Lithology Identification of Lithium Minerals Based on TL-FMix-MobileViT Model

In lithium mineral exploration, rapid and accurate identification of lithium-related rock lithologies is critical. Traditional manual methods are time-consuming and have limited accuracy, whereas some deep learning models, despite offering high precision, suffer from high computational complexity and low inference speeds, limiting their practical application. To address these issues, this study proposes a lightweight deep learning method based on a transfer learning-based Fourier-space mixed sample data augmentation mobile vision transformer (TL-FMix-MobileViT) to efficiently identify six types of lithium-related rock lithologies. Data from Dahongliutan (Xinjiang, China), Portugal, and Spain were used for model training. The model integrates the inverted residual blocks of MobileNetV2, reducing computational cost and accelerating inference with depth-wise separable convolutions, along with a lightweight vision transformer that extracts both local and global features while lowering complexity. Transfer learning with pretrained models reduces the training time and resource usage, while the FMix data augmentation method improves the generalization ability and accelerates convergence. Among three TL-FMix-MobileViT variants (extra-extra small, extra small, and small), the small version performed best, with strong stability and generalization ability, although all variants offer benefits for different scenarios. Compared with seven classic models, TL-FMix-MobileViT achieved the highest classification performance, with over 99% accuracy and reliable inference. Visual comparisons showed that the model effectively captured features at rock boundaries, thereby providing superior classification of mixed rock features compared with other models. This lightweight model provides an efficient and accurate method for lithium-related rock lithology identification, demonstrating its potential for lithium exploration.