Transferring Learning Between Machine Learning and Physics-Based Approaches to Improve Characterization of Watershed Behavior.

Abstract:

Streamflow estimation is very important in the economic and human development of a country. For example, it is used in the design of new hydraulic infrastructure, risk quantification, quick response again flooding, etc. For this reason, learning how to improve our estimates must be one of the aspirations of any surface hydrologist. Chile has an extensive stream gauge network, which is part of the new CAMELS-CL database. This database includes several static attributes for each one of the 516 catchments studied which gives us a valuable database to develop physics-based and data-based models.

Recent work has shown that Machine Learning (ML) can provide better predictive performance than traditional physics-based (PB) models. In hydrology, Kratzert et al. (2019), Nearing et al. (2020), and others have reported similar results when comparing an ML-based model with the extensively studied and calibrated SAC-SMA and other benchmark models over the USA. This creates the opportunity to bridge the gap between the ML-based and PB models by transferring what the ML model has learned to the PB model(s). One of the challenges of any PB model is the structure of it, however, ML models have a flexible structure and processes, which are learned from the data and that allows us to use the ML model as a benchmark to check the structure and process of our PB model. With this in mind, we implemented the GR4J lumped catchment model, the Random Forest (RF) ML approach, and the Long-Short Term Memory (LSTM) ML approach on 322 selected Chilean catchments. Following an experimental perturbations procedure such as the application of a uniform and constant precipitation to analyze the streamflow response, we can investigate how the RF and LSTM machine-learning response against linearity, time of concentration and others, and how different the PB model performances, which can then be used to help improve the physics-based model. Finally, the acquired knowledge is found to be related not only to the meteorological forcings but also with static variables such as aridity which emerges as an important variable to characterize the behaviors of different catchments.