Journal Search Engine
Search Advanced Search Adode Reader(link)
Download PDF Export Citaion korean bibliography PMC previewer
ISSN : 1225-6692(Print)
ISSN : 2287-4518(Online)
Journal of the Korean earth science society Vol.39 No.5 pp.473-482
DOI : https://doi.org/10.5467/JKESS.2018.39.5.473

Application of the Empirical Orthogonal Functions on the GRACE Spherical Harmonic Solutions

Jooyoung Eom1, Ki-Weon Seo2*
1Department of Earth Science Education, Kyungpook National University, Daegu 41566, Korea
2Department of Earth Science Education, Seoul National University, Seoul 08826, Korea
Corresponding author: seokiweon@snu.ac.kr Tel: +82-2-880-4070
August 2, 2018 September 20, 2018 October 2, 2018

Abstract


During the period of 2002 to 2017, the Gravity Recovery And Climate Experiment (GRACE) had observed time-varying gravity changes with unprecedented accuracy. The GRACE science data centers provide the monthly gravity solutions after removing the sub-monthly mass fluctuation using geophysical models. However, model misfit makes the solutions to be contaminated by aliasing errors, which exhibits peculiar north-south stripes. Two conventional filters are used to reduce the errors, but signals with similar spatial patterns to the errors are also removed during the filtering procedure. This would be particularly problematic for estimating the ice mass changes in Western Antarctic Ice Sheet (WAIS) and Antarctic Peninsula (AP) due to their similar spatial pattern to the elongated north-south direction. In this study, we introduce an alternative filter to remove aliasing errors using the Empirical Orthogonal Functions (EOF) analysis. EOF can decompose data into different modes, and thus is useful to separate signals from noise. Therefore, the aliasing errors are effectively suppressed through EOF method. In particular, the month-to-month mass changes in WAIS and AP, which have been significantly contaminated by aliasing errors, can be recovered using EOF method.



초록


    Korea Institute of Marine Science & Technology Promotion

    Introduction

    Understanding of the global water cycle, the exchange of water mass among oceans, atmosphere and land, is essential to decipher Earth's climate change. The global water and ice mass redistribution occurs via precipitation, evapotranspiration, sublimation, surface/subsurface water flow and ice flow. Continuous hydrologic, meteorologic and cryospheric observations for theses phenomena are important to understand and predict climate changes and to manage corresponding natural hazards (IPCC, 2014). Despite much efforts for the observation, sparse in-situ stations hinder understanding the water and ice mass balance. This limitation is particularly true for the Antarctica and Greenland.

    The Earth’s gravity monitoring from space can be an alternative observation of water cycle because the redistribution of water mass can cause changes in the gravity field (Cazenave and Chen, 2010). In particular, the Gravity Recovery and Climate Experiment (GRACE) (2002-2017) mission has provided unprecedented observations for changes in the gravity field. The GRACE mission consists of co-orbiting twin satellites apart from about 220 km at altitude of ~450 km (Tapley et al., 2004). Onboard K-band microwave interferometers measure inter-satellite distance at accuracy of ~μm (Dunn et al., 2003). The measurements are used to determine the gravity field and further estimate water mass variations with accuracy of less than 1 cm equivalent water height. This is about 2 orders of magnitude better than those of EGM96, the pre-GRACE gravity model (Wahr et al., 2004). Since its launch, the monthly gravity anomalies provided by GRACE have been widely used in hydrologic (Alsdorf et al., 2010;Eom et al., 2017a;Reager et al., 2014), atmospheric (Kim et al., 2016;Seo et al., 2014), oceanographic (Leuliette and Miller, 2009;Llovel et al., 2014;Makowski et al., 2015) and cryospheric (Eom et al., 2017b;Seo et al., 2015) studies with global to regional scale.

    However, there are some limitations of the monthly GRACE solutions for a region with small spatial scale less than a few hundred kilometers (Longuevergne et al., 2010). Monthly gravity solutions are determined by integrated along-track observations in combination with a background gravity model during the observation epoch. During the monthly sampling, sub-monthly gravity fluctuations cause aliasing error including spurious longer period variabilities and peculiar meridional stripes (Wahr et al., 2004). To suppress the aliasing error, numerical models are incorporated into GRACE’s data processing in order to remove sub-monthly fluctuation of atmospheric surface pressure and ocean bottom pressure (Bettadpur, 2012). However, those models are insufficient to reduce such aliasing error: high-frequency gravity residuals remain in GRACE observation resulting in significant aliasing error.

    To remedy the aliasing error, spatial filters are commonly applied. After filtering, the error and noise can be remarkably reduced, leading to reveal the pattern of water mass signal but showing broader spatial features with attenuation. To overcome such shortcomings, several approaches have been developed. For example, Velicogna et al. (2005) used an empirical scale factor to compensate the signal loss to nearby oceans when estimating Greenland ice mass loss. More recently, Chen et al. (2013) used a forward modeling method that iteratively adjusts mass fields until its smoothed mass fields are close to the GRACE’s observation after spatial filtering. These signal recovering methods are efficient to estimate basin-averaged mass changes, but limited to understand details of spatial pattern because the former method does not alter the signal pattern and the latter suffers from spurious signal patterns.

    The empirical orthogonal functions (EOF) method can be an alternative for GRACE data to separate water mass signal from aliasing errors. The method decomposes a space- and time-variable dataset into spatial patterns and associated time coefficients describing each spatial pattern’s temporal variability (Hannachi et al., 2007). In atmospheric science, therefore, EOF analysis has been frequently used to find some dominant climate modes such as AO and MJO in climate data (Hurrell and Van Loon, 1997;Madden and Julian, 1971). Based on this feature, EOF method can be also used as a filter by exclusions of non-significant modes. In this study, we suggest the way to use EOF method as a filter for GRACE monthly solutions. We examine the EOF method in the Western Antarctic Ice Sheet (WAIS) and the Antarctic Peninsula (AP) where the conventional GRACE filtering is particularly problematic to recover accurate ice mass loss pattern.

    Data and Methods

    GRACE data

    The monthly GRACE solutions are freely available from the ftp server at the NASA PODAAC (ftp:// podaac.jpl.nasa.gov). Conventionally, a global gravity model is represented by spherical harmonic (SH) function because the gravitational potential field outside Earth’s surface satisfies Laplace’s equation. Therefore, the geoid height observed by GRACE is also described by coefficients of SH function, known as Stoke’s coefficients:(1)

    Δ N ( θ , φ , t ) = R e l = 0 m = 0 P ˜ l m [ Δ C l m t cos( m φ)+ Δ S l m t sin( m φ) ]
    (1)

    where ΔN is geoid height variations at colatitude, θ, and longitude, ϕ at a given month t. Re is the Earth's radius and P ˜ l m is normalized associated Legendre functions. The subscripts l and m represent the degree and order of SH coefficients, respectively. ΔClmt and ΔSlmt are Stokes coefficients after removing mean values representing static geoid. The solutions used in this study are provided by the Center for Space Research (CSR) at University of Texas at Austin, one of the three official GRACE data processing centers. Recently, CSR released a new version of gravity solutions, RL06. A solution consists of SH coefficients up to degree and order 60. In this study, 149 monthly solutions from January 2003 to August 2016 are used and linearly interpolated 15 omitting months completing full-time series for more than 13 years.

    There are several procedures to derive surface mass distribution from GRACE geoid height data. Table 1 outlines the post-processing flow-chart for GRACE SH solutions, which are commonly adopted by GRACE studies (Velicogna and Wahr, 2013). First, there are four geophysical data reductions as follows. (i) GRACE’s range-rate observation using on-board interferometers is not sensitive to the center of mass changes, which are represented by degree 1 SH coefficients, ΔC10, ΔC11 and ΔS11. Commonly these three coefficients are synthesized with the other highorder SH coefficients assuming that ocean mass changes are passively determined by terrestrial water and ice mass variations (Swenson et al., 2008). It is necessary to add the estimated monthly values of the degree 1 coefficients to GRACE SH solutions. (ii) GRACE ΔC20 shows significant aliasing errors due to S2 and K2 tides (Chen and Wilson, 2008), and are replaced with Satellite Laser Ranging (SLR) observations (Cheng and Tapley, 2004). (iii) The effect of viscoelastic mantle responses to the last glaciation, glacial isostatic adjustment (GIA), is corrected by the model based on ICE-5G ice history (A et al., 2013). (iv) Time-varying Stokes coefficients (ΔClmt and ΔSlmt) representing geoid changes are converted to SH coefficients for surface mass density ( Δ C ˜ l m t and Δ S ˜ l m t ) considering elastic loading response of solid Earth (Wahr et al., 1998).

    Second, because the reduced SH coefficients for surface mass density suffer from the aliasing error and random noise, two successive filtering procedures are usually practiced. Swenson and Wahr (2006) found that the aliasing error evidently exhibits correlated patterns and hence developed a de-correlation filter to suppress the error. In addition, random noise is suppressed by applying a kernel function for spatial average (Jekeli, 1981). The shape of the spatial kernel is a Gaussian distribution, and the kernel’s radius, r, was defined at which the kernel’s amplitude drops to a half. The radius range of about 300-500 km is empirically proven to effectively reduce most residual noise after de-correlation filter (Velicogna and Wahr, 2013).

    Therefore, there are three different reduced SHs for surface mass density; ones without any spatial filtering, the other with the de-correlation filtering only and another with the de-correlation filtering followed by 300-km Gaussian filtering (Table 1). Finally, these three SHs representing global surface density fields are localized on our study region including WAIS and AP (Harig and Simons, 2012). We named these datasets ‘no-filtered’, ‘de-correlated’ and ‘two-step’ (represented by gray-, blue- and greencolored boxes in Table 1, respectively) for simplicity.

    2.2. Empirical Orthogonal Functions

    EOF is an appropriate method for multivariate geophysical data field with space-time domain. The method is also known as principal component analysis or loading factor analysis. Although all methods have different nomenclatures and approaches of data interpretation, they all share a common mathematical technique, singular value decomposition (SVD). SVD decomposes a given dataset into orthogonal matrices; it is useful for EOF to separate signal and errors from GRACE observations.

    The monthly GRACE solutions can be arranged into a data matrix with size of n×p. For our GRACE data, n is the number of months (164 in this study), and p is the number of SH coefficients equals to 3721 when the maximum degree is 60. For a given month (t), SH mass anomaly coefficients ( Δ C ˜ l m t and Δ S ˜ l m t ) can be rearranged into a column vector:

    d t = [ Δ C ˜ 00 t Δ S ˜ 11 t Δ C ˜ 10 t Δ C ˜ 11 t Δ S ˜ 22 t Δ S ˜ 21 t Δ C ˜ 20 t Δ C ˜ 21 t Δ C ˜ 22 t ] T .
    (2)

    To construct a data matrix, all column vectors for the study period are merged into an n×p matrix(3):

    D = [ d 1 d 2 d t d n ] ,
    (3)

    where D is a data matrix consisting of monthly solutions according to Eq. (2).

    A data matrix is normally weighted by the number of coefficients at each degree prior to SVD. This procedure is equivalent to the latitude weighting for equal-spaced grid data, which is non-uniformly distributed on the Earth’s surface (i.e., denser poleward). Similarly, the degree weighting for the data matrix consisting of SH coefficients attains uniform contribution of each coefficient into the variance of a decomposed mode. Since the number of coefficients for the degree l is 2l+1, the diagonal weighting matrix W is(4)

    W = diag ( 1 , 3 , 3 , 3 , 5 , 5 , 5 , 5 , 5 , , 2 l + 1 , , 2 60 + 1 )
    (4)

    Then the weighted data matrix is

    D w = W D .
    (5)

    In the EOF analysis, spatial patterns and associated time series of dataset are decomposed by SVD technique:

    D w = U S V T ,
    (6)

    where both U and V are orthogonal matrices. Each column vector in V and U represents spatial pattern and associated time variability, respectively. S is a n×p rectangular matrix whose diagonal elements represent singular values of D, which equal the explained variance of each mode (spatial pattern and associated temporal variation). There are many highlevel computing languages to perform SVD calculation such as Fortran, Python and MATLAB®. For example, we can easily obtain U, S and V by typing ‘[U,S,V] =svd(D)’ in MATLAB® command line. The simplicity of SVD computation makes the EOF filter a powerful filtering strategy for GRACE SH solutions.

    Results

    Long-term variations in GRACE data

    Firstly, we examine long-term variations in ice mass change using the three GRACE datasets (‘no-filtered’, ‘de-correlated’ and ‘two-step’). Since the correlated striping errors are random in time, they do not significantly affect the trends over regions where mass changes steadily increase or decrease. This implies that in our study region, WAIS and AP, the no-filtered dataset can provide accurate long-term variations. On the other hand, the long-term variations from the filtered date sets such as ‘de-correlated’ and ‘twostep’, are likely underestimated due to spatial filtering. This examination highlights signal smoothing effects when filters are applied on GRACE data.

    There are many previous studies reporting that WAIS and AP have experienced rapid decreases in ice mass with acceleration in recent decades (IMBIE, 2018). Our three datasets also show similar results, but there are minor differences depending on filtering strategies. Figure 1 shows the annual rate of ice loss (a-c) and its acceleration (d-f) in each GRACE dataset. In Fig. 1a, large ice loss is observed in the Southern AP and the Amundsen sector of WAIS. A positive anomaly may represent ice mass gain in the center of WAIS, which is likely associated with the stagnation of ice flow in Kamb ice stream (Retzlaff and Bentley, 1993). The acceleration term in the nofiltered dataset (Fig. 1d) shows similar pattern with Fig. 1a in the Amundsen sector, but AP have a dipole structure with ice loss deceleration in the Southern region and acceleration in the Northern region.

    Figures 1b and 1e show results of de-correlation filter on long-term variations in GRACE estimates, which are close to those of ‘no-filter’ cases (Figs. 1a and 1d). However, there are some differences in spatial pattern and its amplitude of ice mass changes. The positive/negative structure in WA region is not clearly separated, compared to those of no-filtered dataset. In AP region, both of trend and acceleration are slightly reduced compared to Figs. 1a and 1d, respectively. These differences in spatial pattern and signal intensity arise because de-correlation filtering removes north-south stripes without distinguishing between signals and aliasing errors. Therefore, the signals with spatial pattern elongated in north-south direction are likely suppressed during de-correlation filtering.

    Figures 1c and 1f show the results when a Gaussian averaging filter is additionally applied to the GRACE data after de-correlation filtering. Amplitudes of the estimated trend and acceleration (Fig. 1c and 1f) are much more reduced than those shown in the previous results. The maximum values of ice loss rate and its acceleration in this case are 201.4 Gton/year and -0.5 Gton/year2, respectively. The recovered estimates by the two-step filter are about 50 % smaller than those of no-filtered dataset in WAIS. This reduction agrees well with numerical results of Velicogna and Wahr (2013). Through this analysis, we conclude that additional caution is necessary when GRACE data is reduced by both de-correlation and Gaussian filtering.

    EOF decomposition and reconstruction

    Before applying EOF filter, we remove trend and acceleration components from the no-filtered dataset. As described above, the long-term variations are likely recovered from the no-filtered dataset with high accuracy. Indeed, the ice loss rate and its acceleration shown in Figs. 1a and 1d generally agreed with the estimates from recent study (IMBIE, 2018). During GRACE era, there has been a consensus about the long-term ice mass variations over the Antarctic continent (IMBIE, 2018). However, there are still large uncertainties about seasonal ice mass variations with a sub-basin scale (Velicogna et al., 2014;Xiaolei et al., 2014). To examine this, we apply EOF filtering method to the de-trended dataset.

    As explained in Section 2, the EOF filtering is divided into two procedures: SVD decomposition of a given data matrix and data reconstruction with selected modes including signals. We therefore decompose the de-trended and no-filtered dataset using SVD. Figure 2 shows time series of the leading four modes (the kth column of U in Eq. (6), hereafter referred as to PCk). The PC1 and PC2 represent apparent inter-annual variations, which are likely associated with ice mass variations. In contrast, PC3 and PC4 show random variations resulting from observational noise and aliasing error. Higher modes also show the similar white noise (not presented here). This indicates that only the leading two modes from SVD decomposition of the dataset represent ice mass signals.

    A similar conclusion also be led by visual inspections of spatial modes. Figure 3 shows results for the first 4 EOF spatial modes (hereafter the kth column of V referred as to EOFk). Before presenting spatial modes, V is restored by the effect of degree weighting in Eq. (5) (i.e., W−1V). Figure 3a (EOF1) reveals a spatially coherent structure of ice mass loss. The first mode explains 89% of the total variance. In addition, EOF1 displays the similar spatial patterns shown in Fig. 1a and 1d. This indicates that a large fraction of ice mass fluctuations arises in a region where ice loss and its acceleration are dominant. EOF2 shows a dipole pattern separating between AP and WAIS, implying that different phases of ice mass variations in AP and WAIS. Although the second mode accounts only for 4.5 % of total variance, larger than three times of the explained variance of 3rd mode, it plays an important role in spatial and temporal variability of ice mass change in the study area.

    The EOF patterns in higher modes shows characteristics of the aliasing errors (Figs. 3c and 3d). In contrast to the first two EOFs, these represent distinctive patterns of meridional stripes in the association with noise and aliasing error. It is reasonable to conclude that these noisy modes should be excluded before examining the ice mass changes in the area. Note that third and higher modes account for 5.6 % of the total variance, which is larger than that of the second mode. This implies that amplitudes of some important ice mass variability in the area are smaller than those of aliasing errors so that the signal and error are hardly separated from each other using conventional filters.

    Based on the inspections of temporal and spatial modes, we conclude that the first two modes include most of ice mass signals, and the signal can be reconstructed with only the two modes. The reconstruction is achieved by multiplying the three matrices with the third and higher columns to be zero in the second term of Eq. (6). This simple algebra produces the EOF-filtered dataset (red box in Table 1).

    Month-to-month mass changes in GRACE data

    We calculate monthly difference of ice mass changes using the four GRACE datasets (‘no-filtered’, ‘de-correlated’, ‘two-step’ and ‘EOF reconstructed’). These represent month-to-month ice mass variability at sub-basin scale except the contribution of the trend and acceleration components. Figure 4a shows the monthly differences in ice mass from the no-filtered dataset. The map is dominated by distinctive northsouth stripes over whole study region, which are similar with EOF3 and EOF4 shown in Fig. 3. Note that the scale bars in Fig. 4 differ for each plot, which indicates that aliasing errors are significantly amplified when ice mass changes are subtracted between successive months. It further implies, during the considered time period, signal variations are relatively small compared to the aliasing error and noise.

    Figure 4b shows the result from the de-correlated dataset. Most of stripes disappear due to the effect of de-correlation filter, but pattern of ice mass changes much differs from what we expected in the region. Small scale positive and negative anomalies are likely artifacts. Figure 4c is a smoothed version of Fig. 4b with the Gaussian spatial averaging. Although the amplitude is significantly reduced, the spatial pattern of ice mass change is still similar with Fig. 4b. Finally, Figs. 4d shows month-to-month variations from the EOF-filtered dataset. The spatial pattern is very similar to those of Figs. 1a and 1d, which represents regions arising ice mass changes. If we presume that regions experinecing steady ice loss for decades also exhibit fluctuations in seasonal ice mass changes, we could conclude that the similarity in spatial pattern between Figs. 1a and 4d provides that EOF filter is most effective in suppressing aliasing errors without signal loss.

    Summary

    GRACE SH solutions are contaminated by aliasing errors due to sub-monthly atmospheric and oceanic fluctuations, which cannot be removed by numerical models during pre-processing procedures. A conventional filtering strategy to reduce the errors is a two-step procedure consisting of decorrelation and Gaussian filters. Although it is effective, the conventional method may also cause signal loss and hinder examination over a sub-basin scale. The limitation is particularly true in WAIS and AP because ice mass signal has a similar pattern of aliasing errors.

    In this study, we suggest the EOF filtering method to overcome the limitations of conventional filtering methods. The feature of EOF analysis decomposing a given dataset into two orthogonal matrices can be applicable to geophysical multi-variable data such as GRACE gravity solutions. Using SVD decomposition, signals and noise are separated into different modes, and we simply reconstruct a filtered-data with signal dominant modes. The comparison with conventional filters over the WAIS and AP regions shows the excellence of EOF filter to reduce aliasing errors.

    In May 22, 2018, the GRACE Follow-On (FO), a successor to GRACE, was successfully launched at Vandenberg, California. GRACE FO is the first satellite-to-satellite gravity mission equipped with the laser ranging interferometer instrument, and the accuracy is expected to be increased by about ten times compared with the GRACE mission. The gravity solutions provided by GRACE FO are expected to continue its wide applications for hydrologic, oceanographic, atmospheric and cryospheric studies. Therefore, EOF analysis on the SH solutions will also be useful for GRACE FO.

    Acknowledgments

    This study was supported by Korea Institute of Marine Science and Technology Promotion (KIMST) researchgrant KIMST20140410.

    Figure

    JKESS-39-473_F1.gif

    Maps of trend (a-c) and its acceleration (d-f) in ice mass change by different GRACE filtering strategies. (a) and (d) show the results for the no-filterd dataset. (b) and (e) are filtered data by de-correlation filter. The results of the two-step filter are shown in (c) and (f). Red line depicts our study region, Western Antarctic Ice Sheet (WAIS) and Antarctic Peninsula (AP). Dashed line represents a boundary of Slepian basis functions used in localization of SH coefficients.

    JKESS-39-473_F2.gif

    The first four temporal modes of EOF decomposition for the de-trended and no-filtered dataset. Each curve has a mean of zero, but curves are offset for clarity. The PC1 (blue) and PC2 (red) represent apparent inter-annual variations, which are likely associated with ice mass variations. In contrast, PC3 (green) and PC4 (black) show random variations resulting from observational noise and aliasing error. All curves have no physical unit because they are patterns of the variance in time after EOF decomposition.

    JKESS-39-473_F3.gif

    The first four spatial modes of EOF decomposition for the de-trended and no-filtered dataset. Each map represents distinct patterns: (a) a spatially coherent structure of ice mass signal over the study area, (b) a dipole pattern over AP and WAIS and (c) and (d) stripes resulting from aliasing errors. The distributions have no physical unit because they are patterns of the variance in space after EOF decomposition.

    JKESS-39-473_F4.gif

    Maps of monthly ice mass changes from September to October in 2014. (a) shows the result from the no-filtered dataset, (b) from the only de-correlated, (c) from the two-step filter and (d) from the EOF reconstructed.

    Table

    The flow chart representing the order of GRACE post-processing steps and resulting three datasets, “no-filtered”, “only de-correlation”, “two-step filtered”, and “EOF-reconstructed”. Dashed arrows mean each dataset is localized by Slepian basis over the study region including WAIS and AP with buffer zone of 0.5 degree. Red line in Fig. 1 enclose the WAIS and AP. Colored boxes in the bottom row are datasets compared each other in Section 3.

    Reference

    1. A, G., Wahr, J., and Zhong, S.,2013, Computations of the viscoelastic response of a 3-D compressible Earth to surface loading: an application to glacial isostatic adjustment in Antarctica and Canada, Geophysical Journal International, 192(2), 557-572, doi:.
    2. Alsdorf, D.E. , Han, S.C. , BatesP. , and MelackJ. , 2010, Seasonal water storage on the Amazon floodplain measured from satellites , Remote Sensing of Environment, 114(11), 2448-2456, doi:.
    3. Bettadpur, S. , 2012, CSR level-2 processing standards document for level-2 product release 05.
    4. Cazenave, A. , and Chen, J. , 2010, Time-variable gravity from space and present-day mass redistribution in the Earth system , Earth and Planetary Science Letters, 298, 263-274, doi:.
    5. Chen, J.L. , Wilson, C.R. , 2008, Low degree gravitational changes from GRACE, Earth rotation, geophysical models, and satellite laser ranging . Journal of Geophysical Research, 113, B06402. doi:
    6. Chen, J. L. , Wilson, C.R. , and Tapley, B.D. , 2013, Contribution of ice sheet and mountain glacier melt to recent sea level rise , Nature Geoscience, 6(7), 549-552, doi:.
    7. Cheng, M. , and Tapley, B.D. , 2004, Variations in the Earth's oblateness during the past 28 years , Journal of Geophysical Research, 109, B09402, doi:.
    8. Dunn, C , Bertiger, W. , Bar-Sever, Y. , Desai, S. , Haines, B. , Kuang, D. , Franklin, G. , Harris, I. , Kruizinga, G. , Meehan, T. , Nandi, S. , Nguyen, D. , Rogstad, T. , Thomas, J.B. , Tien, J. , Romans, L. , Watkins, M. , Wu, S.C. , Bettadpur, S. , Kim, J. , 2003, Instrument of GRACE: GPS augments gravity measurements . GPS World, 14, 16-28.
    9. Eom, J. , Seo, K.-W. , Lee, C.-K. , and Wilson, C.R. , 2017a, Correlated error reduction in GRACE data over Greenland using extended empirical orthogonal functions , Journal of Geophysical Research: Solid Earth, 122, doi:.
    10. Eom, J. , Seo, K.-W. , and RyuD. , 2017b, Estimation of Amazon River discharge based on EOF analysis of GRACE gravity data , Remote Sensing of Environment, 191, 55-66, doi:.
    11. Hannachi, A. , Jolliffe, I.T. , and StephensonD.B. , 2007, Empirical orthogonal functions and related techniques in atmospheric science: A review , International Journal of Climatology, 27(9), 1119-1152, doi:.
    12. Harig, C. , and SimonsF.J. , 2012, Mapping Greenland's mass loss in space and time , Proceding of the National Academy of Science of the United States of America, 109(49), 19934-19937, doi:.
    13. Hurrell, J.W. , and Van Loon, H. , 1997, Decadal variations in climate associated with the North Atlantic Oscillation, in Diaz, H.F. , Beniston, M. , and Bradley, R.S. (eds.), Climatic Change at High Elevation Sites, Springer Netherlands, Dordrecht, 69-94, doi:.
    14. IMBIE, 2018, Mass balance of the Antarctic Ice Sheet from 1992 to 2017 , Nature, 558(7709), 219-222, doi:.
    15. IPCC, 2014, Mitigation of climate change, Contribution of Working Group III to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change, Cambridge, UK and New York, USA.
    16. Jekeli, C. , 1981, Alternative methods to smooth the Earth's gravity field, Report No.327, Department Of Geodetic Science, Ohio State University, Ohio, USA, 28 p.
    17. Kim, B.-H. , Eom, J. , Seo, K.-W. , and Wilson, C.R. , 2016, Spurious barometric pressure acceleration in Antarctica and propagation into GRACE Antarctic mass change estimates , Geophysical Journal International, 206(2), 1306-1314, doi:.
    18. Leuliette, E.W. , and Miller, L. , 2009, Closing the sea level rise budget with altimetry, Argo, and GRACE , Geophysical Research Letters, 36, L04608, doi:.
    19. Llovel, W. , Willis, J.K. , Landerer, F.W. , and Fukumori, I. , 2014, Deep-ocean contribution to sea level and energy budget not detectable over the past decade , Nature Climate Change, 4(11), 1031-1035, doi:.
    20. Longuevergne, L. , Scanlon, B.R. , and Wilson, C.R. , 2010, GRACE hydrological estimates for small basins: evaluating processing approaches on the High Plains Aquifer, USA , Water Resources Research, 46, doi:.
    21. Madden, R.A. , and Julian, P.R. , 1971, Detection of a 40-50 day oscillation in the zonal wind in the Tropical Pacific , Journal of the Atmospheric Sciences, 28(5), 702-708, doi:.
    22. Makowski, J. K. , Chambers, D.P. , and Bonin, J.A. , 2015, Using ocean bottom pressure from the gravity recovery and climate experiment (GRACE) to estimate transport variability in the southern Indian Ocean , Journal of Geophysical Research: Oceans, 120(6), 4245-4259, doi:.
    23. Reager, J. T. , Thomas, B.F. , and Famiglietti, J.S. , 2014, River basin flood potential inferred using GRACE gravity observations at several months lead time , Nature Geoscience, 7(8), 588-592, doi:.
    24. Retzlaff, R. , and Bentley, C. , 1993, Timing of stagnation of Ice Stream C, West Antarctica, from short-pulse radar studies of buried surface crevasses . Journal of Glaciology, 39(133), 553-561. doi:
    25. Seo, K.-W. , Eom, J. , Kwon, B.-D. , 2014, Refinement of GRACE gravity model including Earth s mean mass variations , Journal of Korean Earth Science Society, 35(7), 537-542. (in Korean)
    26. Seo, K.-W. , Wilson, C. , Ted, S. , Kim, B.-M. , Waliser, D. , Tian, B. , Kim, B.-H. , and Eom, J. , 2015, Surface mass balance contributions to acceleration of Antarctic ice mass loss during 2003-2013 , Journal of Geophysical Research: Solid Earth, 120, 3617-3627, doi:.
    27. Swenson, S. , and Wahr, J. , 2006, Post-processing removal of correlated errors in GRACE data , Geophysical Research Letter, 33, L08402, doi:.
    28. Swenson, S. , Chambers, D. , and Wahr, J. , 2008, Estimating geocenter variations from a combination of GRACE and ocean model output , Journal of Geophysical Research, 113, B08410, doi:.
    29. Tapley, B.D. , Bettadpur, S. , Watkins, M. , and Reigber, C. , 2004, The gravity recovery and climate experiment: 482 Jooyoung Eom and Ki-Weon Seo Mission overview and early results , Geophysical Research Letter, 31, L09607, doi:.
    30. Velicogna, I. , and Wahr, J. , 2005, Greenland mass balance from GRACE , Geophysical Research Letter, 32, L18505, doi:.
    31. Velicogna, I. , and Wahr, J. , 2013, Time-variable gravity observations of ice sheet mass balance: Precision and limitations of the GRACE satellite data , Geophysical Research Letters, 40(12), 3055-3063, doi:.
    32. Velicogna, I. , Sutterley, T.C. , and vanden Broeke, M.R., 2014, Regional acceleration in ice mass loss from Greenland and Antarctica using GRACE time-variable gravity data , Geophysical Research Letters, 41(22), 8130-8137, doi:.
    33. Wahr, J. , Molenaar, M. , and Bryan, F. , 1998, Time variability of the Earth s gravity field: Hydrological and oceanic effects and their possible detection using GRACE , Journal of Geophysical Research, 103, 30205- 30229.
    34. Wahr, J. , Swenson, S. , Zlotnicki, V. , and Velicogna, I. , 2004, Time-variable gravity from GRACE: First results , Geophysical Research Letter, 31, L11501, doi:.
    35. Xiaolei, J. , Yunzhong, S. , and Zizhan, Z. , 2014, GRACE RL05-based ice mass changes in the typical regions of Antarctica from 2004 to 2012 , Geodesy and Geodynamics, 5(4), 57-67, doi:.