Linking explainable artificial intelligence and soil moisture dynamics in a machine learning streamflow model

Abstract

Machine learning algorithms are increasingly applied in hydrological studies with promising results. However, these algorithms generally lack the ability for easy interpretability of the results by users. In this study, we compare six different explainable artificial intelligence (XAI) algorithms that help understand the effect of input data on the simulation results. The methods are explored on two distinct approaches for streamflow modeling using a long short-term memory (LSTM) model: a single model approach using only meteorological forcing data and a regional approach including also static catchment attributes. To gain further insight into the internal dynamics of the LSTM models, the relationship between cell states and soil moisture is investigated. A strong correlation suggests that the LSTM models inherently capture the concept of soil moisture as a catchment-scale storage mechanism. The XAI methods are applied to derive a timestep of influence, revealing how many days of input data are relevant for the model output. All XAI methods result in similar seasonal patterns in the timestep of influence, suggesting that the methods are comparable. Setting soil moisture dynamics in context to seasonal development of the timestep of influence suggests resetting LSTM as soon as soil moisture saturation occurs. HIGHLIGHTS Two long short-term memory (LSTM) models were trained for streamflow modeling.; Internal cell states of LSTMs correspond well with ERA5-Land soil moisture dynamics.; Six explainable artificial intelligence algorithms show similar results.; LSTM resets as soon as soil moisture saturation is reached.;

Keywords

• explainable artificial intelligence
• long short-term memory
• machine learning
• soil moisture dynamics
• streamflow modeling