Automatic Selection of Initial Parameters Based on a Modified Convolutional 1 Neural Network to Improve the Accuracy of Hyperspectral Image Classification

Abstract:

In artificial intelligence, deep learning is a very advanced recognition method, which in recent years has recognized a great improvement, especially by applying convolutional neural networks. These networks are applied in several research domains around the world, in recognition of shapes and objects, identifying identities (in Facebook), predictive search (in Google), classification of large satellite  images, etc. Despite all the progress, there are still some changes to make. In this paper, we are interested in the classification of hyperspectral  imaging by Convolutional Neural Networks. First, we proposed a  parameters initialization method using a clustering algorithm (CKmeans). Second, a change in the pooling layer is made. The results of the experiments, on three different hyperspectral images, showed an increase in the accuracy of the classification.