Regional hydrological modelling with deep convolutional-recurrent neural networks: A case study in Western Canada

With the recent success of Long Short-Term Memory (LSTM) neural networks applied to hydrological modelling problems (e.g. Kratzert et al. 2018), there is opportunity to explore the suitability of additional deep learning architectures, as well as a need to understand why these models are successful. In this study, we design a spatio-temporal deep learning model which both: 1) performs well according to traditional hydrological modelling metrics, and 2) learns physically realistic features.

We sequentially combine a Convolutional Neural Network (CNN) with an LSTM in order to encode both space and time into a regional hydrological model. Our study region spans a majority of the Canadian provinces of British Columbia and Alberta where a range of streamflow regimes exist, including rainfall-dominated streams along the Pacific coast, snow- and glacier-melt dominated streams in high-elevation mountain ranges, and relatively arid Prairie streams. We use the past 365 days of weather as a model input, constructed as `video' from climate reanalysis data (with dimensions of 365 x latitude x longitude x channels). Each frame in the video is a day of reanalysis data with precipitation and temperature channels. The model output is the next day of streamflow at over 100 stream gauge stations throughout our study region.

We use the CNN-LSTM model to predict streamflow simultaneously at all stream gauge stations studied. We also perturb the spatio-temporal model input in order to understand where in space the model is most sensitive when making its predictions. Here we develop new metrics to evaluate how physically realistic the model's learning is, based on where the model is sensitive for prediction at each station.

We find that different streamflow regimes are simulated with varying degrees of success, with mountain-runoff dominated streams having the best performance. We show that the model tends to be most sensitive in the areas near the stream gauge stations being predicted for each streamflow regime, providing evidence that the model is automatically learning the areas which are physically relevant for streamflow prediction. This work marks a step towards the development of interpretable deep learning hydrological models.