China global Land Surface Air Temperature dataset (C-LSAT) is a global land surface air temperature (LSAT) product that provides both station-based and gridded data. It integrates a total of 14 data sources, including three global datasets (CRUTEM4, GHCNm, and BEST), three regional datasets, and eight national datasets. A major advancement of this dataset lies in its substantially improved station coverage across most Asian countries, particularly in China and its surrounding regions (Xu et al., 2018; Li et al., 2021; Wei et al., 2025). The dataset was assessed in the Intergovernmental Panel on Climate Change (IPCC) Sixth Assessment Report (AR6), released in August 2021, as “fully meeting the IPCC requirements,” and was accordingly incorporated and utilized in the report (IPCC, 2021). C-LSAT2.1 represents its latest version (Wei et al., 2025).

Building upon this foundation, the China global Merged Surface Temperature (C-MST3.0, Yun et al., 2019; Sun et al., 2021, 2022; Li et al., 2020, 2021) is constructed by merging the land surface air temperature dataset C-LSAT2.1 with the Extended Reconstructed Sea Surface Temperature dataset ERSSTv6 (Huang et al., 2025a, b). According to the spatial extent of the reconstructed Arctic sea ice surface air temperature regions, three variants are defined: C-MST3.0-Nrec, C-MST3.0-Imin, and C-MST3.0-Imax (Li et al., 2026b).

In this study, C-LSAT2.1 was used as the land surface air temperature (LSAT) component, while the 200-member HadSST4.2.0.0 sea surface temperature (SST) ensemble provided by the UK Met Office Hadley Centre (Kennedy et al., 2019) was used as the SST component. These datasets have a monthly temporal resolution and span the period from 1850 to 2024. A total of 105 ensemble members from Phase 6 of the Coupled Model Intercomparison Project (CMIP6; Eyring et al., 2016) historical simulations (Table S1 in the Supplement) were used as training samples for the PConv model.

To better reconstruct realistic climate conditions over Antarctica, we further employ monthly Antarctic station data. Most stations originate from SCAR READER (Turner et al., 2004), GHCNm v4 QCF and QFE (Menne et al., 2018). Specifically, the Byrd station data were obtained from the OSU Polar Meteorology Group (Bromwich et al., 2025b), the Dumont d'Urville station data from Météo-France (Météo-France, 2025), the Amundsen-Scott station data from the University of Wisconsin-Madison (South Pole Meteorology Office, 2025), and the Scott Base station data from the National Institute of Water and Atmospheric Research Ltd (NIWA, 2025). Necessary gap filling was then performed for the Antarctic station records following the method described by Bromwich et al. (2025a). Stations with more than 25 valid years during 1961–1990 are selected, yielding a temporal span of 1957–2024. Subsequently, homogenization tests are performed to eliminate discontinuities caused by station relocations, sensor changes, or other non-climatic shifts. Station observations are mapped to the corresponding 5°×2.5° grid cells (based on station longitude and latitude) and treated as valid data constraints for the reconstruction. Details of the station metadata and temperature anomaly time series are shown in Table S2 and Fig. S1 in the Supplement. Table 1 presents information on the different types of data products used in this study.

In this study, we adopted the data merging approach of Yun et al. (2019) to merge LSAT from C-LSAT2.1 separately with SST from ERSSTv6 and HadSST4. The land–ocean mask used in this process was obtained from the NCAR Command Language (NCL) “landsea.nc” mask file (download at: http://www.ncl.ucar.edu/Applications/Data/cdf/landsea.nc, last access: 23 July 2025). The mask file contains five categories: 0 for ocean, 1 for land, 2 for lake, 3 for small island, and 4 for ice shelf. In this study, land, lake, small island, and ice shelf were treated as land, while the other category was considered ocean. The Antarctic land mask used in this study at a spatial resolution of 5°×2.5° is shown in Fig. S2 in the Supplement.

The PConv method has been successfully applied to the reconstruction of globally missing climate data, including ST anomaly fields, wind fields, and surface solar radiation (Kadow et al., 2020; Zhou et al., 2022; Jiao et al., 2023). The underlying principle is to treat the climate field as a two-dimensional image, with missing regions regarded as irregular holes to be inpainted. PConv is then trained to learn spatial features from a large set of climate data samples, enabling the inference of plausible spatial patterns in the missing regions (Kadow et al., 2020). This study builds upon the work of Kadow et al. (2020), whose primary contribution was the introduction of a PConv-based framework for climate data reconstruction. Kadow et al. (2020) mainly demonstrated the feasibility and accuracy of PConv for climate field reconstruction at the global scale with a resolution of 5°×2.5°, and showed that the method can effectively capture the spatial patterns of ENSO-related SST anomalies. The PConv approach performs well for missing grid points over the low- and mid-latitude areas, but provides limited assessment for polar regions or areas with severe observational deficiencies.

The main contribution of this study is the integration of input observational uncertainty, represented by the HadSST4 ensemble members, and model uncertainty, represented by the Deep Ensemble approach, into a unified reconstruction framework. The law of total variance is then used to decompose and quantify these two sources of uncertainty, thereby providing predictive uncertainty estimates for reconstructed result. In addition to producing reconstructed global historical surface temperature fields, the proposed framework also provides associated uncertainty estimates, offering valuable confidence information for subsequent climate change analyses and the interpretation of reconstruction results.

The reconstruction workflow is illustrated in Fig. 1, and consists of the following steps:

Resolution standardization: The input datasets used for training are uniformly remapped from higher resolution to a spatial resolution of 5°×2.5° using conservative interpolation, resulting in a regular grid composed of 72×72 grid cells.

In deep learning, predictive uncertainty can generally be decomposed into aleatoric uncertainty and epistemic uncertainty (Kendall and Gal, 2017). Aleatoric uncertainty arises from the inherent randomness of the observation process and measurement noise, whereas epistemic uncertainty reflects uncertainty in model parameters due to limited knowledge of the underlying data-generating process (Abdar et al., 2021).

In this study, the 200 ensemble members provided by HadSST4.2.0.0 were used to approximate the observational uncertainty of sea surface temperature (SST). The HadSST4 ensemble members are generated using an observational error model that accounts for multiple sources of uncertainty, including Bucket measurement errors, Engine-room intake measurement errors, and bias-adjustment uncertainties. Each ensemble member represents a physically plausible realization of the SST field constrained by the available observations (Kennedy et al., 2019). The C-LSAT2.1 dataset was merged separately with each HadSST4 ensemble member to construct 200 model input ensembles. Because these ensemble members collectively represent the irreducible uncertainty associated with the observation process, they were regarded as the aleatoric uncertainty of the input data, which was propagated through the reconstruction model via ensemble propagation (Kendall and Gal, 2017).

Epistemic uncertainty, also referred to as model uncertainty, was estimated using the Deep Ensemble approach, which is one of the most effective methods for quantifying epistemic uncertainty in deep learning (Lakshminarayanan et al., 2017). Specifically, five PConv models with identical network architectures, training datasets, and hyperparameters were independently trained using different random seeds for parameter initialization. The model with the best validation performance from each training run was retained to form the ensemble. Owing to different random initializations and stochastic optimization trajectories, the independently trained models converge to different parameter configurations, thereby reflecting uncertainty in model parameter estimation (Lakshminarayanan et al., 2017).

The 200 input ensembles were subsequently fed into each of the five trained PConv models, resulting in a total of 1000 reconstructed temperature fields. This reconstruction ensemble simultaneously incorporates the propagated observational uncertainty from the input data and the prediction variability arising from different model parameterizations, thereby jointly representing both aleatoric and epistemic uncertainty. The ensemble mean of the 1000 reconstructions was adopted as the final reconstructed temperature product. The total predictive uncertainty was decomposed according to the Law of Total Variance (Valdenegro-Toro and Mori, 2022): 1Var(Y)=Eθ[Var(Y|θ)]︸Aleatoric uncertainty+Varθ(E[Y|θ])︸Epistemic uncertainty

Here, θ denotes the parameters of each independently trained PConv model in the Deep Ensemble, and Y represents the reconstructed temperature field. The first term, Eθ[Var(Y|θ)], represents the propagated observational (aleatoric) uncertainty. For each fixed model θ, the variance of the predicted outputs was first calculated from the 200 ensemble predictions, yielding Var(Y|θ). The average of these variances across the five models was then taken to obtain Eθ[Var(Y|θ)]. The second term, Varθ(E[Y|θ]), represents the model (epistemic) uncertainty. Specifically, for each model θ, the ensemble mean prediction E[Y|θ] was first computed from the 200 reconstructed outputs. The variance among the mean predictions of the five independently trained models was then calculated to quantify the uncertainty arising from differences in model parameters.

The sum of these two components yields the total predictive variance. Its square root corresponds to the predictive standard deviation σ, and the 95 % predictive uncertainty interval is reported as 1.96σ.

It should be noted that the total predictive uncertainty estimated in this study represents an approximation of the true predictive uncertainty and is conditional on the adopted ensemble representation. Although the observational uncertainty of the SST input data is propagated through the reconstruction model, the 200-member HadSST4 ensemble primarily represents the uncertainty associated with SST bias corrections and does not explicitly incorporate the spatially correlated error covariance information. Therefore, these spatially correlated covariance errors are not explicitly propagated in our uncertainty quantification framework. The observational errors from C-LSAT2.1 are not explicitly incorporated into the uncertainty quantification framework. According to the uncertainty assessment of the C-LSAT2.1 dataset by Li et al. (2026b), the maximum 1σ station, sampling, and bias uncertainties are approximately 0.065, 0.025, and 0.005 °C, respectively, in 1850, and all decrease to below approximately 0.02 °C after the early twentieth century. Their omission is expected to have only a minor influence on the total predictive uncertainty estimated in this study. Consequently, the estimated uncertainty mainly reflects the contributions from SST input uncertainty and model uncertainty. The resulting total predictive uncertainty can therefore be regarded as a lower-bound estimate of the true predictive uncertainty, and the actual uncertainty over land regions is likely to be larger than that reported in this study.

To ensure comparability across different data sources, all external benchmark ST datasets are first remapped to a spatial resolution of 5°×2.5° consistent with the AI reconstruction inputs and outputs, using bilinear interpolation, resulting in a 72×72 regular grid. All data were converted to ST anomalies relative to the 1961–1990 climatology to eliminate differences in the climate reference among datasets. Subsequently, the performance of the AI reconstruction was systematically evaluated (Sect. S1). The primary evaluation metrics included the Bias, Correlation Coefficient (CC), Root Mean Square Error (RMSE) and Mean Absolute Error (MAE), which were used to assess the consistency and bias of the reconstructed fields relative to reference datasets (Figs. S5–S10 in the Supplement). Based on the model validation results (Sect. S1, Figs. S5–S10), this study ultimately adopted the merged product of C-LSAT2.1 and HadSST4.2.0.0 as the observational dataset, trained the AI model using CMIP6 samples, and estimated reconstruction uncertainty by incorporating the 200-member HadSST4.2.0.0 ensemble.

To comprehensively assess the global applicability and reliability of the AI reconstruction, it was compared against multiple representative global temperature datasets, including Berkeley Earth, HadCRUT5, NOAAGlobalTempv6, C-MST3.0-Imax, and the latest HadCRUT5 reconstruction results of Kadow et al. (2020) trained on the 20CR dataset. The comparison included global annual mean temperature anomaly time series, linear trend estimates, and their statistical significance. In the Antarctica, where observational coverage is sparse, targeted evaluation of the AI reconstruction was performed. Reference datasets included the ERA5 reanalysis, Berkeley Earth, HadCRUT5, DCENT-I Kadow et al.(2020), and the reconstruction by Bromwich et al. (2025a). Cross-validation among these multiple data sources was conducted to assess the reliability and consistency of the Antarctic reconstructions after 1961.

The GMST series of the benchmark datasets used in this study follow the standard products released by their respective official sources. All regional annual mean temperature series are first computed using area-weighted spatial averaging, followed by temporal averaging on an annual basis, in accordance with the World Meteorological Organization (WMO) methodology. Linear trends are estimated using ordinary least squares, and their statistical significance is assessed using a two-sided t-test. The Ocean Niño Index (ONI) is defined following the NOAA Climate Prediction Center standard, as the 3-month running mean of temperature anomalies over the Niño 3.4 region (5° S–5° N, 170–120° W), calculated relative to a centered 30-year base periods that is updated every 5 years.

Prior to the 19th century, the C-LSAT2.1 dataset contained substantial missing data in Asia, South America, Africa, and Antarctica due to the sparse distribution of land-based stations, and this situation was improved considerably after the 19th century (Wei et al., 2025). In addition, Antarctic observational stations began recording data progressively after the 1957/58 International Geophysical Year; however, observational coverage in Antarctica remains extremely sparse (Wang et al., 2023). As shown in Fig. 2, the ocean coverage of NOAAGlobalTempv6 is higher than that of the unreconstructed C-AIRST_M, which is attributed to the ocean component included in the former (ERSSTv6) employs an ANN reconstruction method, which allows for more accurate and stable reconstruction of SST in sparsely observed regions, reduces excessive smoothing, and better preserves spatial variability (Huang et al., 2025a, b). In contrast, HadSST4 primarily relies on statistical interpolation and ensemble-based methods, which aim to correct biases and estimate uncertainties within grid cells, without performing large-scale spatial reconstruction (Kennedy et al., 2019).

During the strong El Niño event in September 1877, the AI model is able to reasonably reconstruct a coherent warming anomaly pattern over the equatorial western Pacific under data-sparse conditions (Fig. 2b1). This pattern is characterized by pronounced warm anomalies in the tropical central and eastern Pacific, with alternating warm and cold anomalies distributed zonally along the equator, reflecting a typical ENSO signal.

For the ENSO reconstructions in September 1893, February 1917, and March 1941 (Fig. 2b2–b4), the spatial patterns produced by the AI model are generally consistent with those from HadCRUT5 (Fig. 2c2–c4) and Berkeley Earth (Fig. 2d2–d4), both of which also use HadSST-based ocean components. However, the spatial distribution of warm and cold anomalies along the equatorial Pacific, namely the classic “warm–cold tongue” structure, is less pronounced than in ERSSTv6 (Fig. 2e2–e4). This discrepancy primarily arises from inherent differences between HadSST and ERSSTv6.

AI model is able to effectively capture the spatial patterns of ENSO, while also reproducing historical ENSO events. As shown in Fig. 3, the positive and negative phases of the ONI, which serves as an indicator of ENSO variability, are generally consistent with those from NOAAGlobalTempv6 and HadCRUT5. The AI-based reconstruction also fills the gaps in the ONI during 1910–1920 in HadCRUT5 infilled, where incomplete spatial coverage over the Niño 3.4 region led to biases or missing values. In addition, the reconstruction identifies the documented El Niño events of 1914–1915, and 1918–1919 (Yu and Kim, 2013), which are also evident in NOAAGlobalTempv6 (Fig. 3c). These results also corroborate the findings of Qian et al. (2026), who investigated historical ENSO variability and reconstruction using long-term Niño indices and their associated spatial patterns.

A few weak El Niño events show slight discrepancies compared with HadCRUT5 due to small differences in the amplitude of the Niño 3.4 index, indicating minor estimation biases in the AI-reconstructed index. It is also noteworthy that the ENSO amplitude in NOAAGlobalTempv6 is generally stronger than that in HadCRUT5 prior to 1950, and this discrepancy is primarily attributable to differences in the SST bias correction methods and spatial infilling approaches employed by the two datasets (Morice et al., 2021; Huang et al., 2025a, b). Consequently, the C-AIRST_M based on HadSST are overall more consistent with HadCRUT5. In summary, the AI models perform well in representing the spatial structure of ENSO, they are also capable of reproducing historical ENSO events through the ONI.

The Global mean surface temperature (GMST) time series reconstructed by AI is shown in Fig. 4, reflecting the long-term variations of global mean temperature from 1850 to 2024. From 1850 to the early 20th century, global mean temperature remained generally stable (Fig. 4a), with occasional short-term cooling signals associated with enhanced volcanic activity and slight decreases in solar radiation (Fang et al., 2022). From the 1910s to the 1940s, the global mean temperature experienced the first rapid and sustained warming phase of the 20th century, with a pronounced increase in magnitude. From the mid-20th century to the mid-1970s, the warming trend weakened, followed by a renewed and more pronounced warming phase from the late 1970s onward, which continues to the present, reflecting the dominant influence of anthropogenic greenhouse gas emissions (IPCC, 2021).

The C-AIRST_M show only minor differences in long-term reconstructed series and trends, and are generally consistent with the HadCRUT5 record (Fig. 4b, Table 2), with overall biases within 0.15 °C. After 1960, differences between the AI reconstructions and other datasets further decrease to within 0.05 °C. During 1930–1940, C-AIRST_M was approximately 0.1 °C warmer than the other datasets, whereas it was slightly cooler during the 1940–1945 World War II period (Fig. 4b). These differences mainly arise from the revised historical SST bias corrections implemented in HadSST4.2.0.0 (Sandford and Rayner, 2026). In particular, the systematic warm bias introduced by the widespread use of Engine Room Intake (ERI) measurements during World War II was corrected (Chan and Huybers, 2021; Kent and Kennedy, 2021). In addition, a truncation error in the Kobe SST records resulted in an approximately 0.4 °C cold bias during 1930–1940, which was corrected in HadSST4.2.0.0 (Sandford and Rayner, 2026). This SST correction also contributed to a higher estimate of the GMST in C-AIRST_M during the same period. The long-term warming trends reconstructed by the AI model for different periods during 1850–2024 (Table 1) are generally consistent with the estimates from other global temperature datasets, including Berkeley Earth, HadCRUT5, and NOAAGlobalTempv6. The AI reconstruction also yields a Global Warming Level (GWL) of 1.49 °C, with only small differences relative to the other datasets, indicating that global warming in 2024 had approached the 1.5 °C warming threshold relative to the pre-industrial period (1850–1900). Overall, the AI reconstruction series shown in Fig. 4 reasonably reproduce the phase-dependent characteristics of global climate evolution over the past 175 years at interannual scales.

Regarding the uncertainty estimation of GMST (Fig. 5), the uncertainty quantified in this study is not completely comparable to that of mainstream observational datasets, which generally consider multiple sources of uncertainty, including measurement errors, sampling errors, bias adjustments, and incomplete spatial coverage. The predictive uncertainty estimated here incorporates the propagation of input uncertainty represented by the 200 ensemble members of HadSST4.2.0.0, as well as the model uncertainty arising from the deep learning framework itself (see Sect. 2.4 for details on uncertainty quantification). Despite the differences in uncertainty sources, the estimated uncertainty in GMST is generally comparable to that of existing mainstream datasets. The 1.96σ uncertainty gradually decreases from approximately 0.2 °C in the 1850s to approximately 0.05 °C after the 1960s (Fig. 5a). In addition, uncertainty exhibits temporary increases during the two World War periods (the 1910s and 1940s), likely associated with reduced global observational coverage.

The spatial distribution of uncertainty (Fig. 5b) shows that the polar regions have substantially higher uncertainty than the low- and mid-latitude regions due to the long-term scarcity of observational records. The uncertainty in the polar regions generally exceeds 0.3 °C, whereas that in most low- and mid-latitude regions remains below 0.2 °C. Notably, Antarctica exhibits substantially larger uncertainty than the Arctic, both in terms of the magnitude of uncertainty and the spatial extent of high-uncertainty regions. This difference is primarily related to the much sparser observational network and insufficient historical observations in Antarctica compared with the Arctic.

In the zonal temperature comparison across different datasets (Fig. 6), it can be seen that within the 60° S–60° N band, where observational sampling is relatively sufficient, the differences between each dataset and C-AIRST_M generally remain within ±0.1 to ±0.2 °C prior to the early 20th century (Fig. 6b and c). The Northern Hemisphere (NH), characterized by a larger land fraction and denser observational networks, is substantially better constrained by observations than the predominantly ocean-covered Southern Hemisphere (SH). Consequently, after 1900, the inter-dataset differences within the equator–60° N region decrease to within 0.1 °C. In contrast, within the equator–60° S region, this convergence is more delayed, with differences only gradually reducing to within 0.1 °C after 1960s. The 95 % uncertainty intervals estimated by C-AIRST_M in both regions remain within ±0.1 °C, indicating relatively low uncertainty. Notably, only NOAAGlobalTempv6 exhibits a persistent cold bias of approximately 0.2 °C relative to the other benchmark datasets during 1920–1960 in this region, likely reflecting weaker constraints over the Southern Ocean.

In the Arctic region (Fig. 6a), prior to 1920, both C-AIRST_M and Kadow et al. (2020) exhibit a cold bias of approximately 0.6 °C relative to HadCRUT5, NOAAGlobalTempv6 and Berkeley Earth. It is also observed that the AI reconstructions (Fig. 6j–l) show varying degrees of cold bias in the Arctic compared with other datasets before 1920. After 1930, as Arctic observations increase, the differences among datasets rapidly decrease to within 0.2 °C. In the Antarctic region (Fig. 6d), during 1850–1960, the estimates from different datasets are not entirely consistent. The Kadow et al. (2020) product and C-AIRST_M remain approximately 0.5 °C colder than NOAAGlobalTempv6, which provides full Antarctic coverage, while the larger variability in HadCRUT5 and Berkeley Earth is likely due to the lack of full spatial extrapolation south of 60° S prior to 1960. As Antarctic observations increase around 1961, differences between the two AI reconstructions rapidly converge to within 0.2 °C.

The above analyses suggest that the reconstructions exhibit some systematic cold bias in the early Arctic (before 1920) and Antarctica (before 1960), where observational constraints are relatively weak. Correspondingly, the 95 % uncertainty ranges in these periods are also relatively large, reaching approximately ±0.5 °C in the Arctic (Fig. 6a) and ±0.8 °C in Antarctica (Fig. 6d). Consequently, the reconstructed temperatures for these periods and regions should be interpreted as exploratory and treated with appropriate caution.

Under the context of global warming, the spatial pattern of warming level from C-AIRST_M is generally consistent with those of other datasets (Fig. 7). The main discrepancies are concentrated south of 60° S, where the PConv-based reconstruction (Fig. 7a and f) exhibit spatial discontinuities in warming relative to NOAAGlobalTempv6. This may be attributed to the inherently less smooth spatial temperature fields produced by the AI reconstruction during 1850–1900 (Fig. 2b1 and b2).

The C-AIRST_M and reconstruction from Kadow et al. (2020) based on HadSST data generally show higher consistency in warming patterns with HadCRUT5 and Berkeley Earth, which also use HadSST, compared to NOAAGlobalTempv6, which is based on ERSSTv6. In particular, near 110° W in the Southern Ocean adjacent to Antarctica, NOAAGlobalTempv6 exhibits a larger cooling magnitude than the other datasets. Although all datasets capture the North Atlantic “cold blob,” the stronger cooling north of this region in NOAAGlobalTempv6 may represent an artifact of its underlying analysis procedure (Chan et al., 2026), a feature not observed in the other datasets.

The latitudinal warming amplitudes further reveal the consistency and differences among the reconstructed products in representing polar warming (Fig. 7g). C-AIRST_M and the reconstruction from Kadow et al. (2020) exhibit slightly stronger warming amplitudes near 80° N than the other datasets. In Antarctica, the warming amplitude estimated by C-AIRST_M is higher than that of NOAAGlobalTempv6 by approximately 0.7 °C in the polar regions, respectively. These differences mainly arise from variations in the estimated baseline climatology for 1850–1900 among different datasets.

Furthermore, the zonal distribution of warming (Fig. 7g) shows a pronounced Arctic amplification effect, with warming in the Arctic substantially exceeding the global mean level. During 2005–2020, relative to the pre-industrial baseline period of 1850–1900, the cumulative warming in the Arctic region reached approximately 2.6–4.0 °C, markedly higher than the corresponding global mean warming over the same period.

Figure 8 presents the spatial distribution of long-term warming trends from 1850 to 2023 across different datasets, as well as the warming trend patterns for three sub-periods: 1850–1910, 1911–1970, and 1971–2023. The results from C-AIRST_M is generally consistent with the reconstruction by Kadow et al. (2020). The spatial gradient of LSAT trends reconstructed by these two datasets is less uniform compared to other datasets, particularly across the continents during 1911–1970. A notable difference between the AI-based reconstructions in this study and NOAAGlobalTempv6 lies in the 1850–1910 period. Specifically, NOAAGlobalTempv6 shows a widespread warming trend over the African continent, whereas the warming magnitude and spatial extent in other datasets. All datasets consistently indicate that, under the background of global warming, the magnitude of the warming trend during 1971–2023 is clearly higher than those in the earlier periods of 1850–1910 and 1911–1970.

Overall, the AI-based reconstructions are able to reasonably reproduce the magnitude and spatial distribution of global warming. However, caution is still warranted in regions covered by sea ice and subject to extremely sparse observational constraints, particularly in the polar regions.

One of the key focuses of this study is the reconstruction of global land surface data (Fig. 9, Table 3). To evaluate the reconstruction performance, we also present the effective data coverage of the original datasets across different regions (Coverage in Fig. 9). In addition, Table 3 summarizes the linear trends of annual mean surface temperature over 1850–2024 for different regions, along with their 95 % confidence intervals. The performance of the Antarctic land reconstruction is further evaluated in detail in Sect. 4, “Antarctic Reconstruction Results.”

Figure 9 shows that across different land regions, the differences between the C-AIRST_M and other datasets gradually decrease to within 0.1 °C as observational coverage exceeds approximately 50 %. This feature indicates that in regions where observational data become progressively denser, the interannual differences among AI-based reconstructions tend to stabilize. In contrast, during earlier periods or in regions with sparse observations and low coverage, discrepancies among reconstruction products are mainly concentrated in Africa and South America, where early observational records are limited. When effective coverage is below 40 % before 1900, C-AIRST_M exhibits differences of varying magnitudes relative to other datasets, ranging from 0.4 to 1.0 °C (Fig. 9c, d, and h), which also leads to lower long-term trends in these regions for the former in Table 3.

Furthermore, during 1860–1870, when effective coverage in Africa, South America, and Oceania (Fig. 9c, d, and f) drops below a critical threshold (∼20 %), and considering the strong spatial sampling heterogeneity and lack of inland temperature information in these continents (Wei et al., 2025), the C-AIRST_M shows varying degrees of deviation from NOAAGlobalTempv6 and HadCRUT5. The most pronounced discrepancy occurs in South America (Fig. 9d), where values are approximately 0.8 °C higher than NOAAGlobalTempv6 and 0.5 °C higher than HadCRUT5. However, around 1920, a systematic negative bias of about 0.3 °C appears. Over the full 1850–2024 period (Table 3), the long-term trends of C-AIRST_M across different regions are broadly consistent with those of HadCRUT5. However, C-AIRST_M exhibits a weaker warming trend than other datasets over Oceania, which can be attributed to inherent temperature biases in early land surface products prior to 1900.

As shown in Table 3, C-AIRST_M indicates the strongest warming over Europe during 1850–2024, with trend of 0.090 ± 0.014 °C per decade, while Oceania shows the weakest warming, at 0.039 ± 0.008 °C per decade. The warming rate in the SH is significantly lower than that in the NH, reflecting hemispheric asymmetry driven by the large land–ocean contrast and the strong heat uptake by the Southern Ocean, which enhances climate inertia in the Southern Hemisphere (Hansen et al., 2010).

From the integrated analysis of regional surface temperature anomaly series, it is evident that Northern Hemisphere land areas contribute most significantly to global land warming. Within the Northern Hemisphere, Europe, Asia, and North America are the dominant contributors. Under conditions of extremely low data coverage, both reconstructions inevitably exhibit biases; however, as coverage increases after the early 20th century, the AI-based reconstruction results show improved consistency and stability across these regions.