Using long-term bare earth composite image and machine learning in lithological mapping of Adrar Souttouf mafic complex (Oulad Dlim massif, Southern Morocco)

Abstract

The success of geological mapping is mainly dependent on the best delineation of the lithological spatial features, among others. To this end, it is common to use remote sensing imagery supported with visual interpretation. The human visual perception is more capable of detecting and differentiating the colored compositions, however, it is not able to capture the information from the multispectral information. To overcome this issue, it is used to reduce the multidimensional information to three dimensions to be employed in colored visualizations. So far, the most tested methods are the linear ones (i.e. principal component analysis (PCA) and canonical correspondence analysis (CCA)) which were applied to a single date image. In this study, we explored innovative methods for lithological mapping to address the following questions: Can machine learning (ML) algorithms enhance the discrimination of key lithological features? Furthermore, can the identified patterns contribute to producing improved map that aids in resolving the two competing hypotheses—subduction or intracontinental rift—proposed for the Adrar Souttouf mafic complex in Moroccan Saharan domain. In this regard, we explored the potential of new ML methods and a composite image of the bare earth reflectance generated from Landsat-8/OLI image time series over ten years (from 2013 to 2023). We selected two new nonlinear methods which are Uniform Manifold Approximation and Projection (UMAP) and autoencoder (AE). The results of the visual interpretation were validated by an extensive field survey. The findings revealed that the linear methods (PCA and CCA) perform better in capturing the local details while the nonlinear methods (UMAP) were performant at the global patterns detection. Surprisingly, the AE was similar to PCA and CCA in local pattern discrimination. We also note that the nonlinear methods are powerful in capturing the whole information from the source data, contrary to the linear methods. These results could be suitable to serve as a basis for geological mapping in the studied massif. We also suggest that the developed methodology could be applied globally to other areas where the generation of barren land is possible.