Convolutional Neural Networks Applied to Antimony Quantification via Reflectance Spectroscopy Using Soils from Northern Portugal: Opportunities and Challenges

1. Introduction

Antimony (Sb) is currently considered a critical raw material (CRM) to the European Union (EU), being strategic to its economy, in a scenario where the global market of Sb is dominated by China. The element Sb is included in group 15 (VA) of the periodic table, located in the second long period of the table between tin (Sn) and tellurium (Te). It is classified as a non-metal or metalloid and may exhibit a valence of +5, +3, 0, or -3, with metallic characteristics in the trivalent state [1]. Sb and its mineral sulfides are reported to have been used by humans at least since 4000 B.C. One of its reported uses in more ancient times is as a main ingredient of a black paste, named kohl, used for colouring eyebrows and lining eyes by Egyptians and others in early biblical times [1]. An ornamental vase found at Tello, Chaldea, is reported to be cast Sb and dates to 4000 B.C. The given name to the metal, stibium, is attributed to Pliny the Elder (50 AD), while “antimonium” is reported to be referred to by an Arabian alchemist, Geber, living in the eighth century [2] [1]. A scientific treatise about the element Sb was written by Nicolas Lemery (1645–1715), containing results of his investigations about the proprieties and different preparations of metal Sb, which was believed to be an important component in the alchemical lore, actuating as a magnet for extracting mercury, a key component for making the Philosopher Stone [3].

Nowadays the main uses of Sb in the EU remain in the textile industry as flame retardants, and as Sb-based semiconductor materials such as lead-acid batteries, lead alloys, catalysts, and stabilisers for plastics, and in the glass and ceramic industry [4]. In the context of the growing demand for electric vehicles, Sb is also being studied as an alternative anode for use in lithium–ion batteries [5]. Today, graphite is mainly employed as an anode in lithium-ion batteries and sodium-ion batteries, although Sb is also considered due to its structure, with a potential for a much better electrical conductor [6].

Several authors have applied machine learning (ML) techniques to detect the presence of heavy metals in soils indirectly through reflectance spectroscopy. Multiple Linear Regression (MLR) and Artificial Neural Network (ANN) were used by Kemper and Sommer [7] to predict the contents of As, Cd, Cu, Fe, Hg, Pb, S, Sb, and Zn, using samples from an area flooded with sludge that resulted from a break of a mine tailings dam in Spain. This accident resulted in a very contrasting mineralogy between the background soil and the contaminated zone. They obtained good predictions for As, Fe, Hg, Pb and Sb. Multiple regression equations were used by Nanni and Demattê [8] to effectively predict Fe₂O₃ and TiO₂ between other soil properties. Partial Least Squares Regression (PLSR) was employed by Cheng et al. [9] to estimate Cd, Pb, As, Cr, Cu and Zn and by Rodríguez-Pérez, et al. [10] for estimating Mn, soil nutrients and other properties such as pH and electrical conductivity. Cheng et al. [9] examined the feasibility of using soil reflectance spectra to estimate the concentrations of the metals in a suburban area of Wuhan City, Hubei Province, China. The concentration of the metals Pb, As, Cr, Cu and Zn in the soil samples were determined by Inductively coupled plasma atomic emission spectroscopy (ICP-AES). They observed how different preprocessing treatments interfere with the results of the prediction model and applied statistical analysis to ascertain the estimation mechanism, focusing on relationships between soil reflectance spectra and concentrations of Soil Organic Matter (SOM), Fe, and heavy metals (Cd, Pb, As, Cr, Cu, and Zn). They concluded that employing Savitzky-Golay spectral pre-treatment yielded favorable PLSR models, however, additional studies were needed to establish internal relationships between the heavy metal concentrations and spectrally active elements, as the authors did not achieve satisfactory results. Rodríguez-Pérez et al. [10] were able to obtain a good performance only for phosphorus, pH and electrical conductivity. Pyo et al. [11] used a CNN and compared it to ANN and Random Forest Regression (RFR) to estimate As, Cu and Pb soil samples taken from a mining area, located in the Geum River watershed of South Korea, and obtained the best results with the CNN model, but also achieving reasonable soil heavy metal estimation accuracy with the other machine learning models. RFR was used by Guo et al. [12] to successfully infer the Zn and Ni concentrations based on the relationship between heavy metals with soil criteria and clay. RFR together with PLSR and Support Vector Machine (SVM) were used to predict Mn, Cu, Zn, Pb, Cr and Ni contents, obtaining the best results with RFR.

CNNs are well-established in various domains, including object detection, image classification, and spectral analysis [13]. By leveraging sparse local connections and weight sharing, CNNs have proven to be effective in learning and extracting local and abstract features from raw spectral data. By stacking multiple convolutional and pooling layers, the CNN model can efficiently capture intricate patterns within the data, making it well-suited for soil content prediction tasks [13,14].

The objective of this study was to verify the possibility of identifying Sb mineralization through the spectral signature of soils with the application of a CNN, a deep learning technique. The major advantage of developing ML and deep learning (DL) methods to quantify and qualify heavy metals in soils is the possibility of analyzing large-scale zones faster, avoiding the high cost and time demanded that is implicit in the traditional approach of geochemical analysis. Reflectance spectroscopy is a remote sensing technique, non-invasive, capable of identifying targets for mineral exploration reducing costs and avoiding environmental impacts. The soils samples for this study were obtained and analyzed in previous studies for characterizing the Sb distribution in the former mining areas of Ribeiro da Serra and Tapada, in Northern Portugal. The CNN model shows promising results, but in this study the overfitting of the model couldn’t be avoided. Despite not having achieved good results for the Sb predictions, this study provides insights about which strategies could be incorporated into future studies.

2. Background

2.1. Study area

The study area encompasses the two former Sb-Au mining concessions of Ribeiro da Serra and Tapada, located in northern Portugal, approximately 30 km East from the city of Porto (Figure 1). The Ribeiro da Serra and Tapada mines were opened in 1880 and 1881, respectively, producing thousands of tonnes of antimonite concentrates annually for exportation [15,16]. The exploitation of Sb hit its peak in the 19^th century, but in the first years of the 20^th century, the competition with the Asiatic Countries led to the closure of the Portuguese Sb mines. During the Second World War, there was an increase in mining activity, and since the 1960s some prospecting campaigns and reconnaissance studies were executed [17]. Nowadays the mining structures are abandoned (Figure 2) and many waste piles and tailings remain in the zone.

Geologically, those Sb deposits are located on the western flank of a Variscan structure, the Valongo Anticline, in the Dúrico-Beirão Mining District, situated within the Iberian Central Zone [16,18,19,20]. The lithostratigraphic succession consists of Cambrian/Pre-Cambrian (pre-Ordovician) rocks with very low-grade metamorphism; Ordovician and Carboniferous sequences, composed of schists and some quartzites, and the Upper Carboniferous formation comprising breccias, conglomerates, and intercalated quartzites. The lithologies vary from east to west, with ages corresponding to Lower Ordovician, Middle and Upper Ordovician, Lower Carboniferous, and pre-Ordovician.

The Sb mineralization occurs in low volume in discontinuous quartz veins that are mainly hosted in Silurian schists and greywackes [20]. The quartz veins have a hydrothermal nature and are associated with Variscan granitic intrusions, or due to fluid mixing of CO₂-rich metamorphic fluids by surface-derived H₂O–NaCl fluids [15]. The dominant directions of the country rocks range from N to NW dipping to W. The most productive veins occur in the EW direction dipping to N (Tapada) and NS dipping to W (Ribeiro da Serra).

Figure 2. Ribeiro da Serra mine infrastructures picture in the late 19^th century [21] (a) and nowadays ruins (b).

Figure 2. Ribeiro da Serra mine infrastructures picture in the late 19^th century [21] (a) and nowadays ruins (b).

2.2. Spectral reflectance

Reflectance spectroscopy offers a means to extract multiple soil properties, both direct and indirect, as well as metal contents [22]. The abundant data generated by soil spectroscopy, whether in the form of point measurements or images, necessitates the implementation of data-modelling procedures.

Materials may reflect or absorb electromagnetic radiation at varying wavelengths, governed by factors such as surface absorption, emissivity, and reflectance characteristics. The spectral range employed for soil reflectance analysis encompasses the Visible and Near-Infrared to Short-Wave Infrared (VNIR-SWIR) region, spanning from 400 to 2,500 nm. This range is further divided into two sub-ranges: VNIR (400 to 1,100 nm) and SWIR (1,100 to 2,500 nm). The interactions between light and matter are intricately tied to the wavelength. Though pure metals don't exhibit absorption within the VNIR-SWIR region, their presence can be indirectly detected through associations with organic matter (OM), interactions with compounds like hydroxides, sulphides, carbonates, or oxides that manifest detectable properties, or adsorption to light-absorbing clays [22].

2.2. Convolutional Neural Networks

The field of geosciences is having a slow but crescent incorporation of ML and DL techniques. ML data analysis methods can automate the creation of analytical models and perform tasks such as classification, regression, or clustering [23,24,25,26] that are useful in the geosciences. The improvement of the model is given by its training with a data set that will condition the results, and evaluation metrics are implemented to quantify the performance of the model with the given data set. Additional samples of data often can improve the model performance and the aim is to obtain a good model that can generalize to new data [23]. The difficulty in the availability of labelled data is exactly what makes the implementation of ML and DL remote sensing applications to geosciences a challenging task. Frequent problems of this nature are the limitation of possibilities in the data collection; the large number of physical variables associated with a limited number of samples; and the difficulty of obtaining high-quality measurements of several geoscience variables, that can only be taken by overly expensive or time-consuming techniques[27]. The heterogeneity of the data and the multi-resolution are often associated with the already challenging multiconnected nature of the different processes in geosciences [27]. Despite those challenges, the capacity for ML and DL methods to actuate in geosciences tasks that were only possible to be executed with dispendious human work, and even in tasks that were not possible to be executed before, make their implementation increasingly attractive.

The capacity of CNN to capture nonlinear behaviors makes it suitable for geological problems. CNN demonstrated powerful abstraction capabilities in the field of geosciences, with the first applications in the field focused on seismic interpretation, more recently being incorporated in broader geosciences applications such as global water storage modeling, landslide prediction, and earthquake arrival time picking [14] [23]. The constitution of CNNs primarily involves three types of layers: convolutional layers, pooling layers, and fully connected layers [13]:

1. Convolutional Layer: In this layer, input features are convolved with learnable kernels, generating various output feature maps. Each kernel has a fixed length and slides over the input feature map with a stride of 1, performing convolution with local regions where the kernel overlaps the input feature map. Non-linear activation functions are then applied to the convolution results to produce output feature maps. This layer's key advantage lies in its ability to learn local patterns while significantly reducing the number of model parameters through weight sharing.

2. Pooling Layer: After convolution, the pooling layer is employed to progressively reduce the spatial size of the output feature maps generated by the convolutional layer.

3. Fully Connected Layer: The fully connected layer comprises neurons that connect to all activations extracted from the convolutional and pooling layers. This layer plays a crucial role in generating the final output results.

3. Materials and Methods

3.1. Soil collection and preparation

The present study benefited from the soil samples collected in the scope of the AUREOLE project-ERA-MIN/0005/2018 (https://aureole.brgm.fr/; access on 25 January 2024). The soil sampling campaigns were carried out covering the area where the underground works in Ribeiro da Serra and Tapada took place and the surrounding areas. The soil sampling aimed to test the mineralization distribution and identify possible new mineralized structures and soil contamination distribution. The first soil sampling campaign was carried out in the Ribeiro da Serra Sb-Au mine zone in 2021, with 157 samples collected, while the second soil sampling campaign, in the Tapada mine zone took place in 2022, with 152 samples collected. The sampling campaigns follow a plan on a grid of 50x50 meters in the area that with an orientation according to the mining structures. Soils were collected from horizon B and horizon C when there was no horizon B which is relatively common in the study area. From the totality of samples, 54 samples from Ribeiro da Serra and 53 samples from Tapada were sent to Bureau Veritas laboratory in Vancouver (Canada) to be analyzed by ICP-MS (Inductively coupled plasma mass spectrometry analysis). Before sending the samples to ICP-MS analysis, the samples were dried in a muffle furnace or at least 48 hours and at a temperature of 55 °C. With the samples dry, after rifling the soil samples were grounded to a size < 200 μm. From the samples sent to ICP-MS analysis, only 99 had their spectral signature collected, due to some samples no longer being available.

3.2. Soil reflectance measuring

The raw soil, previously dried with gravel and pieces of plants removed, was used for taking the reflectance measurements. The soil was spread in a watch glass above a black surface and five different points of the sample were measured, resulting in five spectra per sample. Each spectrum collected is a result of an average of several measurements. For this work an average of 5 measurements was used, each measurement resulted in 40 scans.

The FieldSpec 4 standard resolution spectroradiometer equipment (ASD Inc., Boulder, CO, USA) was used to collect the spectral data. The proceeding included “heating” the equipment for 30 minutes before starting to measure. The spectroradiometer has three sensors, one VNIR and two SWIR sensors. The time of “heating” is necessary to ensure that the three sensors are at the same temperature [28]. Normalization with a perfect albidum (white plate) also is needed every time a new measurement project is started and every two hours of work. The software used was Indo Pro [28].

3.2. Spectral preprocessing

The first step was removing from the dataset samples with Sb content above 1000 ppm, that were considered outliers, for being mostly associated with contaminated areas that don’t represent the Sb concentrations found related with the natural occurrence of mineralization (see section 3.1), common spectral preprocessing was implemented to improve the results obtained by eliminating noise and highlighting spectral features. The wavelengths before 400 nm were removed due to the excessive noise in this zone of the spectra. The pre-processing steps included: convert the reflectance to absorption; remove the continuum; smoothing the signal and calculate the first and second derivatives; convert the waveform spectrogram.

The order of the spectral preprocessing is given in the diagram (Figure 3).

3.2.1. Continuum removal using convex hull

Continuum removal is a technique that allows the extraction of characteristic absorption bands on the reflectance spectrum curves [29]. The convex hull forms a polygon connecting the outermost points within the sample while ensuring that all the internal angles of this polygon are less than 180 degrees, forming the smallest convex shape that encloses all the points in a given set [30]. The continuum-removed data was obtained using a Python script by [28].

3.2.2. Reflectance to absorption, Smoothing the data and calculate the First and Second derivatives

The absorption was calculated from the data, after the continuum removal, using a function implemented in Python, as mentioned in [7], based on reflectance using the logarithmic relationship. After the conversion to absorption, the data was smoothed by applying the Savitzky-Golay filter from the savgol_filter function in the scipy.signal module in Python. To smooth the signal data, the Savitzky-Golay filter calculates a polynomial fit of each window based on polynomial degree and window size. The obtained smoothed data is used to calculate the First and Second derivatives by applying the np.gradient function from the NumPy library.

3.2.3. Convert waveform to spectrogram

This step consists on the application of the Short-Time Fourier Transform (STFT) to the equal_length tensor using tf.signal.stft. A new dimension is added to the spectrogram tensor using spectrogram [..., tf.newaxis]. This dimension is included to the data to make the spectrogram suitable as input with convolution layers in the CNN. This step is essential to convert the two-dimensional data in a three-dimensional data, such structure is required for the application of a CNN. This methodology can be replicated by the application of the Python script available in the supplementary materials (Code S1).

3.3. Application of Convolutional Neural Network

In the present work, to deal with limited computational capacity, a MobiletNet model was used. MobileNets are a class of highly efficient CNN models, built upon a streamlined architecture that leverages depth wise separable convolutions, being a deep neural network with significantly reduced computational demand [31]. The model was implemented using the open-source libraries of TensorFlow and Sklearn and is available in the supplementary materials (Code S2). The model was tested for Sb and other elements, As, Pb, Mn, Zn, that also are not directly detected in the SWIR-VNIR spectral range and have exhibit different Person’s correlation with Sb in the study area (Table 2). The metrics for evaluate the model performance consist in the analyzes of the R² and RMSE obtained.

4. Discussion

Despite achieving relatively good R² values, the presence of significant overfitting weakens the reliability and generalizability of the model's predictions in the present study. Additionally, the observation that, incorporating additional elements into the model training process did not lead to improved results or mitigate overfitting, further underscores the challenge of addressing this issue. This suggests that simply increasing the complexity of the model or incorporating more features does not necessarily yield better performance and may exacerbate overfitting instead. Kemper and Sommer [7] to resolve the overfitting issue, used a methodology to degrade the spectra considering the band center and the full-width half-maximum. The band center refers to the central wavelength or position of a spectral feature or band of interest. This approach was not possible to replicate in the current study, because the band center for the target element is unknown. Also, in their study area, they have a big contrast between the soil of the region, and the contaminated soil, which was from a mine dump and had a high concentration of heavy metals. In the present study, there is no big contrast between the soils that present Sb and the soils that don’t, and the Sb contents are relatively discrete. Wu, et al. [33] found the correlation with total Fe, active and residual, was a major predictive mechanism for heavy metals in soils. Also, OM and clay have a correlation. The soil analysis in the current study didn’t include those proprieties, which can be a way to obtain better results.

As the soil sampling campaigns executed focused in capturing the general distribution of the Sb, they do not capture the progressive increment in Sb in mineralized zones. Moreover, many soil samples capturing anomalous values of Sb are sourced from the mine tailings existent in the region, and their properties may not be representative. Another sampling methodology, focusing on the soils near the known Sb veins and on the progressive contents of Sb in the soils associated with the veins could be more appropriate for this study and can work as a solution for the overfitting in the CNN model. And, also, obtaining organic matter and clay data from the soils can be a different approach that can help to better understand the features in the spectral signature of soils containing Sb that can be employed for its identification.

5. Conclusions

This study found varying concentrations of Sb in the sampled area, with the higher values of Sb influenced by the historical mining activities and potential contamination. The implementation of a CNN with low computational demand, MobileNet model, for predicting Sb values shows promising results with a good fit for the training data, but with issues in to generalize to new data. However, challenges emerged regarding its ability to generalize to new data. Notably, preprocessing steps remain essential for enhancing model performance. Alternative sampling methodologies and the increment in the dataset available, as the incorporation of another soil proprieties such as OM and clay into the analysis could provide more insides to the topic. This study provides insights into the application of deep learning models to predict Sb concentrations using spectral data, while there are still challenges to overcome.