Design and implementation of an Automatic Deep Stacked Sparsely Connected Convolutional Autoencoder (ADSSCCA) neural network for remote sensing lithological mapping using calculated dropout

The accurate mapping of lithological units in tropical environments characterised by dense forest, persistent cloud cover, and limited bedrock exposures from remote sensing data is very challenging. In this paper, we propose the design and implementation of an Automatic Deep Stacked Sparsely Connected Convolutional Autoencoder (ADSSCCA) neural network for lithological mapping without data filtering. The model integrates convolutional neural networks architecture with Principal Component Analysis for feature extraction and dimensionality reduction. We also combine 3D and 2D convolutions to generate feature maps, and the fully connected block is configured using the ADSSCA philosophy with dropout calculated as a function of model capacity. The dropout serves to randomly drop hidden computing units and their connections during training, thus reducing overfitting and improving the model’s generalization ability. The proposed classification model was tested on non-processed Landsat-8 images from the Yaounde-Sangmelima region in southern Cameroon. The proposed ADSSCCA mapped the five dominant lithological classes namely, gneisses, schists, micaschists, biotite-granites and pyroxene-granites with an impressive overall accuracy of 99.85%. The novelty in this study is the removal of the pre-processing stage which results in a simplified work flow, reduces overhead and preserves data integrity. Secondly, the introduction of dynamically calculated dropout based on neuron and layer count enhances generalization. It is expected that; this new finding will enhance lithological mapping and analysis in the forest zones and areas with similar characteristics.