The WoKaS-Iso database compiles observations of oxygen (δ18O) and hydrogen (δ2H), expressed in delta (δ) notation relative to the Vienna Standard Mean Ocean Water (VSMOW), from 241 karst springs and 74 caves accommodating a total of 365 drip observation points (e.g., different speleothems). In addition, WoKaS-Iso includes associated karst spring discharge, cave drip rate, precipitation amount observations and precipitation isotope observations for karst springs and cave sites. Table 1 details the number of time series included in the database for various variables.

Among the 241 karst springs, 122 (51 %) are accompanied by discharge data. Of these, 95 springs include both discharge and output δ18O and δ2H data, while 27 springs possess discharge data paired with either δ18O or δ2H measurements. The remaining 119 springs (49 %) include only output isotope data: 53 springs with both δ18O and δ2H data, and 66 with either δ18O or δ2H data. Within the 74 caves, one cave includes input data alone but lacks output data. The remaining 73 caves comprise 365 drip sites, among these, 162 drip sites (44 %) combine drip rate with drip water isotope data, of these, 155 drip sites have drip rate along with δ18O and δ2H measurements, while 7 drip sites have drip rate and only δ18O data. Additionally, 25 drip sites (7 %) were measured containing drip rate only. Furthermore, 178 drip sites (49 %) contain drip water isotope data alone, 125 drip sites with both δ18O and δ2H data, and 53 drip sites with δ18O only. In addition, the database includes 127 time series of rainwater isotopes and 119 time series of rainfall associated with karst springs, as well as 66 time series of rainwater isotopes and 58 time series of rainfall associated with cave sites. The complete list of karst springs and caves included in the database, along with their coordinates and referenced literature, is presented in Tables S1 and S2 in the Supplement.

Figure 2 illustrates the sources of the karst spring and cave drip data, respectively. For the karst springs, 90 % of the isotope data are raw observations, while 10 % are digitized (extracted from a reference paper). The corresponding karst spring discharge data consist of 86 % raw data from collaborators, the WoKaS discharge database, or online open-access national databases. For the rainwater isotopes, 8 % of the data are from local measurements, 45 % from weather stations, and 47 % from global isotope model products. Besides, 50 % of the rainfall data are sourced from local observations and meteorological station data. In terms of cave drip waters, 96 % of drip water isotopes derive from direct field measurements, with only 4 % digitized. A parallel distribution is observed for drip rate records, where 72 % are sourced from observations. In the case of rainwater, the majority of rainwater isotopes are obtained from both local sampling campaigns and meteorological stations, with approximately 45 % originating from local observations. Likewise, 78 % of rainfall data are acquired from local records and meteorological station datasets.

Karst Springs: As summarized in the left branch of Fig. 1, the time series of δ18O and δ2H output data were collected from (1) published data in peer-reviewed literature, scientific reports, and academic theses; (2) research collaborators within the network. Published data: The publications containing stable isotope time series data for karst springs were identified through systematic searches of academic databases and web search engines using targeted keywords (e.g. “karst spring isotopes”). Some isotope datasets were obtained directly from the supplementary documents accompanying these publications. For the data that were not readily accessible, requests were sent to the corresponding authors to solicit contributions to the database. In cases where no response was received, the datasets were digitized from figures in scientific articles, reports and theses using WebplotDigitizer (Rohatgi, 2023), a web-based open-source tool for data extraction. To maximize digitization accuracy, high-quality figures in articles were retrieved programmatically. The extracted datasets were then visually validated by replotting and comparing them with the original figures to ensure the consistency of the trends. Previous methodological studies have shown that carefully extracted graphical data can closely reproduce original numerical data and generally have limited influence on downstream statistical analyses, although extraction accuracy depends on graph quality, axis scaling, symbol clarity, and the size of plotted features (Van der Mierden et al., 2021; Turner et al., 2023). Following the approach used in the WoKaS database, digitization uncertainty is treated qualitatively through metadata documentation and data quality attribution rather than through fixed numerical error estimates (Olarinoye et al., 2020). All digitized records are explicitly identified in the metadata, allowing users to distinguish them from directly collected observations and to account for this additional uncertainty in subsequent analyses. Research collaborators: To further enrich the WoKaS-Iso database, additional data were acquired through communications with research project partners and institutional collaborations. Advertisements were disseminated to relevant academic departments, organizations and researchers. Furthermore, calls for contributions were made to the karst community through poster and oral presentations at international conferences and workshops. In addition to the sources above, the corresponding karst discharge data were obtained from online national databases such as the U.S. Geological Survey (USGS)'s National Water Information System, and Austria's Bundesministerium für Nachhaltigkeit und Tourismus (eHYD) database.

Cave Monitoring: As illustrated in the right branch of Fig. 1, the compilation of cave drip water isotopes and drip rate for the WoKaS-Iso database followed a structured workflow analogous to that of karst springs. In collaboration with SISAL (Kaushal et al., 2024), data acquisition primarily drew upon the SISAL_mon_v1 (Treble et al., 2026), which provided a key foundation for this component. GNIP station data excluded from SISAL_mon_v1 (as noted in the SISAL workbooks) are retained in WoKaS-Iso to preserve the completeness and representativeness of the original monitoring datasets. For some caves, SISAL_mon_v1 includes only the drip water isotope data provided by the authors, while the original drip rate measurements were not available. To address this, WoKaS-Iso digitized the drip rate information from figures in the publications and incorporated these data into the database, thereby enabling quantitative analyses of cave hydrology, drip water dynamics, and speleothem formation processes. Furthermore, additional cave monitoring sites, independent of SISAL_mon_v1, have been incorporated from published studies and research collaborators. All cave drip water datasets were systematically standardized, and additional quality control according to the WoKaS-Iso protocol (see Chap. 4) was applied. For datasets originating from SISAL_mon_v1, this quality control focused on additional harmonization and alignment with the WoKaS-Iso structure, building on the quality control already performed in Treble et al. (2026). The datasets were then organised within a structure specifically designed to ensure interoperability with the broader WoKaS-Iso. To ensure transparency and facilitate cross-referencing, the corresponding SISAL_mon_v1 cave identifiers are provided in the WoKaS-Iso metadata tables where applicable. Additional information on the cave drip water data collection methodology and site selection in SISAL_mon_v1 is provided in Treble et al. (2026).

The spatial distributions of springs and caves within the database are shown in Fig. 3 together with the World Karst Aquifer Map (WOKAM) (Chen et al., 2017). The locations in the database are primarily situated across key global karst regions, including China, Europe, Australia, the Middle East, and the USA. We summarized the detailed statistics of data type combinations of 241 springs and 339 cave drip sites with isotopic data records in Fig. 4, stratified by the predominant karst areas identified in Fig. 3. Springs are analysed for China, the USA, and Europe, while cave drip sites are grouped by China, Australia, and Europe. The stacked histograms display the distribution of dataset types across these regions and align with the regional distribution trends described in the preceding section. Figure 5 further illustrates regional differences in isotope dynamics by presenting example δ18O time series from three karst springs and three cave drip water systems across North America, Europe, and Asia.

Most karst spring records are collected in Europe, with approximately half of them accompanied by discharge measurements (Fig. 4a). Similar proportions are observed in China, whereas only one-third of spring records in the USA include discharge data. In contrast, more than 80 % of spring records from other regions are associated with discharge measurements. Regarding stable isotope observations, datasets containing both δ18O and δ2H are most common in China (86 %), followed by the USA (70 %) and other regions (62 %). For cave drip water systems (Fig. 4b), drip rate observations are most frequently available in China and are present in roughly half of the records from Australia and Europe. In addition, more than 95 % of cave drip water records in China and Australia include both δ18O and δ2H measurements, while the proportion exceeds 70 % in Europe and other regions.

Stable isotope dynamics generally differ among regions due to differences in climatic conditions, precipitation isotope inputs, and karst hydrological processes. Therefore, three karst springs and three caves with overlapping observation periods were selected from North America, Europe, and Asia as illustrative examples to show the temporal dynamics of δ18O. Figure 5a shows considerable differences in both isotope values and temporal variability among the three springs. BS1 Spring (North America) exhibits relatively enriched δ18O values (approximately -5 ‰ to -2 ‰), whereas Anyak Spring (Europe) remains comparatively stable around -10 ‰. In contrast, Luxi No.1 Spring (Asia) displays larger temporal fluctuations, with δ18O values ranging from approximately -14 ‰ to -9 ‰. Similar regional differences are also evident in cave drip water records (Fig. 5b). Larga Cave (North America) is characterised by relatively enriched isotope values around -3 ‰, while Beke Cave (Europe) and Yangkou Cave (Asia) exhibit more depleted values between approximately -10 ‰ and -7 ‰. Moreover, multiple drip sites within individual caves show common temporal behaviour but differ in absolute δ18O values.

Stable isotopes in spring discharge and drip waters can show strong seasonal variability due to fluctuations in precipitation and temperature under changing climatic conditions. Adequate temporal resolutions are necessary to quantify these seasonal dynamics in detail and capture short-term hydrological patterns during rainfall or snowmelt events. The regional maps in Fig. 3a–f provide the spatial context of isotope sampling resolution, showing that monthly sampling is common across the main data-rich regions, particularly for cave drip water records. Higher resolution isotope records are mainly available for karst springs in Europe, whereas finer than biweekly resolution is rare for cave drip water records in China, Australia, and Europe. Figure 6 provides a summary of temporal resolution for hydrological and isotope measurements in the WoKaS-Iso. Spring discharge datasets generally have the highest temporal resolution, with nearly half available at daily or finer intervals, compared with just over 12 % of cave drip rate datasets (Fig. 6a). Cave drip rate records are mainly concentrated in the biweekly-to-monthly category, reaching approximately 74 %, whereas spring discharge records are more evenly distributed across the resolution classes. For stable isotope observations, resolution between biweekly and monthly is the dominant category for both springs and cave drip waters, accounting for around 64 % of cave drip isotope datasets and more than 40 % of spring isotope datasets (Fig. 6b). In addition, 36 % of spring isotope datasets have a resolution finer than biweekly, while 27 % of cave drip water isotope records are sampled at intervals coarser than monthly.

The workflow for compiling the input datasets is summarized in the central panel of Fig. 1. The WoKaS-Iso database includes precipitation amount and precipitation δ18O and δ2H observations for karst springs and cave sites, compiled from the same sources described in Sect. 2.1.2 (also see Table 1 and Fig. 2). To ensure adequate spatial and temporal coverage, karst springs and cave sites lacking local rainfall and rainwater isotope records were supplemented with the data from Global Network of Isotopes in Precipitation (GNIP) database. Firstly, the geographic coordinates of karst springs, cave sites, and GNIP stations within the related countries were imported into ArcGIS where the closest GNIP station to each karst spring and cave site was identified. Secondly, the distances between the springs or caves and their nearest GNIP stations were measured using ArcGIS. With the aim of reflecting the representativeness of the local climatic and isotopic characteristics for karst springs and caves, only GNIP stations within 25 km were retained. This threshold was selected to be consistent with the spatial resolution of the Isoscape and Sine Curve datasets (5 arcmin, approximately 10 km), corresponding to roughly twice the grid-cell diagonal (∼28 km). It is therefore a conservative choice that is also compatible with the coarser ERA5 dataset (∼31 km). Although this approach cannot fully account for topographic differences between GNIP stations and karst sites, it provides a consistent spatial matching criterion for all products evaluated in this study.

For karst springs, 20 sites possess local rainfall measurements, and 99 springs derive precipitation data from nearby meteorological stations, including GNIP stations, while the remaining 122 springs lack rainfall data. Regarding precipitation isotope observations, only 18 springs have auxiliary local isotope data, 109 springs use data from the meteorological or GNIP networks, and 114 springs have no related isotope records. For cave sites, 21 caves include locally measured rainfall, 37 caves obtain rainfall data from adjacent weather stations, and 16 caves lack rainfall records. As for precipitation isotopes, 33 caves are associated with local measurements, 33 caves rely on weather station-based isotope datasets or GNIP-derived data, and 8 caves lack isotope observations. To offer comprehensive metadata for input observations, the distances from the precipitation and precipitation isotopes sampling sites to cave/spring sites were calculated where geographic coordinates were available and then documented in the metadata tables for spring input and cave input datasets.

Since datasets on precipitation amount and precipitation δ18O and δ2H, observed locally or provided by GNIP, were not available at all output locations or were limited in their temporal extent, the WoKaS-Iso database includes complementary datasets from six global products for continuous rainfall, precipitation δ18O and δ2H, air temperature as well as both actual and potential evapotranspiration. The information on these global products used in this study is listed in Table 2, including relevant modelled variables, time span, spatial and temporal resolution and the source for the construction of models.

Precipitation data were obtained from MSWEP v2.8 (Beck et al., 2017, 2019a, b), a multi-source ensemble product offering daily and monthly estimates at 0.1° resolution from 1979–2020. MSWEP improves global precipitation accuracy by combining gauge, satellite, and reanalysis data using a weighted ensemble based on network density and local performance (Liu et al., 2019). Evaporation estimates were sourced from GLEAM v4.1a (Hulsman et al., 2023; Koppa et al., 2022; Martens et al., 2017; Miralles et al., 2011, 2025; Zhong et al., 2022), which integrates satellite data and physical modelling with machine learning to estimate daily and monthly actual and potential evaporation at 0.1° resolution from 1980–2023. Air temperature time series were extracted from ERA5, the fifth-generation atmospheric reanalysis developed by ECMWF (Hersbach et al., 2020, 2023), which provides consistent daily climate data since 1940 at 0.25° resolution using the 4D-Var data assimilation technique and the CY41R2 forecasting system. To characterize precipitation isotope variability, we used monthly δ18O and δ2H outputs from IsoGSM (Kanamitsu et al., 2002a, b; Nan et al., 2021; Yoshimura et al., 2008; Bong et al., 2024), a general circulation model incorporating isotopic fractionation and nudged with NCEP (National Centers for Environmental Prediction) reanalysis data, covering 1979–2021. Additionally, long-term spatial isotope patterns were derived from the Isoscape product (etopo5, 2025; Bowen et al., 2005; Bowen and Revenaugh, 2003; Bowen and Wilkinson, 2002), which interpolates GNIP data using topographic and geographic predictors at a 5 arcmin resolution. Finally, seasonal isotope dynamics were represented using the sinusoidal precipitation isotope model (Allen et al., 2019), which fits annual sine curves to global precipitation isotope records and maps the resulting parameters to derive daily and monthly isotope predictions from 60° S to 90° N. Together, these products provide a consistent climatological and isotopic framework for interpreting site-level karst and cave hydrology.

While global precipitation and evapotranspiration data were evaluated against observations in their original studies (see Table 2), an evaluation of the isotopic composition of rainfall provided by the three precipitation isotope models – IsoGSM, Isoscape, and the Sine Curve model – is still required. For that, we compared monthly simulated precipitation δ18O and δ2H against site-specific observations from both cave drip water and karst spring locations. For precipitation δ18O comparison, a total of 55 sites were analysed, comprising 26 cave locations, 15 spring locations, and 14 nearby meteorological stations to compensate for the limited availability of isotopic measurements at spring locations. For precipitation δ2H, 40 sites were assessed, including 25 caves and 15 springs. For sites with daily precipitation, the monthly amount-weighted precipitation isotopes were calculated by weighting the available isotope measurements with the corresponding daily precipitation amounts and aggregating them to monthly values. For sites lacking daily precipitation, the stepwise interpolation of precipitation isotopes was applied to derive monthly means.

We quantified the agreement between the global products and the observed isotopic composition of precipitation using the root-mean-square error (RMSE) and the coefficient of determination (R2). Figure 7 presents the performance metrics in predicting δ18O. Based on RMSE, IsoGSM displays a lowest median RMSE value (∼2 ‰) with a narrow interquartile range and minimal outliers, indicating the most consistent predictions across both caves and springs. The Isoscape model exhibits a comparable median RMSE, though with greater variability and higher outliers, reflecting less robustness. In contrast, the Sine Curve model yields the highest median error (>3 ‰) and broader distributions, highlighting significantly inferior performance. Concerning R2, IsoGSM achieves the highest median R2 (∼0.6); however, its R2 values span a wide range. The Isoscape model, having a slightly lower median R2 value (∼0.5), shows a tighter distribution. The Sine Curve model performs worst, with a median R2 of less than 0.2 and a concentration of low values, indicating poor explanatory power.

Figure 8 presents similar performance trends for δ2H predictions. IsoGSM achieves the lowest median RMSE (∼17 ‰) with a relatively narrow spread, reflecting the most accurate and consistent predictions. The Isoscape model follows closely with a slightly higher RMSE (∼19 ‰) and greater variability. The Sine Curve model shows the highest median RMSE value (∼29 ‰) and the broadest RMSE distribution. For R2, there is a slight shift in performance ranking compared to δ18O. The Isoscape model has the highest R2 value (∼0.5). The IsoGSM model has a lower R2 (∼0.4) with a more balanced distribution. The Sine Curve model exhibits the poorest fit, with a median R2 below 0.2.

To evaluate the different performances of the global datasets, we compared time series of observed monthly δ18O and δ2H with model outputs from the Sine Curve, IsoGSM, and Isoscape models across all selected locations. Figure 9 shows the δ18O comparisons at four representative sites: Liangfeng Cave, Yongxing Cave, Chaoshuidong Spring and the Met-II.uz meteorological station near Anyak Spring (see Sect. S1 in the Supplement for a complete comparison of other 51 sites). Overall, the Sine Curve model shows limited ability to reproduce the temporal course of the observations. In contrast, the IsoGSM model closely aligns with the observed δ18O dynamics at the four sites, yielding high R2 values (0.85, 0.81, 0.94, and 0.78 for Liangfeng Cave, Yongxing Cave, Chaoshuidong Spring and the Met-II.uz meteorological station near Anyak Spring, respectively) alongside relatively low RMSE values, reflecting robust performance in both correlation and prediction accuracy. The Isoscape model attains moderate predictions, with R2 exceeding 0.5 at all four sites. Although it outperforms the Sine Curve model, it tends to be less reliable than IsoGSM in capturing the isotopic variability.

Figure 10 presents δ2H time series comparisons between observations and simulations from three models at four sites: Heshang Cave, Yongxing Cave, Krbavica Spring and Laolongshui Spring (comparisons for other sites are shown in Fig. S2 in the Supplement). Similarly, the Sine Curve model shows the poorest agreement with observations at Heshang Cave, Yongxing Cave, and Krbavica Spring, as reflected by low R2 values of 0.37, 0.57, and 0.1 respectively, and high RMSEs (up to 50.11 ‰). Although it shows relatively higher R2 values of 0.85 at Laolongshui Spring, its corresponding large RMSE value of 46.55 ‰ reveals substantial absolute predictive errors. The IsoGSM model shows the best performance at Heshang Cave and Yongxing Cave, with the highest R2 (0.8 and 0.82) and lowest RMSE (15.53 ‰ and 10.56 ‰). These results highlight the IsoGSM model's strong predictive accuracy in tracking δ2H variability at both sites. At Krbavica Spring, IsoGSM and Isoscape exhibit comparable performance in terms of correlation, with R2 values of 0.71 and 0.7. However, IsoGSM shows a lower RMSE (21.3 ‰) compared to Isoscape (35.21 ‰), indicating that although both models reproduce the temporal variation similarly, IsoGSM offers more precise estimates of δ2H magnitudes. At Laolongshui Spring (Fig. 10d), the Isoscape model outperforms the IsoGSM model and Sine Curve model, with the highest R2 and lowest RMSE values.

Overall, the comparisons indicate that IsoGSM shows slightly superior performance over Isoscape and pronouncedly superior performance over the Sine Curve model. Different phase shifts between δ18O and δ2H in the Sine Curve model indicate a poor representation of their physical covariance, which is given through the global/local meteoric water line. Based on these results, IsoGSM is recommended as the preferred product for cave and spring locations where no local observations are available.