Artificial Neural Network Estimation of Sea Surface Temperature from Oceanographic Variables

Abstract

Sea surface temperature (SST) sits at the centre of the climate system, but estimating it accurately is hard because the oceanographic processes that drive it interact in complex, non-linear ways. In this study we build and test a feedforward artificial neural network (ANN) that estimates SST from three physically meaningful inputs, sea surface salinity and the eastward and northward components of ocean current velocity, using monthly reanalysis fields from the Copernicus Marine Service (January 2024 to October 2025, 22 months). The predictors and the target are contemporaneous monthly means, so the model estimates the SST field rather than forecasting it at a lead time. We varied the number of hidden neurons from 10 to 100 in steps of ten, trained each configuration from 30 random initializations with the Levenberg–Marquardt algorithm, and evaluated it on a temporally held-out test period (the final six months) using four metrics computed on the unscaled values in degrees Celsius: mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE). In-sample, every network of 30 neurons or more interpolated the training months almost perfectly, so in-sample accuracy carried no information about generalization. On the held-out months, test RMSE fell from 1.467 °C at 10 neurons to 1.162 °C at 100 neurons, but all configurations beyond 30 neurons performed within one standard deviation of one another: extra width bought fit to the training data, not generalization. The network clearly beat the training-mean climatology (test RMSE 1.892 °C) and a linear regression on the same predictors (1.811 °C), confirming that salinity and velocity fields carry genuine SST information, although persistence from the previous month (0.869 °C) remained stronger on this seasonal series. The study's contribution is a protocol-correct architecture evaluation, held-out testing, repeated initializations, and baseline comparison, and its clearest lesson is that single-run in-sample sweeps can suggest a capacity–performance story that does not survive validation.