Abstract
The conventional approach for principal component analysis (PCA) and its variants applied to remote sensing involves the computation of the input data covariance/correlation matrix and/or that of noise and application of diagonalization procedures for extracting the eigenvalues and corresponding eigenvectors. When the data dimension grows significantly, the matrix computations and manipulations become practically inefficient and inaccurate due to round‐off errors. In addition, all the eigenvalues and their corresponding eigenvectors have to be evaluated. These deficiencies make the conventional scheme inefficient for remote sensing applications. For that we propose here a neural network model that performs the PCA and its variants directly from the original data without any additional non‐neuronal computations or preliminary matrix estimation. Since the end user is usually not a neural network specialist, the neural network model as well as its execution are carefully designed in order to be automated as much as possible. This includes both the design of the network topology and the input/output representation as well as the design of the training algorithms. The global convergence of the model is studied. Its application has been realized on Landsat Thematic Mapper (TM) multispectral data. The obtained results show that the model performs well.