Assessing SWIR and MWIR Hyperspectral Imaging for Rapid Estimation of P2O5 Distribution in Sedimentary Phosphate Drill Cores

Sedimentary phosphate rocks are crucial for global food security, contributing to over 90% of the fertilizer industry's needs. However, their exploration and mining face significant challenges due to substantial horizontal and vertical variations in phosphorus concentrations within the strata. Traditional characterization methods are time-consuming and costly, requiring complex sample preparation, which often limits the spatial resolution of measurements across the ore body. On the other hand, infrared hyperspectral core scanning has emerged as a proven technique for rapid characterization of mineral assemblages along drill cores, which by leveraging advanced machine learning algorithms, offers a powerful tool for predicting geochemical variations. In this context, our study aims to assess the ability of hyperspectral infrared imagery to rapidly quantify the distribution of P₂O₅ in phosphate drill cores using a non-destructive methodology. For this, a ~65-meter drill core from the phosphatic series of Ben Guerir (Morocco) was analyzed. P₂O₅ measurements were acquired using a Thermo Fisher XL5 portable XRF (pXRF), and hyperspectral images were collected using a SPECIM SisuROCK core-scanner with SWIR (1000–2500 nm) and MWIR (2700–5200 nm) cameras. To predict P₂O₅ concentrations from infrared spectra recorded in hyperspectral imagery, we explored a direct method, using high-performing machine learning algorithms trained on a ~5-meter drill core dataset. When applied to the whole drill core dataset, the machine learning algorithms—Random Forest Regressor, KernelRidge Regressor, Gradient Boosting, Support Vector Regressor, and K-Nearest Neighbors— reported good predictive performance with strong correlations of 78%, 78.2%, 67.1%, 74.9%, and 68.7% in the SWIR region and 81.2%, 83%, 80.2%, 83.24%, and 82% in the MWIR region, respectively. Direct estimation of P₂O₅ using the Support Victor Regression model on MWIR imagery thus represents a more effective approach, offering significant potential for P₂O₅ chemical mapping and improved phosphorus resource estimation with a low mean absolute error of 3.29. Further improvements could be achieved by employing a larger training dataset and deep learning algorithms.