Robertson, PhyllisMcCaleb, Karen N.2022-08-162022-08-162022-04-15Robertson, P., McCaleb, K.N. and McFarland, L.A., 2022. Preparing All Educators to Serve Students with Extensive Support Needs: An Interdisciplinary Approach. The New Educator, 18(1-2), pp.87-109.https://hdl.handle.net/1969.6/93826Temporal gaps within the Gravity Recovery and Climate Experiment (GRACE) (gap: 20 months), between GRACE and GRACE Follow-On (GRACE-FO) missions (gap: 11 months), and within GRACE-FO record (gap: 2 months) make it difficult to analyze and interpret spatiotemporal variability in GRACE- and GRACE-FO-derived terrestrial water storage (TWSGRACE) time series. In this study, an overview of data and approaches used to fill these gaps and reconstruct the TWSGRACE record at the global scale is provided. In addition, the study provides an innovative approach that integrates three machine learning techniques (deep-learning neural networks [DNN], generalized linear model [GLM], and gradient boosting machine [GBM]) and eight climatic and hydrological input variables to fill these gaps and reconstruct the TWSGRACE data record at both global grid and basin scales. For each basin and grid cell, the model performance was assessed using Nash–Sutcliffe efficiency coefficient (NSE), correlation coefficient (CC), and normalized root-mean-square error (NRMSE), a leader model was selected based on the model performance, and variables that significantly control leader model outputs were defined. Results indicate that (1) the leader model reconstructed the TWSGRACE with high accuracy over both grid and local scales, particularly in wet and low anthropogenically active regions (grid scale: NSE = 0.65 ± 0.20, CC = 0.81 ± 0.13, and NSE = 0.56 ± 0.16; basin scale: NSE = 0.78 ± 0.14, CC = 0.89 ± 0.07, and NRMSE = 0.43 ± 0.14); (2) no single model was flawless in reconstructing the TWSGRACE over all grids or basins, so a combination of models is necessary; (3) basin-scale models outperform grid-scale models; (4) the DNN model outperforms both GLM and GBM at the basin scale, whereas the GBM outperforms at the grid scale; (5) among other inputs, the Global Land Data Assimilation System (GLDAS)-derived TWS controls the model performance on both basin and grid scales; and (6) the reconstructed TWSGRACE data captured extreme climatic events over the investigated basins and grid cells. The developed approach is robust, effective, and could be used to accurately reconstruct TWSGRACE for any hydrologic system across the globe.en-USAttribution 4.0 Internationalhttp://creativecommons.org/licenses/by/4.0/GRACEGRACE-FOgap fillinggrid scalebasin scalemachine learningArticlehttps://doi.org/10.3390/rs14071565