FOSS4G-Asia 2023 Seoul

Hydrological models play a pivotal role in mathematically simulating the complexities of real-world hydrological processes. However, accurately capturing the intricate physical nature of basin hydrology through parameterized representations remains a formidable challenge, particularly in regions with limited or unavailable gauge sites, a prevalent scenario in developing nations and remote locales. A notable trend has recently emerged where remote sensing inversion products or reanalysis data, such as evapotranspiration, soil water, and leaf area index, are increasingly used as substitutes for ground truth observations in the calibration of hydrological models. But this introduces a considerable risk when directly employing them for hydrological modeling as these products inherently harbor unknown errors.

This study intends to address the above challenges by proposing an adaptive procedural fitting algorithm to support multivariate calibration, employing an iterative stepwise approach. The algorithm simplifies the hydrological model into a directed graph, with each node representing a subprocess in the hydrological model. This allows for quantifying the impact of a parameter on each subprocess by considering the traversal relationship between the subprocess to which the parameter is attached and other subprocesses. Consequently, the iterative calibration process enables the updating of parameters at different levels or learning rates. Furthermore, two categories of methods were introduced to evaluate the outputs of subprocesses relative to their calibrated datasets: ground truth observations and non-observation data with unknown errors. One category employs standard metrics to evaluate the simulated data against observations, while the other utilizes multiple collocation methods to evaluate the simulation with the data with unknown errors. The proposed multivariate calibration framework offers an innovative approach to hydrological modelling, facilitating the effective utilization of increasingly abundant remote sensing products and reanalysis data to enhance the accuracy and reliability of hydrological models.

The presented approach has been realized through the utilization of an open-source GIS solution, integrating tools such as Python, xarray, geopandas, multiprocessing and SWAT to form an efficiently processed parallel program. The demonstration was applied in a case study conducted within the Malian river Basin situated in Gansu Province, China, involving the integration of remotely sensed ET products and gauged streamflow datasets to showcase the applicability and effectiveness of the proposed method.

Remote sensing evapotranspiration (RS-ET) products have been the ubiquitous datasets of echo-hydrology, increasingly employed in precise hydrological modeling. However, these diverse ET products often offer considerable uncertainties, posing substantial challenges when incorporating them into hydrological models. This study aims to investigate the effects of integrating remote sensing evapotranspiration (RS-ET) products into the calibration of hydrological models. Specifically, we address two main questions: (1) among those RS-ET products considered, which offers the highest hydrologic predictability for hydrological modeling? (2) What is the optimal scale for incorporating RS-ET datasets into the hydrological model to achieve maximum benefits for enhancing hydrological modeling accuracy?
We conducted an in-depth exploration of these questions by implementing a hydrological modeling experiment using the Soil and Water Assessment Tool (SWAT) within the Meichuanjiang Basin in Jiangxi Province, China. This research employed four distinct RS-ET products: MOD16, GLASS, SSEBop, and ETMonitor, along with a merged product. Additionally, we examined three distinct scales for aggregating RS-ET datasets: subbasin (SUB), land use (LU), and hydrological response unit (HRU).
Our findings revealed discernible variations in the evapotranspiration (ET) simulations of SWAT models across the study area, upon the calibration with five distinct products. Notably, GLASS and the fusion datasets demonstrated superior ET simulation performance with best scores, evaluated by a Kling-Gupta Efficiency (KGE) both surpassing 0.6. However, SSEBop yielded the lowest KGE of approximately 0.5. This study further highlighted that while the accuracy of ET simulations was enhanced, this improvement incurred a marginal cost in terms of a slight reduction in streamflow precision. However, this compromise was deemed acceptable due to its role in contributing to a more accurate spatial representation of catchment behavior while mitigating equifinality.
Similarly, at different scales of the same ET dataset used in calibration, although the differences in simulation results might be minimal, each dataset seems to have a specific optimal input scale. Notably, during the simulation of the evapotranspiration hydrological component, there is a particularly significant observation: the land use (LU) or subbasin (SUB) scale often proves to be more effective. The rationale for selecting the LU scale lies in the strong correlation between ET and various land use types. By analyzing each land use type separately, we can more accurately capture their distinct ET patterns, leading to improved simulation outcomes. Conversely, the enhanced effectiveness of the SUB scale can be attributed to the scale aggregation effect.
Considering that different RS-ET datasets and aggregation scales introduce uncertainty in observation data for calibration, we infer that the uncertainty in observation data gives rise to simulation uncertainty in inversing SWAT, subsequently backpropagating into parameter uncertainty. These findings shed light on the critical significance of carefully considering the quality of RS-ET datasets and the chosen aggregation scale in the context of hydrological modeling.