Deep learning for hyperspectral prediction of rare earth oxide grades in drill cores: A 1D CNN-BiLSTM framework applied to the Weishan Deposit, China

• Developed a 1D CNN-BiLSTM framework for predicting TRE 2 O 3 grades directly from hyperspectral drill core data. • The model attained a high prediction accuracy on the test set (R 2 = 0.89), attesting to its robustness. Its practical value was further demonstrated through validation on an independent set of borehole samples, where it maintained strong performance (R² = 0.88), thereby confirming its generalizability. • Identified REE-sensitive spectral bands using SHAP analysis, enhancing model interpretability and mineralogical understanding. • Detected previously unrecognized medium-grade zones, revealing new mineralization types and missed-ore potential. Rare earth elements (REEs) play a vital role in the advancement of clean energy, electronics, and defense technologies. However, conventional geochemical analytical methods for determining REE contents in drill-core samples are often constrained by high costs, low efficiency, and coarse sampling intervals, making high-density and continuous analysis difficult to achieve. To overcome these limitations, this study develops a deep learning framework that integrates a one-dimensional convolutional neural network (1D CNN) with a bidirectional long short-term memory network (BiLSTM) to continuously predict total rare-earth oxide (TRE 2 O 3 ) contents from hyperspectral data. To enhance model performance and reliability, a comprehensive data-preprocessing workflow was established, incorporating multi-scale feature extraction and data-augmentation strategies. The hybrid model achieved R 2 values of 0.99, 0.92, and 0.89 for the training, validation, and test sets, respectively—substantially outperforming traditional machine-learning approaches and demonstrating excellent generalization and stability. Furthermore, to improve model interpretability, the SHapley Additive exPlanations (SHAP) method was employed to identify the most influential spectral bands. The critical wavelengths at 520, 580, 795, and 1550 nm correspond to characteristic absorption features of Sm3 + /Eu3 + , Nd3 + , and Pr3 + ions, thereby revealing the physical basis underlying the model’s spectral discrimination of REE minerals. Notably, the model also detected five previously unrecognized medium-grade REE anomalies within intervals initially classified as barren, which were later verified through mineralogical analysis as newly identified mineralized zones. Overall, the proposed deep learning-based hyperspectral prediction framework enables accurate, rapid, and cost-effective assessment of REE resources under non-destructive conditions, offering a robust and scalable approach for intelligent mineral exploration and automated ore-body delineation.