A Cluster-Stacking-Based Approach to Forecasting Seasonal Chlorophyll-a Concentration in Coastal Waters

Abstract

Prediction of Chlorophyll-a (Chl-a) concentration is significant for marine ecology and environmental protection. This paper presents an integrated approach to forecast seasonal Chl-a concentration in coastal waters. Before modeling, feature construction procedures, such as simplification, combination, and normalization, are conducted to identify the potentially significant features. The feature extraction method based on Random Forest (RF) and eXtreme Gradient BOOSTing (XGBoost) is applied to select relevant variables. Then, we propose a Cluster-stacking-based approach which includes a station-oriented clustering model and a stacking-based regression model. The former model is used to divide the observation stations into several groups, thus partitions the study region into several sub-regions and the study dataset into several subsets according to the corresponding stations. In each subset, single regression models including K-Nearest Neighbor (KNN), Support Vector Regression (SVR), Multi-Layer Perceptron regression (MLP) and XGBoost are established in level 0 space and integrated by RF in level 1 space via stacked generalization. We compare the performance of the Cluster-stacking model with that of Cluster-KNN, Cluster-SVR, Cluster-MLP, Cluster-XGBoost and the regression stacking model without cluster. The model evaluation shows that the Cluster-stacking-based approach outperforms others in forecasting Chl-a concentration with a coefficient of determination (R2) of 0.848 and a mean absolute error (MAE) of 0.665 ug/l.