Artificial Neural Network and Random Forest Approaches for Modeling of Sea Surface Salinity

Ocean salinity is a key parameter in oceanic and climate studies, and the accurate estimation of sea surface salinity (SSS) of coastal water is of great scientific interest. This paper reports on a modeling study of SSS using artificial neural network (ANN) and random forest (RF) algorithm. Hong Kong Sea, China was used as case study. Sea biochemistry and sea physical parameters were collected. Sea surface temperature (SST), pH, chlorophyll-a (Chl-a) and total inorganic nitrogen (TIN) were selected as input variables of models. The assessment models were based on a back propagation (BP) neural network and RF algorithm. The results showed that an optimum BP neural network prediction model has 4-20-4-1 network architecture with gradient descent learning algorithm and an activation function including the sigmoid tangent function in the input layer, a hidden layer and linear functions in the output layer. While the optimum RF model was obtained, when RF algorithm had a mtry value of 32 with ntree=2000 and nodesize=4. Optimum BP and RF models for estimating SSS performed well at prediction, regardless of training or testing sets with R 2 above 0.8. Compared with the BP model, RF model was usually slightly stable in models’ performance with respect to different models’ parameters. This research verified that the BP model and RF algorithm could provide an effective and faithful estimation of SSS of coastal water based on sea biochemistry and physical parameters.