Toward Essential Union Between Evolutionary Strategy and Data Assimilation for Model Diagnostics: An Application for Reducing the Search Space of Optimization Problems Using Hydrologic Genome Map


Multi-objective evolutionary algorithms (EAs) and data assimilation (DA) methods are usually employed independently in water resources modeling. EAs are widely recognized to possess stochastic and adaptive capabilities but they lack temporal evolution of model behavior when applied in hydrology. However, DA methods have gained credibility to account for observation and model uncertainties, and to temporally update model states. While these independent capabilities are naturally needed in hydrology, there are limited methodologies to combine EAs with DA methods. Consequently, this study demonstrates the evolutionary data assimilation (EDA) approach as a unified framework to combine the capabilities of multi-objective EM and DA. A unique feature of the EDA is its provision of biological genome-like data leading to further development of a hydrologic genome map with diagnostic and predictive descriptions of the model. The developed genomic map is shown to be temporally persistent connecting distinct landmarks in initial model states, parameters, and input forcing variables. The approach is demonstrated for the assimilation of daily streamflow into the Sacramento Soil Moisture Accounting (SAC-SMA) model in Fairchild Creek catchment in southern Ontario, Canada. The findings show that the genomic map reduced the original search space by 75% and the updated bound by 63%. The diagnostic capability of the genome map is supported by its identification of robust and sensitive model parameters/variables leading to a better description of the model behavior to changes in decision space. In terms of streamflow estimation, the genome map was found to be 84% as accurate as the updated streamflow estimate, 36% more accurate than the calibration output, and 54% more accurate than the open loop estimate. When evaluated for streamflow forecasts for up to 30 days ahead, the genomic map was found to be 99% accurate as the updated estimate and 16% more accurate than the calibrated forecasts. These demonstrated diagnostic and predictive capabilities of the genome map provide an appealing framework towards a better understanding of water resource modeling to improve planning and management outcomes.