No abstract

[1] In August 2007, three long-lived anticyclonic eddies (ACE1, ACE2, and ACE3) were detected by both satellite sea level anomaly (SLA) map and in situ measurements in the northern South China Sea (SCS). ACE3 had a two-core (ACE3(1) and ACE3(2)) structure. In situ stations along 18°N almost cut through the centers of ACE2 and ACE3 (2). Near the centers of ACE2 and ACE3(2), mean temperature and sound velocity are ∼0.65°C and ∼2 m s −1 larger than those in their surrounding areas, respectively, while mean salinity and density are ∼0.02 psu and ∼0.15 m 3 s −1 smaller than those in their surrounding areas due to downwelling near the eddy cores. The depths of maximum and minimum salinity near the eddy cores are ∼65 m and ∼35 m larger than those in their surrounding areas. The vertical depth with current speed larger than 0.05 m s −1 can reach ∼900 m. Their detailed evolutionary processes were depicted based on the variation of geostrophic currents and the trajectories of five drifting buoys. ACE1 lasted 147 days, while ACE2 and ACE3 lasted 168 days and 210 days, respectively. ACE1 had a smaller mean SLA (18.8 cm) in its lifetime than ACE2 (21.8 cm) and ACE3 (25.3 cm) but had a larger negative mean vorticity (−7.7 × 10 −6 s ). One short-lived anticyclonic eddy that split from ACE2 and another one that merged with ACE3 both had a smaller SLA, negative vorticity, and diameter than ACE2 and ACE3, respectively.

Kuroshio intrusion into the South China Sea (SCS) has different forms. In this study, a Kuroshio SCS Index (KSI) is defined using the integral of geostrophic vorticity from 118°to 121°E and from 19°to 23°N. Three typical paths (the looping path, the leaking path, and the leaping path) were identified based on the KSI derived from the weekly satellite Absolute Dynamic Topography from 1993 to 2008. The KSI has a near normal distribution. Using ±1 standard deviation (σ) as the thresholds, the leaking path is the most frequent form with the probability of occurrence at 68.2%, while the probabilities of occurrence for the looping path and the leaping path are 16.4% and 15.4%, respectively. Similar analysis is also conducted on the daily Hybrid Coordinate Ocean Model (HYCOM) Global Analysis from 2004 to 2008. The results are generally consistent with the KSI analysis of the satellite data. The HYCOM data are further analyzed to illustrate patterns of inflows/outflows and the maximum/minimum salinity as representatives of the subsurface/intermediate waters. The Kuroshio bending and the net inflow through the Luzon Strait reduce from the looping path to the leaking path to the leaping path. However, the Kuroshio subsurface water intrudes farthest into the SCS for the leaking path. Vorticity budget associated with the different intrusion types is then analyzed. The tilting of the relative vorticity, the stretching of the absolute vorticity, and the advection of planetary vorticity are important for the change of vorticity, whereas the baroclinic and frictional contributions are three orders smaller.

[1] Oceanic eddies are active and energetic southwest of Taiwan. The formation and propagation of eddies in this area were investigated using 17 year satellite altimeter data. Cyclonic eddies (CEs) and anticyclonic eddies (ACEs) often coexisted, but there were more CEs than ACEs generated during the period from October 1992 to October 2009. ACEs were stronger and, in general, lived longer than CEs. ACEs occurred more often in winter than in other seasons, while CEs were more frequent in summer. Compared with the direct local wind forcing, the Kuroshio path variability appears to be a dominant factor for eddy formation in this area. A conceptual model of an eddy-Kuroshio interaction is proposed. In summer, there exists an outflow northwest of Luzon Island, and the Kuroshio likely leaps across the Luzon Strait. To the north of the outflow and left of the Kuroshio axis, CEs are often formed, which in turn induce ACEs to the west of CEs. In winter, under the influence of northeasterly monsoon, the Kuroshio Current Loop (KCL) appears southwest of Taiwan more frequently than in other seasons, and ACEs are frequently shed from the KCL. Most of the ACEs propagate westward, and, as a result, CEs are often spun up to the east of the ACEs. The surface South China Sea outflow in summer and the KCL in winter are, however, likely related to the monsoons. Therefore, the indirect effects of monsoon winds are also evident in the seasonal variations of eddy occurrence.

Inferred from the satellite and in situ hydrographic data from the 1990s and 2000s, the Kuroshio intrusion into the South China Sea (SCS) had a weakening trend over the past two decades. Associated with the weakened Kuroshio intrusion, the Kuroshio loop and eddy activity southwest of Taiwan became weaker, whereas the water above the salinity minimum became less saline in the northern SCS. The sea surface height southwest of Taiwan increased at a slower rate compared to other regions of the SCS because of the weakened Kuroshio intrusion. Simulations using the Regional Ocean Modeling System (ROMS) Pacific model show that the strength of the Kuroshio intrusion into the SCS decreased from 1993 to 2010 with a negative trend, −0.24 sverdrups (Sv) yr−1 (1 Sv ≡ 106 m3 s−1), in the total Luzon Strait transport (LST). Although wind-induced Ekman transport through the Luzon Strait became weaker, the magnitude at 0.001 Sv yr−1 was too small to compensate for the negative trend of the LST. On the other hand, the piling up of the water induced by monsoon winds was an important mechanism for changing the pressure gradient across the Luzon Strait and eventually affecting the LST. The sea level gradient between the western Pacific and the SCS had a negative trend, −0.10 cm yr−1, corresponding to a negative trend in the geostrophic transport at −0.20 Sv yr−1. The Kuroshio transport east of Luzon Island also had a negative trend, which might also be linked to the weakening Kuroshio intrusion.

Ocean salinity changes play an important role in modulating ocean and climate variability.Analyses of the repeating observations along PN, TK, and 137 E sections reveal that both surface and subsurface salinity in the Northwest Pacific Subtropical Gyre (NPSG) had clear quasi-decadal oscillation (QDO) of $10 year and a sustained freshening trend during 1987-2012. Surface salinity in the NPSG troughed in 1989, 1999, and 2008, and peaked in 1993, and 2003. The peak-to-tough range of surface salinity oscillation can reach 0.3 psu. Meanwhile, surface salinity decreased about 0.10 psu from 1987 to 2012 with a freshening trend of 20.0042 psu yr 21 . These surface salinity anomalies were subducted into the subsurface layer mainly in the ventilated zone along the Kuroshio Extension, and advected over the NPSG. The QDO of the subsurface salinity maximum (S max ) lagged that of the surface salinity by about 1$2 years. Both the peakto-tough range of the S max oscillation (0.15 psu) and its freshening trend (20.0036 psu yr 21) are smaller than those of the surface salinity. Salinity changes in the NPSG likely began in the mid-1970s associated with the North Pacific regime shift during 1976/1977. Analyses of mixed layer salinity budget indicated that air-sea freshwater flux change in the NPSG is the leading factor controlling the surface salinity anomalies, while change of large-scale ocean circulation (Geostrophic advection) also plays a minor role. Salinity and air-sea freshwater flux changes in the NPSG are all closely related to the Pacific Decadal Oscillation, indicating the large-scale ocean-atmosphere interaction.

No abstract

Recent neural language generation systems often hallucinate contents (i.e., producing irrelevant or contradicted facts), especially when trained on loosely corresponding pairs of the input structure and text. To mitigate this issue, we propose to integrate a language understanding module for data refinement with selftraining iterations to effectively induce strong equivalence between the input data and the paired text. Experiments on the E2E challenge dataset show that our proposed framework can reduce more than 50% relative unaligned noise from the original data-text pairs. A vanilla sequence-to-sequence neural NLG model trained on the refined data has improved on content correctness compared with the current state-of-the-art ensemble generator. * Contribution during internship at Microsoft.