Discovery and Interpretation of Hydrological Process Representation Within Convolutional Long Short-Term Memory Neural Networks

Recently, progress has been made towards regional hydrological modelling using deep machine learning. In particular, convolutional long short-term memory (CNN-LSTM) neural networks have been used to map spatiotemporal meteorological information to streamflow at multiple stations throughout a region. However, key knowledge gaps exist around understanding why models make decisions, how physical processes are represented within a model, and how to discover the representation of such processes within deep networks. We use an ensemble of CNN-LSTM models trained to predict streamflow throughout southwestern Canada, a hydrologically diverse region with a range of climate and topographic complexity. We demonstrate the automatic learning of physically relevant and interpretable hydrological processes that vary in space and time, depending on streamflow regime. We focus on the following two aspects: What weather triggers a modelled flow increase? We identify weather patterns that cause modelled streamflow to increase in order to characterize the learned runoff generating mechanisms. We find, for example, that the models have automatically learned that spring streamflow behaviour is dominated by melt-driven flows in nival rivers as compared to rain-driven flows in pluvial rivers. How are physical processes represented by model parameters? We discover a set of LSTM cell states that we interpret as learned representations of glacier melt, which is a key driver of flow at some stream gauge stations but not others. These cell states are strongly connected to glacierized stream gauge stations and are correlated with daily temperatures in late summer, indicating their role as a temperature-controlled summer streamflow source unique to glacier-fed rivers.