ExplainabilityTransparencyTrustworthinesscs.LG
NeuralHydrology -- Interpreting LSTMs in Hydrology
Frederik Kratzert, Mathew Herrnegger, Daniel Klotz, Sepp Hochreiter, Günter Klambauer
March 19, 2019
Explainable AI
109 citations
12 influential
Key Takeaways
-
LSTMs can learn physically meaningful representations: The study demonstrates that LSTM networks trained on rainfall-runoff prediction internally develop patterns consistent with established hydrological science, suggesting the models learn genuine system dynamics rather than spurious correlations.
-
Interpretability is crucial for environmental AI adoption: The authors identify difficulty in interpreting neural network internals as a significant barrier to using LSTMs in environmental sciences, highlighting that technical performance alone is insufficient—stakeholders need to understand how models make decisions.
-
Memory mechanisms align with domain knowledge: The paper shows that LSTM memory cells naturally represent dynamic water storage and reservoir behavior, which are core concepts in hydrology, indicating the architecture's design principles match real-world environmental processes.
-
Catchment-specific validation matters: By testing interpretability across two different catchments (one with snow influence, one without), the research demonstrates that responsible AI deployment in environmental science requires validation across diverse conditions to ensure models generalize meaningfully rather than overfitting to specific datasets.
Abstract
Despite the huge success of Long Short-Term Memory networks, their applications in environmental sciences are scarce. We argue that one reason is the difficulty to interpret the internals of trained networks. In this study, we look at the application of LSTMs for rainfall-runoff forecasting, one of the central tasks in the field of hydrology, in which the river discharge has to be predicted from meteorological observations. LSTMs are particularly well-suited for this problem since memory cells can represent dynamic reservoirs and storages, which are essential components in state-space modelling approaches of the hydrological system. On basis of two different catchments, one with snow influence and one without, we demonstrate how the trained model can be analyzed and interpreted. In the process, we show that the network internally learns to represent patterns that are consistent with our qualitative understanding of the hydrological system.