Evaluate Machine Learning Models Used for Upscaling Surface Ocean CO2 Measurements

Abstract

Upscaling measurements from ground-based sites or underway monitoring to a regional or global scale provides important information to policy makers for environmental management and to researchers looking for a better understanding of relevant issues. We used the reconstruction of global surface ocean CO2 as an example to evaluate the performance of four machine learning models: Random Forest (RF), Support Vector Machine (SVM), Feedforward Neural Network (FNN), and Self-Organization Map (SOM). The results show the performance from high to low as RF, SVM, FNN, and SOM. However, the overall differences of modelled CO2 among the four models are insignificant. Considering the discrete characteristics of RF, it is recommended to use SVM or FNN when the number of data point is not large and continuous estimations are expected. RF has an advantage particularly when the number of data points is very large and the data include categorial variables.