Estimation of Spatiotemporal Solar Data using Artificial Neural Networks and Gradient Pattern Analysis

M. Conceição de Andreade, R.R. Rosa, A. Rios Neto, H.S. Sawant, F.C.R. Fernandes, and R.A. Syck (Bra


: neural networks, spatiotemporal estimation, gradient pattern analysis, phenomenological prediction, solar physics.


In this paper we intend to evaluate the use of Multilayer Perceptron neural networks for a spatiotemporal patterns estimation problem and to compare the performance of the Kalman Filtering with the Backpropagation and the Levenberg-Marquardt training algorithms. The study consists of applying Multilayer Perceptron for estimation of the solar active regions evolution using sequential soft X-ray images observed by the YOHKOH solar satellite telescopes. In this application, the performance test is done by using the mean squared error, image visualization and the Gradient Pattern Analysis (GPA) techniques. The last one is based on the operator for characterization of Amplitude Asymmetric Fragmentation (AAF). The AAF operator is being used for the first time for a performance test with an Artificial Neural Network (ANN) applied in spatiotemporal patterns estimation. The results confirm the efficiency and efficacy of ANN as a tool to estimate spatiotemporal patterns in this kind of application. The tests indicate that although the Kalman Filtering showed an efficacy to learn the patterns comparable to those of the Backpropagation and the Levenberg-Marquardt algorithms, it is inefficient from the computational viewpoint in the sense that it takes a longer processing time for real-time applications. In addition, the ANN performance validation tests confirm the utility of the AAF operator for the performance characterization of spatiotemporal patterns estimation algorithms.

Important Links:

Go Back