Dynamic Weighted Ensemble with a Game Theory Perspective for Ship Classification

Abstract

The transportation industry has seen a rapid increase in ships used for inland waterway transportation. Ship classification is required to avoid inland waterway collisions. The machine learning model uses an LightGBM classifier for classification. The transfer learning (TL) model uses Inception-ResNet-v2, DenseNet121 and EfficientNetB0. A novel dynamic weighted ensemble classifier based on the game theory approach is proposed for the fusion of the outputs of the TL model and the LightGBM classifier model. Comparing the ensemble approach against individual deep learning (DL) and feature-based models, it produces excellent performance and significantly increases the accuracy of ship classification.