Learning hydrologic dynamics from big data with sensitivity analysis and differentiable hydrology
Abstract
The training of a deep, recurrent neural network model on rainfall-runoff data from a large number of catchments lead to a major break-through in hydrological modelling. New avenues are emerging to use big data machine learning not solely as a predictive tool but also as a knowledge discovery tool. Here we explore using sensitivity analysis as well as process learning to provide insights. First, the use of a large number of catchment attributes that affect runoff directly or indirectly bears the problem of getting a good fit for only partly the right reason. For example, DL algorithms could incorrectly fit to certain features, reducing the robustness of future projections under climate change. We use sensitivity analysis to study the extent of this problem with current long short-term memory (LSTM) models in comparison with process-based ones, and to which degree this problem can be reduced by limiting the considered catchment attributes. Further, we included more model structures to understand how such structural priors can reduce false sensitivities and improve model robustness.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.H35S1260S