AN MLPEL machine learning model for bathymetry retrieval based on ensemble learning

Machine learning models have made rapid progress in recent years. As a feedforward neural network, the multilayer perceptron (MLP) is widely used in bathymetry because of its simple structure and good nonlinear fitting ability. However, the classical multilayer perceptron method is faced with a potential problem, which is easy to fall into local minima and overfitting, resulting in failure of model training. In order to solve the problems of overfitting and weak ability to capture complex environments of a single model, this paper propose a multilayer perceptron as the meta -model Ensemble Learning Model (MLPEL) In this model, CatBoost, MLP and Random Forest (RF) are selected for the base model to make preliminary prediction and generate new dataset. MLP model is selected as the meta -model to receive new dataset and predict water depth. To validate the feasibility of the model, we conducted experiments in two study areas, Ganquan Island and Vieques. In Ganquan island, the MSE of MLPEL model is 1.30 m 2 , which is 0.6–2.4 m 2 lower than that of Log-Ratio (LR) model and MLP model. In Vieques island, the MSE of the MLPEL model is 1.54 m 2 , which is 0.5–1.5 m 2 lower than the LR model and the MLP model. The results show that this method can solve the local minimum problem of multi-layer perceptron model well, and obtain the water depth value with high robustness and accuracy.