Feature selection of spectral dimension by hyperspectral remote sensing images based on genetic algorithm and support vector machine

An algorithm is presented for deriving an optimal features classified with a support vector machine. The approach is based on direct objective optimization which is approximated by the selection of appropriate features as the SVM learning predictor in a regularized learning framework. To process the regularized learning, a genetic method provides a learning rule for in an outer loop of an iteration, while at each iteration training predictor model using gradient descent is to gradually added the feature into improving the existing model. The inner loop is heuristic to perform support vector machine training and provide support vector coefficients on which the gradient descent depends. The experiment was conduced on the Airborne Visible/Infrared Imaging Spectrometer(AVIRIS) data for classification. The result shows that the feature selection of spectral dimension and support vector machine are jointly optimized.