A Systematic Quantitative Analysis of Machine and Deep Learning Architectures for Hyperspectral Mineral Mapping

Hyperspectral remote sensing has become a core technology for mineral exploration due to its ability to capture diagnostic spectral absorption features of surface materials. Over the past decade, algorithmic approaches for hyperspectral mineral mapping have evolved from classical machine learning models to deep neural representation architectures. However, systematic quantitative comparison across architectural generations and sensor platforms remains limited. This study presents a structured analysis of 169 peer-reviewed publications to evaluate the temporal evolution, performance characteristics, and operational constraints of machine learning (ML), convolutional neural network (CNN), and transformer based frameworks. Aggregated results indicate performance progression from 82–90% overall accuracy in classical ML models to 94–97% in recent attention-based architectures, primarily driven by improved spectral–spatial feature coupling. Despite these gains, cross-sensor generalization, computational scalability, and geological interpretability remain critical limitations. This review distinguishes itself from prior surveys by providing the first structured meta-analytic comparison across four algorithmic generations (classical ML, hybrid ML, CNN based DL, and transformer/attention architectures) using standardized accuracy metrics extracted from 132 studies reporting quantitative results. A unified mineral mapping framework integrating representation learning with domain adaptation and geological validation is proposed. The findings provide a decision-support reference for selecting algorithmic strategies in operational hyperspectral mineral exploration.