Towards Learning Universal, Regional, and Local Hydrological Behaviors via Machine-Learning Applied to Large-Sample Datasets (1907.08456v2)

Abstract: Regional rainfall-runoff modeling is an old but still mostly out-standing problem in Hydrological Sciences. The problem currently is that traditional hydrological models degrade significantly in performance when calibrated for multiple basins together instead of for a single basin alone. In this paper, we propose a novel, data-driven approach using Long Short-Term Memory networks (LSTMs), and demonstrate that under a 'big data' paradigm, this is not necessarily the case. By training a single LSTM model on 531 basins from the CAMELS data set using meteorological time series data and static catchment attributes, we were able to significantly improve performance compared to a set of several different hydrological benchmark models. Our proposed approach not only significantly outperforms hydrological models that were calibrated regionally but also achieves better performance than hydrological models that were calibrated for each basin individually. Furthermore, we propose an adaption to the standard LSTM architecture, which we call an Entity-Aware-LSTM (EA-LSTM), that allows for learning, and embedding as a feature layer in a deep learning model, catchment similarities. We show that this learned catchment similarity corresponds well with what we would expect from prior hydrological understanding.

• The paper presents LSTM-based hydrological modeling that enhances regional rainfall-runoff predictions using 531 basin datasets.
• It introduces the EA-LSTM model with an embedding layer that learns catchment similarities, outperforming both regional and locally calibrated models.
• Benchmark results demonstrate a significant mean NSE improvement (0.67 vs. 0.44), highlighting the method's scalability and accuracy.

A Review of "Towards Learning Universal, Regional, and Local Hydrological Behaviors via Machine-Learning Applied to Large-Sample Datasets"

The paper "Towards Learning Universal, Regional, and Local Hydrological Behaviors via Machine-Learning Applied to Large-Sample Datasets" by Kratzert et al. explores the use of Long Short-Term Memory (LSTM) networks for hydrological modeling, focusing on regional rainfall-runoff simulations. The research confronts the longstanding challenge of hydrological sciences: achieving spatially continuous modeling across large regions without compromising accuracy, a scenario where traditional models often falter.

Main Contributions

1. Regional Model Calibration with Machine Learning: The authors propose using LSTMs trained on extensive datasets encompassing 531 basins from the CAMELS dataset. The approach leverages not only meteorological time series data but also static catchment attributes to enhance the prediction accuracy of runoff models.
2. Entity-Aware LSTM (EA-LSTM): A novel adaptation of the LSTM architecture, EA-LSTM, is introduced. This model variation incorporates an embedding layer capable of learning catchment similarities. The results demonstrate that the EA-LSTM can effectively differentiate between diverse hydrological behaviors, reflecting catchment-specific attributes.
3. Benchmarking Against Traditional Models: The LSTM and EA-LSTM models are benchmarked against several hydrological models, both regionally and locally calibrated. The findings reveal that the EA-LSTM not only surpasses regionally calibrated models but also exceeds the performance of some models individually calibrated for specific basins.

Numerical Results and Implications

• The LSTM-based models showed robust performance improvements, achieving significantly higher mean and median Nash-Sutcliffe Efficiency (NSE) values across the basins compared to traditional models.
• Specifically, the EA-LSTM demonstrated a mean NSE of 0.67, outperforming regionally calibrated models (mHM with an NSE of 0.44) and reducing basin-specific prediction failures.
• Notably, the EA-LSTM can learn and generalize hydrological similarities across multiple basins, which is a significant advancement over traditional models that require specific basin calibration.

Implications for AI and Hydrological Sciences

The implications of this work are substantial. By integrating data-driven methods with hydrologic modeling, the paper shows promise for achieving scalable and transferable hydrological models that do not require recalibration for each new catchment area. Such models can potentially offer a unified framework for regional hydrological modeling while preserving the ability to account for local variations in catchment behavior.

Future Directions

The paper opens avenues for further research into dynamically incorporating evolving catchment attributes over time. Future work could involve using real-time remote sensing data to update static inputs, potentially leading to even more responsive and accurate models. Moreover, investigating the EA-LSTM's capability with more diverse and expansive datasets could further affirm its utility in real-world applications.

Overall, the paper makes a significant contribution to the intersection of machine learning and hydrology, providing a promising methodology for addressing one of the field's enduring challenges.