Evidential Reasoning for Ship Classification: Fusion of Deep Learning Classifiers

Maritime Situational Awareness (MSA) is becoming critical in a world dominated by complex maritime civilian or military activities, diplomatic and ecological challenges for all coastal countries. Ship tracking, detection and classification through automated surveillance systems help improve the MSA. This paper proposes a new approach to ship classification from an airborne surveillance platform, such as the Aurora CP140. The proposed approach is based on the combination of several deep learning classifiers, using evidential reasoning (Dempster-Shafer theory) to better take into account the uncertainty at the last layer of the classifier. The traditional softmax layer is replaced with more adequate layers to model the uncertainty. Such layers are based on the min-max or ReLu scalings, jointly with additional modeling of the uncertainty. Results obtained from maritime observation videos are compared: the evidential fusion approach provides better classification results than the initial Bayesian classifier.