Neural network sensor fusion: Creation of a virtual sensor for cloud-base height estimation

Abstract

<p>Sensor fusion has become a significant area of signal processing research that draws on a variety of tools. Its goals are many, however in this thesis, the creation of a virtual sensor is paramount. In particular, neural networks are used to simulate the output of a LIDAR (LASER. RADAR) that measures cloud-base height. Eye-safe LIDAR is more accurate than the standard tool that would be used for such measurement; the ceilometer. The desire is to make 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, nonlinear environment. The result is an accurate prediction of cloud-base height. Thus, a virtual sensor is created. A total of four different learning algorithms were studied; two global and two local. In each case, the very best state-of-the-art learning algorithms have been selected. Local methods investigated are the regularized radial basis function network, and the support vector machine. Global methods include the standard backpropagation with momentum trained multilayer perceptron (used as a benchmark) and the multilayer perceptron trained via the Kalman filter algorithm. While accuracy is the primary concern, computational considerations potentially limit the application of several of the above techniques. Thus, in all cases care was taken to minimize computational cost. For example in the case of the support vector machine, a method of partitioning the problem in order to reduce memory requirements and make the optimization over a large data set feasible was employed and in the Kalman algorithm case, node-decoupling was used to dramatically reduce the number of operations required. Overall, the methods produced somewhat equivalent mean squared errors indicating that the descriptive capacity of the data had been reached. However, the support vector machine was the clear winner in terms of computational complexity. As well, through its ability to determine its own dimensionality it is able to relate information about the physics of the problem back to the user. This thesis, contributes to the literature on three fronts. First, it demonstrates the concept of creating of a virtual sensor via sensor fusion. Second, in the remote-sensing field where focus has typically been on pattern classification tasks, this thesis provides an in-depth look at the use of neural networks for tough regression problems. And lastly, it provides a useful tool for the meteorological community in creating the ability to add large-scale, cloud-field information to predictive models.</p>