Bayesian weighted Dynamic Distribution of Samples-based Remote Sensing Feature Selection Method

Feature selection is basic and important research for extracting valuable information from high-dimensional remote sensing features. However, high-dimensional data contains redundancy and noise that weakens the recognition results. Classical feature selection methods rely on feature discrimination measures and ignore the risk of subsequent recognition errors and waste feature discrimination performance by statically evaluating each feature using all samples. This paper proposes a Bayesian weighted Dynamic Distribution of Samples Feature Selection (DDSFS) algorithm using the Bayesian risk criterion. The algorithm is a forward search that uses the Parzen window to estimate the mutual information between features and categories. Then, the sample distribution is dynamically adjusted according to the recognized risk until the feature subset is selected. Using remote sensing images collected from Google Earth, the experimental results illustrate that the DDSFS has better performance than other typical feature selection algorithms for different classifiers, which illustrates the superiority of the approach.