This study builds on C-LSAT 2.0 station data (Xu et al., 2018; Yun et al., 2019), combined with additional observations integrated from various countries, regions, and global sources, covering the period from 2013 to 2023. Compared to version 2.0, the C-LSAT 2.1 station data significantly increased the number of observation stations (Tavg increased from 15 936 to 25 085 stations, Tmax increased from 13 648 to 25 086 stations, and Tmin increased from 13 629 to 25 083 stations, as shown in Fig. 1 of Xu et al., 2025).

Various data sources commonly assign different station IDs to the same station. Therefore, matching the data from various sources with the corresponding stations in the C-LSAT station data is a problem that requires urgent resolution. Most stations have a core five-digit ID. For example, the core ID for the “JAN MAYEN” station is “01001”. In GSOD, this appears as “01001099999”; in the CLIMATE Report, this appears as “01001”; and in C-LSAT station data, this appears as “601001001000”. For stations lacking a consistent core ID, we employ the station name or identify nearby stations to locate the corresponding stations and complete the updates. Notably, when the sequence of a station is derived from multiple data sources, there may be homogenization discrepancies. In such cases, the application of calibration procedures for the specific station is necessary.

When updating data from multiple sources, duplicate stations are inevitable. They arise either because different data sources assign distinct IDs to the same station or because iterative updates generate new duplicates. Duplicate stations can affect the interpolation of both the climatology and anomaly fields, causing deviations in the interpolation results. To address this issue, it is essential to eliminate duplicate stations. Based on the same core IDs and similar station names, the C-LSAT 2.1 station data are filtered to identify and select duplicate stations. Subsequently, time series from each duplicate and its corresponding update sources or nearby stations are plotted for comparison. A reference station with the longest and most consistent record is then chosen. The data from the duplicate stations are selectively merged with the reference station or retained unmodified, ensuring the retention of a single representative station for each group of duplicates (Rennie et al., 2014; Xu et al., 2018).

The Tavg variable contains climatology (1961–1990) in the C-LSAT 2.1 station data including 13 746 stations (Fig. 1). Among these, 11 975 stations calculate Tavg using the average of Tmax and Tmin. The remaining 1771 stations, which lack either Tmax or Tmin, are primarily derived from datasets such as CRUTEM4, HISTALP, and SCAR. Compared to other datasets, the C-LSAT 2.1 station data demonstrate substantial improvements in station coverage in multiple regions, especially in East Asia. Figure 1 illustrates the C-LSAT 2.1 station data updates; compared to version 2.0, the number of stations increased significantly for Tmax, Tmin, and Tavg, particularly after the 1970s. These additional stations substantially expand spatial coverage, thereby enhancing the accuracy of data and reducing uncertainty after gridding. For temporal coverage, the majority of stations provide data for 50–80 years, with a few covering 80–100 years (Table S1 in the Supplement).

Data quality control is a crucial step to ensure the accuracy and reliability of datasets. By identifying and eliminating outliers, invalid data, and measurement errors, this process reduces the influence of observational biases, ensuring the consistency and integrity of the data.

First, when updating station data, if a station has a data record exceeding 15 years, its new observations are subjected to quality control. Any anomaly – defined as the difference between the updated data and the prior monthly mean – that exceeds 5 times the standard deviation is classified as an outlier and set to missing.

Subsequently, quality control is performed on all station data during gridded-data generation. This study follows the methods proposed by Lawrimore et al. (2011) and Menne et al. (2009) to implement the necessary quality control steps for C-LSAT 2.1 station data. The number of data values excluded during the quality control procedure is shown in Table 1.

Climatic outlier check. Stations with monthly records exceeding 10 years were selected, with the period from 1961 to 1990 as the climatology. The monthly climatological mean value was subtracted from the selected stations to calculate anomalies for each station. The standard deviation (SD) for each month during the climatology period was subsequently calculated. Any anomaly exceeding 5 times the SD for the corresponding month was flagged as an outlier and excluded.

Data homogenization is crucial for understanding climate change. Although its influence on a global or large scale may be limited, its impacts on local regions are often substantial (Peterson et al., 1998; Ribeiro et al., 2016). Homogenization removes data discontinuities caused by non-climatic factors such as station relocations, instrument changes, and environmental transformations (e.g. urbanization), ensuring that the data accurately reflect signals of climate change (Eccel et al., 2012; Jiao et al., 2023). Homogenized data enhance reliability and reduce error propagation.

The homogenization process of C-LSAT station data follows the work of Xu et al. (2025). Using the method proposed by Peterson and Easterling (1994), a reference series was constructed by selecting three to five neighbouring stations with correlation coefficients greater than 0.8 relative to the target station. Based on the spatial distances of these stations, a reference LSAT series was generated through a weighted average of first-order differences. Subsequently, the RHTest V4 software was used to detect and correct discontinuities in the target series (Wang and Feng 2013). The PMTred algorithm (derived from the penalized maximal t test, PMT) in RHTest V4 served as the primary algorithm to detect discontinuities in the target station's monthly average Tmax and Tmin series at a significance level of 5 %. For any confirmed breakpoints, the differences between the target series and the reference series were uniformly allocated using the mean adjustment (Wang et al., 2008a, b). According to this procedure, 726 breakpoints (in 420 stations) for the 25 086 Tmax stations and 1276 breakpoints (in 754 stations) for the 25 083 Tmin stations of the C-LSAT station data were adjusted. The homogenized Tmax and Tmin data were then combined into the LSAT and DTR datasets (Table 2).

TPS interpolation was used to generate the climatology fields (1961–1990) of LSAT and DTR in this study. This was originally proposed by Wahba (1990) and later optimized and improved by Hutchinson et al. (Hutchinson, 1991, 1995, 1998a, b; Hutchinson and Gessler, 1994), evolving into the partial TPS, which integrates covariate-dependent interpolation, extending the previous method that was limited to calculations based on independent variables. Based on the TPS, Hutchinson (1991) designed and developed the software ANUSPLIN, which enables multivariable data interpolation. This software has been widely adopted for meteorological data interpolation. This study employed ANUSPLIN for climatology field interpolation.

Due to the strong correlation between temperature and elevation, longitude, latitude, and elevation were selected as variables for interpolating LSAT and DTR climatology field. The elevation data used in this study were obtained from the ETOPO2022 published by NOAA (National Oceanic and Atmospheric Administration) (available at https://www.ncei.noaa.gov/products/etopo-global-relief-model, last access: 3 January 2024). This dataset integrates topography, bathymetry, and coastline data from regional and global datasets, providing a comprehensive and high-resolution representation of the Earth's geophysical features (MacFerrin et al., 2025).

Because of the Earth's spherical shape, the TPS cannot provide a globally consistent surface interpolation. Thus, the globe was divided into regions for separate interpolation. This study refers to the global partitioning scheme from the CRU (New et al., 1999) and WorldClim2 (Fick and Hijmans, 2017) datasets, dividing the globe into 20 regions for interpolation. The spatial distribution is shown in Fig. 3. In terms of station density, the highest density is observed around 40° N and 40° S, while the lowest density occurs at the poles and the Equator. After interpolating the data for each region, the data from the 20 regions are merged into the global dataset. A known limitation of ANUSPLIN interpolation is the occurrence of larger errors near regional boundaries. To address this, when interpolating the 20 regions, the boundaries of each region are extended (by 5° latitudinally and 10° longitudinally). After interpolation, the extended areas are clipped, and then merged into the global dataset. This approach helps minimize boundary-related errors of the dataset.

Longitude and latitude are typically used as independent variables for meteorological interpolation. However, whether elevation should be treated as an independent variable or a covariate demands careful evaluation. There are three main indicators for evaluating the interpolation accuracy of the climatology field: the square root of generalized cross-validation (RTGCV), the mean square residual (RTMSR), and the data error variance estimate (RTVAR). The RTGCV quantifies the overall error of data fitting during the cross-validation process, measuring the model's generalization capability. The RTMSR reflects how well the model fits the input data, and the RTVAR evaluates the uncertainty in the data. Another indicator, the signal-to-noise ratio (SNR), is typically used to indicate the complexity of the fitted surface. It represents the ratio between the signal and the error value in the ANUSPLIN software output file. This value generally needs to be less than 1 to indicate that the chosen interpolation experiment is feasible.

Parameter schemes (Table 3) and their corresponding results (Table 4, Figs. 4 and 5) are presented below. Overall, DTR interpolation errors exceed those for LSAT, and interpolation errors (not included in Table 4) increase markedly at spline order 4 compared to orders 2 and 3. As a result, all order-4 experiments (A3 and B3) are excluded. For the Antarctic (region 20), the four error metrics of LSAT demonstrated substantial increases, indicating high uncertainty in this area. This is attributed to the notably low and uneven station distributed for the Antarctic. Considering the increased error mentioned before, both LSAT and DTR for the Antarctic are excluded from this study. Thus, the subsequent contents of this study exclude the Antarctic (region 20). After excluding region 20, the SNR, RTGCV, RTMSR, and RTVAR are compared across the remaining 19 regions. For LSAT, experiment B2 demonstrates optimal performance; however, with respect to the DTR, although experiment B2 achieves the highest overall effectiveness, it produces negative values in some regions and is therefore excluded, leading to the adoption of experiment B1 (Table 4).

Based on the cross-validation results, the mean absolute error (MAE), the root mean squared error (RMSE), and the distance between indices of simulation and observation (DISO) are computed for C-LSAT HRv1 and C-LDTR HRv1 climatology field evaluation (Fig. 6). DISO is a comprehensive performance evaluation index that combines the correlation coefficient (r), absolute error (AE), and RMSE, and the closer its value is to 0, the better the agreement between the simulation and observation (Hu et al., 2019, 2022; Zhou et al., 2021). For C-LSAT HRv1, the MAE, RMSE, and DISO for the Southern Hemisphere fall below the global means, whereas Northern Hemisphere values exceed the global means. In contrast, the C-LDTR HRv1 MAE and RMSE are higher for the Southern Hemisphere than globally. However, the comprehensive index DISO confirms that the Southern Hemisphere's overall performance still surpasses that of both the Northern Hemisphere and the global average. For high-altitude and complex terrain regions, the Tibetan Plateau is selected for validation. The results indicate that all LSAT indices in this region surpass their global and hemispheric levels, whereas DTR performance remains consistent (Fig. S1 in the Supplement). This discrepancy can be attributed to a combination of factors, including a limited observational network, significant topographic variations, land use, and so on. To improve data reliability, future work should refine spatial resolution and optimize variable selection.

The LSAT anomaly variations of C-LSAT HRv1 and C-LSAT 2.1 from 1901 to 2023 for the globe, Northern Hemisphere, and Southern Hemisphere are presented in Fig. 10. The anomaly trends obtained in C-LSAT HRv1 are largely consistent with those of C-LSAT 2.1, with warming rates of 0.132±0.015, 0.140±0.017, and 0.106±0.011°Cdecade-1 for the globe, Northern Hemisphere, and Southern Hemisphere, respectively. The LSAT change trends for the globe and Northern Hemisphere demonstrate comparable patterns, with warming being predominantly concentrated in two periods: the 1900–1930s and the 1970–2020s, with accelerated warming in the latter period. A slight cooling trend occurs in the middle period, from the 1940s to the 1960s. The warming for the Southern Hemisphere is relatively slower and continues throughout the entire 1901–2023 period without experiencing the cooling trend observed for the global and Northern Hemisphere during the 1940–1960s. Its warming rate also undergoes a pronounced acceleration after the 1970s.

Annual warming rates for C-LSAT HRv1 (Table 5) are lowest for 1901–1950, rise sharply after 1951 to peak in 1979, and then decline moderately by 1998. This slowdown corresponds to the 1998–2014 warming hiatus; although no cooling is detected, the warming rate is reduced.

Spatially, the LSAT change trend indicates continuous warming globally from 1901 to 2023, with the fastest warming occurring in regions such as northern North America, eastern South America, eastern Europe, and eastern Asia (Fig. 11). Regarding different periods, the fastest warming was observed between 1998–2023 (particularly in areas north of 60° N), while the slowest warming occurred during 1901–1950 (Fig. 12).

At the continental scale, both C-LSAT HRv1 and C-LSAT 2.1 show a warming trend across all six continental domains since the early 20th century, with this trend intensifying after the 1970s and manifesting regional differences (Fig. 13). The warming is pronounced in Asia, Europe, and North America, whereas it remains comparatively moderated in South America, Africa, and Oceania, reflecting the different responses of the climate system to global warming. Both datasets are consistent in terms of their long-term trends; however, differences in short-term fluctuations may stem from variations in spatial resolution and processing methods.

The DTR anomaly changes of C-LDTR HRv1 for the global mean, Northern Hemisphere, and Southern Hemisphere from 1901 to 2023 are presented in Fig. 16. During 1950–2010, C-LDTR HRv1 remains highly consistent with C-LDTR, with change rates of -0.031±0.006, -0.038±0.006, and -0.011±0.011°Cdecade-1 for the globe, Northern Hemisphere, and Southern Hemisphere, respectively. However, there are notable discrepancies before 1950 and after 2010. From 1901 to 1950, the station number is limited, which resulted in greater uncertainty. Consequently, the differences between the two datasets are more pronounced. This is particularly apparent for the Southern Hemisphere, where the DTR fluctuations and the differences between the two datasets are significantly larger than those for the globe and Northern Hemisphere. After 2010, the reduction in DTR (or Tmax and Tmin) station data leads to the differences between C-LDTR HRv1 and C-LDTR, which is further reflected in other DTR datasets (Xu et al., 2025). The DTR is stable during the 1900s–1940s and 1980s–1990s, declines during the 1950s–1970s, and shows a slight increase after the 2000s.

Table 6 shows the DTR change rates of C-LDTR HRv1 for different periods. The change rate is stable from 1901 to 1950, then initiates a decline in 1951, stabilizes again in 1979, and peaks at 1998. The DTR change rates for the Southern Hemisphere are more pronounced than those for the globe and Northern Hemisphere after 1979.

It is noteworthy that there is no obvious spatial pattern in the changes in DTR. During the period of most significant change, 1998–2023, the regions with the most rapid DTR increases are North America, Europe, and Oceania, whereas other regions, including Northeast Africa, South Asia, and the Middle East, demonstrate a pronounced downward trend (Figs. 17 and 18).

Based on the C-LDTR HRv1 and C-LDTR datasets, Fig. 19 illustrates the complex variation characteristics and significant regional differences of DTR across six continents between 1901 and 2023. DTR anomalies in Asia, Africa, and South America show the downward trend, whereas the changes in Europe, North America, and Oceania remain relatively stable. Europe demonstrates a general upward trend throughout the entire 1901–2023 period, while DTR for the remaining five continents declines before the 1970s but rebounds after 2010.