Polar-orbiting Partnership (S-NPP) satellite provides clear-sky TIR observations at 13:30 LT (local time). The VNP21 LST was retrieved using a TES algorithm using observations from the M14–M16 bands as input. Recent temperature-based (T-based) evaluation showed that its RMSE is 1.79 K during nighttime and 2.79 K during the day (Na et al., 2024a). Here, VNP21 LST values with the VZA < 5° were employed for cross-evaluation of Tnadir as done by Wei et al. (2025). Meanwhile, the MODIS-referenced FY-4A LST product was temporally interpolated to match the exact observation time of VIIRS LST. The resulting near-nadir VNP21 LST was first resampled to a spatial resolution of 0.05°, and then bias-corrected to MODIS-referenced FY-4A LST using nighttime matchups (consistent with the bias correction applied between MODIS and FY-4A LST products). A total of 12 379 672 matchup pairs were collected over the full disk of FY-4A in 2020, yielding a regression slope of 0.9478 and an intercept of 13.9510. This relationship was subsequently applied to the daytime VIIRS near-nadir LST. After bias correction, the root mean square difference (RMSD) between the VIIRS LST and MODIS-referenced FY-4A LST decreased from 6.21 to 4.53 K, while the mean bias difference (MBD) was reduced from -4.04 to -1.85 K. Finally, the RMSD, MBD, and coefficient of determination (R2) were used as three evaluation indicators for the cross-validation. The extracted VNP21 LST results for four typical days are shown in Fig. 1. Only several narrow strips were extracted due to the limitation of VIIRS field-of-view (VZA < 5°). VNP21 LST shows significant seasonal variation. It is relatively lower in 1 March and 1 December 2020 than that in 1 June and 1 September 2020 in the northern hemisphere and the situation is reversed in the southern hemisphere.

T-based evaluation (i.e., directly comparing satellite-derived LST with in situ LST) is the most widely used validation method and should be performed whenever possible (Guillevic et al., 2018; Li et al., 2014, 2023a; Na et al., 2024a). The in situ pyrgeometer measures both upward and downward hemispherical longwave radiation, which can be converted into hemispherical LST to perform the T-based validation for the normalized Themi products. As shown in Fig. 2, 5 sites from the Heihe Watershed Allied Telemetry Experimental Research (HiWATER) experiment within the Heihe River Basin (HRB) and 3 sites from the Terrestrial Ecosystem Research Network (TERN) OzFlux network in 2020 were selected, which have been used in LST evaluation studies (Beringer et al., 2016; Che et al., 2019; Li et al., 2025, 2020). More detailed information on the in situ sites is listed in Table 2. In addition, the median standard deviation (SD) of Landsat 8 LST within a 0.05° window and the RMSE between nighttime FY-4A LST and in situ LST are also provided, as these two metrics were used to assess the spatial representativeness of the selected validation sites (see Sect. 5.3).

As listed in Table 2, five land cover types are represented across the 8 validation sites: 1 cropland (CRO) site, 3 grassland (GRA) sites, 1 barren/sparse vegetation (BSV) site, 2 savanna (SAV) sites, and 1 evergreen broadleaf forest (EBF) site. The temporal resolution of the in situ observations varies by network: 10 min for the HiWATER sites, and 30 min for the OzFlux sites. Linear interpolation was applied to align in situ measurements with the exact satellite observation time. The in situ LST (Tinsitu) is calculated using the Stefan–Boltzmann law, as shown in Eq. (4): 4Tinsitu=SULRinsitu-1-εbbSDLRinsituεbbσs4 where SULR_insitu and SDLR_insitu are the surface upward and downward longwave radiation measured by the in situ pyrgeometer, respectively. εbb is the broadband emissivity derived from the FY-3B MuSyQ product. Outliers were identified and removed using the “3σ-Hampel identifier” method for each site (Davies and Gather, 1993; Pearson, 2002). The RMSE, mean bias error (MBE), and R2 were used as evaluation metrics to validate the LST products.

The FY-4A LST disk has different viewing geometries for each pixel, and these values need to be normalized to a reference direction before subsequent applications. The KDM offers a solution by simulating the angular distribution of LST through a linear combination of several kernel functions. However, traditional KDM typically requires three or four simultaneous clear-sky LST observations to calibrate the unknown kernel coefficients, which cannot be satisfied by FY-4A satellite. To enhance its applicability, Qin et al. (2023, 2025) proposed a time-evolving KDM (TEKDM) by coupling a diurnal temperature cycle (DTC) model (i.e., depicting LST temporal variation) and a KDM model (i.e., depicting LST angular variation) as shown in Eq. (5). Based on this TEKDM model, Na et al. (2024b) further estimated daytime nadir LST using GOES-16 and MODIS LST products, and obtained a significantly improved accuracy. 5Tdirt,θs,θv,Δφ=fiso(t)×1+α⋅KGapFractionθv+β⋅cos⁡θs⋅KHotspotθs,θv,Δφ,W where Tdir(t,θs,θv,Δφ) represents the directional LST observed at the local time of t, with the solar zenith angle (SZA) of θs, the VZA of θv, and the relative azimuth angle (RAA) of Δφ. The term fiso(t) is the isotropic kernel coefficient. The coefficients α and β are unknowns corresponding to the gap fraction kernel (KGapFraction) and hotspot kernel (KHotspot), respectively. W is another unknown parameter related to the hotspot width. The temporal dynamics of isotropic kernel coefficient (i.e., fiso(t)) during daytime can be effectively modeled using a DTC model as given in Eq. (6): 6fiso(t)=T0+Ta×cos⁡πω⋅t-tm where T0 is the temperature at sunrise, Ta represents the amplitude of daily LST variation, ω is the length of the daytime period, tm is the local time at which the LST reaches its maximum. The nighttime LST was not processed here, as the angular effect is generally negligible during nighttime. In this study, the LSF kernel and the Chen kernel were selected to represent KGapFraction and KHotspot respectively, as shown in Eqs. (7)–(9). 7KGapFractionθv=KLSFθv=1+2cos⁡θv0.96+1.92cos⁡θv-14⋅cos⁡θv1+2cos⁡θv+0.15⋅1-e-0.75cos⁡θv-1.03048KHotspotθs,θv,Δφ,W=KChenθs,θv,Δφ,W=e-ξπ⋅W9ξ=arccos⁡cos⁡θv⋅cos⁡θs+sin⁡θv⋅sin⁡θs⋅cos⁡Δφ where ξ is the angular distance between the viewing direction and the solar direction. The TEKDM involves seven unknown parameters to be calibrated: T0, Ta, ω, tm, α, β, and W, therefore, at least seven clear-sky daytime observations are required as inputs. To enhance the angular information content of the input data, two daytime clear-sky MODIS observations (i.e., TERRA and AQUA) were utilized. Then, the TEKDM can be solved under the condition with ≥5 clear-sky FY-4A observations. A “trust-region-reflective” algorithm was employed for the non-linear optimization of Eq. (5), with initial values and parameter boundaries set according to Qin et al. (2025). Then, a 17 d moving average (i.e., covering 8 d before and after the target day) was applied to the parameters of α, β, and W for each pixel for two purposes: filling gaps in pixels with fewer than seven observations, and reducing the impact of outliers in the estimated TEKDM parameters. As shown in Eq. (10), the LST at any direction (i.e., the Tdir(t,θs,θv,Δφ)) can be simulated with the input of FY-4A direction LST – i.e., Tdir(t,θs,θvFY,ΔφFY) – and averaged α, β, and W parameters. 10Tdirt,θs,θv,ΔφTdirt,θs,θvFY,ΔφFY=1+α⋅KLSFθv+β⋅cos⁡θs⋅KChenθs,θv,Δφ,W1+α⋅KLSFθvFY+β⋅cos⁡θs⋅KChenθs,θvFY,ΔφFY,W Specifically, the nadir LST can be derived when setting θv=0° as shown in Eq. (11): 11Tnadirt,θs,0°,Δφ=Tdirt,θs,θvFY,ΔφvFY⋅1+α⋅KLSF(0°)+β⋅cos⁡θs⋅KChenθs,0°,Δφ,W1+α⋅KLSFθvFY+β⋅cos⁡θs⋅KChenθs,θvFY,ΔφvFY,W. Correspondingly, the normalized hemispherical LST (Themi) could be calculated through integration as shown in Eq. (12). 12Themit,θs=1π⋅∫02π∫0π/2Tdir4t,θs,θv,Δφ⋅cos⁡θv⋅sin⁡θvdθvdφv14=Tdirt,θs,θvFY,ΔφvFY⋅1π⋅∫02π∫0π/21+α⋅KLSFθv+β⋅cos⁡θs⋅KChenθs,θv,Δφ,W4⋅cos⁡θv⋅sin⁡θvdθvdφv141+α⋅KLSFθvFY+β⋅cos⁡θs⋅KChenθs,θvFY,ΔφvFY,W To accelerate the computation of Themi, the integral in the numerator of Eq. (12) was approximated using a third-order polynomial fit based on the input variables of α, β, W, θs and KChen at nadir. More details about the integration procedure are provided in Appendix A. As shown in Fig. A1, this polynomial approximation substantially reduces computational cost while introducing an uncertainty of less than 0.1 K.

The all-weather LST estimation method consists of two main sub-steps. First, a hypothetical clear-sky LST was predicted using a Bayesian optimization-based CatBoost model. Second, the CRF correction was applied to estimate the cloudy-sky LST, ultimately producing cloud-filled 0.05° all-weather directional, nadir, and hemispherical LST.

ML-based LST reconstruction methods have been widely used because they can accurately simulate the non-linear relationship between the hypothetical clear-sky LST and vegetative, meteorological, and topographical parameters (Li et al., 2024a; Ma et al., 2024a; Zhang et al., 2024). The CatBoost is a recently developed model based on the traditional gradient-boosting decision tree (GBDT) framework with several enhancements: (1) ordered target statistics, which encode categorical features without target leakage by using permutation-based strategies; (2) ordered boosting, which minimizes gradient bias by employing permutation-driven training to reduce overfitting; (3) oblivious trees, which apply the same splitting criterion across all nodes at each level, thereby improving speed and regularization (Prokhorenkova et al., 2018). These improvements enhance the model's robustness and efficiency, especially when handling large-scale datasets containing categorical variables, and have proven effective in LST reconstruction studies (Dai et al., 2025). Here, the CatBoost model was used to capture the complex non-linear relationship (i.e., fclr) between hypothetical clear-sky LST and the auxiliary parameters, as represented in Eq. (13). 13Tclr=fclrDOY,θs,Lat,Lon,LCT,CT,FVC,DEM,TATC,d1,TATC,d2,TATC,n1,TATC,n2,Rin,TERA5-skin,T2m,D2m where the input features include the geolocation and temporal parameters, such as the DOY, θs, latitude (Lat), longitude (Lon). The other parameters in Eq. (13) have been introduced in Table 1. It should be noted that the reanalysis variables are introduced to provide spatial information that is not captured by satellite observations, rather than to replace satellite information. The primary observational constraint remains the official FY-4A LST products. The clear-sky LST values remain unchanged and also act as the training labels for cloudy-sky LST estimation.

The target labels consisted of daytime Tdir, Tnadir, and Themi, along with nighttime FY-4A LST from 1 January 2018 to 31 December 2023. A total of six ML models were trained on a year-by-year basis for each target temperature, rather than using a single model across multiple years. Each model is applied exclusively to data from its corresponding year, without temporal extrapolation to other periods. Considering the large volume of label data, a two-step approach was employed to determine the number and geolocation of sample points as introduced in Appendix B. As shown in Fig. B1, 37 000 locations distributed over the FY-4A disk were selected for each year. Then, the clear-sky LST and corresponding input features were extracted to train the CatBoost model in Eq. (13). To better align with the 0.25° spatial resolution of ERA5, a FY-4A pixel was strictly considered as clear-sky only when all surrounding pixels within a 5×5 neighbourhood were under clear-sky conditions. Six CatBoost models were trained for each year from 2018 to 2023. The average number of extracted samples per year was 138 817 085, which were randomly divided into training, testing, and Bayesian optimization sets in a ratio of 6:3:1. The Bayesian optimization method adjusted the CatBoost hyperparameters by reconstructing the posterior distribution of the cost function, defined as the average fitting RMSE from two-fold cross-validation. After the ML training, the CatBoost model can predict the hypothetical clear-sky Tdir, Tnadir, and Themi under cloudy-sky conditions, which requires CRF correction.

The core of the CRF correction lies in establishing the relationship between net radiation changes and CRF-induced LST changes, based on the surface radiation budget equation. The CRF-corrected LST is equal to the sum of hypothetical clear-sky LST with CRF-induced LST variation (i.e., ΔTCRF). Considering that our aim is not to fully resolve the SEB equation, a simplified analytical correction equation (Eq. 14) proposed by Zhang et al. (2024) was adopted for the CRF correction of Themi. 14γ⋅ΔTCRF=ΔRin-εbbσsTclrhemi+ΔTCRF4-Tclrhemi4 where ΔTCRF is the correction value for hemispherical LST; γ is an energy transfer parameter, which could be estimated by Eq. (15). A temporal median filter with a 17 d window and a spatial median filter with a 3×3 window were applied to replace outlier values in the calculation of γ. ΔRin is the change in surface incoming radiation induced by the CRF effect, which could be estimated by Eq. (16); Tclrhemi is predicted hypothetical clear-sky hemispherical LST using Eq. (13). 15γ=Rintn-Rintsr-εbbσsTclrhemitn4-Tclrhemitsr4Tclrhemitn-Tclrhemitsr16ΔRin=(1-Albedo)⋅SDSRclr-SDSRcld+εbb⋅SDLRclr-SDLRcld where Rin is the clear-sky surface incoming radiation calculated by Eq. (2). tn and tsr denote the noon time and sunrise time. The subscripts “clr” or “cld” indicate that the parameter is under hypothetical clear-sky or actual cloudy-sky conditions. After the determination of γ and ΔRin, the only unknown parameter ΔTCRF in Eq. (14) can be estimated. Here, it was analytically solved through neglecting its quartic and cubic terms (see Eq. 17). 176εbbσsTclrhemi2ΔTCRF2+γ+4εbbσsTclrhemi3ΔTCRF-ΔRin=0 After the pixel-by-pixel estimation of ΔTCRF, the mean (μCRF) and standard deviation (σCRF) of ΔTCRF were calculated for each image. To reduce the impact of extreme values, the maximum and minimum limits of ΔTCRF were set to μCRF±3σCRF. In the end, the cloudy-sky Tdir, Tnadir, and Themi are obtained by summing the CatBoost-estimated hypothetical clear-sky LST (i.e., using Eq. 13) and ΔTCRF (i.e., using Eq. 17). It should be noted that Themi represents the physically most defensible all-weather LST quantity, whereas cloudy-sky Tdir and Tnadir are model-derived estimates based on two independent assumptions: the assumed hypothetical clear-sky LST–ancillary variable relationship and the application of the same CRF correction derived for hemispherical LST.

To meet the 1 km spatial resolution and 1-hour temporal resolution requirement of GCOS, downscaling methods have garnered increasing attention for enhancing the spatial textures of low-resolution LST products (e.g., FY-4A LST). Fusion-based downscaling approaches typically require high-resolution LST data as input, whereas kernel-based methods require LST-related auxiliary data (i.e., regression kernels). Combining fusion-based and kernel-based downscaling methods could achieve higher accuracy than using either method alone (Li et al., 2023b; Xia et al., 2019). Dong et al. (2023b) recently proposed a simple and effective downscaling (SED) method that utilizes clear-sky, high-resolution Landsat 8 LST at an initial time to downscale low-resolution MODIS data at a target time. However, this requirement is difficult to satisfy over the full-disk region due to the frequent presence of clouds. To address this limitation, an IHDA method is proposed by leveraging the ATC-fitted spatiotemporal trend surface of LST (Liang et al., 2025). First, the ATC model is adopted to estimate gap-free LST at the initial time of 13:30 LT (TATC,d2) using Eq. (1) with the driven data of MYD11A1 LST products. Although the ATC model may introduce additional uncertainty, recent studies have shown that incorporating meteorological data can further reduce the modeling residuals (Liu et al., 2019; Yang et al., 2024). Here, a CatBoost model is employed to establish the non-linear relationship (i.e., fdownscale) between the resampled 0.05° regression kernels and the 0.05° LST bias (i.e., ΔLST0.05=T0.05-TATC,d2, where T0.05 is the 0.05° all-weather LST at the target time to be downscaled) as shown in Eq. (18). 18ΔLST0.05=fdownscaleTATC,d2,MASTd2,YASTd2,dxd2,MASTn2,YASTn2,dxn2,T2m,DEM The ATC parameters of daytime (13:30 LT) and nighttime (01:30 LT) MODIS LST were included as regression kernels to characterize surface thermal properties. Additionally, ERA5-Land air temperature (i.e., T2m) was incorporated to capture cloud information, whereas the DEM was introduced to represent the topographic information. For model training, two typical days (i.e., the mid-month of January in winter and July in summer) were initially used to tune the hyperparameters of the CatBoost model using a Bayesian optimization approach. Then, the fdownscale was established individually for each image, covering 157 680 images in total (i.e., 6 years × 365 d × 24 h × 3 types of LST). For each image, 70 % of the samples were used to train the CatBoost model, while the remaining 30 % were reserved for testing. The 0.01° LST difference (ΔLST0.01pred) was then predicted using the trained CatBoost model with 0.01° regression kernels as input: 19ΔLST0.01pred=fdownscaleTATC,d20.01,MASTd20.01,YASTd20.01,dxd20.01,MASTn20.01,YASTn20.01,dxn20.01,T2m0.01,DEM0.01 where the “0.01” means the input features are in 0.01° resolution. Finally, a residual redistribution process was applied to produce the final ΔLST_0.01 as shown in Eq. (20). 20ΔLST0.01=ΔLST0.01pred+ΔLST0.05-ΔLST0.05pred0.01 where ΔLST_0.05 represents the difference between the 0.05° all-weather LST and the ATC-simulated LST. ΔLST0.05pred means the 0.05° LST difference obtained by aggregating the predicted ΔLST0.01pred. The second term on the right-hand side of Eq. (20) is the CatBoost modeling residual to be redistributed. A 3×3 median filter was applied to reduce the impact of extreme outliers in the residuals. The notation (⋅)0.01 indicates the process that the 0.05° residual (i.e., ΔST0.05-ΔLST0.05pred) was resampled to 0.01° using the widely used bilinear interpolation. Finally, the downscaled 0.01° all-weather directional, nadir, and hemispherical LST (T0.01) could be calculated by summing the 0.01° ATC simulated LST at initial time (TATC,d2 (0.01°)) and the Eq. (20) estimated ΔLST_0.01 as below. 21T0.01=TATC,d2(0.01°)+ΔLST0.01

Figure 4 shows the spatial distributions of FY-4A directional LST (Tdir), FY-4A VZA, normalized nadir LST (Tnadir), the LST difference between Tnadir and Tdir, hemispherical LST (Themi), and the difference between Themi and Tdir on 24 June 2020 at 03:00 UTC (i.e., the Beijing time of 11:00 BJT). As shown in Fig. 4c, Tnadir is 1.5 K larger than Tdir on average. It closely resembles Tdir (with a small value of Tnadir-Tdir) when the VZA is within 40°, and becomes significantly higher than Tdir as VZA increases, with a correction value of Tnadir-Tdir exceeding 5 K near the edge of the FY-4A disk as shown in Fig. 4d. When the VZA is small, the observation angle of FY-4A is close to the nadir direction, resulting in a small correction value from Tdir to Tnadir. However, more cool vegetated elements were viewed at large VZA (i.e., the well-known gap fraction effect), which resulted in a much lower Tdir than Tnadir and further led to a larger angular correction value. For hemispherical LST, the Themi (Fig. 4e) is only 0.2 K higher than Tdir on average. It is approximately 2 K lower than Tdir (Fig. 4a) in Australia, while being very close to Tdir in northern China. At the edge of the FY-4A disk, Themi is around 3 K higher than Tdir (where the VZA exceeds 60°, see Fig. 4b and f). This spatial pattern can be partially explained by the value of the hemispherical equivalent angle (i.e., the VZA at which Tdir=Themi), which has been reported to range from 44 to 53° in relevant studies (Hu et al., 2023; Zhang et al., 2025). When the VZA is smaller than 44°, Tdir tends to be higher than Themi because more hot soil components are viewed. Conversely, for large VZAs over 53° (e.g., at the edge of FY-4A disk), Tdir becomes lower than Themi as more cool vegetation components are viewed.

There will be significant inconsistencies in directional LST between FY-4A and MODIS due to the TRD effect. Figure 5a quantifies this discrepancy by presenting the RMSD between FY-4A and MOD11A1 directional LST. After correcting both LST products to nadir values, the corresponding RMSD is shown in Fig. 5b. The average RMSD across the full FY-4A disk decreases from 3.26 to 2.12 K, representing a reduction of 1.14 K (35.0 %). The extremely high values near the western edge of the disk are substantially reduced after correction. However, relatively high values over the western Qinghai–Tibetan Plateau persist, which may be partly attributed to limitations of the employed LSF-Chen model, as it is derived under the assumption of flat terrain. Overall, the TEKDM demonstrates strong performance in reducing angular inconsistencies between LST products as expected.

The cross-validation between FY-4A Tdir, Tnadir, and the VNP21A1 near-nadir LST (i.e., the VZA < 5°) in 2020 is shown in Fig. 6. The root mean squared difference (RMSD) and mean bias difference (MBD) are 4.53 and -1.85 K for Tdir as shown in Fig. 6a. After angular normalization, the RMSD and MBD for Tnadir is 2.56 and 0.06 K as shown in Fig. 6b, with a 1.97 K (i.e., 43.5 %) reduction in RMSD and an almost unbiased MBD. Figure 6c shows the RMSD values of Tdir and Tnadir across different VZA intervals with a step of 10°. The RMSD difference between them remains within 0.2 K when VZA is less than 40°. However, when VZA exceeds 40°, the RMSD for Tdir increases substantially (from 3.4 to 6.5 K), while the RMSD for normalized Tnadir remains lower than 3.0 K for large VZA. The higher RMSD of Tdir at large VZA can be partially attributed to its significant underestimation as shown in Fig. 6d. The MBD of Tdir exceeds -3.8 K when the VZA is between 60 and 70°, whereas that of the normalized Tnadir is 0.9 K. The gap fraction effect could explain this angular dependence of MBD, i.e., a larger VZA leads to a smaller Tdir compared with nadir LST.

The evaluation of 0.05° FY-4A Tdir and Themi during daytime over 8 stations is shown in Fig. 7a and b. Results show that the RMSE decreased from 2.22 to 2.00 K, with an improvement of 0.22 K. The MBE changed from 0.19 to -0.54 K. For the temperatures between 280 to 300 K with the highest density, the scatter points become closer to the 1:1 line after the angular normalization. Figure 7c and d shows the RMSE and MBE of FY-4A Tdir and Themi at different local times. Tdir exhibits a pronounced overestimation with higher RMSE in the morning, and an underestimation with smaller RMSE in the afternoon. After the angular normalization, the Themi shows significantly improved performance in the morning, with an RMSE reduced by approximately 0.4–0.8 K (Fig. 7c). As shown in Fig. 7d, the MBEs of Themi are much smaller than those of Tdir in the morning, with a maximum reduction of 1.7 K at 11:00 LT.

The prediction accuracy of fclr in reconstructing hypothetical clear-sky Tdir, Tnadir, and Themi over the test set from 2018 to 2023 is summarized in Table 3. The average annual number of samples is 41 312 444. The RMSE ranges from 2.22 to 2.46 K for Tdir, from 2.28 to 2.52 K for Tnadir, and from 2.23 to 2.46 K for Themi. RMSE values remain relatively stable from 2018 to 2021, followed by an increase of approximately 0.2 K during 2022–2023. On average, the RMSE for Tdir is 0.06 K lower than that for Tnadir, and is the same as that for Themi, indicating comparable fitting performance across different CatBoost models. The MBE is zero for all models, indicating that no significant systematic bias was introduced in the predicted hypothetical clear-sky LST.

Spatial heterogeneity in LST can lead to variations in model performance. The pixel-level RMSE of the CatBoost model for 2020 was evaluated by comparing the CatBoost-predicted hypothetical nadir LST with the normalized FY-4A nadir LST. The nadir configuration is well suited for quantifying the performance of ML model, because it minimizes systematic uncertainties associated with angular effects. As shown in Fig. 8, the mean RMSE of the cloud-filling results is 2.49 K. Higher uncertainties are observed over regions such as the Qinghai–Tibet Plateau and the Indonesian Peninsula, which can likely be attributed to persistent cloud cover. In such regions, the reduced availability of clear-sky observations limits the training samples, leading to increased uncertainty in the machine learning model performance.

The estimated 0.05° all-weather Themi results were further evaluated by the in situ hemispherical LSTs from 8 in situ stations (Fig. 9). The samples are divided into clear-sky (Fig. 9a) and cloudy-sky (Fig. 9b) conditions. The “3σ-Hampel identifier” method was applied to minimize the influence of outliers at each site. Both daytime and nighttime Themi values were considered here. Under clear-sky conditions, 0.05° Themi shows an RMSE of 2.48 K and an MBE of -0.83 K. The distribution of residuals (ΔLST=Themi-Tinsitu) approximately follows a normal distribution. Under cloudy-sky conditions, Themi shows a higher RMSE of 3.65 K and a more significant negative MBE of -1.43 K. It is challenging to maintain the same level of accuracy under cloudy-sky conditions, as uncertainties in both the CatBoost-predicted hypothetical LST and the estimated ΔTCRF are introduced.

Figure 10a and b shows the temporal consistency of FY-4A official LST, Themi, and in situ LST at the DM and SDQ sites throughout 2020 at 04:00 and 16:00 UTC (i.e., 12:00 and 00:00 BJT). Overall, the all-weather Themi agrees well with the in situ LST. At the DM site during daytime (Fig. 10a), the clear-sky RMSE (MBE) of FY-4A official LST is 3.20 K (2.43 K), whereas the corresponding values for Themi are 2.15 K (0.46 K), indicating that the systematic overestimation is significantly mitigated after angular normalization. Similarly, at the SDQ site during daytime (Fig. 10b), the clear-sky RMSE (MBE) decreases from 4.77 K (4.02 K) for the FY-4A official LST to 3.60 K (2.55 K) for Themi. During nighttime, the clear-sky Themi is the same as FY-4A official LST, showing lower RMSE values than during daytime. The RMSE (MBE) values are 2.38 K (-0.94 K) at the DM site and 2.2 K (-0.17 K) at the SDQ site. Under cloudy-sky conditions, the RMSE (MBE) values are 3.54 K (-0.53 K) at the DM site and 4.09 K (1.97 K) at the SDQ site during daytime, and 4.58 K (-3.77 K) at the DM site and 3.46 K (-1.36 K) at the SDQ site during nighttime. In summary, the all-weather LST could effectively capture the characteristics of the annual temperature cycle.

In the spatial resolution downscaling process, the CatBoost model was used to establish the relationship between regression kernels and the LST difference (e.g., the satellite LST minus ATC-simulated LST) at 0.05° resolution. The prediction accuracy of each month over the test set from 1 January 2018 to 31 December 2023 for daytime and nighttime conditions is shown in Fig. 11. The nighttime RMSE ranges from 1.8 to 2.4 K, which is consistently lower than the daytime RMSE (2.6–3.1 K). This discrepancy can be partially attributed to the fact that daytime LST tends to be more heterogeneous than nighttime LST, leading to greater modeling errors during the day. The daytime RMSE increases from January to May and then decreases. In contrast, the nighttime RMSE decreases from January to July and then rises. The RMSE difference among directional, nadir, and hemispherical LST are within 0.2 K. Overall, the results demonstrate the reliable capability of the CatBoost model in the spatial resolution downscaling.

The spatial distribution of prediction accuracy for the machine learning model during the downscaling process is presented in Fig. 12, where the RMSE was calculated between the CatBoost-estimated LST and the input all-weather nadir LST. The overall RMSE is 2.28 K, with higher values predominantly observed in mountainous regions, likely due to enhanced surface heterogeneity and elevation-induced variability. It should be noted that clear-sky pixels remain unchanged in the all-weather LST product; therefore, the ANCFDS-LST for these pixels is affected only by uncertainties introduced during the downscaling step. In contrast, for cloud-contaminated pixels, the final LST incorporates uncertainties arising from both the cloud-filling and downscaling processes.

Downscaled LST is expected to exhibit enhanced spatial texture, which is quantitatively evaluated in this study. The local variance ratio (LVR) and the structural similarity index (SSIM) were employed as metrics to compare the spatial textures of bilinearly interpolated LST (derived from 0.05° all-weather LST) and the downscaled 0.01° nadir LST. Values of LVR and SSIM closer to 1 indicate a higher similarity between the reconstructed and reference spatial textures. The reference texture was reconstructed using the ATC model, which also provided the prior thermal texture information during downscaling. Therefore, the LVR and SSIM metrics mainly evaluate the consistency and preservation of the prescribed ATC-based spatial texture (Jia et al., 2024), rather than serving as an independent assessment of sub-pixel LST accuracy. As shown in Fig. 13, the median LVR increases from 0.41 to 0.97, while the median SSIM improves from 0.83 to 0.91, indicating that the downscaled LST better preserves the imposed spatial texture characteristics than bilinear interpolation.

T-based validation results for the all-weather Themi product before and after downscaling are presented in Fig. 14. The RMSE remained unchanged at 3.34 K for both the 0.05 and 0.01° resolutions, while the MBE changed only marginally from -1.27 to -1.28 K. These nearly identical validation statistics suggest that the downscaling procedure introduces minimal additional uncertainty. This consistency can be attributed to two main factors. First, the downscaling framework employs a residual redistribution strategy, which preserves the large-scale thermal information of the original 0.05° LST product. As a result, the accuracy of the downscaled 0.01° Themi product remains strongly constrained by the performance of the original coarse-resolution LST. Second, the validation sites were rigorously screened to ensure strong spatial representativeness at both the 0.05 and 0.01° scales. Consequently, the retrieved LSTs at the two spatial resolutions are expected to exhibit comparable validation accuracy.

The spatial distributions of all-weather Tnadir and Themi at 0.05 and 0.01° resolutions on 24 June 2020 at 04:00 UTC (12:00 BJT) are presented in Fig. 15. Two representative regions are highlighted for detailed comparison: the western Tibetan Plateau (red rectangles in Fig. 15a–d) and the Greater Khingan Mountains (blue rectangles in Fig. 15a–d). As shown in Fig. 15a–d, the large-scale spatial patterns of LST are well preserved after downscaling, with both Tnadir and Themi exhibiting clear latitudinal and elevational gradients (i.e., lower temperatures are consistently observed in high-latitude and high-altitude regions). Themi integrates directional LST values over all viewing angles and thus generally shows smoother spatial variability than Tnadir. At finer scales (Fig. 15e–l), the downscaled results (0.01°) reveal substantially enhanced spatial details compared to the coarse-resolution counterparts. In particular, complex terrain features become more distinguishable, such as ridge–valley structures over the Tibetan Plateau and the heterogeneous forest–grassland patterns in the Greater Khingan Mountains. The contrast between sunlit and shaded slopes, as well as between higher and lower elevations, is more pronounced in the downscaled LST fields. These improvements in spatial texture and terrain-related variability demonstrate the effectiveness of the IHDA downscaling method in capturing fine-scale thermal heterogeneity.

In this study, a temporal aggregation method spanning 8 d before and after the target day (a total of 17 d) was employed to aggregate the KDM parameters (i.e., α, β, and W). It not only filtered out extreme values in the TEKDM results but also filled the gaps for pixels where TEKDM could not be solved. The 17 d compositing method was comprehensively compared with three other aggregation approaches: (1) 9 d composition: averaging values from four days before and after the target day (a total of 9 d). (2) Monthly composition (i.e., ∼30 d in total): calculating the monthly average of KDM parameters, assuming they remain constant throughout each month. (3) Yearly composition (i.e., ∼365 d in total): calculating the yearly average of KDM parameters, assuming they remain constant throughout the whole year. After calibrating the KDM parameters using each approach, FY-4A directional LST could be normalized to Tnadir. Then, the 2020 Tnadir results were cross-validated using VNP21A1 near-nadir LST (i.e., VZA < 5°), following the same procedure as in Fig. 6.

Figure 16a–e shows the cross-validation results between the Tdir/Tnadir and VNP21A1 near-nadir LST, whereas Fig. 16f shows the valid pixel fraction (i.e., the ratio of pixels with valid KDM parameters to the total number of clear-sky land surface pixels) for the above temporal aggregation methods. The RMSD (MBD) of the FY-4A official LST is 4.53 K (-1.85 K), while the RMSD (MBD) for the four compositing methods ranges from 2.53 to 2.66 K (-0.12 to 0.07 K). The RMSD values in ascending order are: monthly method < 17 d method < 9 d method < yearly method. In terms of valid pixel fractions, the order is the 9 d method (89.3 %) < 17 d method (93.6 %) < monthly method (95.9 %) < yearly method (99.5 %). The significant lower valid pixel fractions for the 9 and 17 d methods in August 2020 (Fig. 16f) are attributed to the absence of MYD11A1 LST products between 16 and 31 August 2020 (induced by a Formatter-Multiplexer Unit/Solid State Recorder (FMU/SSR) error, more details can be found at https://atmosphere-imager.gsfc.nasa.gov/issues/l1b#Issue01, last access: 16 June 2026). The yearly method was not adopted since it has the largest RMSD, whereas the 9 d method was not used since it has the lowest valid pixel fraction. The 17 d and monthly methods could achieve comparable and acceptable accuracy. The 17 d method was ultimately selected with reference to parameter aggregation approaches commonly used in the optical remote sensing domain (Liu et al., 2013).

The relative feature importance of the CatBoost model was determined based on each feature's contribution to the reduction of the loss function. We employed the logarithmic value of (1 + scorenorm) to quantify the importance of each model (where score_norm is the max-min normalized feature scores). The average relative feature importance across all CatBoost models is shown in Fig. 17. In the estimation of hypothetical clear-sky LST (Fig. 17a), TERA5-skin exhibited the highest importance. The T2m and Rin were also identified as influential predictors, aligning with the findings of Zhang et al. (2024). The θs accounted for a significant 9.9 % of the importance, reflecting the strong influence of solar radiation on LST. Most other features had relatively low importance (below 5 %), and ATC_d1, CT, FVC, and LCT contributed less than 1 %. In the LST downscaling task from 0.05 to 0.01° using CatBoost model (Fig. 17b), the ATC-simulated LST at the initial time had the highest importance of 30.8 %. The air temperature (T2m) and MAST_d2 also had high importance exceeding 10 %. The YAST_d2, DEM, and dxn2 showed moderate importance between 5 % and 10 %, while the dxd2, MAST_n2, and YAST_n2 contributed less than 5 %. Eliminating those features with low importance has the potential to enhance computational efficiency and reduce dependency on related auxiliary datasets, which should be carefully assessed in future studies.

Tnadir was evaluated against near-nadir LST from VNP21A1, whereas the 0.05° Themi product was validated using in situ measurements. The T-based validation of Themi requires in situ sites with high spatial representativeness. To this end, the spatial heterogeneity of 11 HiWATER sites, 11 OzFlux sites, and 6 sites over the Tibetan Plateau in 2020 was assessed (Beringer et al., 2016; Che et al., 2019; Ma et al., 2024b). Following Zhang et al. (2024), two metrics were adopted: (1) the median standard deviation (SD) of Landsat 8 LST within a 0.05° window, and (2) the RMSE between nighttime FY-4A LST and in situ LST. Sites that substantially deviated from the median values of these two indicators were excluded, as illustrated in Fig. 18. Ultimately, 8 out of 28 candidate sites were retained, all satisfying the criteria of nighttime RMSE < 3.57 K and SD < 2.17 K.

The inter-comparison between the generated all-weather Themi product and four existing all-weather LST products against in situ LST measurements in 2020 is shown in Fig. 19. Although the absolute accuracy is influenced by the spatial representativeness of in situ sites as analyzed in Sect. 5.3, the relative accuracy of different LST products is comparable. Two of the employed datasets are reanalysis products: ERA5-Land LST at 0.1° resolution and China Land Surface Data Assimilation System (CLDAS) LST at 0.0625° resolution. Additionally, two remote sensing products were included: a global all-weather 0.05° LST dataset generated by Jia et al. (2023), and a regional 0.02° LST product for East Asia region developed by Dong et al. (2022). Jia's method first estimates hypothetical clear-sky LST using a Kalman filter with ERA5, Himawari-8 and MODIS LSTs as inputs. An iterative CRF correction is then applied to produce the cloudy-sky LST. Dong's approach first retrieves clear-sky LST from Himawari-8 using a TES algorithm, and then estimates cloudy-sky LST using a multiresolution Kalman filter based on CLDAS LST. For consistency, all four products were resampled to 0.05° resolution to match the 0.05° Themi for inter-comparison, and only HiWATER sites were used due to the limited spatial extent of CLDAS LST. Under clear-sky conditions (Fig. 19a), ERA5 and CLDAS show similar accuracy with RMSE values of 4.2–4.5 K during the day and 4.2 K at night. Our Themi product has superior accuracy with RMSE values of 2.3 K during both day and night. Dong's LST shows accuracy comparable to Themi, whereas Jia's LST exhibits RMSE values between those of the reanalysis products and Themi. Under cloudy-sky conditions (Fig. 19b), ERA5 LST has the lowest accuracy with an RMSE of 5.2 K during the day and an RMSE of 4.1 K at night. CLDAS LST has RMSEs of 5.1 K during daytime and 4.0 K at night, which performs worse than remote sensing-based LST. The RMSE of Themi is comparable to that of Jia's LST, but lower than that of Dong's LST. The overall performance of the Themi product under both clear-sky and cloudy-sky conditions demonstrates its strong potential for subsequent applications.

The evaluation results of the generated ANCFDS-LST demonstrate its high accuracy and detailed texture. Several limitations remain to be addressed in future work:

The TEKDM used in this study is developed for vegetated canopies, which is dominated by the gap fraction and hot spot effects. However, the applicability of this model to other land covers, such as bare soil, snow, and urban areas, requires further exploration and validation (Du et al., 2023, 2024; Wang et al., 2024a; Xiong et al., 2024). Moreover, the reliability of TEKDM in complex terrain also demands in-depth studies (Zhan et al., 2025).