Building up Trustworthiness of Deep Learning-based Emulator for Environmental Process-based Model
Abstract
Deep learning (DL) models have been widely employed as emulators to process-based models (PBMs) in environmental systems due to their fast computation with high accuracy. Nevertheless, the black-box nature of the DL models prevents us from evaluating the trustworthiness of an emulator. That is, the extent to which DL-based emulators can capture the interdependencies encoded in the PBM. To this end, sensitivity analysis provides an opportunity to open these black boxes by evaluating the interdependencies of a DL-based emulator. Here, we conducted a proof-of-concept study to assess the trustworthiness of DL-based emulators for emulating ten years daily discharges simulated by the Soil Water Assessment Tool (SWAT) watershed model for the American River Watershed. Two different DL-based emulators are developed, taking in the same set of standardized SWAT parameters but simulating discharges using different transformations (i.e., standardization (DNN-SD) and power transformation (DNN-PT)). We assessed the interdependencies of the emulators by comparing their parameter sensitivities with that of SWAT, computed by mutual information (MI). The result shows that while DNN-SD underestimates the parameter sensitivity during low flow period, DNN-PT overestimates their sensitivity for high flow. This distinction further facilitated developing a mixture emulator, which shows the improved performance by taking the weighted sum of the two emulators. The findings in this study revealed various emulation biases using different DL models. The biases in predictions suggest the significance of assessing their interdependencies before further usage and highlight the potential of a mixture or ensemble model in developing emulators for PBMs.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.H52C..01J