Dynamic Wide and Deep Neural Network for Hyperspectral Image Classification

Recently, deep learning has been successfully and widely used in hyperspectral image (HSI) classification. Considering the difficulty of acquiring HSIs, there are usually a small number pixels as training instances. Therefore, it is hard to fully use advantages networks; for example, very layers with large parameters lead overfitting. This paper proposed dynamic wide neural network (DWDNN) HSI classification, which includes multiple efficient sliding window subsampling (EWSWS) networks can grow dynamically according complexity problems. The EWSWS DWDNN was designed both direction transform kernels hidden units. These extract features from low high level, because they extended direction, learn more steadily smoothly. windows stride were reduce dimension each layer; therefore, computational load reduced. Finally, all weights only connected layer, iterative least squares method compute them easily. tested several data including Botswana, Pavia University, Salinas remote sensing datasets different numbers instances (from big). experimental results showed that had highest test accuracies compared typical machine methods such support vector (SVM), multilayer perceptron (MLP), radial basis function (RBF), recently 2D convolutional (CNN) 3D CNN