Hyperspectral imagery has seen a great evolution in recent years. Consequently, several fields (medical, agriculture, geosciences) need to make the automatic classification of these hyperspectral images with a high rate and in an acceptable time. The state-of-the-art presents several classification algorithms based on the Convolutional Neural Network (CNN) and each algorithm is training on a part of an image and then performs the prediction on the rest. This article proposes a new Fast Spectral classification algorithm based on CNN, and which allows to build a composite image from multiple hyperspectral images, then trains the model only once on the composite image. After training, the model can predict each image separately. To test the validity of the proposed algorithm, two free hyperspectral images are taken, and the training time obtained by the proposed model on the composite image is better than the time obtained from the model of the state-of-the-art.