Neural networks for sensor fusion in remote sensing

Cloud base height is a continuous variable that falls within the range of zero to fourteen kilometers and is useful for understanding the Earth's radiation budget. Advances in LIDAR (laser radar) technology have provided accurate cloud base height measurements. However, new sensor development and deployment are costly processes. The paper is motivated by a desire to make LIDAR output of cloud base height information available at a network of ground based meteorological stations without actually installing LIDAR sensors. To accomplish this, fifty-seven sensors ranging from multispectral satellite information to standard atmospheric measurements such as temperature and humidity, are fused in what can only be termed as a very complex, non-linear environment. The result is an accurate prediction of cloud base height. Thus, a virtual sensor is created. This fusion is performed via neural network architectures. More specifically the choices of learning algorithms reflect the state-of-the-art in neural network design and include; as local methods, the regularized radial basis function (RBF) network and the support vector machine (SVM). Global methods include the node decoupled extended Kalman filter trained multi-layer perception (NDEKF-MLP), and as a benchmark, the venerable backpropagation algorithm. Overall, the support vector machine has shown itself to be the method of choice especially when complexity was considered, Excessive storage requirements occurred in the RBF case and the global methods required large committee machines to overcome the effects of local minima.