Statistical Applications to Downscale GRACE-Derived Terrestrial Water Storage Data and to Fill Temporal Gaps

Abstract

The Gravity Recovery and Climate Experiment (GRACE) has been successfully used to monitor variations in terrestrial water storage (GRACETWS) and groundwater storage (GRACEGWS) across the globe, yet such applications are hindered on local scales by the limited spatial resolution of GRACE data. Using the Lower Peninsula of Michigan as a test site, we developed optimum procedures to downscale GRACE Release-06 monthly mascon solutions. A four-fold exercise was conducted. Cluster analysis was performed to identify the optimum number and distribution of clusters (areas) of contiguous pixels of similar geophysical signals (GRACETWS time series); three clusters were identified (cluster 1: 13,700 km2; cluster 2: 59,200 km2; cluster 3: 33,100 km2; Step I). Variables (total precipitation, normalized difference vegetation index (NDVI), snow cover, streamflow, Lake Michigan level, Lake Huron level, land surface temperature, soil moisture, air temperature, and evapotranspiration (ET)), which could potentially contribute to, or correlate with, GRACETWS over the test site were identified, and the dataset was randomly partitioned into training (80%) and testing (20%) datasets (Step II). Multivariate regression, artificial neural network, and extreme gradient boosting techniques were applied on the training dataset for each of the identified clusters to extract relationships between the identified hydro-climatic variables and GRACETWS solutions on a coarser scale (13,700–33,100 km2), and were used to estimate GRACETWS at a spatial resolution matching that of the fine-scale (0.125° × 0.125° or 120 km2) inputs. The statistical models were evaluated by comparing the observed and modeled GRACETWS values using the R-squared, the Nash–Sutcliffe model efficiency coefficient (NSE), and the normalized root-mean-square error (NRMSE; Step III). Lastly, temporal variations in GRACEGWS were extracted using outputs of land surface models and those of the optimum downscaling methodology (downscaled GRACETWS) (Step IV). Findings demonstrate that (1) consideration should be given to the cluster-based extreme gradient boosting technique in downscaling GRACETWS for local applications given their apparent enhanced performance (average value: R-squared: 0.86; NRMSE 0.37; NSE 0.86) over the multivariate regression (R-squared: 0.74; NRMSE 0.56; NSE 0.64) and artificial neural network (R-squared: 0.76; NRMSE 0.5; NSE 0.37) methods; and (2) identifying local hydrologic variables and the optimum downscaling approach for individual clusters is critical to implementing this method. The adopted method could potentially be used for groundwater management purposes on local scales in the study area and in similar settings elsewhere.

Century‐Scale Reconstruction of Water Storage Changes of the Largest Lake in the Inner Mongolia Plateau Using a Machine Learning Approach

“…Jing et al (2020) reconstructed the GRACE TWSA by using the RF regression model combined with an SMW structure, the LR model, and the GLDAS, and found the RF model outperforms the LR model. Sahour et al (2020) applied the multivariate regression (MR), ANN, and XGBoost approaches to fill the GRACE in the time series data over the Lower Peninsula of Michigan during 2002-2016 and found that the XGBoost model outperformed the MR and ANN.…”

Dynamic Mapping of Subarctic Surface Water by Fusion of Microwave and Optical Satellite Data Using Conditional Adversarial Networks

“…However, empirical approaches using machine learning (ML) have been used for a flexible fusion of datasets with highly different features [28,29]. The fundamental process involves ML training with matched pairs between different types of data and then using the training results to predict spatially high-resolution but temporally low-resolution data from counterpart data (temporally high-resolution but spatially low-resolution).…”

Improving the Resolution and Accuracy of Groundwater Level Anomalies Using the Machine Learning-Based Fusion Model in the North China Plain

“…Although data assimilation methods remain consistent in the physical process, some shortcomings still require to be considered [ 25 ]. The implementation of data assimilation is relatively complicated [ 26 ], and its accuracy is subject to the full error covariance matrix of GRACE observations and hydrological models [ 27 , 28 ].…”

“…Artificial neural networks (ANNs) have the capabilities of simulating complex hydrological characteristics to an arbitrary degree of accuracy [ 34 , 35 ]. This makes ANN becomes an attractive measure in the downscaling researches, which have been applied to some typical regions, e.g., the Northern High Plains [ 36 ], California’s Central Valley [ 34 ], and the Lower Peninsula of Michigan [ 26 ]. Similarly, some tree-based machine learning algorithms (e.g., random forest (RF) and gradient boosting decision tree (GBDT)) become popular in regression tasks with the advantages of simplicity and effectiveness.…”