Iron Ore Grade Estimation Using Visible–Near-Infrared Hyperspectral Imaging and Convolutional Neural Network

Hyperspectral imaging has been widely used in geoscience from space satellites to laboratory equipment due to its huge potential in qualifying and quantifying minerals. However, hyperspectral imaging has also experienced difficulties in detailed ore grade estimations due to its spectral complexity. To address this complexity, this paper presents an iron ore grade prediction method that combines hyperspectral imaging and a deep learning-based Convolutional Neural Network (CNN) prediction model. Hyperspectral images of different grades of iron ore were acquired, followed by training the prediction model using these hyperspectral images. The CNN model was trained using 100,000 spectra. Post-training prediction results show that the CNN model outperformed conventional models in all evaluation metrics which were 0.974, 2.714, and 11.226 in R2, validation RMSE, and test RMSE respectively. The analysis results of the trained CNN model furthermore suggested that the model successfully extract spectral features. Lastly, the results demonstrate an outstanding prediction capability for the iron ore grade by utilizing a CNN and Visible Near-Infrared (VNIR) hyperspectral imaging for the artificially prepared iron ore samples in the laboratory. Hence, the CNN-based model offers the potential in hyperspectral imaging to efficiently predict ore grade in geological remote sensing applications.