A comprehensive review of machine learning and deep learning applications in remote sensing for lithological mapping

Abstract

Lithological mapping is critical for resource management, environmental assessment, and geohazard mitigation in regions where climate change and rapid land use require robust geoscientific tools. Integrating remote sensing (RS) with machine learning (ML) and deep learning (DL) provides objective and reproducible alternatives to traditional field surveys. This quantitative review synthesizes 444 independent case studies across 154 peer-reviewed publications to evaluate algorithm performance across diverse RS and geological settings. Kruskal–Wallis testing confirms statistically significant performance disparities among classification algorithms and architectures (H = 48.02, p < 0.001, ε²= 0.11). Categorical meta-regression also reveals that geological context and sensor configurations significantly control the performance. For single-source RS data, conventional ML algorithms yield stable baseline metrics (mean overall accuracy (OA) ≈ 0.84—0.85). Conversely, hybrid DL architectures excel in high-dimensional settings, with multisource fusion (particularly multispectral + geophysical + geochemical data) achieving highest predictive accuracy (mean OA ≈ 0.90—0.95, kappa > 0.90). Categorical analyses demonstrate that mixed sedimentary–igneous–metamorphic contexts impose statistically significant accuracy penalties (p < 0.01) compared to igneous settings, though DL models better maintain accuracy in sedimentary settings (mean OA > 0.86, kappa > 0.81). Environmental constraints remain pervasive, with dense vegetation causing accuracy drops up to 5% in OA and 8% in kappa across all models. Egger’s regression (p = 0.6) confirms the absence of severe publication bias despite high methodological heterogeneity. Future research should prioritize label-efficient techniques, context-aware multimodal integration, lightweight edge deployment, and standardized multisensor benchmarks.