Hourly surface temperature observations spanning 1981–2024 were compiled through the Integrated Surface Database (ISD), a global database that consists of hourly and synoptic surface observations compiled from numerous sources into a single common ASCII format and common data model (Smith et al., 2011). It is important to note that NOAA formally terminated updates to the ISD in 2025, transitioning to the new Global Historical Climatology Network hourly (GHCNh) framework. However, our evaluation indicates that GHCNh currently lacks several critical meteorological elements present in the legacy ISD – specifically the 12 and 24 h temperature extremes. Given that these explicitly reported extremes are essential for our algorithm, the historical ISD remains the only viable source for reconstructing the high-fidelity 1981–2024 baseline presented in this dataset.

The ISD-derived temperature parameters encompass five specific temporal components: discrete sub-daily instantaneous temperatures (which are utilized to derive the daily mean temperature, denoted as Tave), 24 h explicitly reported maxima and minima (Tmax-24 h and Tmin-24 h), and 12 h explicitly reported maxima and minima (Tmax-12 h and Tmin-12 h).

Figure 1 visualizes the global distribution of sub-daily temperature data volumes across the 24 h UTC cycle spanning the period from 1981 to 2024. The analysis reveals striking temporal discrepancies among different temperature parameters. For the discrete hourly temperature observations (utilized to derive Tave), the data volume is continuously distributed across all hours, characterized by a highly robust multi-peak pattern. The primary peaks align perfectly with the standard 6-hourly synoptic times (00:00, 06:00, 12:00, and 18:00 UTC), complemented by secondary peaks at the intermediate 3-hourly intervals (03:00, 09:00, 15:00, and 21:00 UTC). This dense and temporally consistent distribution provides a solid foundation for calculating highly representative daily mean temperatures. Analysis of the discrete sub-daily reporting frequency reveals that the global network primarily operates on a combination of 3-hourly and 6-hourly cadences, manifesting as two dominant 6-hourly regimes offset by 3 h: 00:00/06:00/12:00/18:00 UTC and 03:00/09:00/15:00/21:00 UTC.

In stark contrast, the explicitly reported extreme temperatures (Tmax and Tmin) exhibit extreme temporal concentration. The 24 h extremes (Tmax-24 h and Tmin-24 h) are overwhelmingly anchored at just two specific reporting times: 06:00 and 18:00 UTC. Similarly, the 12 h extremes present a highly asymmetric, diurnal-driven reporting pattern. Specifically, Tmin-12 h reaches its absolute volumetric peak at 06:00 UTC, capturing the nighttime cooling, whereas Tmax-12 h overwhelmingly peaks at 18:00 UTC, corresponding to daytime warming. It should be noted that the overwhelming concentration of explicit 24 h extreme reports at specific hours (e.g., 06:00 and 18:00 UTC) is primarily an artifact of international data exchange standards, rather than a reflection of the longitudinal/geographic distribution of stations. WMO synoptic reporting protocols require member states to aggregate and submit extreme records at standardized global synoptic hours (WMO, 2015). Consequently, the ISD archive structurally “anchors” the vast majority of explicit global extreme reports to these specific UTC timestamps. These distinct structural characteristics explicitly demonstrate that while hourly observations offer continuous sub-daily coverage, explicit extreme reports are highly sparse outside of a few specific synoptic hours. This temporal fragmentation structurally mandates and quantitatively justifies the necessity of our secondary fallback strategy, which utilizes the high-frequency hourly observations to robustly reconstruct daily extremes when explicit records are absent.

The multi-panel maps in Figs. 2 to 6 illustrate the spatiotemporal evolution of sub-daily temperature data volumes from 1981 to 2024. The analysis reveals a striking contrast in data availability: while discrete hourly temperatures exhibit continuous and stable growth in spatial coverage and reporting frequency over the four decades, explicitly reported extremes (12 and 24 h Tmax/Tmin) suffer from severe geographic fragmentation and decadal volatility, notably experiencing a pronounced global decline between 1990 and 2010. These stark spatiotemporal discrepancies visually highlight the limitations of relying exclusively on explicitly reported extremes and quantitatively justify the necessity of utilizing high-density hourly data as a secondary fallback strategy.

In summary, sub-daily observations exhibit irregular observation intervals compared to Tave. Moreover, Tmax and Tmin do not always appear in pairs as expected, indicating that date shifts and inhomogeneities may not only exist between regions but also between the raw Tmax and Tmin records themselves.

The legacy GSOD dataset evaluated herein was accessed on 2025-03-01, reflecting the version that relied exclusively on the ISD as its input. The Global Surface Summary of the Day (GSOD, NOAA NCEI, 1999), maintained by the U.S. National Centers for Environmental Information (NCEI), generates daily meteorological summaries from over 9000 global weather stations. Derived through systematic processing of the ISD, this product provides continuous records spanning 1929 to present, with post-1973 data exhibiting optimal completeness. Readers should be aware that with the retirement of the ISD, the legacy GSOD is slated for replacement by a new product, the Surface Summary of the Day (SSOD). Because the SSOD has not yet been officially released, a direct comparison with the true successor to GSOD is not currently possible. Furthermore, because GHCNh does not currently support our methodological requirements, real-time operationalization of GLBD-FED is temporarily paused.

The quality of the data source is one of the most critical factors influencing the final data product. Notably, there were numerous misrecorded values for Tmax and Tmin over 12 and 24 h periods in the global hourly data, particularly in South America.

Figure 8 presents a case study from San Antonio Oeste, Argentina (877840-99999), covering 5 January to 12 January 2023. The red and blue lines represent the derived upper limits of temperature from Tmax-12 h and Tmax-24 h, respectively, while the black circles indicate hourly Tave. Panel (a) displays results from the raw data, revealing that San Antonio Oeste tends to record Tmax-24 h at 12:00 UTC and Tmax-12 h at 00:00 and 12:00 UTC. The blue lines (derived from Tmax-24 h) are consistently lower than the hourly Tave, suggesting that these Tmax-24 h values are likely erroneous due to incorrect recording.

The case study at San Antonio Oeste is highly representative of a systematic structural artifact within the legacy archive. A comprehensive global assessment of this specific mis-recording error from 1981 to 2024 reveals that it affected numerous stations worldwide, with concentrated occurrences in Western Europe, Russia, China, Canada, and South Africa. The primary driver of this anomaly is the misclassification of high-frequency sub-daily reports (e.g., 1 h summaries) as 24 h extremes. The spatial distribution and frequency of the stations affected are detailed in the Appendix (Fig. A1). To ensure high-throughput scalability across the global network, these corrections were coded as automated conditional rules integrated directly into our processing pipeline. The specific programmatic procedures are as follows:

Tmax-24 h records at hh:00 were compared with the highest hourly Tave from the previous 12 h (Tave-12 h high) and 24 h (Tave-24 h high). If Tmax-24 h <Tave-24 h high - 0.5 °C and Tmax-24 h ≥ Tave-12 h high, it indicates that the Tmax-12 h at hh:00 was incorrectly recorded as Tmax-24 h and should be corrected.

Aiming to obtain a set of global daily temperature data standing for the air temperature status under a prospective dateline, new algorithms were developed.

Given the massive volume of the global dataset and our objective to provide a rapid, scalable “first estimate” baseline, the entire quality control procedure is fully automated. Quality controls were implemented for both the input hourly data and the produced daily data. Global hourly temperature data quality tests include a spike value test, a stuck value test (which identifies prolonged sequences of the same value in the data series), and an inner consistency test (assessing the relationships between Tmax, Tave, and Tmin). The daily data underwent these same tests, along with additional temporal and spatial consistency tests. Specifically, the temporal consistency test identifies values that deviate excessively from the station's long-term historical distribution, while the spatial consistency test compares a station's record with simultaneous observations from neighboring stations within a 100 km radius to detect localized anomalies. This specific suite of quality control tests was selected because it represents the established international standard, strictly adhering to WMO guidelines (WMO, 2011, 2018) and the comprehensive automated quality assurance protocols developed for major global datasets such as GHCN-Daily (Durre et al., 2010; Menne et al., 2012) and the ISD (Lott, 2004). Details of these tests for daily data are provided in the Appendix.

Throughout the processing pipeline, data quality results are evaluated via programmed conditional checks at each step and automatically flagged into three distinct categories: credible, suspicious, and erroneous. The quality test results at each stage are compiled into a final assessment of data quality levels for each individual data point (i.e., specifically for each individual daily record). A final quality level for a given observation is flagged as credible if there is no more than one suspicious test result and no erroneous test results across all applied tests. Conversely, a value is flagged as erroneous if more than one erroneous test result is found; otherwise, the final quality level is classified as suspicious.

It is important to emphasize that GLBD-FED is fundamentally designed as a near-real-time, first-hand operational product rather than a retrospectively homogenized benchmark dataset. Its primary utility lies in rapid, temporally accurate evaluations of regional synoptic weather events and validating numerical weather prediction models. Caution should be exercised if applying it directly to long-term decadal climate trend detection without further statistical homogenization.

Furthermore, while the GLBD-FED processing framework was built for continuous near-real-time updates, its current operationalization is constrained by upstream data source transitions. Following the recent retirement of the ISD archive, we evaluated its successor, NOAA's GHCN-Hourly (GHCNh). Our assessment revealed that the meteorological elements contained in GHCNh have been significantly reduced, lacking the specific sub-daily extreme reports necessary to robustly support our daily Tmax and Tmin reconstruction methodology. Therefore, while real-time streaming is currently paused, the completed 1981–2024 GLBD-FED archive stands as a highly valuable, temporally aligned 44-year historical first-hand baseline for the global meteorological community.

Figure 12 presents the spatial distribution (panels a1, a2, and a3) and temporal changes (panels b1, b2, and b3) of global daily Tmax, Tave, and Tmin data from 1981 to 2024. The colorful dots in panel (a) indicate the duration of daily data at each site, while the gray and black curves represent the daily data volume and the 15-point smoothing results, respectively.

Tmax, Tave, and Tmin exhibit very similar spatial distributions and temporal changes over the last four decades. Panels (a1), (a2), and (a3) show approximately 17 000 sites with at least one year of daily Tmax data. Sites in China, Japan, and central Europe have extensive time series of daily Tmax (≥30 years, indicated by red dots). Western Europe and the US demonstrate a high spatial density of daily Tmax data.

Notably, the apparent abundance of short time series in the U.S. (indicated by green and blue dots, <20 years of records) is largely an artifact of the legacy ISD station identification system. Under this older system, minor station relocations or administrative updates frequently resulted in the assignment of new station IDs, artificially fragmenting what are physically continuous observations. As recent advancements have shown, processing these records through the unified identifier framework of the newer GHCNh dataset (e.g., merging closely located fragments under a single USW-prefix station ID) effectively resolves these artificial breaks and restores the long-term continuity of the time series. Conversely, countries like China tend to retain a consistent ID across relocations to preserve long-term continuity, which necessitates more rigorous homogenization attention in subsequent applications. Brazil also displays a high spatial density of data, with most stations beginning temperature observations in the last decade. Substantial sites with at least 20 years of daily temperature data can be found in southern Canada, coastal Australia, Russia, and southern Asia. Although the Antarctic and Arctic regions are among the most challenging for meteorological observation globally, dozens of sites have commenced measurements in the current century.

The global daily data volume for Tmax, Tave, and Tmin has increased significantly over the last four decades. As shown in panels (b1), (b2), and (b3), the global daily data volume rose from approximately 3000 entries per day in the 1980s to around 10 000 entries per day in recent years.

It should be noted that the occasional sharp drops observed in the time series are real features of the raw data stream, corresponding to specific periods of severe disruption in global data transmission caused by operational instabilities (e.g., telecommunication outages) within the WMO Global Telecommunication System (GTS) and national exchange networks.

Furthermore, the quasi-periodic fluctuations (dips) observed in the early Tmin record (Fig. 7, panel b3) are artifacts of historical data scarcity in the underlying ISD archive. Our reaggregation algorithm is designed to primarily utilize explicitly reported extremes (Tmin-12 h/24 h), relying on high-frequency hourly observations as a secondary fallback. During these earlier years, the raw archive occasionally lacked both explicit extreme reports and sufficient sub-daily observations. When neither data source meets the necessary computational thresholds, valid daily Tmin values cannot be reconstructed, leading to temporary drops in the aggregated data volume.

Figure 13 presents the multi-decadal time series (1981–2024) of global daily temperature differences between GSOD and GLBD-FED. Calculated as the arithmetic mean across the strictly limited spatiotemporal intersection of the two datasets (i.e., only matched records present in both GSOD and GLBD-FED were included), the results demonstrate that GSOD exhibits a warmer daily Tmax (around +0.3 °C), a colder Tmin (around -0.3 °C), and nearly the same daily Tave (around +0.1 °C) relative to GLBD-FED throughout the entire period. This means that rapid, near-real-time extreme weather assessments and operational synoptic evaluations relying on legacy datasets like GSOD would systematically misrepresent the magnitude of extreme events – overestimating warm extremes and underestimating cold extremes – compared to methodologically uniform evaluations based on GLBD-FED.

A detailed examination of the multi-decadal time series (Fig. 8) reveals two notable temporal features: a pronounced seasonal cycle in the biases prior to 1990, and a long-term downward trend in the daily average temperature (Tave) bias (decreasing from approximately +0.25 °C in the 1980s to roughly +0.10 °C in the recent decade). Both features are intrinsically linked to historical data quality artifacts and algorithmic characteristics. In the early era, the seasonal cycle in the Tmax bias was largely driven by GSOD's processing of cold-season data from high-latitude regions. During boreal winters, raw observations in these environments were prone to anomalous positive spikes (unrealistic short-term jumps). GSOD's fallback extraction strategy misclassified these anomalous spikes as valid daily maximums, artificially inflating the winter Tmax and subsequently skewing the overall Tave upward. Simultaneously, the Tmin seasonality was driven by periodic variations in the volume of duplicated records within the GSOD archive. The long-term decreasing trend in the Tave bias reflects the progressive modernization of the global observing network. As automated weather stations and enhanced transmission protocols were widely deployed, the frequency of raw sensor noise dropped significantly. In contrast, GLBD-FED demonstrates higher stability throughout the 44-year period, as its rigorous temporal consistency checks successfully filtered out these anomalous spikes even during the early, noisier era.

Figure 14 illustrates the spatiotemporal heterogeneity of daily temperature discrepancies between GSOD and GLBD-FED from 1981 to 2024. The sites in GSOD with higher daily Tmax (≥0.5 °C) relative to GLBD-FED are primarily located in Southern Africa, Brazil, Argentina, Canada, and the western United States (panel a1), accounting for 21.4 % of all sites (approximately 17 000) (panel a2).

Interestingly, GSOD shows lower Tmax values in China compared to GLBD-FED. This discrepancy arises from several types of reports labeled with the same ID in the ISD. The Metar and Synoptic reports in China are independent and measured at different sites. Metar temperature reports tend to be lower than Synoptic temperature reports since the latter are mostly recorded in urban areas. GSOD prioritized Metar reports as data sources, while GLBD-FED optimized Synoptic reports, leading to the observed negative bias in China.

Meanwhile, sites in GSOD with a lower daily Tmin (≤-0.5 °C) compared to GLBD-FED are predominantly concentrated in Australia, Russia, and Northeast Asia (panel c1), accounting for 17.4 % of all evaluated sites. For eastern Australia in particular, this regional bias is highly likely attributable to structural discrepancies between national reporting practices and global transmission standards. Specifically, the standard national observation time for Tmin in eastern Australia is 22:00 or 23:00 UTC (depending on the season), whereas international synoptic reports sometimes stamp these observations at 00:00 UTC. This temporal misalignment leads to Time of Observation Bias (TOB) causing double-counting or the shifting of minimum temperatures across the 00:00 UTC boundary in the legacy GSOD dataset. Daily Tave from GSOD and GLBD-FED exhibits much greater consistency compared to Tmax and Tmin (panel b1), with 95.3 % of sites maintaining differences within ±0.5 °C (panel b2).

GLBD-FED and GSOD employ different algorithms to produce daily temperature data, resulting in discrepancies in data properties and biases between the two datasets, particularly for Tmax and Tmin. Table 1 outlines three main differences in the key processes of daily Tmax and Tmin data production between GLBD-FED and GSOD.

First, GLBD-FED and GSOD have distinct definitions for daily Tmax and Tmin data. GLBD-FED considers records that could or almost represent the highest or lowest temperature in a 24 h period as daily Tmax and Tmin, while GSOD selects the highest and lowest records within the 24 h.

Second, although both GLBD-FED and GSOD utilize sub-daily data shared globally through the Global Telecommunication System (GTS) as their data source, they apply this data differently. GLBD-FED prefers Synoptic reports over Meteorological Aerodrome (Metar) and other reports when multiple types of sub-daily reports share the same ID, whereas GSOD does the opposite.

Third, GLBD-FED makes efforts to adjust the daily temperature boundary to 00:00 UTC whenever possible, while GSOD retains the highest and lowest values as they appear within the 24 h period.