A weighting function approach for neural network nonlinear time series analysis of satellite remote sensing of rainstorms

Abstract

One of frequently used neural networks, i.e., a radial-based function network (RBFN) with Gaussian activation functions is employed to study the nonlinear time series by carrying out the characterization experiments for a GMS- 5 satellite 11 µm IR observations of rainstorm process. The proposed methodology mainly uses RBFN to approximate the nonlinear time series signal first: then the characteristics of its weighting functions changed with time are analyzed. The difficulty due to the effects of high noise on the signal processing using neural networks is addressed. Thus, finally a more integrated method combining the neural network analysis with wavelet packet decomposition is introduced. The preliminary results show that the proposed approach for nonlinear time series analysis is efficient and promising.