Clasificación Litológica en Afloramientos de Rocas Mediante Técnicas de Aprendizaje Automático Usando Imágenes Satelitales

Abstract

The application of machine learning techniques to perform terrain characterization from satellite images represents a significant advance in the study of the Earth’s surface. Because of the heterogeneity of rock structures and the difficulties associated with abrupt chan ges in topography, conventional classification possesses a large subjective component, which sometimes can affect the accuracy of the process. The desired result of using machine lear ning techniques in Lithological classification is to generate a tool that besides assisting the classification process, also helps to improve the aspects that may limit the results, such as optimizing processing time, increasing classification accuracy, larger data processing (images with higher spectral and temporal resolution) and to cover larger study areas. The metho dology is developed on the geospatial analysis platform with cloud processing Google Earth Engine (GEE) and it is implemented in four main phases: the first one consists of generating additional variables for the supervised classification model using statistical techniques, image transformation, and spectral enhancement, including the calculation of geological spectral indices and Principal Component Analysis (PCA), as well as incorporating topographic in formation through addition of Digital Terrain Model (DEM), adding together texture data through the Gray Level Matrix Co-occurrence (GLMC), and finally incorporate an additional band of the k-means algorithm no supervised classification. The second phase incorporates algorithm training under the supervision of Random Forest (RF), Support Vector Machine (SVM), Classification and Regression CART and Minimum Distance MD, taking into ac count the variables generated in the first phase together with the field reference information or ground thruth to reproduce several classification models that in the third phase, whichever displays the best performance metrics, is chosen leading to a reliable classification model, then, in the last phase, it will be applied to a surface area with unknown characteristics and lithology providing support information to carry out geological cartography without having required presence of people on site