Detection of river ice using relevance vector machine

Sparse kernel methods are very efficient in classification problems and offer advantages such as their capacity to find sparser and probabilistic solutions. This paper presents a river ice detection method based on relevance vector machine (RVM). We investigated how the kernel type and the kernel parameter influence ice detection accuracy and the number of relevant vectors. In addition, experiments were conducted with a varying size of training sets. Accuracies are compared with regular SVM. Experimental results clearly demonstrate that slightly higher detection accuracy is obtained using the RVM-based approach with a significantly smaller relevance vector rate, and, therefore, much faster testing time compared with an SVM-based approach. The RBF kernel approach is more suitable for river ice detection, which requires low complexity and stability for real-time river ice detection.