Soil moisture is a key variable for a range of hydrological and ecological processes, yet capturing its small-scale variability and preferential flow phenomena remains challenging. Recent advancements in deep learning have demonstrated potential in predicting hydrological variables, but conventional data-driven models often struggle to represent small-scale variability effectively. In this study, we integrate Long-Short Term Memory (LSTMs) networks and Gaussian Mixture Models (GMMs) to simulate soil moisture dynamics while explicitly quantifying its associated variability. Unlike deterministic approaches, our probabilistic framework accounts for nonlinear relationships between inputs and outputs while modeling the inherent small-scale variability in soil moisture. We apply this methodology to a comprehensive in-situ soil moisture dataset from the Attert experimental basin, where the experimental design incorporates three replicated soil moisture profiles at each location and depth within a 5-meter radius. These replications are fundamental to our probabilistic framework: they provide direct, co-located observations of the natural spread in soil moisture under identical boundary conditions, allowing the model to learn the statistical structure of small-scale variability. This design enables disentangling sensor noise from genuine spatial heterogeneity and provides an empirical basis for training models that capture both temporal dynamics and local-scale variability.. Our results demonstrate that the proposed model reproduces soil moisture dynamics across multiple depths and scales, achieving an average Kling-Gupta Efficiency (KGE) of 0.52, Rank Correlation of 0.72, and Root Mean Squared Error of 0.036 m3m-3, while also capturing the key aspects of small-scale variability and sensor uncertainty. Furthermore, the modeled distributions offer new insights into the spatiotemporal structure of soil moisture and underscore the value of probabilistic modeling in hydrological approaches. By explicitly incorporating small-scale variability into the modeling process, our approach enhances both the interpretability and reliability of soil moisture predictions. While LSTMs effectively capture temporal dynamics, our findings underscore the necessity of incorporating variability quantification to improve model accuracy and generalization. This study highlights the potential of probabilistic deep learning frameworks in hydrological modeling and supports their broader application for improved soil moisture estimation and variability assessment.

Accurate flood forecasting in small and medium-sized catchments remains a majorchallenge due to rainfall prediction uncertainties, limited hydrological data at thatscale as well as short warning times related to rapid response times of such systems.The joint German research project KI-HopE-DE (KI-gestützte Hochwasserprognosefür kleine Einzugsgebiete in Deutschland) aims to improve flood prediction inGermany by testing new machine learning-based approaches and bringing togethermeteorologists and hydrologists, both from academia and from operational services.Within this framework, the German Weather Service (DWD) contributes by providinghigh-resolution meteorological datasets to support data-driven modelling, as well asby testing training strategies that stronger account for weather model particularities.KI-HopE-De develops a regionally-trained Long Short-Term Memory (LSTM) usingdata from 1,626 catchments across Germany. Besides the classical training based onmeteorological observations, in our study we explore the use of the novel ICON-FORCE (Fine-scale Observation-based Reanalysis for Central Europe) reanalysisdataset for optimizing the LSTM. This approach attempts to leverage the proximity ofthe ICON-FORCE reanalysis data and the ICON-D2 forecast model later used forinference and making forecasts. Meteorological variables derived from ICON-FORCEreanalysis are used as input features to capture spatiotemporal dependencies andinteractions relevant for runoff generation. The study is designed as a large-sampleexperiment to systematically assess the added value of high-resolutionmeteorological reanalysis for data-driven flood prediction, with a particular focus onmodel robustness and applicability across divers catchments.By focusing on the interface between meteorology and hydrology, this workcontributes to ongoing efforts to better integrate atmospheric and hydrologicalinformation in flood forecasting models.

Abstract. Deep learning has become a standard tool for streamflow modeling, but its application to water quality remains challenging due to sparse, irregular, and noisy in-situ observations. Yet, water-quality variables are tightly linked to discharge and to each other through shared hydrological and biogeochemical controls, suggesting that jointly modeling water quantity and quality may help compensate for limited data availability. In this study, we compare single-target, multi-target, and self-supervised LSTM models for the joint simulation of discharge and six water-quality variables (NO₃–N, PO₄–P, DO, DOC, EC, and WT) across 408 German catchments. Our results highlight that extending the baselineLSTM from single-target discharge prediction to jointly predicting discharge and all six water-quality variables does not substantially degrade discharge performance (across all baseline configurations, median KGE' = 0.84–0.87) and yields median KGE' values between 0.35 (PO₄–P) and 0.94 (WT) for the water-quality targets. Interestingly, learning discharge as a co-target consistently outperforms models that use observed discharge as an additional input, indicating that jointly learning water quantity and quality is more effective than using discharge as a predictor. A variable-averaged loss function is key to balance the strongly uneven observation densities of discharge and water-quality variables. Building on this multi-target framework, we further explore whether an alternative training strategy based on self-supervised learning can better exploit the incomplete and heterogeneous nature of environmental observations. Our evaluation reveals that the self-supervisedLSTM yields a level of predictive skill comparable to the multi-target supervised baseline under inference conditions restricted to meteorological drivers, while effectively leveraging cross-variable dependencies to enhance PO₄–P and DO reconstructions when contextual water-quality data are provided. Besides showcasing the strong performance of LSTMs in water-quality simulations with sparse and irregular data, our results demonstrate that multi-target learning provides an effective framework for coupled water quantity–quality modeling, while self-supervised learning offers additional flexibility for exploiting incomplete and heterogeneous environmental observations, yielding predictive skill comparable to, and in some cases exceeding, that of calibrated process-based water-quality models and supporting the use of LSTMs as a scalable alternative for regional water-quality simulation in data-sparse environments.

Abstract. Deterministic model predictions can struggle to adequately capture extreme events such as floods and droughts, which are of particular relevance in hydrology. This limitation arises because deterministic models collapse the conditional runoff distribution to a single point estimate. Probabilistic modeling provides a way to address this issue by explicitly representing uncertainty and assigning non-zero probabilities to a range of possible outcomes, including rare and extreme events, thereby capturing the full range of plausible hydrological responses. Motivated by this perspective, we examine whether probabilistic Long Short-Term Memory (LSTM) models improve the representation of extreme events in rainfall–-runoff simulations across Switzerland. Overall, the probabilistic models show good calibration, although some miscalibration remains for the extremes. Differences between models mainly manifest in how uncertainty is distributed: some approaches produce narrower and lighter-tailed distributions, while others yield broader distributions with heavier tails. These trade-offs highlight that probabilistic models differ not only in sharpness but also in how their calibration for rare events. We observe this tradeoff also in models' accuracy metrics. When evaluating the mean of the probabilistic predictions using the Nash–Sutcliffe efficiency (NSE), none of the probabilistic approaches outperform the deterministic LSTM in terms of average predictive accuracy. However, a clear advantage over the determinsitc models emerges when focusing on the tail of the discharge distribution. For the most extreme events (top 0.1 % of the discharge distribution), the deterministic LSTM underestimates more than 90 % of observed values (since it provides estimates of an expectation), whereas probabilistic predictions can capture a substantially larger fraction (67 %) of these extremes within their upper predictive bounds. Building on the additional information provided by probabilistic runoff predictions, we further show how they can be translated into actionable flood warnings using reinforcement learning. To this end, we introduce a Flood Risk Communication Agent (FRiCA) that operates on probabilistic runoff predictions and learns decision rules for issuing warnings of varying intensity. The FRiCA is implemented as an LSTM-based policy network and is trained by rewarding correct warning levels while penalizing the underestimation of flood severity. Results indicate that the FRiCA outperforms simple fixed heuristics, such as issuing warnings based on the predictive mean or a fixed high quantile (e.g., the 99th percentile). While this behavior already demonstrates the potential of reinforcement learning for improved flood risk communication, it also motivates further exploration of better reward design and policy network definition for context-dependent decision policies that adapt to varying hydrological and societal contexts.

Abstract. Large-sample hydrology has recently been driven by two key developments. First, the introduction of hydrological benchmark datasets such as CAMELS-US and CARAVAN, and second, the emergence of deep‑learning modelling frameworks, particularly LSTM‑based regional models, which have demonstrated performance on par with, and in some cases exceeding, that of process-based models for streamflow prediction in gauged and ungauged settings. Building on these developments, we investigate whether (i) further enhanced LSTM architectures, (ii) new sets of static features, or (iii) a combination of both enable us to significantly improve Predictions in Ungauged Basins (PUB). In this study, we evaluate a state-of-the-art regional LSTM model (base LSTM) against embedded (EMB-LSTM) and cross‑attention enhanced (CA-LSTM) variants, in combination with a suite of newly applied static features, namely MODIS surface reflectance bands, ALPHAEARTH embeddings, DEM-, meteorology- and catchment coordinate-derived auxiliary aggregates, and conventional CAMELS attributes. We tested these model-and-data combinations in pseudo‑ungauged 5‑fold cross‑validation across the 531 CAMELS‑US catchments. Model performance was quantified by the Nash‑Sutcliffe Efficiency (NSE), while latent‑space complexity was assessed via the Shannon effective rank (erank). Results show that the quality of static features is more important than architectural improvements. ALPHAEARTH embeddings attained the highest median NSE, but only in combination with auxiliary static feature data (ALPHAEARTHplus). Architectural refinements yielded only modest improvements. Thereby the relatively simple EMB-LSTM, which allowed the LSTM layer to better ingest ALPHAEAERTHplus static features, outperformed the other architectures. With this combination, we achieved a median performance of NSE = 0.726, significantly improving the state-of-the-art PUB performance (NSE = 0.69) for the CAMELS-US dataset. Auxiliary analysis indicates that further improvement is possible when adding MODIS bands as additional dynamic features to the model. In conclusion, our study indicates that, broadly speaking, (a) better data is more important than better architecture, (b) better architecture is necessary only to accommodate better data, (c) the single layer LSTM remains the most suitable core model as of now, and (d) the Shannon effective rank complexity of the latent space is a useful diagnostic for linking improved PUB performance to improved quality of latent hydrological representation inside the model. Overall, this highlights the need for improved measurement‑derived descriptor datasets, especially for soil and geology.

Abstract. CAMELS (Catchment Attributes and MEteorology for Large-sample Studies) datasets have been a major driver for advances in large-sample hydrology, facilitating regional studies and the development of deep learning methods and hydrological models by providing homogenized data across large domains, typically at the national scale. However, investigating highly dynamic events, such as flash floods, requires sub-daily resolution, which is often hindered by the daily time steps of most existing CAMELS datasets. Here, we present CAMELS-DE-1h, providing hourly time series of discharge and meteorology for 1,611 catchments in Germany, spanning the period from 2001 to 2024. This dataset homogenizes the extensive but deeply fragmented high-resolution hydrological gauge data managed independently by the German federal states, combining it with high-resolution meteorological forcing from the German Weather Service (DWD). With a median catchment area of 132.4 km², CAMELS-DE-1h includes many small-to-medium-sized basins where hydrological responses occur primarily on sub-daily scales. Alongside the time series, the dataset includes comprehensive static catchment attributes covering soil characteristics, land cover, hydrogeology, and human influences. A novel addition to CAMELS-DE-1h is a readily processed archive of operational short-term weather forecasts (ICON-D2, 48-hours lead time), which are used by German flood forecasting agencies in their operational settings. Including both deterministic runs and precipitation ensembles, these operational forecast data are available at a large scale for the first time for the period from 2021 to 2024. This allows for the evaluation of historical weather forecast quality, the testing of hydrological models in realistic operational settings, and the use of ensemble data to investigate the coupling of meteorological and hydrological forecast uncertainties. Finally, we provide baseline performance benchmarks using a regionally trained Long Short-Term Memory (LSTM) network and a conceptual HBV (Hydrologiska Byråns Vattenbalansavdelning) model. These models achieve median Nash-Sutcliffe Efficiencies (NSE) of 0.82 (LSTM) and 0.69 (HBV) for 1496 catchments selected based on their data availability for training / calibration and testing. By combining high-resolution observations with operational forecasts, CAMELS-DE-1h provides a consistent basis for the systematic comparison and development of hydrological and hydro-meteorological models under realistic conditions. CAMELS-DE-1h is available at: https://doi.org/10.5880/fidgeo.2026.045 (Dolich et al., 2026).

Long Short-Term Memory (LSTM) networks have demonstrated state-of-the-art performance for operational flood forecasting at the daily scale (Nearing et al., 2024). Recent advances have extended LSTM-based models to higher temporal resolutions through multi-frequency LSTM architectures (Acuña Espinoza et al., 2025) and introduced robust strategies for handling missing data, such as masked-mean embeddings (Gauch et al., 2025).Building on this work, we introduce an LSTM-based approach that allows producing hourly flood forecasts in an operational setting, while being robust to missing data. Moreover, using masked-mean embeddings plus teacher-forcing (Williams et al., 1989) and noise injection strategies during training, allows the model to integrate observed stream flow observations when available, for enhanced prediction accuracy, while keeping the flexibility to operate without this signal. To evaluate model performance, we benchmarked the new approach against LARSIM, the current operational model in many federal states in Germany. Our results show that the LSTM-based architecture outperforms the LARSIM model in predictive accuracy, while additionally offering robustness to missing inputs and faster inference times.These findings highlight the potential of deep learning–based models for operational flood forecasting at an hourly resolution, while introducing strategies to increase robustness and add valuable information, when available. Reference:Acuña Espinoza, E., Kratzert, F., Klotz, D., Gauch, M., Álvarez Chaves, M., Loritz, R., & Ehret, U. (2025). Technical note: An approach for handling multiple temporal frequencies with different input dimensions using a single LSTM cell. Hydrology and Earth System Sciences, 29(6), 1749–1758. https://doi.org/10.5194/hess-29-1749-2025Gauch, M., Kratzert, F., Klotz, D., Nearing, G., Cohen, D., & Gilon, O. (2025). How to deal with missing input data. Hydrology and Earth System Sciences, 29(21), 6221–6235. https://doi.org/10.5194/hess-29-6221-2025Nearing, 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., & Matias, Y. (2024). Global prediction of extreme floods in ungauged watersheds. Nature, 627(8004), 559–563. https://doi.org/10.1038/s41586-024-07145-1Williams, R. J., & Zipser, D. (1989). A learning algorithm for continually running fully recurrent neural networks. Neural computation, 1(2), 270-280.

In this study, 1,575 runoff series of the CAMELS-DE database are used to investigate the runoff behaviour for intermittency across Germany. The analysis showed that 16 % of the stations had at least one zero-flow period during the investigation period between 1981 and 2020. 20 stations fell dry on average every one to two years, for 3 to 200 days per year, followed by 38 stations with medium intermittent behaviour with zero-flow days every 2 to 5 years. A larger part of the intermittent stations fell dry rarely: 40 stations every 6 to 10 years and 149 every 10 years or more rarely. The remaining fraction, 84 % of all stations of the CAMELS-DE data set, was perennial during the entire investigation period. An assessment of temporal variability of the annual number of zero-flow days revealed an inhomogeneous change behaviour: for each zero-flow frequency category from frequent to rarely intermittent, different temporal change patterns-increase, decrease or no trend-were found in the dataset. For around 76 stations, an increase of zero flow days was detected, which was statistically significant at 21 stations. 52 of the stations exhibited no significant trend over the entire time period, but a clear increase in the years 2018 to 2020. A moderate correlation was detected between the annual number of zero-flow days and the mean spring and summer temperature, as well as between the potential evapotranspiration and the June SPEI-3 drought index for stations with a highly intermittent regime and increasing trend. For the annual number of zero-flow days of all other intermittent stations and climate metrics, correlation was only weak or not detectable. In order to assess potential association between the classification of the runoff regime into intermittent or perennial regimes and the 67 catchment attributes from the CAMELS-DE dataset, random forest analysis was used. It resulted in no clear picture which attributes may explain the tendency towards intermittent runoff. In contrast, a whole range of not directly discharge-related attributes were identified, which included the frequency of dry days and the areal coverages of aquifer, geochemical rock type, land-use and soil parameters. From those identified attributes, catchments with more permeable soils with higher sand and low clay content tended to be characterized by intermittent rather than perennial runoff.

Challenges and risks related to flood prediction and warning, which became evident during and in the aftermath of devastating flood events in recent years in Germany, call for a new level of cooperation between weather and flood forecasting and their respective operational services. While these domains have historically been structurally quite separate in Germany, several collaborative projects are now laying the foundation for increased exchange and joint development.The first phase of the project “Co-Design of Innovations between Weather and Flood Forecasting” (Italia–Deutschland science-4-services, IDEA-S4S), initiated directly in response to the lessons learned from the Ahrtal flood in 2021, addresses multiple topics. One focus is on improving mutual knowledge and understanding of forecast evaluation and verification, especially given the constantly increasing number of weather forecasting systems. Another emphasis has been placed on analyzing and training communication along the entire warning chain down to disaster management authorities, in particular regarding the handling of forecast uncertainty. The upcoming second phase of Co-Design aims to deepen this communication aspect by testing decision-making strategies that incorporate probabilistic criteria in multi-stage alert and readiness plans, in direct collaboration with emergency response units. Moreover, the second project phase will, among other topics, explore ML-based approaches to combine multiple ensemble models with different lead times into a best-guess ensemble as input for a hydrological model, thereby facilitating the use of the growing number of models in downstream flood applications.The joint research project KI-HopE-De aims to test new machine-learning-based approaches and brings together meteorologists and hydrologists from both academia and operational services. First, a new high-resolution hourly hydrometeorological dataset (CAMELS-DE-1h) will be compiled and published for around 1,600 catchments across Germany. Second, various training strategies—from both hydrological and meteorological perspectives—for a regionally trained LSTM flood forecasting model will be developed to provide a benchmark model that is compared with the current operational models in Germany.This contribution provides an overview of ongoing projects and highlights related presentations at the conference. We will share results and provide an outlook on upcoming work, and look forward to exchanging experiences with other initiatives at the intersection of meteorology and hydrology.

Understanding and preparing for extreme events in a warming climate remains challenging, particularly for modelling flash floods in small- to mesoscale catchments. While top-down modelling approaches that describe fluxes at the system scale are often effective for riverine floods driven by saturation-excess runoff, bottom-up approaches are better suited to capturing intensity-controlled runoff generation and associated preferential flow processes. Based on the gradient-conserving simplification of representative hillslopes, a meso-catchment scale spatially distributed, process-based model was applied to simulate a severe summer flood event that occurred in 1994 in southwest Germany. Our approach provides a balance between the complexity required to represent coupled flow processes at the hillslope scale and the practical constraints of scaling these to the mesoscale. Following evaluation against available observations, the model is used to reconstruct flood magnitudes in poorly gauged but severely affected headwater regions in the catchment. The results highlight the influence of spatial variability in gradients and land use on runoff generation in these areas. To further explore these findings, we conducted additional simulations across a range of precipitation return periods to examine the sensitivity of flood response under different scenarios. The results suggest that uncertainties are more pronounced at smaller spatial scales, likely due to data limitations. Finally, simplified nature-based solution (NbS) scenarios were implemented at the hillslope and headwater scales to explore their potential influence on downstream flood response. This study contributes to improved understanding of overland flow responses over mesoscale catchments, a critical scale for flood management, particularly under increasing convective extremes as a result of anthropogenic climate change.

CAMELS datasets have been primary accelerators for Large-Sample Hydrology (LSH), providing extensive, harmonized hydro-meteorological data and establishing benchmarks that have fundamentally changed how data-driven models in Hydrology are developed and evaluated. However, to date, these efforts have predominantly focused on daily resolution. While the overall performance of deep learning models for daily rainfall-runoff modelling has reached a high standard - often plateauing with "vanilla" LSTMs - significant challenges remain. These include the accurate representation of flood peaks, drought dynamics, performance under non-stationary conditions, and the capturing of rapid events in small catchments. Although initial LSH studies have explored hourly data, fully exploiting sub-daily information remains an open and pressing challenge. The shift to high-resolution datasets offers the potential to improve modeling extreme floods and their dynamics and to capture runoff generation processes also in smaller catchments. However, this transition requires a reassessment of the current state-of-the-art: do the limitations of daily modelling persist at the hourly scale, are they resolved by higher resolution data, and which entirely new challenges arise?To address these questions and facilitate the transition to sub-daily LSH, we introduce CAMELS-DE-1h, a comprehensive hourly dataset for Germany. It covers 1626 catchments with streamflow and meteorological forcing data spanning 2001 - 2024. Uniquely, CAMELS-DE-1h includes historical short-term meteorological forecasts (ICON-D2, 48 hours lead time) from 2021 - 2024, both as deterministic and ensemble forecasts. This novelty enables rigorous research regarding the propagation of meteorological uncertainty into hydrological predictions and the development of deep learning models for operational settings. With CAMELS-DE-1h, we provide open-source LSTM benchmarks for both discharge simulation and forecasting, and use these benchmarks to evaluate the transition from daily to hourly simulations. Specifically, we analyze how the transition to hourly resolution alters model behavior regarding peak flow timing and hydrograph shape, and discuss the challenges such as computational costs and the need for evaluation metrics adapted to sub-daily Large-Sample Hydrology.

Deep learning models, particularly long short-term memory networks (LSTMs), have set new standards in streamflow prediction but require extensive data and computational resources. This raises a practical question: which parts of the data are truly indispensable, especially when computational budgets are limited or when only sparse observations are available? To address this, we examine which type of data is most essential in model training. We systematically ablate the CAMELS-US dataset using four families of sampling strategies - hydrological extremes, event rarity/statistical representativity, temporal context, and spatial representativity - and use the ablated datasets in training. Among all tested approaches, sampling based on statistical representativity via random sampling consistently outperformed more targeted strategies, achieving strong performance (NSE > 0.7) and good representativity on as little as 10% of the data. Sampling hydrological extremes is the second-most efficient strategy, particularly when high-flow and low-flow extremes are sampled jointly, but with the largest performance gains stemming from high-flow events. Concerning temporal context, surprisingly short sequence lengths (<3 weeks) and training periods (<2 years) were sufficient for competitive performance (NSE > 0.7). These findings provide practical guidance for efficient data selection in data-driven modeling and provide groundwork for future studies on training strategies.

Hydrological models differ in the way how hydrological processes are implemented. A rigorous comparison of different hydrological model structures is needed to disentangle the link between similarities and differences in process representations and simulated hydrological processes, states and fluxes. A major challenge in model comparison is to identify effects of individual processes. To move a step in this direction, we developed controlled experiments and compared three hydrological models (HBV, mHM, SWAT+) in nine German catchments (400-3000 km²) along an elevation gradient. We aim at presenting a framework for a consistent comparison of process representations in model structures consisting of three steps: (1) A model comparison protocol was developed for a detailed comparison of process representations in model structures. Consistency was achieved by using the same input data for all models. By grouping the processes in a standardized way, differences and similarities between the models were identified. (2) To investigate the dominant model components, a daily parameter sensitivity analysis was carried out for the three models with different hydrological variables as target variables (e.g. actual evapotranspiration, soil moisture, snow and discharge). The dominant model parameters and associated processes vary more between the models than between the catchments. This also applies to the temporal variability of the parameter sensitivity. (3) The model performance was analysed for a set of different performance criteria. The optimal parameter values differ greatly depending on which performance criteria were selected. This is in particular true for soil and evapotranspiration parameters. Typical patterns can be derived between catchments of different landscapes. The joint analysis of these three methodological steps demonstrates the benefit of a detailed process analysis in model structures for a better understanding of suitable process representations. Therefore, it shows the potentials for improving model structures.

CAMELS datasets are recognized in the hydrological community as consistent and comprehensive benchmark datasets for hydrological and meteorological analyses. CAMELS stands for "Catchment Attributes and MEteorology for Large-sample Studies”. CAMELS datasets link landscape and catchment attributes (e.g. land use, geology, soil properties), hydrological time series (e.g. water level, discharge) and meteorological time series (e.g. precipitation, air temperature) in a large number of catchment areas. They clearly indicate the uncertainties and processing of individual variables and thus enable the comparison of models and data in different landscapes, but also contribute to the general understanding of hydrological processes across landscapes. This is crucial for assessing the consequences of the climate crisis and improves the basis for water resource management decisions. Although CAMELS datasets are intensively used in other countries, such a dataset is still lacking for Germany.This contribution highlights the crucial importance of consistent and easily accessible benchmark datasets for hydrological research and education. We discuss both the challenges faced so far in compiling the dataset and the future ambitions of the project. In addition, an overview is given of the scope of the first version of the CAMELS-DE data set, which will include around 2,000 measuring stations with daily time series of discharge and water level with an average length of nearly 50 years in mainly small and medium-sized catchments. Also included are the landscape and catchment attributes as well as meteorological time series. A key focus is on the easy availability and straightforward import of data into programming environments. We discuss how such benchmark datasets not only increase efficiency in the use of environmental data, but also play a key role in ensuring the reproducibility of research results. Especially in the age of machine learning learning, they form an indispensable basis for modern, data-driven hydrology. By integrating CAMELS-DE into the research landscape, we want to emphasize that data publications and benchmark datasets are much more than a by-product of a doctoral thesis, but rather the basis and key to modern environmental science.

In this presentation, we explore the application of Long Short-Term Memory networks (LSTMs) to predict hourly tree-level sap flow across Europe, utilizing the comprehensive SAPFLUXNET database. This study emphasizes the potential of deep learning in estimating transpiration and understanding forest water use dynamics and plant-climate interactions. By developing LSTM models with varied training sets, we assess their capability to perform in previously unencountered conditions. Our research reveals that these models achieve an average Kling-Gupta Efficiency of 0.77 when trained on 50% of the time series across all forest stands, and 0.52 for models trained on 50% of the forest stands without prior gauging. These continental-scale models not only meet but often exceed the performance of specialized and baseline models across all tree genera and forest types. In this submission, we will discuss the methodologies employed, the challenges faced, and the insights gained from this research. The presentation will also highlight the broader implications of this study for ecohydrological investigations, particularly the enhanced capacity of deep learning models to generalize sap flow data, thereby improving our understanding of ecohydrology from individual trees to a continental scale.

While measured streamflow is commonly used for hydrological model evaluation and calibration, an increasing amount of data on additional hydrological variables is available. These data have the potential to improve process consistency in hydrological modeling and consequently for predictions under change, as well as in data-scarce or ungauged regions. Here, we show how these hydrological data beyond streamflow are currently used for model evaluation and calibration. We consider storage and flux variables, namely snow, soil moisture, groundwater level, terrestrial water storage, evapotranspiration, and altimetric water level. We aim at summarizing the state-of-the-art and providing guidance for the use of additional hydrological variables for model evaluation and calibration. Based on a review of the current literature, we summarize observation methods and uncertainties of currently available data sets, challenges regarding their implementation, and benefits for model consistency. The focus is on catchment modeling studies with study areas ranging from a few km 2 to ~500,000 km 2 . We discuss challenges for implementing alternative variables that are related to differences in the spatio-temporal resolution of observations and models, as well as to variable-specific features, for example, discrepancy between observed and simulated variables. We further discuss advancements required to deal with uncertainties of the hydrological data and to integrate multiple, potentially inconsistent datasets. The increased model consistency and improvement shown by most reviewed studies regarding the additional variables often come at the cost of a slight decrease in streamflow model performance.

Long Short-Term Memory (LSTM) networks have shown strong performance in rainfall–runoff modelling, often surpassing conventional hydrological models in benchmark studies. However, recent studies raise questions about their ability to extrapolate, particularly under extreme conditions that exceed the range of their training data. This study examines the performance of a stand-alone LSTM trained on 196 catchments in Switzerland when subjected to synthetic design precipitation events of increasing intensity and varying duration. The model’s response is compared to that of a hybrid model and evaluated against hydrological process understanding. Our study reiterates that the stand-alone LSTM is characterised by a theoretical prediction limit, and we show that this limit is below the range of the data the model was trained on. We show that saturation of the LSTM cell states alone does not fully account for this characteristic behaviour, as the LSTM does not reach full saturation, particularly for the 1-day events. Instead, its gating mechanisms prevent new information about the current extreme precipitation from being incorporated into the cell states. Adjusting the LSTM architecture, for instance, by increasing the number of hidden states, and/or using a larger, more diverse training dataset can help mitigate the problem. However, these adjustments do not guarantee improved extrapolation performance, and the LSTM continues to predict values below the range of the training data or show hydrologically unfeasible runoff responses during the 1-day design experiments. Despite these shortcomings, our findings highlight the inherent potential of stand-alone LSTMs to capture complex hydro-meteorological relationships. We argue that more robust training strategies and model configurations could address the observed limitations, ensuring the promise of stand-alone LSTMs for rainfall–runoff modelling.

It is widely assumed that data-driven models achieve good results only with sufficiently large training data, whereas process-based models are usually expected to be superior in data-poor situations. To investigate this, we calibrated several process-based and data-driven hydrological models using training datasets of observed discharge that differed in terms of both the number of data points and the type of data selection, allowing us to make a systematic comparison of the learning behaviour of the different model types. Four data-driven models (conditional probability distributions, regression trees, artificial neural networks, and long short-term memory networks) and three process-based models (GR4J, HBV, and SWAT+) were included in the testing, applied in three meso-scale catchments representing different landscapes in Germany: the Iller in the Alpine region, the Saale in the low mountain ranges, and the Selke in the transition between the Harz and central German lowlands. We used information measures (joint entropy and conditional entropy) for system analysis and model performance evaluation because they offer several desirable properties: they extend seamlessly from uni- to multivariate data, they allow direct comparison of predictive uncertainty with and without model simulations, and their boundedness helps to put results into perspective. In addition to the main question of this study - to what extent does the performance of different models depend on the training dataset? - we investigated whether the selection of training data (random, according to information content, contiguous time periods, or independent time points) plays a role. We also examined whether the shape of the learning curve for different models can be used to predict the achievable model performance based on the information contained in the data and whether using more spatially distributed model inputs improves model performance compared to using spatially lumped inputs. Process-based models outperformed data-driven ones for small amounts of training data due to their predefined structure. However, as the amount of training data increases, the learning curve of process-based models quickly saturates, and data-driven models become more effective. In particular, the long short-term memory network outperforms all process-based models when trained with more than 2-5 years of data and continues to learn from additional training data without approaching saturation. Surprisingly, fully random sampling of training data points for the HBV model led to better learning results than consecutive random sampling or optimal sampling in terms of information content. Analysing multivariate catchment data allows predictions about how these data can be used to predict discharge. When no memory was considered, the conditional entropy was high. However, as soon as memory was introduced in the form of the previous day or week, the conditional entropy decreased, suggesting that memory is an important component of the data and that capturing it improves model performance. This was particularly evident in the catchments in the low mountain ranges and the Alpine region.

Data-driven techniques have shown the potential to outperform process-based models in rainfall–runoff simulation. Recently, hybrid models, which combine data-driven methods with process-based approaches, have been proposed to leverage the strengths of both methodologies, aiming to enhance simulation accuracy while maintaining a certain interpretability. Expanding the set of test cases to evaluate hybrid models under different conditions, we test their generalization capabilities for extreme hydrological events, comparing their performance against long short-term memory (LSTM) networks and process-based models. Our results indicate that hybrid models show performance similar to that of the LSTM network for most cases. However, hybrid models reported slightly lower errors in the most extreme cases and were able to produce higher peak discharges.