Methodological Advances in Urban Remote Sensing: Machine Learning for Geological Mapping

Geological maps play a pivotal role in urban planning providing essential information about surface and subsurface conditions to ensure safe and sustainable infrastructure development. This paper examines the use of machine learning and deep learning techniques in geological mapping for urban environments, focusing on both regional analysis and detailed assessments. For regional mapping, medium-resolution data like Landsat, ASTER, and Sentinel-2 were processed with classical ML algorithms, such as Support Vector Machines (SVM), to classify lithological units and detect geological features, offering valuable insights during the preliminary phases of urban planning. In contrast, detailed geological mapping requires high-resolution datasets. Advanced DL models, including convolutional neural networks, recurrent neural networks, and Graph Convolutional Networks, address challenges like data complexity, noise, and class imbalance. A case study conducted on the northern and western margins of Tunis highlights the practical application of various techniques. Lithological unit classification was performed using SVM. The choice of SVM was based on a comparative analysis with multiple supervised classifiers. The data utilized included ASTER VNIR-SWIR and SRTM DEM data, and the results demonstrated an overall accuracy of 78.44%. Despite its strengths, the study encountered challenges related to imbalanced datasets and temporal variability in the remote sensing data. Potential solutions, such as synthetic data generation via Generative Adversarial Networks and weighted loss functions, are proposed to improve model performance.