Estimating iron content in ores : an investigation with machine learning methods and hyperspectral images.
Mineral processing processes such as mineral exploration activities can gain from more agile processes for characterizing ores. The characterization, carried out by traditional methods in the laboratory, is very accurate, but in general it is time deficient. The analysis of hyperspectral images can bring faster results than traditional laboratory analysis, however, the accuracy of the characterization is still a challenge to be investigated. These difficulties are related to environmental factors such as lighting and humidity, sample factors such as grain size and homogeneity, and modeling factors, such as choice of spectral bands, resolution of images and types of models for characterization. Considering the challenges mentioned, this research aimed to answer questions related to the modeling factors and therefore, we investigated machine learning methods to estimate the iron content in iron ore samples based on wavelengths of hyperspectral images in the Visible and near infrared region (VNIR) between 400 and 1000 nm; most relevant attributes for the model and we validate the results with the use of statistical evaluation metrics. The performance of the models showed constant results, which present low variance and dispersion and with iron dosage estimation accuracy above 90% using Random Forests (RF) and Multilayer Perceptrons (MLP).