Authors: Andy Wood – Colorado School of Mines
Title: Never train a process-based hydrology model on a single basin? Applying lessons from deep learning in hydrology to the calibration of traditional land/hydrology models.
Presentation Type: Poster
Abstract: Model calibration is arguably the most critical capability needed for hydrologic modeling and applications, but remains a long-standing challenge over large and heterogeneous domains like CONUS. Recent advances in deep learning (DL) modeling in hydrology demonstrate that calibration benefits from training on large collections of hydro-climatically varied basins. We have now shown that this lesson is true for traditional process-based (PB) land/hydrology models, which may be far more complex and expensive. We demonstrate a DL based calibration strategy (termed a ‘large-sample emulator’ LSE) as applied to a CONUS watershed-based implementation of SUMMA, a process-based hydrology model, coupled with the mizuRoute channel routing model. We tested a variety of DL methods and choices in developing a DL model emulation and parameter optimization scheme to calibrate and regionalize SUMMA parameters. We find that jointly training the LSE using catchment attributes, parameters and performance metrics outperforms traditional individual basin calibration, localizing the influence of parameters and providing support for model implementation in out-of-sample catchments. This presentation summarizes the methodologies, results, and challenges in this effort, undertaken for USACE and Reclamation to build national modeling resources and water security initiatives. The DL-based approach offers a compelling contrast to existing calibration approaches being applied to NextGen models.