Estimation of the barrier layer thickness in the Indian Ocean based on hybrid neural network model

Abstract

Accurately and effectively estimating of the barrier layer thickness (BLT) is essential for the research of ocean thermodynamics, ocean dynamics, and air-sea interaction. Artificial intelligence model provides an effective means for accurately estimating BLT from sea surface and gridded Argo data. The present study focuses on the application of a hybrid particle swarm optimization-based artificial neural networks model (PSO-ANN) for estimating the BLT in the Indian Ocean. The input variables of the hybrid model include sea surface temperature (SST), sea surface salinity (SSS), sea surface height (SSH) and precipitation (P), and the output variable is the BLT value. Multivariate satellite and gridded Argo data collected from the Indian Ocean between January 2012 and December 2019 (i.e., a database consists 239,568 training datasets and 34,224 testing datasets) are provided to prepare the training and testing datasets for the model. The parameters of ANN model, such as network parameter, network weights, and dropout rates are optimized using the PSO algorithm to achieve the best estimation model. R-squared (R2) and root mean square error (RMSE) are used to evaluate the performance of the model. Two groups of comparative experiments (Case 1 and Case 2) on the performance of the PSO-ANN model demonstrate that the model in Case 2 can better capture the complex features of the BLT in the ocean region. The performance of the PSO-ANN model in Case 2 is further compared with the data-driven estimation models such as the traditional ANN model and the known multilinear regression model (MRM), as well as the CESM2-WACCM dynamic model from CMIP6. The comparison results show that the dynamic model has the worst performance among the four models. Moreover, the annual average RMSE value for the PSO-ANN model is 1.83 m, which is 12% and 84% lower than that of the traditional ANN and MRM, respectively. The R2 value for the model of 0.85 is improved by 4% and 40% compared to that of two models. Furthermore, three regions with significant seasonal fluctuations of the BLT in the Indian Ocean are selected to further evaluate the estimation accuracy of the hybrid model in 2019; the Southeast Arabian Sea (SEAS), the Bay of Bengal (BoB), and the Eastern Equatorial Indian Ocean (EEIO). As a result, the hybrid model is capable of reflecting seasonal variation trends in these regions, but there is room for improvement in the estimation accuracy. In addition, the correlation analysis between BLT and sea surface parameters indicates that there exist significant correlations between the BLT and SSS, P. The results of this study show that the proposed hybrid model can be used to estimate and analyze BLT in certain regions with complex ocean dynamics processes. Moreover, the model can be extended to estimate other key ocean parameters and provide effective technical support for studying the internal ocean parameters.