Machine Learning Approaches for Real-Time Mineral Classification and Educational Applications

The main objective of the present study was to develop a real-time mineral classification system designed for multiple detection, which integrates classical computer vision techniques with advanced deep learning algorithms. The system employs three CNN architectures—VGG-16, Xception, and MobileNet V2—designed to identify multiple minerals within a single frame and output probabilities for various mineral types, including Pyrite, Aragonite, Quartz, Obsidian, Gypsum, Azurite, and Hematite. Among these, MobileNet V2 demonstrated exceptional performance, achieving the highest accuracy (98.98%) and the lowest loss (0.0202), while Xception and VGG-16 also performed competitively, excelling in feature extraction and detailed analyses, respectively. Gradient-weighted Class Activation Mapping visualizations illustrated the models’ ability to capture distinctive mineral features, enhancing interpretability. Furthermore, a stacking ensemble approach achieved an impressive accuracy of 99.71%, effectively leveraging the complementary strengths of individual models. Despite its robust performance, the ensemble method poses computational challenges, particularly for real-time applications on resource-constrained devices. The application of this methodology in Mineral Quest, an educational Python-based game, underscores its practical potential in geology education, mining, and geological surveys, offering an engaging and accurate tool for real-time mineral classification.