Mitigating the influence of topographies on hyperspectral images for the sorting of iron ore based on variational mode decomposition and spectral property

Abstract

Hyperspectral imaging (HSI) can be applied for the purpose of sorting iron ores with varying total iron (TFe) content. However, the surface topographies of iron ore can negatively impact its hyperspectral images, consequently affecting the sorting effectiveness for each individual piece of iron ore. To address this issue, this study conducted an experiment on hyperspectral image of the iron ores. Firstly, variational mode decomposition (VMD) was employed to decompose the spectra of iron ore into trend signals and peak-valley signals. The two signals were then used for the calculation of three spectral properties (baseline-shift level, steepness level, and amplitude level). To identify the critical property capable of distinguishing normal spectra and the topographies-affected, the difference of the three spectral properties between the topographies-affected and the normal spectra were analyzed. According to the analysis results, a VMD-based correction model (VMDC) was proposed, which aims to identify the topographies-affected spectra in an ore and then correct them to resemble the normal spectra. Finally, the coefficient of variation (CV) for the correction was assessed and a test was conducted for TFe content prediction and sorting of each piece of iron ore. There are three primary conclusions summarized as follows:1. The mean Jeffries–Matusita distance (JMD) of baseline-shift level, steepness level, and amplitude level between normal spectra and topographies-affected spectra were 1.889, 1.307, and 1.170, respectively. This implies that baseline-shift level can be regarded as a key property for distinguishing whether a spectrum is topographies-affected or not. 2. After correction, there was a noticeable improvement in the topographies-affected areas of the ore samples, and the topographies-affected spectra showed the consistence with the normal spectra. Upon examination, the mean CV for VMDC is 0.115, whereas the uncorrected data has a mean CV of 0.315, indicating a reduction in spectral variability after VMDC correction. The results demonstrate the effectiveness of VMDC in mitigating the impact of topographies. 3. In the TFe content prediction and sorting, VMDC consistently demonstrates the best performance, for example when using RF as the regressor, coefficient of determination (R2) and accuracy of Original Data-RF, MSC-RF and VMDC-RF were 84.3% and 0.894, 86.7% and 0.903, and 88.7% and 0.921. The evaluation results confirm that VMDC can improve iron ore sorting. Overall, the proposed method shows potential for practical applications in iron ore sorting.