Evaluations of the machine learning models for the coastal habitat classification

Abstract

A ‘short-neck clam’ (Ruditapes philippinarum) is one of the most important commercial shellfish. The amount of the shellfish production has been severely reduced due to the unexpected invasion of the ‘Japanese mud shrimp’ (Upogebia major) in some Korean tidal flats. Thus, it is highly required to know the habitat suitability for both organisms. In this study, the diverse simulations of the habitat classification were carried out using the available habitat data of the U.major and R. philippinarum. Supervised learning methods such as decision tree, k-Nearest Neighbor (kNN), Support Vector Machine (SVM) and Artificial Neural Network (ANN) were used with the three optimal clusters defined by R package ‘NbClust’. The decision trees were applied ‘bagging’ and ‘adaboost’ algorithms. Based on the simulation results, the prediction accuracies of each model in case of using the test data are estimated to be about 55-65%. This is considered to be due to outlier effects, and the overfitting problem due to the relatively small number of samples. In many biological data, these are still challenging problems.