This dissertation investigated the integration of multi-sensor remote sensing and machine learning (ML) techniques for lithological mapping, hydrothermal alteration detection, and artisanal and small-scale mining (ASM) monitoring in the Red Sea Hills (RSH) region of northeast Sudan. This study was motivated by the need to develop efficient, low-cost, and replicable geospatial approaches for geological and environmental assessments in remote, data-scarce, and geologically complex terrains. The Red Sea Hills, part of the Arabian–Nubian Shield (ANS), host diverse lithological assemblages, significant mineralization zones, and rapidly expanding ASM activities, yet remain underexplored using advanced earth observation methods. This dissertation thus aimed to enhance understanding of the region’s geological evolution and anthropogenic impacts through the application of machine learning algorithms to multi-sensor optical and hyperspectral datasets. The literature review highlighted the evolution of remote sensing techniques for lithological and mineral exploration, emphasizing the role of multispectral and hyperspectral data in discriminating lithological units, detecting alteration zones, and monitoring ASM dynamics in the study area. Previous studies on the ANS and RSH primarily focused on regional tectonics and conventional geological mapping using medium-resolution satellite data. However, the application of high-resolution sensors and advanced ML algorithms for integrated lithological and ASM mapping remains limited. This review synthesizes findings from prior work on supervised and ensemble classifiers, particularly Random Forest (RF), Support Vector Machine (SVM), and Multivariate Adaptive Regression Splines (MARS), and discusses challenges related to spectral similarity, limited reference data, and validation accuracy. Emerging methods, such as pseudo-labelling (PL) and data augmentation have been identified as promising strategies for addressing data scarcity. The review also contextualizes ASM within the broader socioeconomic and environmental landscape, emphasizing its dual role as both an economic lifeline and a source of significant ecological degradation and geological disturbance. This study utilized an array of remote sensing datasets including Landsat-8, Landsat-9, Sentinel-2, PlanetScope, PRISMA, and EnMAP imagery, complemented by field observations, petrographic analysis, and existing geological maps. The study area encompassed representative sites within the RSHs, characterized by diverse lithological units, including ophiolites, metavolcanics, granitoids, marble, altered rocks, superficial deposits, and active ASM operations. Pre-processing involved radiometric and atmospheric correction, followed by image enhancement techniques such as false colour composites (FCC), Band Ratios (BR), and Principal Component Analysis (PCA) to optimize spectral separability. Supervised machine learning classifiers (RF, SVM, Naïve Bayes, and MARS) were applied to generate lithological and ASM maps, with model optimization achieved through hyperparameter tuning, oversampling, and cross-validation. Pseudo-labelling was employed to augment training data by incorporating high-probability samples, and a region-growing segmentation algorithm was developed to further enhance training datasets. Accuracy assessment was conducted using both point-based and area-based approaches, computing metrics such as overall accuracy (OA), user’ and producer’ accuracy, F1-score, and Kappa coefficient (Foody, 2002). This study also introduced a spatially explicit evaluation framework to assess the extent and spatial consistency of ASM expansion. The integration of multispectral (Landsat-8/9, and Sentinel-2) and hyperspectral (PRISMA) datasets demonstrated that both RF and SVM classifiers are highly effective in lithological discrimination, achieving OAs between 0.90 and 0.96. Sentinel-2 outperformed Landsat data in term of spectral separability, particularly in distinguishing metavolcanic, marble, and altered rock units. PRISMA hyperspectral imagery produced the most accurate classification (OA = 0.96; κ = 0.95) when combined with RF, followed by Naïve Bayes (OA = 0.92; κ = 0.90). The results are consistent with field observations and petrographic analyses. Misclassifications primarily occurred between granitoids, altered rocks, and superficial deposits owing to their mineralogical similarity and weathering-induced spectral overlap. The study confirmed that ensemble algorithms, such as RF, provide stable and robust outputs, even with heterogeneous training data. The detection of hydrothermal alteration zones using PRISMA data successfully identified gossan occurrences associated with volcanogenic massive sulphide (VMS) mineralization. Spectral analysis and image processing revealed diagnostic signatures of Fe- and OH-bearing minerals, including hematite, goethite, kaolinite, and sericite, which were validated by field and petrographic evidence. Alteration assemblages, such as ferrugination, silicification, sericitization, chloritization, and carbonatization, were spatially correlated with structural lineaments and ASM sites, demonstrating that artisanal mining often exposes and exploits hydrothermally altered rocks. These findings highlight the potential of hyperspectral remote sensing for mapping mineralization indicators and guiding exploration in arid environments in the future. PlanetScope, Sentinel-2, and Landsat datasets revealed substantial temporal expansion of ASM activities between 2003 and 2024. Binary and multiclass RF classifications consistently indicated that the ASM areas increased by 150–300% across the monitored sites. For example, at Site 3, PlanetScope data showed ASM growth from 50 ha in 2016 to 125 ha in 2024, while Sentinel-2 data recorded an increase from 46.4 ha in 2017 to 85.2 ha in 2024. Change detection analysis confirmed that the expansion primarily occurred over the meta-andesite and, to a lesser extent, the meta-basalt units. These lithologies, rich in iron and magnesium, are closely associated with gold mineralization, which explains the miners’ spatial targeting patterns. The findings also demonstrated that ASM expansion led to significant geomorphological and environmental impacts, including drainage alteration, surface degradation, and destruction of structural features such as faults and orebody continuity. Experiments on pseudo-labelling demonstrated that this technique can improve model stability and reliability when ground-truth data are limited. However, improvements in accuracy were not consistent across classifiers or lithologies. The RF model benefited moderately from PL, whereas MARS displayed over-optimistic probability outputs that did not correspond to the true accuracy levels. The region-growing segmentation method, developed to augment the training data, improved the classification accuracy for ASM mapping, particularly when small but representative additional samples were used. These results highlight the importance of expert-guided validation and adaptive data augmentation in geological remote sensing. This dissertation contributes to the fields of remote sensing and geological mapping in several ways. -Methodological Advancement: It integrates multi-sensor optical and hyperspectral data with machine learning and data augmentation to overcome challenges of limited reference data in arid, inaccessible regions. -Novel Application: This study represents one of the first comprehensive applications of PRISMA hyperspectral and Planet data for ASM and lithological mapping in the RSH. -Spatial Insight: It provides the first detailed spatiotemporal assessment of ASM expansion in northeast Sudan, linking mining activity to specific lithological and alteration zones. -Validation Framework: It combined point- and area-based accuracy assessment approach was introduced, improving the spatial interpretation of the classification performance. -Geological Understanding: It enhances the understanding of the relationship between hydrothermal alteration, ASM activity, and geology within the RSHs, supporting more sustainable mineral exploration and land-use planning. The findings confirm that the integration of multi-sensor remote sensing and machine learning offers a powerful and cost-effective framework for lithological mapping and ASM monitoring in complex and data-scarce environments. Hyperspectral data, particularly PRISMA data, provide superior performance in detecting subtle mineralogical variations, whereas ensemble learning algorithms such as RF offer robust and interpretable results. The study also underscores the potential of adaptive data augmentation (pseudo-labelling and region-growing) to improve classification accuracy under reference data limitations. Future research should expand the application of advanced hyperspectral missions (e.g., EnMAP and CHIME) for mineral exploration and investigate the use of deep learning architectures (CNNs, transformers) to improve spatial feature extraction. The integratation of InSAR and LiDAR data can further enhance the detection of geomorphological and structural deformations associated with mining. Finally, the regional-scale implementation of the developed methodology could support the sustainable management of ASM activities and guide mineral resource policy in Sudan and other parts of the Arabian–Nubian Shield.