Recurrent-spectral convolutional neural networks (RecSpecCNN) architecture for hyperspectral lithological classification optimization

Lithological mapping (LM) is critical to geologic and environmental studies and has traditionally relied on labor-intensive field investigations. Hyperspectral imagery (HSI) provides a powerful dataset for data-driven alternative techniques by capturing detailed spectral information about the Earth’s materials. This research explores deep learning for accurate LM with HSI. We introduce in this study the Recurrent Spectral-3D-CNN (RecSpecCNN) model, which utilises both spectral and spatial information in HSI data alongside state of art models (SVM, 1D-CNN, 2D-CNN, 3D-CNN, and HybridSN). The EnMAP hyperspectral image covering the Eastern Kerdous inlier in Morocco was used for LM. The performance of the proposed model was evaluated together with the other models for the classification of 12 litho-units of extended geological age from Paleoproterozoic to Cenozoic represented by a Quaternary sediment. The analysis of the results allows the selection of the more robust deep learning architecture for LM and then produces an updated map for the Ameln Valley region. The proposed model shows better performance in LM, achieving an overall accuracy of 98.46% and a kappa accuracy of 98.25% using the EnMAP dataset. The analysis of 2-, 3-, 5- and 10-fold cross-validation ultimate performances demonstrated that over the EnMAP dataset, RecSpecCNN improved the accuracy of five benchmark models using the EnMAP dataset. RecSpecCNN improves the classification of hyperspectral images by integrating 3D-CNNs and Bidirectional Long Short-Term Memory (BILSTM) to capture detailed spectral dependencies. This hybrid architecture effectively models complex spectral relationships to improve classification accuracy in the study area. Accordingly, RecSpecCNN architecture can be adapted to different hyperspectral classification tasks. The results of the current study show that using the proposed model and the new EnMAP images for LM tasks provides more accurate results. The use of EnMAP data with the RecSpecCNN model represents an efficient and cost-effective tool that offers a considerable alternative to airborne spectroscopic surveys in the Moroccan Anti-Atlas Belt as well as in similar semi-arid regions worldwide.