Abstract
Automated mineral and extraterrestrial body identification, including asteroids and meteorites, is crucial for geology, planetary science, and resource management developments. The project here provides a deep learning system for the classification of a heterogeneous dataset of more than 5,600 mineral and meteorite RGB images featuring diverse mineral classes and rock types. This work rigorously test a set of up-to-date convolutional neural networks (CNNs) and transformer models, e.g., DenseNet-121, Swin Transformer, MobileNetV3, and BEiT. Standard preprocessing and data splitting are used to train and validate models, and their performance is measured in terms of accuracy, precision, recall, and F1-score. DenseNet-121 has the best accuracy of 98.58%, showing excellent performance in fine-grained mineral classification, followed by transformer models like Swin Transformer, which also show competitive performance, especially for intricate textures. These findings emphasize the ability of deep learning to eliminate mineralogical and meteoritic image classification, facilitating rapid analysis in geological exploration, mining operations, and space missions. Future development involves incorporating metadata and applying lightweight models to utilize in-field, real-time analysis.