Honey, I blew up the forcings: Evaluating the robustness of Deep Learning-based rainfall-runoff models
Aldo Tapia1 and Andrew Bennett1
1Department of Hydrology and Atmospheric Sciences. The University of Arizona, Tucson, AZ
Hydrological models are fundamental for understanding and predicting streamflow dynamics, especially under varying management strategies and climate scenarios. In recent years, Deep Learning rainfall-runoff models have demonstrated a high predictive skill. However, the robustness to forcing variables changes, especially when these changes are out of the training distribution, is still under discussion. In this study, we systematically evaluate both physically-based and multiple DL approaches (LSTM, CNN, Transformer, and MLP-Mixer) for streamflow prediction across the CAMELS basins in the Contiguous United States under a wide range of conditions.
We designed and conducted perturbation experiments of precipitation and temperature to explore the sensitivity of the model structures. An initial perturbation analysis includes applying Directional Expectation across models and computing the rate of change in streamflow as a function of precipitation and temperature perturbations. We also include a range of perturbations based on a pseudo-global warming experiment derived from downscaled CMIP6 outputs and compare with the baseline. Both analyses are performed with different datasets, while the rate of change obtained from the Directional Expectation test will be evaluated under non-stationary conditions.
We compare the performance and insights of physics-based models with those of DL methods by analyzing key drivers of streamflow variability. We use the physically-based model to set a baseline for the expected results caused by our range of perturbed inputs, providing a weak benchmark for determining DL model robustness to changes.