Predicting Temperature–Salinity Profiles at Unobserved Stations and Times in a Glacier-Fed Antarctic Bay: Neighbour Information, Physical Constraints and Leakage-Aware Validation

Campaign hydrography in glacier-fed bays is sparse, making prediction at unsampled stations and dates difficult. We evaluated spatial and temporal transfer and physical constraints using 1,982 quality-controlled YSI EXO profiles from 24 Admiralty Bay stations, Antarctica (2018–2022). A graph-attention model with GRU-D-style temporal decay predicted temperature and salinity anomalies; a residual multilayer perceptron added static-stability and temperature–salinity (T–S) hull penalties. We compared five validation schemes with climatology, a spatial generalised additive model (GAM), gradient boosting, inverse-distance weighting (IDW), and persistence. In leave-one-site-out tests with same-day neighbours, IDW achieved a temperature root-mean-square error (RMSE) of 0.110 °C, versus 0.249 °C for climatology. In buffered leave-one-day-out tests (±7-day buffer), the lagged graph model achieved skill scores of 0.45 for temperature and 0.32 for salinity; graph parameters were fixed at leave-one-site-out medians. The apparent advantage of gradient boosting under station withholding disappeared when same-day data were excluded (temperature skill = −0.14), revealing campaign-identifier leakage. On a 12-station subset, increasing the stability-penalty weight from 0 to 3 reduced the share of profiles with density inversions from 37% to 18%, close to the base model’s 20%, while increasing temperature RMSE by 0.004 °C. A summer T–S hull excluded 56% of same-day IDW predictions for June–November. Conformal coverage was near nominal across stations but deteriorated under seasonal shift for several models; same-day IDW remained near nominal. The stability penalty improved plausibility relative to the unconstrained residual model but not accuracy; the hull penalty had little effect on leave-one-site-out accuracy. Recent neighbouring profiles contained most of the transferable information.

Authors

Institutions

Publication Details

Journal
World Scientific News
Published
2026-11-19
DOI
https://doi.org/10.65770/mldt4430
Primary Topic
Oceanographic and Atmospheric Processes
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Predicting Temperature–Salinity Profiles at Unobserved Stations and Times in a Glacier-Fed Antarctic Bay: Neighbour Information, Physical Constraints and Leakage-Aware Validation

Love Cherukei Nnoka, Chuck Howard
World Scientific News
Oceanographic and Atmospheric Processes
article

Predicting Temperature–Salinity Profiles at Unobserved Stations and Times in a Glacier-Fed Antarctic Bay: Neighbour Information, Physical Constraints and Leakage-Aware Validation

Love Cherukei Nnoka, Chuck Howard
article en

Abstract

Campaign hydrography in glacier-fed bays is sparse, making prediction at unsampled stations and dates difficult. We evaluated spatial and temporal transfer and physical constraints using 1,982 quality-controlled YSI EXO profiles from 24 Admiralty Bay stations, Antarctica (2018–2022). A graph-attention model with GRU-D-style temporal decay predicted temperature and salinity anomalies; a residual multilayer perceptron added static-stability and temperature–salinity (T–S) hull penalties. We compared five validation schemes with climatology, a spatial generalised additive model (GAM), gradient boosting, inverse-distance weighting (IDW), and persistence. In leave-one-site-out tests with same-day neighbours, IDW achieved a temperature root-mean-square error (RMSE) of 0.110 °C, versus 0.249 °C for climatology. In buffered leave-one-day-out tests (±7-day buffer), the lagged graph model achieved skill scores of 0.45 for temperature and 0.32 for salinity; graph parameters were fixed at leave-one-site-out medians. The apparent advantage of gradient boosting under station withholding disappeared when same-day data were excluded (temperature skill = −0.14), revealing campaign-identifier leakage. On a 12-station subset, increasing the stability-penalty weight from 0 to 3 reduced the share of profiles with density inversions from 37% to 18%, close to the base model’s 20%, while increasing temperature RMSE by 0.004 °C. A summer T–S hull excluded 56% of same-day IDW predictions for June–November. Conformal coverage was near nominal across stations but deteriorated under seasonal shift for several models; same-day IDW remained near nominal. The stability penalty improved plausibility relative to the unconstrained residual model but not accuracy; the hull penalty had little effect on leave-one-site-out accuracy. Recent neighbouring profiles contained most of the transferable information.

World Scientific News
University of Africa Toru-Orua (NG), Rivers State University (NG)
Oceanographic and Atmospheric Processes
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon