Development of a model identification framework for runoff estimation in ungauged basins using earth observation data and application to three Greek basins

Hydrological modeling provides a mathematical framework that supports water resources management, prediction of responses to hydroclimatic extremes, efficient infrastructure design and understanding of environmental processes among other applications. Effective use of these models conventionally depended on reliable, continuous, and sufficiently long time series of runoff at the outlet of the basin under analysis to serve as a reference dataset for parameter calibration. However, reliable in situ runoff observations are often limited by the spatial coverage of flow gauge networks, discontinuities in monitoring records, and measurement and rating-curve errors. To avoid the limitations posed by runoff data availability, alternative approaches that leverage the increased accessibility of remotely-sensed/reanalysis data can be used. In the present study, a sensitivity- and uncertainty-informed parametrization and calibration framework is proposed and tested on the Evinos, Mornos and the Boeotikos Kephisos river basins. Model parametrization was informed by an extended sensitivity and uncertainty analysis of the outputs of the conceptual hydrological model Zygos using Morris’s screening method, the Extended Fourier Amplitude Sensitivity Test (eFAST) and Regional Sensitivity Analysis (RSA), in which a Kling-Gupta Efficiency (KGE) score threshold was used to classify parameter sets as behavioral or non-behavioral. This workflow was repeated for the simulated time series of runoff (Q), soil moisture (SM) and total evapotranspiration (ET) and for six different model configurations. The SM and ET results were used to assist in the selection of the appropriate model for each river basin and to restrict the feasible parameter space for calibration whereas Q results were retained as a reference for subsequent verification. Parameter calibration was performed under 10 different scenarios, with Q-based calibration serving as the reference scenario and the remaining nine scenarios using SM and/or ET as calibration targets. KGE, unbiased root mean square error (ubRMSE), normalized root mean square error (NRMSE), and the Pearson correlation coefficient (𝑟) were used as objective functions. Single-objective optimization was conducted on six scenarios utilizing the Shuffled Complex Evolution-University of Arizona (SCE-UA) algorithm. The remaining four cases were optimized using multi-objective optimization with the ε-dominance Non-dominated Sorting Genetic Algorithm II (ε-NSGA II). The results showed that the sensitivity and uncertainty patterns of SM and ET did not always match those of Q, indicating that the two EO variables could not fully constrain the parameters governing Q generation. The parametrization procedure supported the selection of the 7P and 5P configurations for Evinos and Mornos, respectively, whereas the absence of behavioral SM and ET solutions in Boeotikos Kephisos prevented parameter range restriction and led to the selection of 9P. The selected EO-based models achieved validation KGE values for Q of 0.708, 0.310 and 0.419 in the three basins, respectively, compared with 0.812, 0.776 and 0.802 for the corresponding Q-calibrated reference (Base) models. Although certain Q characteristics were reproduced satisfactorily, considerable differences remained in peak magnitude, timing and low-flow behavior, particularly in Mornos and Boeotikos Kephisos. The results also showed that satisfactory SM and ET calibration did not necessarily lead to satisfactory Q predictions, while the choice of calibration objectives influenced the simulated hydrological response. Generally, the results showed that EO data can be useful for Q estimation, although several developments would still be necessary in order for the proposed framework to be operationally applied in ungauged basins.