Sedimentary phosphate classification based on spectral analysis and machine learning
Introduction
Phosphate deposits represent a major asset for Morocco. With approximately 70% of the total world reserve (Ober, 2018), Morocco is the 1st exporter country of phosphate. It is also the 2nd phosphate producer worldwide, a ranking obtained due to a relentless extraction strategy implemented by the OCP group, which relies on a continuous process exploiting the rich phosphate layers. The fuzzy aspect of the limits of the layers introduces impurities during the extraction process, which implies inserting preprocessing steps along the production line to obtain consumable phosphate. Multispectral and hyperspectral remote sensing could prove useful in this process by reducing the fuzziness impact.
Spectral analysis and imaging, particularly hyperspectral imaging (HSI), have been deeply incorporated in both agriculture and mineralogical based studies. The resulting high dimensional data represents the responses of the studied elements over narrow spectral bandwidths. Each element gives a distinct spectral response according to its physiochemical construction and reaction to electromagnetic energy. As a result, the features that best differentiate the elements can be identified (Gupta, 2017). Studies have been conducted to identify spectral signatures of pure minerals by combining both X-ray diffraction and spectrophotometry. Thus, reliable and reference data sets were constructed, namely the United States Geological Survey (USGS) and the ECOSTRESS spectral libraries (Baldridge et al., 2009; Kokaly et al., 2017; Sucich et al., 2018).
For sedimentary phosphate, the difficulty resides in the presence of complex mixtures. In addition to the existing horizontal and vertical variabilities, the discontinuity of the geological repository and the occurrence of infiltration phenomena enrich the heterogeneity of the deposits. Moreover, the color shade similarities between phosphate class and other mineral types referred to as sterile, such as limestone (Fig. 1) restrain the use of supervised classification methods (Nash et al., 2004; Maloy and Treiman, 2007). These observations raise the question of whether spectral analysis can be used to identify the ambiguous limits between the phosphate and the sterile layers and select adequate features for their discrimination.
In this context, we undertook this study to:
- 1.
Explore the use of spectral information, without any further experimental analysis, to algorithmically select major discriminating features between phosphate rocks and sterile.
- 2.
Pave the way for the inspection of spectral imaging techniques in sedimentary phosphate classification and mapping, a field poorly investigated so far.
Essentially, the aim is to assess the feasibility of identifying spectral wavelengths, using spectroscopic analysis, over which accurate phosphate classification is possible. The results should guide the selection of an adequate multispectral inspection sensor to use on-site for an optimized phosphate extraction process.
Under this perspective, we propose a new classification approach that focuses on spectral analysis and feature selection prior to using a machine learning classifier. First, representative samples are collected from a major Moroccan phosphate deposit. The samples are then prepared to undergo a spectral analysis whereby reflectance responses between 250 and 2500 nm are collected. The spectral library contains representative spectra of both phosphate and sterile classes. Secondly, PCA is applied to define descriptive spaces with characterizing wavelengths of each class. The union of both spaces is then subjected to a separability evaluation using the Bhattacharyya distance to find the best features differentiating the two classes. Finally, the K-Nearest Neighbors (KNN) algorithm is used to evaluate whether our feature reduction approach can improve the phosphate classification accuracy. Other classifiers are tested for comparison as well.
The rest of this paper is organized as follows: Section 2 provides a brief overview of conducted studies in spectral analysis. Section 3 describes the working environment and the preprocessing steps applied to the collected samples. Section 4 details the proposed analytical approach for this classification problem. Experimental results and discussion are the subjects of Section 5, and the main conclusions are drawn in the last section.
Section snippets
Related work
Studying the behavior of earth and soil elements over a growing spectral range is a trending research field. The spectral response contains meaningful information that can be used for several purposes, especially with the growing accessibility to various equipment like air born, satellite-based, or laboratory fitted appliances like spectroscopes.
Many studies of minerals and rocks enabled the construction of the USGS library, the largest -still not exhaustive- spectral database available (Kokaly
Study area and samples collection
Samples were retrieved from the Ben Guerir deposit, a major extraction site of phosphate in Morocco. Situated in the Gantour basin, it goes 25–30 km from east to west, and 10–20 km from north to south, with an average altitude of 430 m containing phosphate series from the Lutetian age to the Maastrichtian age (Daafi et al., 2014). From a cutting face of this deposit, alterations of phosphate beds (layers, furrows, and bundles) and sterile to slightly phosphatic levels or interlayers can be
Proposed approach
In our study, we faced two main challenges: first, the studied site was not previously explored for spectral analysis ends, and no existing reference data of the deposit was available. Second, the complex composition of the soil samples is likely to result in similar reflectance behavior over some particular wavelengths and negatively impact the discrimination outcome, leading to errors in the classification model. In response to these challenges, we established an approach with three main
Spectral data set construction
Spectral signatures were collected for the gathered samples. The latter were selected in a way to ensure the representativeness of the deposit. The binary ground truth labeling of the samples was performed by experts, in conformity with the standards and protocols of the field.
Representative signatures from the constructed spectral library are displayed (Fig. 5 to Fig. 8). The spectral reflectances are grouped according to the geological ages of the samples and following the vertical layering
Conclusion
In this paper, we constructed a spectral database and investigated the possibility of analytically identifying spectral features to discriminate between phosphate and sterile samples from the Moroccan Ben Guerir deposit. The obtained spectral signatures were analyzed to highlight the difficulty to distinguish between the spectral responses and the raised challenges to develop an accurate discrimination system. The proposed approach enabled the identification and use of highly informative
Computer code availability
The python version used in this article is 3.6.4 and is freely available for download (https://www.python.org/downloads/release/python-364/. All source code developed is open source and available at GitHub site: https://github.com/Crajaa/spec_phos_clf.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgment
The Authors would like to acknowledge the support through the R&D Initiative -Appel à projets autour des phosphates APPHOS- sponsored by OCP (OCP Foundation, R&D OCP, Mohammed VI Polytechnic University, National Center of Scientific and technical Research CNRST, Ministry of Higher Education, Scientific Research and Professional Training of Morocco MESRSFC) under the project entitled *Feasibility Study of a system discriminating between phosphate and sterile during the mining extraction process
References (55)
- et al.
Gasoline classification using near infrared (NIR) spectroscopy data: comparison of multivariate techniques
Anal. Chim. Acta
(2010) The ASTER spectral library version 2.0
Rem. Sens. Environ.
(2009)Hyperspectral classification based on spectral–spatial convolutional neural networks
Eng. Appl. Artif. Intell.
(2018)Identification and classification of polymer e-waste using laser induced breakdown spectroscopy (LIBS) and chemometric tools
Polym. Test.
(2017)Geology and mine planning of phosphate deposits: benguerir deposit Gantour basin–Morocco
Procedia Eng.
(2014)- et al.
Discriminating ore and waste in a porphyry copper deposit using short-wavelength infrared (SWIR) hyperspectral imagery
Miner. Eng.
(2017) Spectral–spatial hyperspectral image classification based on KNN
Sens. Imag.
(2016)- et al.
An automated mineral classifier using Raman spectra
Comput. Geosci.
(2013) Assessing the ability to combine hyperspectral imaging (HSI) data with Mineral Liberation Analyzer (MLA) data to characterize phosphate rocks
Int. J. Appl. Earth Obs. Geoinf.
(2018)Discrimination of the geographical origin of Codonopsis pilosula using near infrared diffuse reflection spectroscopy coupled with random forests and k-nearest neighbor methods
Vib. Spectrosc.
(2012)