Application of MLP and RBF networks to cloud detection

This paper compares the performances of multi-layer perceptrons (MLPs) and radial basis function (RBF) networks on detecting clouds in NOAA/AVHRR images. The main results show that the RBF networks are able to handle complex atmospheric and oceanographic phenomena while the conventional rule-based systems and MLPs can not. In particular, the experimental evaluations show that the RBF networks can converge to global minima while the MLPs can only achieve this occasionally, and that classification errors made by the RBF networks decrease dramatically when the number of basis functions increases. In addition, these errors are almost identical when the number of basis functions reaches a threshold. Only on a few rare occasions when the backpropagation algorithm attains an optimal solution and the classification errors made by the MLPs be comparable to (but still larger than) the ones made by the RBF networks. However, the results show that achieving such optimal solutions is difficult. It is, therefore, concluded that the RBF networks are better than the MLPs for cloud detection.