Sedimentary Phosphate Classification Based On Spectral Analysis And Machine Learning

The process of phosphate extraction can significantly benefit from the advances in spectral analysis and Artificial Intelligence to reduce the cost of the drilling operation. The ambiguities caused by the apparent similarities between different layers and by the existing mineralogical alterations complexify the delineation of phosphate layers with conventional vision systems. In this paper, we established a spectral signature database of representative samples collected from the Ben Guerir deposit in Morocco, over the 250?2500 nm spectral range (midultraviolet to mid-infrared). The aim is to build a spectral database of the samples and select an optimal waveband set capable of good discrimination, which can be used in a multispectral inspection system for an accurate delimitation of the soil layers during the drilling operation. First, the reflectance signatures of the extracted soil samples were collected. Second, principal component analysis (PCA) loadings were investigated to determine the most informative feature subspaces. A Bhattacharyya distance (B-distance) separability test was then implemented to select the most dissociating combinations of 3 bands from these subspaces. Finally, a machine learning classification test was used to evaluate the capacity of the selected features to discriminate phosphate samples. The results demonstrate the impact of selecting an informative reduced feature set and show good discrimination rates based on the combined information of wavelengths from the Ultraviolet (UV) or the Near-infrared (NIR) spectral ranges.