Residual prediction to improve the meteorological based sea surface temperature forecasts using ANN

Potential fishing regions (PFR) are largely dependent on local sea surface temperature (SST) pattern, to plan an optimal trip to PFR accurate SST predictions are inevitable. This study features SST prediction near Tohoku region in Japan at short (1Hr), mid-short (8Hr) and long-term (24Hr) horizons using artificial neural network (ANN) with meteorological parameters as input and observed SST as targets.

Errors using such approach shown a noisy structure in predictions due to noise in some of the meteorological inputs, which was reduced after inputs were moving averaged. Resulting SST predictions were still not satisfactory compared to the required accuracy level (0.5 to 0.8°C) needed for fishing purposes. Therefore, a residual analysis was carried out on errors which shown errors to fit as extreme value distribution. This suggested that there was still much of the information left in errors and can be extracted. Thus, a residual correction was adopted by predicting such residuals and adding them back to earlier predictions like a simple data assimilation approach. After residual corrections, errors were best fitting to the Gaussian distributions which is a near ideal case. Results were also confirmed for space invariant (buoy) and space variant (drifter) SST targets.

Errors after residual corrections were within limits of required accuracy, root mean square error in SST prediction at drifter (buoy) for 1 Hr, 8 Hr and 24 Hr lead were noted as 0.12°C (0.05°C), 0.50°C (0.23°C) and 0.84°C (0.38°C) respectively. Thus, it can be stated that residual correction is very powerful for correcting the first predictions especially for cause-effect based ANN approach. It can also be said that such residual correction can be beneficial for other oceanic parameters, as residual corrections were independent of parameter in consideration.