Abstract:

Convolution Neural Network has shown good results, in recent years, in recognition and big data imagery processing. Its strength is accumulated in the deep extraction of information. The great information’ development make human unable to handle all cases manually. Thus, CNN input requires some manual parameters initialization, among which, the kernels number. In this paper, we present a solution to select the input parameters automatically, specifically the kernels number and positions choices. Indeed, our method is done in four main steps, which are: the spectral bands' extraction, the features maps creation, and the CNN layers treatment. The experiment on three different hyperspectral datasets gave excellent results. The comparisons of our method with other similar ones show that our method is the best.