Leveraging spatial autocorrelation information of remotely sensed evapotranspiration for mitigating the impact of data uncertainty on hydrological modeling
Global remotely sensed evapotranspiration (RS-ET) products are increasingly pivotal in enhancing the accuracy and scope of hydrological modeling, particularly in regions where traditional ground-based streamflow data are sparse or non-existent. These products play a pivotal role in understanding the dynamics of the climate-soil-vegetation system, where evapotranspiration constitutes a substantial portion of water loss following precipitation events. Their extensive spatial coverage and accessibility have significantly expanded the capability to predict hydrological dynamics in ungauged basins, offering insights that were previously inaccessible through in-situ observations alone.
Despite their benefits, RS-ET products are tempered by inherent uncertainties, primarily stemming from biases that vary across datasets and geographical regions. These biases manifest as either overestimation or underestimation compared to ground truth measurements, posing challenges for the accurate calibration of hydrological models. Traditional approaches in hydrological modeling commonly utilize absolute ET values directly derived from RS-ET products for model calibration, without accounting for potential biases. However, the reliability of such direct calibrations is contingent upon the quality and accuracy of the RS-ET data, which remains uncertain in many cases.
To address these challenges, this study shifts the focus from absolute ET values to utilizing spatial structural information embedded within RS-ET data, particularly emphasizing spatial autocorrelation, which refers to the tendency of ET values at nearby locations to exhibit similarities. Employing the local Moran's I index, a spatially weighted autocorrelation statistic that is insensitive to biases, we capture the spatial structure of ET data across sub-basins. Additionally, a composite Kling-Gupta Efficiency (KGE) metric, integrating absolute ET values and spatial autocorrelation information in a weighted manner, is employed for calibrating hydrological models.
Three calibration schemes are thus designed to analyze the effectiveness of spatial autocorrelation in hydrological modeling: one focusing solely on absolute ET values, another solely on spatial autocorrelation, and a combined approach. Testing these schemes for hydrological modeling with four RS-ET products in the Meichuan basin—MOD16, GLASS, and SSEBop with large biases, and PMLV2 with minimal bias—the study demonstrates varying effectiveness across these schemes.
For RS-ET products with substantial biases, hydrological modeling using spatial autocorrelation proved to be the optimal solution. It achieved a higher KGE and lower Percent Bias (PBIAS) on simulated streamflow compared to using solely the absolute ET value or the combined approach. Conversely, for RS-ET products with minimal biases, hydrological models calibrated using both the combined approach were considered the preferred solutions. This approach can result in a high KGE, similar to that obtained from spatial autocorrelation information alone, while maintaining a reasonable PBIAS. Therefore, we recommend calibrating hydrological models using both absolute ET values and spatial autocorrelation information in regions where ground ET observations are available. This approach enhances the robustness and reliability of hydrological predictions, mitigating the influence of biases inherent in RS-ET products.
In contrast, in scenarios where the quality of RS-ET products is unknown, we suggest calibrating using only spatial autocorrelation information, thereby circumventing potential biases and improving model accuracy under such circumstances. Moreover, methodologically, the study contributes by demonstrating the efficacy of the local Moran's I index in capturing the spatial structure of ET data within hydrological sub-basins. This geostatistical measure not only quantifies spatial autocorrelation but also identifies clusters and patterns of ET values, thereby enriching our understanding of spatial variability in hydrological processes.
Furthermore, the comprehensive analysis of the composite KGE index underscores the significant contribution of spatial autocorrelation information to hydrological modeling, surpassing the influence of absolute ET values in enhancing model performance. In conclusion, the spatial autocorrelation-based approach presented in this study represents a significant advancement in the application of global RS-ET products for hydrological modeling. By leveraging spatial structural information and mitigating biases inherent in RS-ET data, this approach not only improves the accuracy of hydrological predictions but also enhances the practical utility of RS-ET products in diverse hydrological contexts. Future research directions may explore additional spatial statistical techniques and incorporate a broader array of RS-ET datasets to further refine and validate these findings across different geographical settings and hydrological conditions.