Data drives our understanding of hydrological processes, supports model development, and enables anticipatory water management. This contribution introduces EARLS: European Aggregated Reconstructions for Large-sample Studies. EARLS offers daily streamflow reconstructions for more than 10 000 basins in Europe including uncertainty estimates, covering the period from 1953 to 2023. The reconstruction is derived from a single Long Short-Term Memory (LSTM) based rainfall–runoff model trained on more than 5000 basins. LSTMs represent the state of the art in rainfall–runoff modeling and are well suited to provide predictions in ungauged basins. We evaluate the quality of the reconstruction through quantitative evaluation on two held-out sets of basins and by conducting a qualitative assessment that compares EARLS-based peak flows and flood timing to previous large-scale hydrological studies. EARLS represents a new generation of datasets that harness the capabilities of Deep Learning to obtain accurate and high-resolution data. EARLS is available at 10.5281/zenodo.13864842.
European Commission101059372Introduction
Data availability is central to hydrological science. It is the basis for advancing our understanding of hydrological processes, building prediction models, and anticipatory water management. However, in many regions and periods, observations are scarce. Practitioners often rely on hydrological model predictions to compensate for these deficiencies e.g.,. The resulting dataset are referred to as reconstructions. More generally, we can think of reconstructed data as datasets that are generated from observational records with the aid of models to address gaps in the data. There exist many challenges associated with making reconstructions. In particular, the non-linearity between drivers and the unique characteristics of each basin make it difficult to use process-based hydrological models for high-quality reconstructions e.g.,. Here, purely data-driven approaches provide a great opportunity. Current approaches based on Machine Learning (ML) are able to simulate a diverse range of rainfall–runoff responses when trained using data from multiple basins . They are able supply high-quality predictions e.g., and are often suitable for simulating even in ungauged basins . These characteristics render data-driven approaches an exceptional but currently underused tool for large-scale hydrological analyses.
Data availability also plays a particularly important role in Large Sample Hydrology (LSH). LSH concentrates on multiple basins, as opposed to conducting detailed studies in individual regions. A central question in LSH is how we can transfer knowledge about runoff responses between different basins, given their hydrological similarities e.g.,. Large-scale datasets allow us to identify patterns and formulate conclusions about hydrological processes across diverse regions. Recent advancements in LSH have led to the development of several large-scale hydrological databases, compiling streamflow records from multiple locations. Notable global databases are: (1) The Global Streamflow Indices and Metadata Archive GSIM;, which contains monthly, seasonal and yearly indices for over 35 000 locations; (2) The Global Runoff Data Center database GRDC;, which provides discharge estimates for over 10 000 locations; and (3) The Caravan project , which incorporates already open-source published streamflow data from various countries.
The requirement to drive rainfall–runoff models with meteorological forcings led to the development of integrated datasets that include meteorological time series such as precipitation and temperature. To our knowledge, the Model Parameter Estimation Experiment MOPEX provided the first openly available large sample dataset of this kind. It contains 431 basins within the United States. Some of the most important contributions in popularizing large-scale datasets after MOPEX stem from the basin Attributes and MEteorology for Large-sample Studies (CAMELS) initiatives , and other derivations such as LamaH and CABra . Each CAMELS dataset is tailored to a specific region, but at its core follows the same logic–connecting meteorological variables and static basin attributes with streamflow data. The Caravan project incorporates already open-source published streamflow data from various CAMELS countries – and with its recent extension also contains the sharable GRDC data . In this study, we however use EStreams as our “raw material” for model building. EStreams also constitutes an integrated dataset, but goes into a different direction: It provides basic data for setting up hydrological models, but also a catalog streamflow data from national data providers. Unlike the CAMELS and Caravan datasets, it focuses only on the European scale. It also offers higher spatial resolution, and is designed to be continuously updated, providing the latest records in both time and space.
Despite these numerous developments in building large-scale databases for hydrology, important sampling gaps exist. This holds in particular with respect to high-quality runoff observations. This limits the usability of said databases for certain scientific applications and decision-making processes at a pan-European level. To address these limitations, here we present a data-driven daily runoff reconstruction product for natural streamflow. We name it EARLS: European aggregated reconstruction for large-sample studies. Our main goal for EARLS is to provide data-driven streamflow reconstructions that enable the analysis of hydrological processes in the style of . The reconstructions represent daily simulations of natural streamflow, are provided in mm, and cover the period from 1953 to 2020.
One can view EARLS as part of a new generation of datasets that leverage ML to achieve highly accurate predictions e.g.,. We employ a rainfall–runoff model based on Long Short-Term Memory LSTM; to create these reconstructions. Recent studies have demonstrated the accuracy of this approach in various contexts e.g,. The model incorporates static attributes that describe basin properties (say, average elevation) and a series of meteorological forcings (say, daily precipitation) to simulate streamflow for a given basin, following the approach introduced by and . We evaluate the resulting simulations in terms of predictive performance for ungauged basins. Like EStreams, EARLS focuses on Europe. In particular, we use a subset of the EStreams basins with CAMELS attributes for training our model (see Sect. ), aiming to provide spatially extensive long-term streamflow reconstructions with uncertainty estimates, including for ungauged basins across Europe.
Method
EARLS provides streamflow reconstructions for 17 043 European basins. The data are in daily resolution, comprise uncertainty estimates, and span for each basin from 1 January 1953 to 30 June 2023. The time period is constrained by the meteorological forcing used (see below). We plan to extend it later as the meteorological dataset gets updated. Gaps in the reconstructions only occur when the dynamic inputs are erroneous – which can happen if the meteorological forcing have gaps at different timesteps. EARLS contains 14 161 basins without data gaps, 2655 with gaps of variable lengths, and we were not able to produce a simulation for 227 basins (these are not part of the 2655 basins wit gaps). For these basins there is so much missing data in the inputs that our modelling approach is not able to make simulations at all. We also keep these in EARLS since the meteorological forcing is regularly revised and we plan to update on a regular basis (Sect. ). Since it is not appropriate for royalty to be alone, we plan to produce EARLS versions with less and less gaps in the future, and eventually even start sister projects with different focal points. The basins for model training are from EStreams (Sect. ). However, we also derive a set of virtual ungauged basins to provide continuous spatial coverage (Sect. ). The model is lumped (which necessitates an aggregation of the inputs at the basin level) and provides not only point estimates but also uncertainty estimates in the form of a distribution prediction. Section provides an overview of the modeling setup.
Training and evaluation data
We spatially aggregate dynamic inputs (i.e., meteorological forcings) and static inputs (i.e., basin-specific static attributes) for each basin. The basin shapes are either derived from EStreams or HydroATLAS . When both shapes are available we use both for the training. All streamflow data is in millimeters per day (mm d−1). We use 4 dynamic inputs (i.e., precipitation, minimum, maximum, and mean temperature) that we aggregate basin wise from version 28 of E-OBS dataset . This means that we do not use all available dynamic inputs, i.e., we do not leverage all information that we could. The EARLS LSTM will not provide the best possible predictions indeed, albeit we do not show the results here, we did make some experiments with more inputs and did obtain better results; see also:. However, we believe this disadvantage is offset by the increase of flexibility in use of the model. E-OBS is a daily-resolution gridded dataset covering the European region (25–71.5° N × 25° W–45° E). This defines the extent of EARLS in both space and time. E-OBS interpolates station data from European National Meteorological Services and other providers, spanning 1 January 1950 to present (in EARLS the modelling period starts, however, in 1953, since we use the first three years as buffer time). It comprises time series of meteorological variables such as daily mean, maximum, and minimum temperature, daily total precipitation and mean sea level pressure. For the static inputs, we use 13 attributes that we aggregate from HydroATLAS following the convention from Caravan;: basin area, the average elevation, the average slopes, the average stream gradient, the average long-term air temperature, the minimum long-term air temperature, the maximum long-term air temperature, a global aridity index , a global climate moisture index , the average fraction of sand the average fraction of clay, the average fraction of silt, and the average organic carbon content. We chose this subset for the sake of simplicity, but, in general, different inputs and combinations therefore are thinkable to create new models and datasets.
The streamflow training data originate from national and regional hydrometric through the EStreams catalog. The streamflow data are generally accessible through the respective agencies; however, they are distributed under heterogeneous licensing conditions (as mention, availability of does not necessarily imply permission for third-party redistribution). Consequently, and in accordance with the data governance approach adopted for EStreams, the raw daily streamflow observations that we use for model training are not redistributed as part of the EARLS dataset. Instead, the full documentation of the original data providers and references to the corresponding sources are available in the Supplement and the data publication . This allows users to obtain the records directly from the responsible agencies under their respective licensing frameworks.
Preprocessing
To get streamflow observation we alleviate EStreams . EStreams contains hydro-climatic variables and landscape descriptors, and references to openly available streamflow records for 17 130 European basins. It includes basin delineations, hydro-climatic signatures, and landscape attributes (topography, soils, geology, vegetation, and land cover), and gives the necessary information to access daily streamflow data from the data providers, which we cannot redistribute (see above). The data quality of EStreams basin does, however, vary considerably. We filter out gauged basins according to the following criteria:
Each basin needs to have high-quality delineations see Table 3 in.
To minimize aggregation errors in the basin mean attributes and simultaneously reduce the effects of channel routing, we only include basins equal or larger than 50 km2 and smaller than 100 000 km2.
We require each basin to have at least 30 years of, not necessarily consecutive, daily streamflow observations.
We exclude basins which, based on the attributes derived in EStreams, include more than four dams or reservoirs within the basin boundary.
We require the presence of meteorological time-series from E-OBS for the basins.
We exclude basins where hydrological signatures indicate potential data problems, based on the following criteria:
The long-term average streamflow needs to be below 10 mm d−1.
The long-term runoff ratio as defined in cannot be larger than 1.
After applying these constraints to the EStreams catalog, we are left with 5786 basins with streamflow observations (blue circles in Fig. ; see also Appendix ). We further partitioned these into 4789 training basins, 500 chosen basins for validation, and 500 for testing. That is, we do not apply any time split for the training and evaluation of the EARLS LSTM. In other words, validation is carried out in space, but not in time.
Spatial distribution of the gauged (blue circles) and ungauged basins (grey triangles).
Our preprocessing is not perfect. Specifically, many of the low basins (a) are influenced by human activities, such as the presence of dams and reservoirs (Portugal and Spain); (b) exhibit extensive canal systems and numerous lakes (Denmark, Sweden, and Norway); (c) contain karstic geology (Central Europe); or (d) are situated regions where the meteorological forcing data are scarce Iberian Peninsula and southern Italy; see. In our data preprocessing we filter out basins that exhibit a high degree of human influence (Sect. ). In Spain and Portugal, anthropization is primarily caused by numerous dams constructed for water supply. The identification of such structures at such a large-scale is challenging . Hence, EStreams may not correctly capture the total number of dams and reservoirs in many regions because of the used data sources as discussed in. Furthermore, the high number of natural lakes and the presence of canalization systems may also negatively influence the model's performance in these areas. The same applies for the presence of karstic systems, which poses a challenge for closing the water balance in some basins. EStreams made significant efforts to label such basins . However, despite the efforts we were not able to produce ex-ante labels for said basins or create an adequate criterion to filter them out (Sect. ). This affects both, model training and evaluation, since some signals are not learnable in the first place.
Rainfall–runoff model
Our LSTM-based rainfall–runoff model uses dynamic and static inputs to estimate streamflow. Since the LSTM is a deep learning architecture we will use concepts and language from machine learning to describe how we set it up. For example, for the model selection we will use a triple split (training, validation, and test set) and we will refer to the selection procedure as training (and not, say, as model calibration as is usual in hydrology). To provide uncertainty estimates, we adapt a simplified version of the approach from . In short, instead of estimating the streamflow directly, the LSTM outputs the three parameters of an asymmetric Laplacian distribution (a double exponential with a parameter) – and is trained using maximum likelihood (Fig. ). Appendix provides a short overview of our approach. For a more detailed and general exposition we refer to . From here on out we refer to this model as the EARLS LSTM. The next two sections describe how we set up the training and evaluation of the EARLS LSTM. Further technical details are available in Appendix .
High-level model conceptualization. For each time step t The LSTM outputs the location parameters μt, the scale parameter bt, and the asymmetry parameter τt to parameterize an asymmetric Laplace distribution for each predicted timestep. For the use and definition of the static and dynamic inputs we refer to Sect. .
Streamflow reconstructions
Overview of the modeling setup for the (a) gauged and (b) ungauged basins. The yellow boxes indicate that we concatenate the same static input for each timestep. The static attributes are just different from basin to basin. In contrast, the dynamic input vary for each timestep (and basin).
Our model setup deviates slightly between gauged and ungauged basins (Fig. ), since for the latter no observations are available (Fig. ). Specifically, we use the gauged basins for training and evaluating the EARLS LSTM (Fig. a). For each gauged basin we use the observations from the EStreams catalog (Sect. ), the static inputs from HydroATLAS, and the dynamic inputs from E-OBS. The ungauged basins (Fig. b) are the basins without corresponding observations. They add a set of virtual basins as additional support basins for EARLS. The goal here is to achieve dense coverage across Europe (gray basins in Fig. ). We delineate the ungauged basins from the union of all upstream level 12 polygons of the level 12 layer of HydroATLAS. Like in the gauged case, the static and dynamic inputs are derived from HydroATLAS and E-OBS, respectively. The resulting “simulation layer” consists of 11 277 basins (some overlapping with the gauged basins), which yields a total of 17 043 EARLS basins when combined with the gauged basins that we use for training.
Evaluation
We corroborate the data quality of EARLS by evaluating the model performance on a large set random test basins. On top of that, we provide extensive appendices that examine the data properties with a mix of quantitative and qualitative measures (Appendix : We examine whether model performance is related to static or dynamic basin characteristics (Appendix ), perform a comparative analysis with a process based model on the basis of 161 separately chosen basins (Appendix ), and conduct a qualitative assessment against published literature results (Appendix ).
For the model evaluation we report the performances for the training, validation and test set (for a technical definition we refer to Appendix ). It normally does not make sense to report training and validation performances in a ML context, since one is generally interested in generalization – and performance measures are biased for the training and validation sets. However, since we also publish simulations for these basins we argue that it is informative for users to also report the respective performances. Specifically, we report Nash–Sutcliffe efficiency (NSE) for each basin and over the time horizon of a given portion. For a given basin it is defined as:
NSE=1-∑t=1T(ot-st)2∑t=1T(ot-o¯)2,
where t=1,2,…,T is the time index for the given basin and time horizon, o the observations, s the simulations, and o¯=1/T∑t=1Tot the sample mean of the evaluated data. Appendix also reports the model performance on the test set with regard to other metrics.
Evaluation results and discussion
In the following, we show the results of our model evaluation. To assess the model performance we compute the NSE for 500 test basins over the full range of the EARLS time horizon (i.e., 1953–2023). The EARLS LSTM achieves a median NSE of 0.66 for the 500 test basins. We view this as a good result, given our splitting strategy, which leads to potentially difficult to predict ungauged basins for the evaluation (Sect. and Appendix ), and the utilization of only a single dynamic input product as, for example, opposed towho examine the use of multiple forcings. On top of that, our pre-filtering strategy is rather coarse (see Appendix ), hence many (ungauged) basins remain in the dataset that can be difficult to handle (e.g., Karst-affected catchments). Still 10 % of the basins exhibit NSE values that are lower than 0.0 (Fig. a).
To give readers a rough comparison: These results are similar to the ones in , who report that 8 % of the basins exhibit negative NSE values. The highest basin has an NSE of 0.93 and the lowest of -25.42. These values are not randomly distributed in space (Fig. b). Appendix provides empirical cumulative distribution functions for other metrics.
We posit that most low accuracy values can be attributed to data quality issues rather than to shortcomings in the LSTM (Sect. ). On top of that it is worth to remember that that there is evidence that the NSE can be rather erratic in arid climatic regimes e.g.,.
Performance evaluation. For both plots we clip the negative values of Nash–Sutcliffe efficiency (NSE) at -0.5 to focus on the important part of the performance distribution. (a) Empirical cumulative density function. Each point on the line represents the model performance for one of the 500 test basins. The red-solid line shows the performance for the test data. The dashed lines show the performance for the training and validation data. (b) The correspondend spatial distribution of the respective NSE values.
Limitations
Our results suggest that EARLS as a dataset is well suited for large-sample hydrological studies in Europe. However, there exist several important limits to EARLS: The simulation quality is restricted by (a) the streamflow observation quality, (b) the input quality, and (c) the capabilities of the EARLS LSTM. The quality of streamflow and input data varies in space (e.g., different measurement standards across countries) and time (e.g., improvements in measurement technique over time). As a matter of fact, some of the observations are highly atypical. It would not be surprising if they cannot be modeled with the available information. The same kind of reasoning applies to the forcings. For example, E-OBS is more accurate in high station-density regions like Germany and Austria, and less so in lower density areas like Spain, Portugal, or Eastern Europe see. These results suggests the existence of non-trivial distribution shifts in the reconstructions. From a machine learning perspective distribution shifts are challenging to model well. Lastly, with regard to (c), a machine learning model – such as the EARLS LSTM – by design, can only capture the signal that is in the data. We chose an LSTM-based approach since it represents the best simulations in gauged and ungauged settings . Nonetheless, we kept the setup simple for ease of use. For instance, we use a limited set of dynamic inputs that are available in many observation-based datasets and use a simple single Laplacian distribution in the output and a single meteorological product for the dynamic inputs. We did not conduct extensive intercomparisons and expect better results with more complex setups. In that case, the current EARLS version will serve as a (strong) baseline.
Further, EARLS incorporates distributional prediction for each time step by providing the conditional parameters of an asymmetric Laplace distribution (Sect. ). This does not represent the full uncertainty of the prediction. On top of that, the conditional distribution assumes no knowledge of the streamflow. Thus, if we naively take samples at a given time step, then this does not account for the autocorrelative nature of the streamflow.
Data and code availability
The project homepage for EARLS is https://earls-dataset.github.io/ (last access: 13 July 2026). We envision the page as a living document for future news and updates. The EARLS data is available at 10.5281/zenodo.13864842. It includes streamflow reconstructions for 17 043 European basins. We build the dataset with extensibility in mind. The idea here is that the current structure becomes a blueprint for potential future expansions. Specifically, the data of EARLS is organized as follows:
The “coordinates.csv” file contains basin outlet information with 5 columns: basin id (idx), type, and estimated latitude (lat) and longitude (lon) of the outlet, and the distance to the nearest gauged station. The type indicates whether the streamflow information of a given basin was used for training. Together with the information of nearest gauging station it can be used by analysts to demarcate whether a reconstruction is well supported by the training data (see Appendix ).
The “license.md” file contains information about the licensing.
The “shapefile” folder includes a shapefile with all basin boundaries.
The “reconstructions” folder contains CSV files. In general, each file is named after the basin id and has at least the columns for date (date) and simulation (sim). The simulations are given in mm d−1 and we use the location parameter of our conditional distribution estimate. To provide a probability density estimate for each reconstruction timestep in EARLS we added two additional columns (namely: tau and b). Together with the location parameter (saved as sim) these fully define an asymmetric Laplacian. Hence EARLS comprises uncertainty estimates for each time step (Sect. ).
The “model-card” folder contains 3 files: “model-card.html”, and “earls-crest.png”. The html document includes the png as logo and renders a model card. A model card is a short summary of the model genesis, designed to increase transparency by communicating information about trained models to broad audiences . We include all three files in the dataset so that future extensions can adapt them with maximal ease. We will also host the markdown files on the main home so that the permanent identifier within the model card can be used to access the data from there.
Additional data/folders are optional, but can be used to provide background information. For instance, to enable benchmarking the current EARLS version also contains an “inputs” folder, which comprises the basin-aggregated dynamic and static inputs (Sect. ):
For the dynamic inputs derived from we use precipitation in mm d−1, daily minimum temperature in °C, daily maximum temperature in °C, and daily mean temperature in °C.
For the static inputs derived from we use basin area in km2, average elevation in m, average slopes in degrees, average stream gradient in dm km−1, average long-term air temperature in °C, minimum long-term air temperature in °C, maximum long-term air temperature in °C, a global aridity index , a global climate moisture index , average fraction of sand in %, average fraction of clay in %, average fraction of silt in %, and average organic carbon content in t ha−1.
The EARLS LSTM is not part of the dataset itself. However, we provide the code for the EARLS LSTM, our experiments, and plots at https://github.com/earls-dataset/paper-code (last access: 16 July 2026) (in addition, a snapshot of the code can be found at 10.5281/zenodo.19107554, ).
Conclusions
We provide a data-driven streamflow reconstruction product for Europe, called EARLS (European aggregated runoff reconstruction for large-sample studies). EARLS is part of a new generation of datasets created using machine learning e.g.,. The main purpose of EARLS is to enable large-sample hydrological streamflow analysis at European scale, for instance to complement or strengthen analyses based on hydrological model simulations . As of now, EARLS consists of reconstruction for 17 043 European basins from 1953 to 2023, at a daily scale. From these, over 11,000 represent ungauged basins from HydroATLAS. This motivates our model-driven approach for the reconstructions. The model also enables us to provide predictions that are distributional in nature. That is, for each time step EARLS provides a conditional uncertainty estimate – which can, for example, be used to compute the likelihood of a given model. This, for example, enables researcher to explore which situations are associated with what kind of uncertainties or to train classical models on top of it using the information in their objective functions.
argued that streamflow is a model-derived variable – a virtual quantity. We like to think that creating reconstructions for thousands of basins brings this observation to a new level. As of yet it remains unclear how valuable such synthetic observatories will be for the community. In our eyes the main value of EARLS-like reconstructions is that the model extract streamflow information from the meteorological input signals. This makes it possible to create a wide and dense net of streamflow observations in space and time so that researchers do not have to rely on the sparsely available streamflow data. Hence, the model can be viewed as a virtual sensor that provides estimations daily streamflow values with their associated aleatoric uncertainties (Sect. ). An alternative view is to look at the reconstructions as support points for interpolation exercises, which allow for more nuanced patterns than just using the spatially sparsely distributed streamflow observations. That said, for us it is important to emphasize that EARLS – like all simulated datasets – is not a replacement for observations . EARLS is only possible due to large amounts of diverse, high-quality data Appendix ;. Given the increase in data availability and advancements in computational approaches, we posit that machine learning will become integral to datasets (in many applications it already is). In the future, entirely new forms of datasets may emerge, inheriting their own advantages and disadvantages. EARLS, and other currently published datasets might then be seen as stepping stones for this new class of dataset.
In the future we would like to extend EARLS into an ensemble of reconstructions by leveraging the different available inputs (static or dynamic) and using different filtering criteria (Sect. ). To give a specific example of how this could look like: EStreams comes with its own set of static inputs and we want to build reconstructions with them. Similarly, we could use and combine, as in other dynamic inputs like the ones from ERA5-Land . We plan to conduct more extensive hyperparameter searches and model comparisons, and to enlarge the scope beyond Europe to a global scale. We encourage the wider community to participate.
Technical details of the modeling process
This appendix lines out the technical aspects of the modeling process.
Training
Ultimately, modeling for EARLS is an exercise in spatial generalization. Gauged and ungauged basins are not randomly distributed (Sect. , Fig. ). Direct evaluation is only possible for the former, but the EARLS LSTM should also generalize to the latter. We choose a data split strategy that reflects this inherent challenge. Intuitively, our goal is to partition the data so that the distribution of the training, validation, and test sets is different enough to estimate an out-of-sample model performance for the ungauged basins. To measure the difference in distribution, we use the Wasserstein-1 distance of the standardized static inputs (standardization is used to prevent that we just measure the differences in feature magnitudes). The Wasserstein or earth mover’s distance W1 is a distributional distance that measures the smallest distance between two sets of samples. It is widely used in machine learning e.g.,. Formally, the Wasserstein-1 distance is defined as
W1(μ,ν)=infγ∈Γ(μ,ν)∫M×M|xi-xj|dγ(x,y),
where Γ(μ,ν) is the set of all couplings of μ and ν. In theory the considered distance can be chosen, but we only consider the absolute difference here. In practice, we sample xi from a dataset Di and xj from a dataset Dj respectively. The sampling is necessary, since the W1 distance expects that we have the same amount of samples from the compared distributions, but we use a different number of basins for each set. Namely: The training set Dtrain with 5386 basins, and validation and test sets, Dval and Dtest, with 500 basins.
We sum the distances between all pairs of the three partitions to get an overall distance dW that expresses how far apart they are from each other:
dW=1K∑kKW1(Atraink,Aval)+1K∑kKW1(Atraink,Atest)+W1(Atest,Aval).
Here, K=20 defines the number of repetitions, and Atraink indicates a random subset of size 500 from the training dataset. This is needed because W1 assumes the same sample size from the distributions. To get the final split, we make create 3×20 random sets and select the partition with the largest distance dW. From an optimization standpoint, obtaining large dW values is a combinatorial problem and the use of random partitioning is suboptimal. Hence, we experimented with clustering-based subsetting and approaches using an optimizer to exchange individual data points during the development. These and similar strategies would align more closely with cluster-based splitting of training, test and validation sets as proposed by or . If dW becomes too large, the training performance will not translate to the validation performance, and the validation performance, in turn, will not indicate the test performance. We defer the design of better separation schemes that optimally emulate the task in question to future work.
Technical specificationsHyperparameters
In order to find good hyperparameters for the EARLS LSTM we use a mixture of manual and grid search. The goal was to find a good trade-off between model simplicity and generalization. Table shows the final parameters for the EARLS LSTM.
Distributional predictions
In order to achieve distributional predictions we adapt the approach from , letting the LSTM parameterize a single asymmetric Laplacian for each time step. To get the runoff estimate within the EARLS we use the location parameter of the distribution. We do the training in normalized space as described in. For the dataset we rescale the location and scale parameters, but leave the asymmetry parameter unchanged.
Hyperparameter settings for the EARLS LSTM.
HyperparameterValueInitial forget gate bias3Hidden size350Batch size3000Number of epochs40 (and choose best validation performance for model)Learning rate schedule (epoch:learningrate)0:0.001, 5:0.0005, 20:0.0001Clipping to gradient norm1.0LossNegative Log-likelihoodSequence length365Standard deviation of the noise added to the streamflow observations0.1A very short introduction to mixture density networks with asymmetric Laplacian distributions
In many hydrological modeling settings, we are interested in representing predictive uncertainty in a flexible, data-driven way. One such approach are mixture density networks (MDNs), which ingest the same input as a neural network would but parametrize a distribution as output. In other words, with MDNs we can provide an distributional estimation of the runoff for each time step. adopted a special form of MDN, which they coined Countable Mixture of Asymmetric Laplacians (CMAL). CMAL replaces the Gaussian components of a standard MDN with asymmetric Laplacian components. The idea here is to make the individual components of the MDN more powerful and move them closer to the conditions encountered in rainfall–runoff modeling. In CMAL, each component can express skewness and has heavier tails than a Gaussian. By mixing several such components, we can model multimodal, skewed uncertainty structures. In EARLS, however, we only use the special case CMAL, where we have a single component: An asymmetric Laplacian distribution (ALD).
An ALD is defined by the probability density function:
ALD(y∣μ,b,τ)=τ(1-τ)bexp-(y-μ)(τ-1{y<μ})b,
where μ is the location parameter (and also the mode), b>0 is the scale parameter (which controls the dispersion), and τ∈(0,1) is an asymmetry parameter (which defines the skewness of the distribution). For each time step t of the prediction the LSTM outputs all three parameters {μ,b,τ} in dependence of the received input. When τ=0.5, the ALD reduces to a symmetric Laplace distribution. For τ≠0.5, one tail becomes longer than the other, producing a skewed shape. This makes ALD components suitable for modeling asymmetric error distributions that are common in flow predictions. In the EARLS dataset we report the μ as the streamflow estimation, but also provide b and τ. Hence, the dataset provides full distributional estimations for each simulated time step.
We train our model by minimizing the negative log-likelihood of observed data:
L=-∑tlogALDyt∣μ(xt),b(xt),τ(xt).
Further data quality checksPost hoc model examination
As a post hoc analysis of the model evaluation we investigate the relationship between model performance and (a) static inputs or (b) streamflow respectively. For this, we use the 500 basins from the test set (Appendix ) and the basin similarity measure from . In their conception Z=Z1,Z2,…,ZM is a collection of different basin attributes where each ZM=[z1(m),z2(m),…,zN(m)] is a vector of a given attribute m (e.g., average elevation) over the different basins M. That is, each entry in the vector is property of interest (e.g., the average elevation) for a given basin. Then, for two basins i and j the “Bertola distance” d is the Euclidean distance:
d(Z,i,j)=∑m=1Mzi(m)-zj(m)sd(Zm),
where sd(Zn) denotes the standard deviation of Zn. We make use of two different choices for Z. The first set consists of the 13 static inputs of our model and the second choice depends on the streamflow only. For the latter, we choose the logarithm of the mean of the annual maximum specific streamflow following . That is, the annual maximum discharge normalized to a basin area of 100 km2see. Formally, we can express these choices as
da(Z=S1,S2,…,S13,i,j)=∑m=1Msi(m)-sj(m)sd(Sm),
and
db(Z=log(Q̃),i,j)=∑m=1Mlog(q̃i(m))-log(q̃j(m))sd(log(Q̃m)),
where Si refers to the static inputs, and log(Q̃) refers to the normalized annual discharges. The distance da does solely depend on the static inputs and is therefore always computeable. In contrast, db does only depend on the runoff and can hence exclusively be used for model diagnosis in gauged basins. To understand whether particularly low or high model performance might be related to static basin attributes or runoff dynamics, we analyze two subsets of basins: basins for which NSE<0 and those for which NSE>0.8. We then compare the following sets of average distances
D<0.0=Ei(d(Z,i,j))|i≠jandNSEi<0.0∀i,j∈1,2,…,M,
and
D>0.8=Ei(d(Z,i,j))|i≠jandNSEi>0.8∀i,j∈1,2,…,M.
If there is a pattern between within the two subsets D<0.0 and D>0.8, then it could be possible to predict whether the model performs well or not – and modelers can use that information to infer whether a model performs well or not for a given basin.
Results
We observe a shift between badly performing basins in D<0,0 and well performing basins in D>0.8 (Fig. a). This shift is absent for the same analysis using static attributes (Fig. b). In fact, for the latter, the mode of the distribution is at a lower value despite comparing higher dimensional entities (i.e., the 13 static attributes). These quantitative results suggest that it is easier to discriminate the model performances with streamflow observation than with static attributes. The results also align with our hydrological justifications for model performance (Sect. ), since anthropogenic factors are not encoded in static attributes but are reflected in the streamflow.
Comparison of the distributions of average “Bertola distances d” (Appendix ) for the subset of basins with low NSE values (D<0.0) and the subset of basins with high NSE values (D>0.8). Plot (a) shows the respecitve distances for the streamflow and plot (b) fo the static inputs (Appendix ).
Model to model comparison
We use the mesoscale hydrological model mHM; as a reference model for the comparative evaluation. Specifically, we run mHM in two configurations for seperate 161 basins that are not part of EARLS. The first configuration is represented by the global parametrization from , hereafter referred to as “mHM default”. For the second configuration, we calibrate the parameters for each basin with 1970–1999 for training and use the years 2000–2020 for testing, hereafter referred to as “local mHM”. In both cases, we use E-OBS for the dynamical inputs (precipitation, average temperature and potential evapotranspiration estimated with the Hargreaves-Samani method and static inputs from Table . Since mHM is intrinsically a semi-distributed model, we set the spatial resolution of the model to 0.25°. For the local mHM, we maximize the NSE using the dynamically dimensioned search algorithm with 1000 iterations. The comparative evaluation with the EARLS LSTM is therefore asymmetric: Firstly, the reconstructions are tested for an ungauged setting, while local mHM – as our reference model – is calibrated using a traditional time-split setting. Secondly, the mHM default is not a result of a specific calibration process for the task at hand and is hence disadvantaged.
Morphological data used for the hydrological model mHM.
DescriptionSourceDigital Elevation Model from U.S. Geological Survey (USGS)Soil map from SoilGridsLand cover from the European Space Agency (ESA)LAI climatology from NASA Global Inventory, Monitoring, and Modelling StudiesResults
In terms of performance, the EARLS LSTM ranks between the mHM default and the locally calibrated mHM (Fig. ). That is, until approximately the 15th percentile of NSE values, the EARLS LSTM performance is close to the mHM default performance, and roughly starting at the 40th percentile, it is close to the local mHM model. Else, it is somewhere in between, and for the best performing basins EARLS LSTM even outperforms the latter. All in all, we argue that these are promising results for an ungauged evaluation – especially if we keep in mind that this is an asymmetric comparison: the EARLS LSTM operates “out-of-sample” (validation in space), while mHM operates “in-sample” in space (validation in time).
Empirical cumulative density functions for the comparative evaluation. The mHM default is a “best-guess” mHM calibration that summarizes many studies; the mHM local are per basin calibrated models evaluated in a traditional time-split fashion; and the EARLS LSTM represents the ungauged performance of the EARLS model.
Qualitative assessment against published literature
We use EARLS to redo the core parts of the flood-timing analysis from and the flood-peak trends analysis from . Both studies use observations from 1960 to 2010 (albeit the full period is not available for all gauges). We select the same timeframe and use the publicly available code from to compute the trends and spatial interpolations of the peak trends. To recreate the results of we adapt it for the flood-timing analysis according to their supplementary material (Appendix provides details of the adoption process).
With the reproduction of the maps we want to provide a visual check for the data quality of EARLS. As far as we know, this also constitutes the first corroboration of the results from with different raw data (since the underlying raw data from the original papers are not available open access). The intrinsic limit of this assessment is that there is is no actual ground truth available. The maps from are derived by interpolating from the sparsely available gauging station data, while we get the maps from EARLS by interpolating from a denser network of gauging station (however, based on reconstructions). Hence, neither of them should be seen as absolute.
Results
The first part of our qualitative assessment revolves around the timing of river floods in Europe. These results corresponds to Figs. 1 and 3 of . We encourage readers to compare our version with these depictions since all key patterns from the original analysis are preserved in the EARLS version. Figure shows the average timing of the yearly streamflow maxima from EARLS. The overall pattern of our this version closely corresponds to the original – but with a larger number of support points and a wider area of analysis. The reproduction of the corresponding analysis of flood timing trends from EARLS (Fig. ), also mirrors the large-scale trends from Fig. 1 from the original publication. The effect of the Pyrenees, the Alps and the Carpathians are clearly visible. The 4 approximate key region with distinct drivers that highlight are also reflected in the EARLS version: (1) Northeastern Europe with earlier snowmelt; (2) the North Sea region with later winter storms; (3) Western Europe along the Atlantic coast with earlier soil moisture maxima; and (4) Parts of the Mediterranean coast (West Spain, South France, Croatia, etc.) with stronger Atlantic influence in winter.
Reproduction of Fig. 3 from with EARLS data. Each arrow represents a basin outlet (either from a gauged or ungauged basin). Color and arrow direction indicate the average timing of floods over the period 1960–2010 (namely: light blue are winter floods; green to yellow are spring floods; orange to red are summer floods; and purple to dark blue are autumn floods). The lengths of the arrows indicate the concentration of floods (0, evenly distributed; 1, all floods occur on the same date).
The plotting data from are openly available. We can therefore directly compare them with our EARLS version. To this end, we made a reproduction of their results (Fig. a) and contrasted it with the corresponding EARLS version (Fig. b). Both show similar large-scale trends, but significant discrepancies exist for large and small-scale patterns. Scandinavia and North-East Europe have the biggest divergence in terms of large-scale patterns. There, the original version shows slightly decreasing trends in floods, while the EARLS version shows no or slightly increasing trends. In absolute terms the differences are not that large, but the extent at which the differences occur is noteworthy. The largest absolute differences, on the other hand, occur in the north of Portugal (where the EARLS version shows major negative trends) and West Russia/Ukraine (where the original analysis depicts substantially larger negative values).
Some differences can be explained by data availability (Appendix ). In general, the EARLS version appears to have more details, which, we posit, is linked to a higher number of data points used for kriging. EARLS lacks reconstructions for the Asian part of Turkey, while have observations there. have limited observations in western Russia and northern Ukraine, showing strong negative trends in flood magnitudes. EARLS has denser coverage, showing negative trends restricted to a specific eastern region. However, E-OBS is based on few observations stations in Eastern Europe the actual number varies from variable to variable; see, e.g., Fig. 6 in. Neither of the two analyses have data in the most northern part of West Russia, resulting in flat spatial trends. Furthermore, in the data from the availability of streamflow observations varies widely from station to station, while EARLS is more homogeneous in this regard, potentially leading to different trend estimates. In summary, we argue that the results of our qualitative assessment show the merit our the EARLS data for scientific inquiry and corroborate the quality of the simulations.
Our EARLS-based remake of the flood peak timing trends from . Increasing trends are depicted in blue, negative trends in red. The most extreme negative trends can be spotted in Spain/Portugal and the strongest decreasing trends in the west of Norway. In general, the overall patterns match the ones from the original publication. However, they do show more detail because the interpolation is made on basis of many more supporting points than available for the original (Appendix ).
Comparison of the trend analysis for flood peaks from 1960 to 2010 derived from different datasets: Our reproduction with data of annual maxima from (a), and (b) EARLS. Increasing trends are in blue, decreasing ones in red. manually binned their data into specific classes. In contrast, we choose a continuous color-scale centered around zero. document the data and the original code for (a), while provides the complete plotting code and the underlying data for the figure.
Other metrics for evaluation
Empirical cumulative density function of the Kling–Gupta efficiency (KGE) for the test evaluation. Each point on the line represents the model performance for one of the 500 test basins.
This appendix shows the results for different evaluation metrics for the evaluation experiment presented in Appendix . Specifically, Fig. a shows the Kling–Gupta efficiency, (b) shows the Pearson's correlation coefficient, and (c) shows the root mean squared error.
Estimation of trend statistics
Our procedure for estimating the long-term trends statistics for yearly flood timing and yearly peak flow trends (Appendix ) follows and respectively. This section partially mirrors the supplementary material and describes the technical details of our implementation.
Yearly peak-flow trends
For the flood trend analysis part we follow : First, we extract a series of observerations that consists of the highest peak discharge recorded in each calendar year (i.e., the annual maximum peak flow) from EARLS. Then, we estimate the trend in each series using a robust approach and interpolate the trend estimates using Kriging. only provide the data for their experiments, while provide data and R-code for their experiments. Hence, for this demonstration we modified their code to stay as close as possible to their results. The robust estimation is achieved by using the Theil–Sen slope β:
β=medianQi-Qji-j|i,j∈Jand i≠j,
where Qz indicates the maximal streamflow of a given year z and J contains the indices of all the years from 1960 to 2010 – which corresponds to the time span that use.
Flood timing analysis and trends
The demonstration of the flood timing analysis follows . Specifically, we reproduce two of their investigations: (1) an examination of long-term trend of the flood-timing and an analysis of the average flood timing over Europe (represented by Figs. 1 and 3 in ; and Figs. and in our contribution). Following their procedure, we first compute for each station the average day D within a year where peak flows have occurred during the observation period. And, to account for the cyclic nature of yearly data all calculations are performed using circular procedures/statistics. As we only take the stations for which the null hypothesis of circular uniformity (which is assessed with Kuiper's test) is rejected with a significance level of α=0.1. For this contribution no code is available. Hence, we use the description in the supplementary material to modify the code from for the trend examination. This means, that code for the average flood timing analysis is largely from ground up (according to the provided documentation from the supplementary material).
We convert date of occurrence Di of a flood in year i into an angular value using:
θ=2πmiwith0≤θi≤2π,
where mi is the number of days for year i, and Di gives the corresponding day of the year so that Di=1 corresponds to 1 January and Di=mi to 31 December.
We compute the average date of occurrence D of a flood at a station as:
D¯=tan-1y¯x¯m¯365forx¯>0,y¯≥0,tan-1y¯x¯+πm¯2πforx¯≤0,tan-1y¯x¯+2πm¯2πforx¯>0,y¯<0.
Here, the arc-tangens tan-1 yields the angle in radians, x and y are the cosine and sine components of the average date, m is the average number of days per year (which we re-calculate to account the fact that some years had no data), and n is the total number of flood peaks at that station. That is:
x¯=1n∑i=1ncos(θi),y¯=1n∑i=1nsin(θi),
and
m¯=1n∑i=1nmi.
Lastly, define concentration R of the date of occurrence around the average date as:
R=x¯2+y¯2with0≤R≤1.
Indeed, the mapping of D¯ and R constitutes our reproduction of the average flood timing analysis from .
Trends in timing
For the timing trend estimation we use the same adjusted Theil-Sen slope estimator as reported by . The computation is similar to the one given by Eq. (), but adds a correction factor k to account for the circularity of the task:
β∘=medianQi-Qj+ki-j|i,j∈Jand i≠jwithk=-m¯ifDj-Di>m¯/2,m¯ifDj-Di<-m¯/2,0 otherwise.
Here, the Theil-Sen slope β∘ has units of days per year. Thus to get an estimate for the 10 year period reported in it has to be scaled accordingly. Lastly, we used the same Kriging approach as we do for the yearly peak-flow trends to get a map of the large-scale spatial patterns within Europe.
Locations
If we compare the “location”/support points for kriging between and EARLS we can see that the latter uses many more location points to supported the interpolation (Fig. ). To provide a rough estimate about the difference in density we count the number of basins within a box around Scandinavia (colored dots in Fig. ). In that case we get approximately 300 for the location points from , and around 2500 simulated stations locations in the same box for EARLS (note: for the former count we rounded up in the decimal, while for the latter we rounded down in the hundreds). However, an analyst might still feel uncomfortable to use reconstructions that are not at gauging station that was used for model training or too far away from such a station station. The “coordinates.csv” file of EARLS (Sect. ) contains two columns to provide assistance for such circumstances: The “basintype” provides a flag for gauged and ungauged basins (the former were used for training, the latter not) and the “nearest_gauged_distance” column provides the Euclidean distance to the nearest gauged station computed from the latitude and longitude. As shown in Fig. this allows analysts to demarcate region where the reconstructions are supported on solid ground truth data – i.e., Western and Northern Europe – from reconstructions that are not – i.e., Eastern and Southeastern Europe.
Comparison of the available of location/support points for the kriging-based interpolation. Plot (a) shows the reference points from and plot (b) the simulated station from EARLS. The colored areas in both (a) and (b) refer to the same area of the data based on a box a given latitude and longitude. For the former it contains approximately 300 points, for the latter around 2500.
Distance to nearest gauged station. Points represent simulation sites, with the color indicating the Euclidean distance to the nearest gauged station. In the plot, gauged sites (that we used for training) and those near them appear near the lower end of the color scale, while points that are far away from stations that have been used during the training on the higher.
The supplement related to this article is available online at https://doi.org/10.5194/essd-18-5485-2026-supplement.
Author contributions
JZ and DK developed the idea, conceptualization, and method of the paper. DK did all LSTM simulations. PM conducted the mHM simulations. TN and FF helped with the EStreams setup and model realizations. MG provided the setup for the ungauged basins as well as additional model simulations, control experiments, and checks. CF contributed extensively to make EARLS more reproducible. All authors were involved in the writing of the paper.
Competing interests
The contact author has declared that none of the authors has any competing interests.
Disclaimer
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. The authors bear the ultimate responsibility for providing appropriate place names. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.
Acknowledgements
We thank Frederik Kratzert for his input with the data and the modeling setup, as well as Rohini Kumar for his help with the mHM model, and Emanuele Bevacqua for discussing intricacies of the E-OBS data quality with us. We acknowledge the E-OBS dataset and the data providers in the ECA&D project (https://www.ecad.eu, last access: 13 July 2026).
Financial support
Daniel Klotz and Jakob Zscheischler acknowledge funding from the Helmholtz Initiative and Networking Fund (Young Investigator Group COMPOUNDX, grant agreement no. VH-NG-1537). Daniel Klotz acknowledges the STARS4Water project funded through the European Union's Horizon Europe research and innovation program under the grant agreement no. 101059372.The article processing charges for this open-access publication were covered by the Helmholtz Centre for Environmental Research – UFZ.
Review statement
This paper was edited by Kirsten Elger and Conrad Jackisch and reviewed by Wouter Berghuijs and Juliane Mai.
ReferencesAddor, N., Newman, A. J., Mizukami, N., and Clark, M. P.: The CAMELS data set: catchment attributes and meteorology for large-sample studies, Hydrol. Earth Syst. Sci., 21, 5293–5313, 10.5194/hess-21-5293-2017, 2017.Addor, N., Nearing, G., Prieto, C., Newman, A. J., Le Vine, N., and Clark, M. P.: A Ranking of Hydrological Signatures Based on Their Predictability in Space, Water Resour. Res., 54, 8792–8812, 10.1029/2018WR022606, 2018.Almagro, A., Oliveira, P. T. S., Meira Neto, A. A., Roy, T., and Troch, P.: CABra: a novel large-sample dataset for Brazilian catchments, Hydrol. Earth Syst. Sci., 25, 3105–3135, 10.5194/hess-25-3105-2021, 2021.Alvarez-Garreton, C., Mendoza, P. A., Boisier, J. P., Addor, N., Galleguillos, M., Zambrano-Bigiarini, M., Lara, A., Puelma, C., Cortes, G., Garreaud, R., McPhee, J., and Ayala, A.: The CAMELS-CL dataset: catchment attributes and meteorology for large sample studies – Chile dataset, Hydrol. Earth Syst. Sci., 22, 5817–5846, 10.5194/hess-22-5817-2018, 2018.Arino, O., Ramos Perez, J. J., Kalogirou, V., Bontemps, S., Defourny, P., and Van Bogaert, E.: Global Land Cover Map for 2009 (GlobCover 2009), PANGAEA [data set], 10.1594/PANGAEA.787668, 2012.Arjovsky, M., Chintala, S., and Bottou, L.: Wasserstein Generative Adversarial Networks, in: Proceedings of the 34th International Conference on Machine Learning, edited by: Precup, D. and Teh, Y. W., vol. 70 of Proceedings of Machine Learning Research, 214–223, https://proceedings.mlr.press/v70/arjovsky17a.html (last access: 13 July 2026), 2017.Beck, H. E., de Roo, A., and van Dijk, A. I. J. M.: Global Maps of Streamflow Characteristics Based on Observations from Several Thousand Catchments, J. Hydrometeorol., 16, 1478–1501, 10.1175/JHM-D-14-0155.1, 2015.Bertola, M., Blöschl, G., Bohac, M., Borga, M., Castellarin, A., Chirico, G. B., Claps, P., Dallan, E., Danilovich, I., Ganora, D., Gorbachova, L., Ledvinka, O., Mavrova-Guirguinova, M., Montanari, A., Ovcharuk, V., Viglione, A., Volpi, E., Arheimer, B., Aronica, G. T., Bonacci, O., Čanjevac, I., Csik, A., Frolova, N., Gnandt, B., Gribovszki, Z., Gül, A., Günther, K., Guse, B., Hannaford, J., Harrigan, S., Kireeva, M., Kohnová, S., Komma, J., Kriauciuniene, J., Kronvang, B., Lawrence, D., Lüdtke, S., Mediero, L., Merz, B., Molnar, P., Murphy, C., Oskoruš, D., Osuch, M., Parajka, J., Pfister, L., Radevski, I., Sauquet, E., Schröter, K., Šraj, M., Szolgay, J., Turner, S., Valent, P., Veijalainen, N., Ward, P. J., Willems, P., and Zivkovic, N.: Megafloods in Europe can be anticipated from observations in hydrologically similar catchments, Nat. Geosci., 16, 982–988, 10.1038/s41561-023-01300-5, 2023. Beven, K., Buytaert, W., and Smith, L. A.: On virtual observatories and modelled realities (or why discharge must be treated as a virtual variable), Hydrol. Process., 26, 1905–1908, 2012.Blöschl, G., Hall, J., Parajka, J., Perdigão, R. A. P., Merz, B., Arheimer, B., Aronica, G. T., Bilibashi, A., Bonacci, O., Borga, M., Čanjevac, I., Castellarin, A., Chirico, G. B., Claps, P., Fiala, K., Frolova, N., Gorbachova, L., Gül, A., Hannaford, J., Harrigan, S., Kireeva, M., Kiss, A., Kjeldsen, T. R., Kohnová, S., Koskela, J. J., Ledvinka, O., Macdonald, N., Mavrova-Guirguinova, M., Mediero, L., Merz, R., Molnar, P., Montanari, A., Murphy, C., Osuch, M., Ovcharuk, V., Radevski, I., Rogger, M., Salinas, J. L., Sauquet, E., Šraj, M., Szolgay, J., Viglione, A., Volpi, E., Wilson, D., Zaimi, K., and Živković, N.: Changing climate shifts timing of European floods, Science, 357, 588–590, 10.1126/science.aan2506, 2017.Blöschl, G., Hall, J., Viglione, A., Perdigão, R. A. P., Parajka, J., Merz, B., Lun, D., Arheimer, B., Aronica, G. T., Bilibashi, A., Boháč, M., Bonacci, O., Borga, M., Čanjevac, I., Castellarin, A., Chirico, G. B., Claps, P., Frolova, N., Ganora, D., Gorbachova, L., Gül, A., Hannaford, J., Harrigan, S., Kireeva, M., Kiss, A., Kjeldsen, T. R., Kohnová, S., Koskela, J. J., Ledvinka, O., Macdonald, N., Mavrova-Guirguinova, M., Mediero, L., Merz, R., Molnar, P., Montanari, A., Murphy, C., Osuch, M., Ovcharuk, V., Radevski, I., Salinas, J., Sauquet, E., Šraj, M., Szolgay, J., Volpi, E., Wilson, D., Zaimi, K., and Živković, N.: Changing climate both increases and decreases European river floods, Nature, 573, 108–111, 10.1038/s41586-019-1495-6, 2019.Bundesanstalt für Gewässerkunde (BfG): GRDC – Global Runoff Data Center, https://www.bafg.de/GRDC/EN/02_srvcs/21_tmsrs/210_prtl/prtl_node.html (last access: 1 December 2023), 2023.Chagas, V. B. P., Chaffe, P. L. B., Addor, N., Fan, F. M., Fleischmann, A. S., Paiva, R. C. D., and Siqueira, V. A.: CAMELS-BR: hydrometeorological time series and landscape attributes for 897 catchments in Brazil, Earth Syst. Sci. Data, 12, 2075–2096, 10.5194/essd-12-2075-2020, 2020.Coxon, G., Addor, N., Bloomfield, J. P., Freer, J., Fry, M., Hannaford, J., Howden, N. J. K., Lane, R., Lewis, M., Robinson, E. L., Wagener, T., and Woods, R.: CAMELS-GB: hydrometeorological time series and landscape attributes for 671 catchments in Great Britain, Earth Syst. Sci. Data, 12, 2459–2483, 10.5194/essd-12-2459-2020, 2020.Danielson, J. J. and Gesch, D. B.: Global Multi-Resolution Terrain Elevation Data 2010 (GMTED2010), U.S. Geological Survey Open-File Report 2011–1073, 26 pp., 10.3133/ofr20111073, 2011.Do, H. X., Gudmundsson, L., Leonard, M., and Westra, S.: The Global Streamflow Indices and Metadata Archive (GSIM) – Part 1: The production of a daily streamflow archive and metadata, Earth Syst. Sci. Data, 10, 765–785, 10.5194/essd-10-765-2018, 2018.do Nascimento, T. V. M., Rudlang, J., Höge, M., van der Ent, R., Chappon, M., Seibert, J., Hrachowitz, M., and Fenicia, F.: EStreams: An integrated dataset and catalogue of streamflow, hydro-climatic and landscape variables for Europe, Sci. Data, 11, 879, 10.1038/s41597-024-03706-1, 2024.Duan, Q., Schaake, J., Andréassian, V., Franks, S., Goteti, G., Gupta, H., Gusev, Y., Habets, F., Hall, A., Hay, L., Hogue, T., Huang, M., Leavesley, G., Liang, X., Nasonova, O., Noilhan, J., Oudin, L., Sorooshian, S., Wagener, T., and Wood, E.: Model Parameter Estimation Experiment (MOPEX): An overview of science strategy and major results from the second and third workshops, J. Hydrol., 320, 3–17, 10.1016/j.jhydrol.2005.07.031, 2006.Duc, L. and Sawada, Y.: A signal-processing-based interpretation of the Nash–Sutcliffe efficiency, Hydrol. Earth Syst. Sci., 27, 1827–1839, 10.5194/hess-27-1827-2023, 2023.Fang, B., Bevacqua, E., Rakovec, O., and Zscheischler, J.: An increase in the spatial extent of European floods over the last 70 years, Hydrol. Earth Syst. Sci., 28, 3755–3775, 10.5194/hess-28-3755-2024, 2024.Färber, C., Plessow, H., Kratzert, F., Addor, N., Shalev, G., and Looser, U.: GRDC-Caravan: extending the original dataset with data from the Global Runoff Data Centre, Zenodo [data set], 10.5281/zenodo.8425587, 2023.Fowler, K. J. A., Acharya, S. C., Addor, N., Chou, C., and Peel, M. C.: CAMELS-AUS: hydrometeorological time series and landscape attributes for 222 catchments in Australia, Earth Syst. Sci. Data, 13, 3847–3867, 10.5194/essd-13-3847-2021, 2021.Ghiggi, G., Humphrey, V., Seneviratne, S. I., and Gudmundsson, L.: GRUN: an observation-based global gridded runoff dataset from 1902 to 2014, Earth Syst. Sci. Data, 11, 1655–1674, 10.5194/essd-11-1655-2019, 2019.Gudmundsson, L., Do, H. X., Leonard, M., and Westra, S.: The Global Streamflow Indices and Metadata Archive (GSIM) – Part 2: Quality control, time-series indices and homogeneity assessment, Earth Syst. Sci. Data, 10, 787–804, 10.5194/essd-10-787-2018, 2018.Hargreaves, G. H. and Samani, Z. A.: Reference Crop Evapotranspiration from Temperature, Appl. Eng. Agr., 1, 96–99, 10.13031/2013.26773, 1985.Haylock, M. R., Hofstra, N., Klein Tank, A. M. G., Klok, E. J., Jones, P. D., and New, M.: A European daily high-resolution gridded data set of surface temperature and precipitation for 1950–2006, J. Geophys. Res.-Atmos., 113, 10.1029/2008JD010201, 2008.Helgason, H. B. and Nijssen, B.: LamaH-Ice: LArge-SaMple DAta for Hydrology and Environmental Sciences for Iceland, Earth Syst. Sci. Data, 16, 2741–2771, 10.5194/essd-16-2741-2024, 2024.Hengl, T., de Jesus, J. M., Heuvelink, G. B. M., Gonzalez, M. R., Kilibarda, M., Blagotić, A., Shangguan, W., Wright, M. N., Geng, X., Bauer-Marschallinger, B., Guevara, M. A., Vargas, R., MacMillan, R. A., Batjes, N. H., Leenaars, J. G. B., Ribeiro, E., Wheeler, I., Mantel, S., and Kempen, B.: SoilGrids250m: Global Gridded Soil Information Based on Machine Learning, PLOS ONE, 12, e0169748, 10.1371/journal.pone.0169748, 2017. Hijmans, R. J., Cameron, S. E., Parra, J. L., Jones, P. G., and Jarvis, A.: Very high resolution interpolated climate surfaces for global land areas, Int. J. Climatol., 25, 1965–1978, 2005.Hochreiter, S. and Schmidhuber, J.: Long Short-Term Memory, Neural Comput., 9, 1735–1780, 10.1162/neco.1997.9.8.1735, 1997.Höge, M., Kauzlaric, M., Siber, R., Schönenberger, U., Horton, P., Schwanbeck, J., Floriancic, M. G., Viviroli, D., Wilhelm, S., Sikorska-Senoner, A. E., Addor, N., Brunner, M., Pool, S., Zappa, M., and Fenicia, F.: CAMELS-CH: hydro-meteorological time series and landscape attributes for 331 catchments in hydrologic Switzerland, Earth Syst. Sci. Data, 15, 5755–5784, 10.5194/essd-15-5755-2023, 2023.Hrachowitz, M., Savenije, H., Blöschl, G., McDonnell, J., Sivapalan, M., Pomeroy, J., Arheimer, B., Blume, T., Clark, M., Ehret, U., Fenicia, F., Freer, J., Gelfan, A., Gupta, H., Hughes, D., Hut, R., Montanari, A., Pande, S., Tetzlaff, D., Troch, P., Uhlenbrook, S., Wagener, T., Winsemius, H., Woods, R.A. Zehe, E., and Cudennec, C.: A decade of Predictions in Ungauged Basins (PUB) – a review, Hydrolog. Sci. J., 58, 1198–1255, 10.1080/02626667.2013.803183, 2013.Jiang, S., Bevacqua, E., and Zscheischler, J.: River flooding mechanisms and their changes in Europe revealed by explainable machine learning, Hydrol. Earth Syst. Sci., 26, 6339–6359, 10.5194/hess-26-6339-2022, 2022.Jiang, S., Tarasova, L., Yu, G., and Zscheischler, J.: Compounding effects in flood drivers challenge estimates of extreme river floods, Sci. Adv., 10, eadl4005, 10.1126/sciadv.adl4005, 2024.Klein Tank, A. M. G., Wijngaard, J. B., Können, G. P., Böhm, R., Demarée, G., Gocheva, A., Mileta, M., Pashiardis, S., Hejkrlik, L., Kern-Hansen, C., Heino, R., Bessemoulin, P., Müller-Westermeier, G., Tzanakou, M., Szalai, S., Pálsdóttir, T., Fitzgerald, D., Rubin, S., Capaldo, M., Maugeri, M., Leitass, A., Bukantis, A., Aberfeld, R., van Engelen, A. F. V., Forland, E., Mietus, M., Coelho, F., Mares, C., Razuvaev, V., Nieplova, E., Cegnar, T., Antonio López, J., Dahlström, B., Moberg, A., Kirchhofer, W., Ceylan, A., Pachaliuk, O., Alexander, L. V., and Petrovic, P.: Daily dataset of 20th-century surface air temperature and precipitation series for the European Climate Assessment, Int. J. Climatol., 22, 1441–1453, 10.1002/joc.773, 2002.Klingler, C., Schulz, K., and Herrnegger, M.: LamaH-CE: LArge-SaMple DAta for Hydrology and Environmental Sciences for Central Europe, Earth Syst. Sci. Data, 13, 4529–4565, 10.5194/essd-13-4529-2021, 2021.Klotz, D., Kratzert, F., Gauch, M., Keefe Sampson, A., Brandstetter, J., Klambauer, G., Hochreiter, S., and Nearing, G.: Uncertainty estimation with deep learning for rainfall–runoff modeling, Hydrol. Earth Syst. Sci., 26, 1673–1693, 10.5194/hess-26-1673-2022, 2022.Klotz, D., Gauch, M., Kratzert, F., Nearing, G., and Zscheischler, J.: Technical Note: The divide and measure nonconformity – how metrics can mislead when we evaluate on different data partitions, Hydrol. Earth Syst. Sci., 28, 3665–3673, 10.5194/hess-28-3665-2024, 2024.Klotz, D., Miersch, P., do Nascimento, T. V. M., Fenicia, F., Gauch, M., and Zscheischler, J.: EARLS: Code and Weights, Zenodo [code], 10.5281/zenodo.19107554, 2026a.Klotz, D., Miersch, P., do Nascimento, T. V. M., Fenicia, F., Gauch, M., and Zscheischler, J.: EARLS: European aggregated reconstruction for large-sample studies, Zenodo [data set], 10.5281/zenodo.13864842, 2026b.Klotz, D., Miersch, P., Nascimento, T., Fenicia, F., Gauch, M., and Zscheischler, J.: EARLS: Code and Weights, Zenodo [code], 10.5281/zenodo.19107555, 2026c.Knoben, W. J. M.: Setting expectations for hydrologic model performance with an ensemble of simple benchmarks, Hydrol. Process., 38, e15288, 10.1002/hyp.15288, 2024.Kraft, B., Schirmer, M., Aeberhard, W. H., Zappa, M., Seneviratne, S. I., and Gudmundsson, L.: CH-RUN: a deep-learning-based spatially contiguous runoff reconstruction for Switzerland, Hydrol. Earth Syst. Sci., 29, 1061–1082, 10.5194/hess-29-1061-2025, 2025.Kratzert, F., Klotz, D., Herrnegger, M., Sampson, A. K., Hochreiter, S., and Nearing, G. S.: Toward Improved Predictions in Ungauged Basins: Exploiting the Power of Machine Learning, Water Resour. Res., 55, 11344–11354, 10.1029/2019WR026065, 2019a.Kratzert, F., Klotz, D., Shalev, G., Klambauer, G., Hochreiter, S., and Nearing, G.: Towards learning universal, regional, and local hydrological behaviors via machine learning applied to large-sample datasets, Hydrol. Earth Syst. Sci., 23, 5089–5110, 10.5194/hess-23-5089-2019, 2019b.Kratzert, F., Klotz, D., Hochreiter, S., and Nearing, G. S.: A note on leveraging synergy in multiple meteorological data sets with deep learning for rainfall–runoff modeling, Hydrol. Earth Syst. Sci., 25, 2685–2703, 10.5194/hess-25-2685-2021, 2021.Kratzert, F., Nearing, G., Addor, N., Erickson, T., Gauch, M., Gilon, O., Gudmundsson, L., Hassidim, A., Klotz, D., Nevo, S., Shalev, G., and Matias, Y.: Caravan – A global community dataset for large-sample hydrology, Sci. Data, 10, 61, 10.1038/s41597-023-01975-w, 2023.Kratzert, F., Gauch, M., Klotz, D., and Nearing, G.: HESS Opinions: Never train a Long Short-Term Memory (LSTM) network on a single basin, Hydrol. Earth Syst. Sci., 28, 4187–4201, 10.5194/hess-28-4187-2024, 2024.Kumar, R., Samaniego, L., and Attinger, S.: Implications of distributed hydrologic model parameterization on water fluxes at multiple scales and locations, Water Resour. Res., 49, 360–379, 10.1029/2012WR012195, 2013.Linke, S., Lehner, B., Ouellet Dallaire, C., Ariwi, J., Grill, G., Anand, M., Beames, P., Burchard-Levine, V., Maxwell, S., Moidu, H., Tan, F., and Thieme, M.: Global hydro-environmental sub-basin and river reach characteristics at high spatial resolution, Sci. Data, 6, 283, 10.1038/s41597-019-0300-6, 2019.Loritz, R., Dolich, A., Acuña Espinoza, E., Ebeling, P., Guse, B., Götte, J., Hassler, S. K., Hauffe, C., Heidbüchel, I., Kiesel, J., Mälicke, M., Müller-Thomy, H., Stölzle, M., and Tarasova, L.: CAMELS-DE: hydro-meteorological time series and attributes for 1582 catchments in Germany, Earth Syst. Sci. Data, 16, 5625–5642, 10.5194/essd-16-5625-2024, 2024.Mai, J., Shen, H., Tolson, B. A., Gaborit, É., Arsenault, R., Craig, J. R., Fortin, V., Fry, L. M., Gauch, M., Klotz, D., Kratzert, F., O'Brien, N., Princz, D. G., Rasiya Koya, S., Roy, T., Seglenieks, F., Shrestha, N. K., Temgoua, A. G. T., Vionnet, V., and Waddell, J. W.: The Great Lakes Runoff Intercomparison Project Phase 4: the Great Lakes (GRIP-GL), Hydrol. Earth Syst. Sci., 26, 3537–3572, 10.5194/hess-26-3537-2022, 2022. Mayr, A., Klambauer, G., Unterthiner, T., Steijaert, M., Wegner, J. K., Ceulemans, H., Clevert, D.-A., and Hochreiter, S.: Large-scale comparison of machine learning methods for drug target prediction on ChEMBL, Chem. Sci., 9, 5441–5451, 2018.Mitchell, M., Wu, S., Zaldivar, A., Barnes, P., Vasserman, L., Hutchinson, B., Spitzer, E., Raji, I. D., and Gebru, T.: Model Cards for Model Reporting, in: Proceedings of the Conference on Fairness, Accountability, and Transparency, FAT* '19, Association for Computing Machinery, New York, NY, USA, 220–229, ISBN 9781450361255, 10.1145/3287560.3287596, 2019.Muñoz-Sabater, J., Dutra, E., Agustí-Panareda, A., Albergel, C., Arduini, G., Balsamo, G., Boussetta, S., Choulga, M., Harrigan, S., Hersbach, H., Martens, B., Miralles, D. G., Piles, M., Rodríguez-Fernández, N. J., Zsoter, E., Buontempo, C., and Thépaut, J.-N.: ERA5-Land: a state-of-the-art global reanalysis dataset for land applications, Earth Syst. Sci. Data, 13, 4349–4383, 10.5194/essd-13-4349-2021, 2021.Nasreen, S., Součková, M., Vargas Godoy, M. R., Singh, U., Markonis, Y., Kumar, R., Rakovec, O., and Hanel, M.: A 500-year annual runoff reconstruction for 14 selected European catchments, Earth Syst. Sci. Data, 14, 4035–4056, 10.5194/essd-14-4035-2022, 2022.Nearing, G., Cohen, D., Dube, V., Gauch, M., Gilon, O., Harrigan, S., Hassidim, A., Klotz, D., Kratzert, F., Metzger, A., Nevo, S., Pappenberger, F., Prudhomme, C., Shalev, G., Shenzis, S., Tekalign, T. Y., Weitzner, D., and Matias, Y.: Global prediction of extreme floods in ungauged watersheds, Nature, 627, 559–563, 10.1038/s41586-024-07145-1, 2024.Nevo, S., Morin, E., Gerzi Rosenthal, A., Metzger, A., Barshai, C., Weitzner, D., Voloshin, D., Kratzert, F., Elidan, G., Dror, G., Begelman, G., Nearing, G., Shalev, G., Noga, H., Shavitt, I., Yuklea, L., Royz, M., Giladi, N., Peled Levi, N., Reich, O., Gilon, O., Maor, R., Timnat, S., Shechter, T., Anisimov, V., Gigi, Y., Levin, Y., Moshe, Z., Ben-Haim, Z., Hassidim, A., and Matias, Y.: Flood forecasting with machine learning models in an operational framework, Hydrol. Earth Syst. Sci., 26, 4013–4032, 10.5194/hess-26-4013-2022, 2022.O, S. and Orth, R.: Global soil moisture data derived through machine learning trained with in-situ measurements, Sci. Data, 8, 170, 10.1038/s41597-021-00964-1, 2021.O, S., Orth, R., Weber, U., and Park, S. K.: High-resolution European daily soil moisture derived with machine learning (2003–2020), Sci. Data, 9, 701, 10.1038/s41597-022-01785-6, 2022.Peters-Lidard, C. D., Clark, M., Samaniego, L., Verhoest, N. E. C., van Emmerik, T., Uijlenhoet, R., Achieng, K., Franz, T. E., and Woods, R.: Scaling, similarity, and the fourth paradigm for hydrology, Hydrol. Earth Syst. Sci., 21, 3701–3713, 10.5194/hess-21-3701-2017, 2017.Salwey, S., Coxon, G., Pianosi, F., Lane, R., Hutton, C., Bliss Singer, M., McMillan, H., and Freer, J.: Developing water supply reservoir operating rules for large-scale hydrological modelling, Hydrol. Earth Syst. Sci., 28, 4203–4218, 10.5194/hess-28-4203-2024, 2024.Samaniego, L., Kumar, R., and Attinger, S.: Multiscale parameter regionalization of a grid-based hydrologic model at the mesoscale, Water Resour. Res., 46, 10.1029/2008WR007327, 2010. Sawicz, K., Wagener, T., Sivapalan, M., Troch, P. A., and Carrillo, G.: Catchment classification: empirical analysis of hydrologic similarity based on catchment function in the eastern USA, Hydrol. Earth Syst. Sci., 15, 2895–2911, 10.5194/hess-15-2895-2011, 2011.Sen, P. K.: Estimates of the Regression Coefficient Based on Kendall's Tau, J. Am. Stat. Assoc., 63, 1379–1389, 10.1080/01621459.1968.10480934, 1968.Senent-Aparicio, J., Castellanos-Osorio, G., Segura-Méndez, F., López-Ballesteros, A., Jimeno-Sáez, P., and Pérez-Sánchez, J.: BULL Database – Spanish Basin attributes for Unravelling Learning in Large-sample hydrology, Sci. Data, 11, 737, 10.1038/s41597-024-03594-5, 2024.Shen, J., Qu, Y., Zhang, W., and Yu, Y.: Wasserstein Distance Guided Representation Learning for Domain Adaptation, Proceedings of the AAAI Conference on Artificial Intelligence, 32, 10.1609/aaai.v32i1.11784, 2018.Sweet, L.-b., Müller, C., Anand, M., and Zscheischler, J.: Cross-Validation Strategy Impacts the Performance and Interpretation of Machine Learning Models, Artificial Intelligence for the Earth Systems, 2, e230026, 10.1175/AIES-D-23-0026.1, 2023.Theil, H.: A rank-invariant method of linear and polynomial regression analysis, Indagat. Math., 12, 173, https://ir.cwi.nl/pub/18446/18446A.pdf (last access: 13 July 2026), 1950.Tolson, B. A. and Shoemaker, C. A.: Dynamically dimensioned search algorithm for computationally efficient watershed model calibration, Water Resour. Res., 43, 10.1029/2005WR004723, 2007.Tolstikhin, I., Bousquet, O., Gelly, S., and Schoelkopf, B.: Wasserstein auto-encoders, arXiv [preprint], 10.48550/arXiv.1711.01558, 2017.Torres, L. C., Pereira, L. M., and Amini, M. H.: A survey on optimal transport for machine learning: Theory and applications, arXiv [preprint], 10.48550/arXiv.2106.01963, 2021.Tucker, C. J., Pinzon, J. E., Brown, M. E., Slayback, D. A., Pak, E. W., Mahoney, R., Vermote, E. F., and El Saleous, N.: An Extended AVHRR 8-km NDVI Dataset Compatible with MODIS and SPOT Vegetation NDVI Data, Int. J. Remote Sens., 26, 4485–4498, 10.1080/01431160500168686, 2005.Zomer, R. J., Xu, J., and Trabucco, A.: Version 3 of the global aridity index and potential evapotranspiration database, Sci. Data, 9, 409, 10.1038/s41597-022-01493-1, 2022.