A Deep Learning Approach of Artificial Neural Network With Attention Mechanism to Predicting Marine Biogeochemistry Data

Abstract

Predicting marine biogeochemical data is an effective method to solve the problem of marine data-scarcity and provides data support for fundamental research in marine science. Machine learning techniques are commonly used to improve the stability and accuracy of predicting biogeochemistry data. However, current methods based on Random Forest (RF) and Artificial Neural network (ANN) often struggle to effectively capture the intricate features of ocean data, resulting in suboptimal prediction accuracy. In this study, we develop a novel deep learning method called artificial neural network with attention mechanism (ANN-att) for predicting marine biogeochemistry data. We compare and evaluate the performance of RF, ANN, and ANN-att based on two widely used ocean data sets in marine biogeochemistry: GLODAP v2.2022 and MOSAIC 2.0. Our results show that the prediction accuracy of the ANN-att method is higher than other methods by 6% for GLODAP v2.2022 and 30% for MOSAIC v.2.0. Additionally, the prediction maps of surface ocean dissolved oxygen and Δ14C in the West Pacific demonstrate that ANN-att has a significant advantage in predicting marine biogeochemistry data with stronger nonlinear characteristics.