Satellite based lithological Mapping: A review of Sensors, Methods, and emerging Machine-Learning approaches

• Review of multispectral and hyperspectral lithological mapping. • A unified taxonomy of ML and DL mapping approaches is proposed. • Deep learning outperforms classical ML in complex geological settings. • Spectral similarity and data scarcity remain key mapping challenges. • Future trends include multi-source data fusion and open science. Understanding the spatial organization of geological formations is essential for mineral exploration, engineering planning, and environmental assessment. While field observations remain indispensable, they are often costly and difficult to conduct across large or inaccessible regions. Over the past decade, advances in satellite remote sensing have progressively filled this gap. Multispectral missions such as Landsat, Sentinel-2 and ASTER, together with the growing availability of hyperspectral sensors including Hyperion, PRISMA mission and EnMAP, now allow subtle mineralogical differences to be detected over broad areas. In parallel, the rapid expansion of machine learning (ML) techniques has reshaped the way geological data are analysed. Classical algorithms such as Support Vector Machines or Random Forests have proven highly effective for heterogeneous terrains, while more recent deep learning (DL) architectures capture both spectral features and spatial context, offering significant gains in mapping accuracy. To provide a structured overview of these developments, this review follows a focused literature-selection strategy combining keyword filtering and cross-database screening. Beyond synthesizing existing results, the review proposes a unified taxonomy of ML and DL approaches for lithological mapping and offers a comparative assessment of their strengths, limitations, and typical use cases. The analysis also highlights persistent challenges spectral confusion, uneven ground-truth availability, and the lack of standardized preprocessing pipelines which continue to limit reproducibility across studies. The review concludes by outlining several research priorities, including improved integration of multisource datasets (hyperspectral, Synthetic Aperture Radar (SAR), Digital Elevation Models (DEMs), geophysics), the adoption of transparent and reproducible workflows, and the development of hybrid spectral–spatial models capable of operating in complex geological settings. These directions provide a foundation for more reliable, scalable and interpretable lithological mapping in future applications.