Long Short-Term Memory networks (LSTMs) have been around since the early 90’s but only in the last few years have LSTMs gained significant popularity in the hydrological sciences. Related publication counts have grown exponentially, and LSTMs power some of the largest-scale operational flood forecasting systems.

In this presentation, I'll look back at my relatively short career as a student and researcher at the intersection of hydrology and machine learning. I don't claim to have introduced LSTMs to hydrology, but I'll share my own experience helping to develop this modeling approach into what it is today. We will look at what I saw in this neural network architecture, and why I thought it was well suited for hydrologic applications.

The tale goes as follows: Once upon a time, in a land (not so) far far away, a (not so young) master student of environmental engineering was teaching himself the dark arts of machine learning (ML). While studying ML for automated fish detection, he stumbled upon the LSTM architecture. Having just concluded a course on the design of conceptual hydrological models, he noticed the underlying similarity between the LSTM and these established approaches — and more generally, the conceptual approach for modeling the water cycle. With one of his dearest colleagues and friends, he started to work night and day (actually more nights than days) to see if the LSTM is indeed suitable for hydrology. From initial attempts at emulating the ABC and HBV models, to first real-world experiments in individual catchments, the LSTM was showing great potential. But it was not until he discovered the CAMELS dataset and started experimenting with large-sample hydrology that he fully understood the potential of LSTMs for applications in hydrology. Equipped with nothing more than his first GPU, he embarked on a quest to explore the wondrous lands of academia. Countless nights were spent on the computer, forging transatlantic friendships, conducting experiments and writing publications. Eventually, he ascended to the ranks of PhDs by defending his research against Reviewer #2 and the high council of the PhD committee. Fast forward in time, today, LSTMs are widely used and among others, power Google’s current operational, global-scale flood forecasting model. And thus, the now (not so) old research scientist lived happily ever after with his wife and his children, and continues, to this day, to do much the same as he had in those earlier years.

If there is one thing that I would like for you to take away from this talk, it is that I hope my presentation will motivate young scientists to stay curious, to follow their own ideas, to not get demotivated by initial pushback and to not be afraid of reaching out to more senior researchers. I want to advocate strongly the importance of open science, of reproducibility, of collaborations, of benchmarking and of open data sharing to advance science.

How to cite: Kratzert, F.: Long Short-Term Memory networks in hydrology: From free-time project to Google’s operational flood forecasting model, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-5863, https://doi.org/10.5194/egusphere-egu25-5863, 2025.

The use of artifical intelligence (AI) and machine learning (ML) approaches in various scientific and engineering disciplines has grown exponentially over recent years. This upsurge also includes applications of physics-guided ML models and explainable AI. However, in addition to the dificulties involved in the identification of relevant model inputs, the advantages, contributions, and credibility of ML models are still open challenges, especially when these models are evaluated against the perceptual hydrologic understanding of the system in question. In this study, we aim to investigate some of these challenges using the case of seasonal streamflow forecasting with lead times up to three months in several hydrologically challenging river basins of prairie provinces of Canada (i.e., Alberta, Saskatchewan, and Manitoba).

Multiple ML techniques, including Random Forest (RF) and Long Short-Term Memory (LSTM) models, are used to produce ensemble forecasts for 135 sub-basins of the Nelson-Churchill River Basin, comprising the vast area from the Rocky mountains up to the Hudson Bay, with the monthly temporal resolution and spatial scales of the order of 200 km² to ~1.0 x10⁶ km², as reflected by drainage areas of all sub-basins. A large set of potential inputs (105 predictors) is used in this study. These potential inputs include hydrometeorological variables derived from the Daymet database, Environment and Climate Change Canada’s hydrometric network, and hydrometeorological forecasts from the European Centre for Medium-Range Weather Forecasts, and various static attributes of all sub-basins.

The Pearson’s correlation coefficient (CC) and Partial Mutual Information (PMI) were used, as model agnostic methods, to analyze the set of potential predictors and identify the most appropriate inputs for seasonal flow forecasting, prior to ML model development. Subsequently, modeling experiments were designed to investigate the ML model performance and test the usefulness of CC and PMI based techniques on modeling results. The model-agnostic and model-dependent findings were compared and analyzed in light of the perceptual understanding of the hydrological system. Furthermore, the Convergent Cross-Mapping (CCM) method was used with selected variables to further explore the causal, rather than correlational, relationships and interpret the results with the aim of developing ethical and responsible ML (ERML) models. We define ERML models as data driven models that are transparent and hydrologically explainable.

The preliminary results of this study indicate that PMI is quite effective in filtering some of the CC-based selections, which might form multiple equifinale sets of predictors. This step is critical for identifying the most relevant and necessary inputs. In spite of the coarse spatial and temporal resolutions, which complicate crisp hydrologic perceptions, the CCM method seems to support the selection of various input variables with hydrologic causality, strengthening the transparency and credibility of ML models.

How to cite: Elshorbagy, A., Nguyen, D.-H., Khaliq, M. N., Akhtar, M. K., and Unduche, F.: Positioning ML Models for Spatial and Temporal Modeling of River Flows Through Causality and Information Content Analyses, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-3093, https://doi.org/10.5194/egusphere-egu25-3093, 2025.