classification systems perform almost equally with the overall accuracies of 53.01% and 49.98% for the first and second models respectively though the keras model has the advantage of not integrating the pre-processing model, hence increasing its efficiency. The application of these two systems to the study area resulted in the generation of an updated geological mapping with six lithological classes detected including the Gneiss, the Micaschist, the Schist and three versions of Granites (Granite1, Granite2 and Granite3).

This article, proposes the design and implementation of a hybrid Deep Artificial Neural Network (DANN) to be used in mapping lithology using non-preprocessed Landsat images collected from tropical heterogeneous environments. In it, a sophisticated stacking of the hidden layers is performed through the introduction of an autoencoder topology where a wise variation of the dropout is defined at the encoder block to face the nonlinearity imposed by the limiting factors of such environments. The decoder however is left over without any dropout in order to ensure the reconstruction of the compressed data collected from the encoder. There is a relationships between the number of neurons per hidden layer of the encoder block, the number of hidden layers of the encoder, the dropout percentage suitable to better model the dataset. The resulting architecture which we call Automatic Deep Stacked Sparsely Connected Autoencoder (ADSSCA) is an optimized hybrid neural network architecture based on well formulated rules providing in advance, the definition of, the network topology, the total number of neurons and the number of hidden layers to be used in an extremely noisy environments. The implementation of the ADSSCA on raw landsat-8 images from an area of southern Cameroon produced an overall accuracy of 92.76%. In addition, five lithological classes where identified with similar individual accuracies.