Figure 1.
GULFSTREAM and OSMOSIS domains.
Figure 1.
GULFSTREAM and OSMOSIS domains.
Figure 2.
One () and 11 day () accumulated along-track nadir and wide-swath SSH pseudo-observations (meters) on 4 August 2013 (a,b) and 5 August 2013 (c,d). (a) Nadir (d = 0), (b) nadir (d = 0) + swot, (c) nadir (d = 5), (d) nadir (d = 5) + swot.
Figure 2.
One () and 11 day () accumulated along-track nadir and wide-swath SSH pseudo-observations (meters) on 4 August 2013 (a,b) and 5 August 2013 (c,d). (a) Nadir (d = 0), (b) nadir (d = 0) + swot, (c) nadir (d = 5), (d) nadir (d = 5) + swot.
Figure 3.
Variance of the observation error as a function of the hourly lag between the observations and the day to estimation time . Blue barplots are the conditional distributions of according to hourly time lag; dotted red lines are their variances and a solid black line is the corresponding parametric fit.
Figure 3.
Variance of the observation error as a function of the hourly lag between the observations and the day to estimation time . Blue barplots are the conditional distributions of according to hourly time lag; dotted red lines are their variances and a solid black line is the corresponding parametric fit.
Figure 4.
Sketch of the iterative fixed-point algorithm.
Figure 4.
Sketch of the iterative fixed-point algorithm.
Figure 5.
Daily spatial nRMSE computed over the four 20-day non-continuous validation periodss for the six supervised/unsupervised FP-GENN configurations. The spatial coverage of 11 days () accumulated along-track nadir (a) expanded with wide-swath SWOT data (b) is provided by the red barplot. (a) Nadir, (b) nadir+swot.
Figure 5.
Daily spatial nRMSE computed over the four 20-day non-continuous validation periodss for the six supervised/unsupervised FP-GENN configurations. The spatial coverage of 11 days () accumulated along-track nadir (a) expanded with wide-swath SWOT data (b) is provided by the red barplot. (a) Nadir, (b) nadir+swot.
Figure 6.
Daily spatial nRMSE computed for the four 20-day non-continuous validation periodss for OI, (post-)AnDA, VE-DINEOF, FP-ConvAE and FP-GENN. The spatial coverage of one day () accumulated along-track nadir and that of wide-swath SWOT data are respectively provided by the red and green barplots. (a) Nadir, (b) nadir+swot.
Figure 6.
Daily spatial nRMSE computed for the four 20-day non-continuous validation periodss for OI, (post-)AnDA, VE-DINEOF, FP-ConvAE and FP-GENN. The spatial coverage of one day () accumulated along-track nadir and that of wide-swath SWOT data are respectively provided by the red and green barplots. (a) Nadir, (b) nadir+swot.
Figure 7.
Taylor diagram and signal-to-noise ratio computed for the four 20-day non-continuous validation periods for OI, (post-)AnDA, VE-DINEOF, FP-ConvAE and FP-GENN computed for both nadir use only and joint assimilation/learning with wide-swath SWOT data. (a) Taylor diagram, (b) signal-to-noise ratio.
Figure 7.
Taylor diagram and signal-to-noise ratio computed for the four 20-day non-continuous validation periods for OI, (post-)AnDA, VE-DINEOF, FP-ConvAE and FP-GENN computed for both nadir use only and joint assimilation/learning with wide-swath SWOT data. (a) Taylor diagram, (b) signal-to-noise ratio.
Figure 8.
Global SSH gradient field reconstruction obtained by OI, (post-)AnDA, VE-DINEOF, FP-ConvAE and FP-GENN using along-track nadir data only. (a) Ground truth (), (b) OI, (c) post-AnDA, (d) VE-DINEOF, (e) FP-ConvAE, (f) FP-GENN.
Figure 8.
Global SSH gradient field reconstruction obtained by OI, (post-)AnDA, VE-DINEOF, FP-ConvAE and FP-GENN using along-track nadir data only. (a) Ground truth (), (b) OI, (c) post-AnDA, (d) VE-DINEOF, (e) FP-ConvAE, (f) FP-GENN.
Figure 9.
Global SSH gradient field reconstruction obtained by OI, (post-)AnDA, VE-DINEOF, FP-ConvAE and FP-GENN for a joint assimilation/learning of along-track nadir with wide-swath SWOT data. (a) Ground truth (), (b) OI, (c) post-AnDA, (d) VE-DINEOF, (e) FP-ConvAE, (f) FP-GENN.
Figure 9.
Global SSH gradient field reconstruction obtained by OI, (post-)AnDA, VE-DINEOF, FP-ConvAE and FP-GENN for a joint assimilation/learning of along-track nadir with wide-swath SWOT data. (a) Ground truth (), (b) OI, (c) post-AnDA, (d) VE-DINEOF, (e) FP-ConvAE, (f) FP-GENN.
Figure 10.
One () and 11 day () accumulated along-track nadir and wide-swath SSH pseudo-observations (meters) on 4 August 2013 (a,b) and 5 August 2013 (c,d). (a) Nadir (d = 0), (b) nadir (d = 0) + swot, (c) nadir (d = 5), (d) nadir (d = 5) + swot.
Figure 10.
One () and 11 day () accumulated along-track nadir and wide-swath SSH pseudo-observations (meters) on 4 August 2013 (a,b) and 5 August 2013 (c,d). (a) Nadir (d = 0), (b) nadir (d = 0) + swot, (c) nadir (d = 5), (d) nadir (d = 5) + swot.
Figure 11.
Daily spatial nRMSE computed for the four 20-day non-continuous validation periods for the six supervised/unsupervised FP-GENN configurations. The spatial coverage of 11 days () accumulated along-track nadir (a) expanded with wide-swath SWOT data, (b) is provided by the red barplot. (a) Nadir, (b) nadir+swot.
Figure 11.
Daily spatial nRMSE computed for the four 20-day non-continuous validation periods for the six supervised/unsupervised FP-GENN configurations. The spatial coverage of 11 days () accumulated along-track nadir (a) expanded with wide-swath SWOT data, (b) is provided by the red barplot. (a) Nadir, (b) nadir+swot.
Figure 12.
Daily spatial nRMSE computed for the four 20-day non-continuous validation periods for OI, (post-)AnDA, VE-DINEOF, FP-ConvAE and FP-GENN. The spatial coverage of one day () accumulated along-track nadir and that of wide-swath SWOT data are respectively provided by the red and green barplots. (a) Nadir, (b) nadir+swot.
Figure 12.
Daily spatial nRMSE computed for the four 20-day non-continuous validation periods for OI, (post-)AnDA, VE-DINEOF, FP-ConvAE and FP-GENN. The spatial coverage of one day () accumulated along-track nadir and that of wide-swath SWOT data are respectively provided by the red and green barplots. (a) Nadir, (b) nadir+swot.
Figure 13.
Taylor diagram and signal-to-noise ratio computed for the four 20-day non-continuous validation periods for OI, (post-)AnDA, VE-DINEOF, FP-ConvAE and FP-GENN computed for both nadir use only and joint assimilation/learning with wide-swath SWOT data. (a) Taylor diagram, (b) signal-to-noise ratio.
Figure 13.
Taylor diagram and signal-to-noise ratio computed for the four 20-day non-continuous validation periods for OI, (post-)AnDA, VE-DINEOF, FP-ConvAE and FP-GENN computed for both nadir use only and joint assimilation/learning with wide-swath SWOT data. (a) Taylor diagram, (b) signal-to-noise ratio.
Figure 14.
Global SSH gradient field reconstruction obtained by OI, (post-)AnDA, VE-DINEOF, FP-ConvAE and FP-GENN using along-track nadir data only. (a) Ground truth (), (b) OI, (c) post-AnDA, (d) VE-DINEOF, (e) FP-ConvAE, (f) FP-GENN.
Figure 14.
Global SSH gradient field reconstruction obtained by OI, (post-)AnDA, VE-DINEOF, FP-ConvAE and FP-GENN using along-track nadir data only. (a) Ground truth (), (b) OI, (c) post-AnDA, (d) VE-DINEOF, (e) FP-ConvAE, (f) FP-GENN.
Figure 15.
Global SSH gradient field reconstruction obtained by OI, (post-)AnDA, VE-DINEOF, FP-ConvAE and FP-GENN for a joint assimilation/learning of along-track nadir with wide-swath SWOT data. (a) Ground truth (), (b) OI, (c) post-AnDA, (d) VE-DINEOF, (e) FP-ConvAE, (f) FP-GENN.
Figure 15.
Global SSH gradient field reconstruction obtained by OI, (post-)AnDA, VE-DINEOF, FP-ConvAE and FP-GENN for a joint assimilation/learning of along-track nadir with wide-swath SWOT data. (a) Ground truth (), (b) OI, (c) post-AnDA, (d) VE-DINEOF, (e) FP-ConvAE, (f) FP-GENN.
Table 1.
Temporal and spectral statistics used to assess the performances of the interpolators in the observation system simulation experiment.
Table 1.
Temporal and spectral statistics used to assess the performances of the interpolators in the observation system simulation experiment.
| Name | Formula |
---|
| nRMSE | nRMSE() = |
| Error variance | = |
| Correlation | COR() = |
Temporal domain | Reconstruction score | R-score = |
| Interpolation score | I-score = |
| Auto-encoder score | AE-score = |
Spectral domain | RAPS | RAPS() = |
Signal-to-Noise Ratio | SNR() = |
Table 2.
Specifications of GENN learning-based strategies.
Table 2.
Specifications of GENN learning-based strategies.
Configurations | Data |
---|
| | Observations | Gap-Free Maps | DUACS OI |
---|
Supervised 1 | Input | | 🗸 | yes/no |
Target | | 🗸 | |
Supervised 2 | Input | 🗸 | | yes/no |
Target | | 🗸 | |
Unsupervised | Input | 🗸 | | yes/no |
Target | 🗸 | | |
Table 3.
Sea surface height (SSH) and SSH gradient field R/I/AE-scores computed for the four 20-day non-continuous validation periods for OI, (post-)AnDA, VE-DINEOF, FP-ConvAE and FP-GENN for both nadir use only and joint assimilation/learning with wide-swath SWOT data.
Table 3.
Sea surface height (SSH) and SSH gradient field R/I/AE-scores computed for the four 20-day non-continuous validation periods for OI, (post-)AnDA, VE-DINEOF, FP-ConvAE and FP-GENN for both nadir use only and joint assimilation/learning with wide-swath SWOT data.
| Model Type | R-Score | I-Score | AE-Score | | Model Type | R-Score | I-Score | AE-Score |
---|
nadir | OI | 87.32 | 72.17 | _ | nadir | | 78.03 | 75.97 | _ |
AnDA | 94.85 | 77.91 | _ | | 85.56 | 79.14 | _ |
VE-DINEOF | 96.11 | 72.72 | _ | | 82.69 | 75.61 | _ |
FP-ConvAE | 87.82 | 76.32 | 82.85 | | 77.80 | 76.81 | 75.89 |
FP-GENN | 91.78 | 84.56 | 93.15 | | 81.05 | 80.56 | 84.24 |
nadir + SWOT | OI | 93.25 | 74.25 | _ | nadir + SWOT | | 73.83 | 75.78 |
AnDA | 96.05 | 83.55 | _ | | 89.89 | 82.88 | _ |
VE-DINEOF | 97.13 | 75.28 | _ | | 88.19 | 76.69 | _ |
FP-ConvAE | 80.63 | 77.51 | 83.26 | | 76.20 | 76.49 | 75.84 |
FP-GENN | 96.49 | 90.13 | 95.58 | | 86.96 | 85.33 | 88.23 |
Table 4.
SSH and SSH gradient field R/I/AE-scores computed for the four 20-day non-continuous validation periods for OI, (post-)AnDA, VE-DINEOF, FP-ConvAE and FP-GENN for both nadir use only and joint assimilation/learning with wide-swath SWOT data.
Table 4.
SSH and SSH gradient field R/I/AE-scores computed for the four 20-day non-continuous validation periods for OI, (post-)AnDA, VE-DINEOF, FP-ConvAE and FP-GENN for both nadir use only and joint assimilation/learning with wide-swath SWOT data.
| Model Type | R-Score | I-Score | AE-Score | | Model Type | R-Score | I-Score | AE-Score |
---|
nadir | OI | 42.05 | 32.11 | _ | nadir | | 48.83 | 47.57 | _ |
AnDA | 58.85 | 47.02 | _ | | 58.78 | 55.17 | _ |
VE-DINEOF | 26.29 | 30.61 | _ | | 33.11 | 35.28 | _ |
FP-ConvAE | 37.20 | 31.67 | 47.77 | | 32.15 | 35.87 | 41.24 |
FP-GENN | 67.94 | 62.52 | 80.40 | | 50.53 | 52.12 | 60.41 |
nadir + SWOT | OI | 54.21 | 47.75 | _ | nadir + SWOT | | 36.83 | 47.30 | _ |
AnDA | 81.15 | 70.91 | _ | | 72.35 | 67.59 | _ |
VE-DINEOF | 69.08 | 32.98 | _ | | 22.08 | 24.90 | _ |
FP-ConvAE | 45.15 | 42.70 | 47.93 | | 38.22 | 43.13 | 42.03 |
FP-GENN | 77.16 | 69.56 | 83.08 | | 56.29 | 59.21 | 67.69 |