Applying Knowledge-Based and Data-Driven Methods to Improve Ore Grade Control of Blast Hole Drill Cuttings Using Hyperspectral Imaging

Abstract

This study introduces a novel method utilizing hyperspectral imaging for instantaneous ore-waste analysis of drill cuttings. To implement this technique, we collected samples of drill cuttings at regular depth intervals from five blast holes in an open pit gold mine and subjected them to scanning using a hyperspectral imaging system. Subsequently, we employed two distinct methods for processing the hyperspectral images. A knowledge-based method was used to estimate ore grade within each sampled interval, and a data-driven technique was employed to distinguish the ore and waste for each sample interval. Firstly, leveraging the mixed mineralogical composition of the samples, the Linear Spectral Unmixing (LSU) technique was utilized to predict ore grade for each sample. Additionally, the Gradient Boosting Classifier (GBC) was used as an efficient data-driven approach to classify ore-waste samples. Both methods rendered accurate results when they were compared with results obtained through laboratory X-ray diffraction (XRD) analysis and gold assay analysis for the same sample intervals. Adopting the proposed methodology in open pit mine operations can significantly enhance the process of grade control during blast hole drilling. This includes reducing costs, saving time, minimizing uncertainty in ore grade estimation, and establishing more precise ore-waste boundaries in resource block models.

1. Introduction

1.1. Ore Grade Control in Open Pit Mines

Mining consists of several linked unit operations, and the performance of the initial stages (drilling and blasting) preconditions the downstream unit operations in the production cycle [1]. In metallic ore deposits with heterogeneous grade distribution, ore grade control is considered a critical process. It ensures that the material fed to the mill is of economic grade and reconciles the ore reserve block model to identify and manage variance between the planned and actual ore grades. The process is essential for production planning, ore blending, operational management, and mineral processing. Ore grade control based on non-representative samples can lead to ore losses and dilution [2], reconciliation problems [3,4], and ultimately increasing project risk and leading to significant economic repercussions [5,6,7].

In conventional practice, 3D block models of ore reserves are typically constructed using diamond drill core samples and reverse circulation (RC) drilling data. The core and RC samples are generally spaced out, so the resulting ore reserve models are too rough to meet day-to-day operational needs for grade control. As a result, blast hole cutting samples are used to refine the reserve model for short-term production planning. The manual collection and laboratory analysis of these samples are time-consuming, costly, and prone to inaccuracies [2,8,9]. This necessitates more efficient and reliable methods for real-time analysis.

Relatively few solutions have been proposed to overcome the limitations of manual blast hole cuttings assay analysis. Most of the proposed methods provide a near real-time blast hole assay analysis and allow for analyzing a larger amount of drill cuttings, thus reducing sampling variance and inaccuracy. Charbucinski et al. [10] employed prompted gamma neutron activation (PGNAA) for in situ estimation of copper grade along blast holes. Most recently, pulsed fast thermal neutron activation (PFTNA) technology has also been applied for in situ elemental analysis of drilled blast holes. However, the application and accuracy of this technique for various types of mineral deposits are still unknown. Ackermann et al. [9] developed a new approach for the real-time analysis of drill cuttings to detect ore through the application of laser-induced breakdown spectroscopy (LIBS). Niemeläinen E. [2] introduced a new sampling and analysis method for the automatic representative sampling of cuttings at each depth and instant assaying using line X-ray-fluorescence analysis. Most recently, Fernandez et al. [11] studied the integration of X-ray fluorescence (XRF) analyses and measurement-while-drilling (MWD) records in an underground mine with different mineralogical characteristics.

Recent advances in ore-waste discrimination techniques based on imaging and image processing have garnered significant attention. RGB color image processing involves analyzing the visible light spectrum (red, green, and blue) to identify mineralogical features and textures. This method is not too expensive and simpler compared to hyperspectral imaging. RGB images can be processed using computer vision techniques to detect color patterns and contrasts, which can be correlated with different ore types. Navarro et al. [12] introduced a combined strength-grade factor to assess rock strength and waste/ore classification based on analyzing rock strength properties, assaying of drilling chips and 3D UAV reconstructions of the post-blast bench face using RGB images. Shatwell et al. [13] presented a real-time, ore-sorting algorithm based on image processing in a gold and silver mine. They employed a high-resolution photography camera to acquire RGB color images. Khomiac and Benndorf [14] explore using RGB images to forecast geometallurgical properties, enhancing the mine-to-mill process by providing a cost-effective alternative to traditional lab analysis. They evaluated various image segmentation methods to automatically extract and analyze ore particles and mineral content from RGB images.

1.2. Hyperspectral Imagining in Ore Grade Control

Hyperspectral sensors can capture reflectance spectra across a wide range of the electromagnetic spectrum. Their application in mining has been expanded across different scales ranging from large-scale, satellite-based imagery to close-range laboratory scale and from mineral exploration to mapping alteration zones on pit wall exposures, detecting acid mine drainage and ore grading and sorting [15,16]. Given the high potential of hyperspectral imaging for identifying various materials based on their distinct spectral signatures, several studies have been conducted to implement this technique in discriminating between ore and waste material [17,18,19,20]. Dalm et al. [17] investigated the usage of NIR-SWIR hyperspectral imagery for discriminating between ore and waste by characterizing alteration mineralogy at epithermal gold deposits. Lypaczewski et al. [20] discovered that detecting the phengitic white mica by hyperspectral images correlates with gold grade and can be used to discriminate between unmineralized, weakly mineralized, and highly mineralized drill core rock samples.

The main goal of this study is to develop a hyperspectral remote sensing approach to better control ore grades in an operational mine using real-time analysis of drill cuttings. This study uses hyperspectral images to examine ore/waste discrimination in blast hole drill cuttings. The primary focus of our research is to propose an innovative, automated, and practical approach to the classification and mineral mapping of drill cuttings with high accuracy.

In this study, data-driven and knowledge-based classification techniques are employed to assess the efficiency and robustness of hyperspectral imaging in differentiating between ore and waste materials. Data-driven methods focus on utilizing the inherent patterns and statistical properties present in the hyperspectral data to extract meaningful information. Abdolmaleki et al. [18] employed supervised and unsupervised learning techniques to classify hyperspectral images obtained from drill cores to distinguish ore and waste samples. Choros et al. [19] investigated the application of hyperspectral imaging to discriminate ore and waste during excavation, utilizing supervised learning methods. Knowledge-based approaches integrate spectral properties and expert insights to analyze hyperspectral data, often relying on spectral libraries and spatial models [21,22,23,24].

This work applies the Gradient Boosting Classifier (GBC) as a data-driven method and Linear Spectral Unmixing (LSU) as a knowledge-based approach to classify ore and waste using hyperspectral images of drill cuttings. Recently, Manian et al. [25] assessed various machine learning methods for classifying hyperspectral images and found that GBC provides superior performance in supervised classification. In another successful application, Loggenberg [26] applied the Extreme Gradient Boosting method to classify hyperspectral images. Boosting algorithms have also performed well in multispectral data for geological mapping and exploration [27]. LSU has also been extensively used for ore and mineral mapping, as demonstrated in various studies [22,23,24].

Our study represents an original contribution to developing and evaluating data-driven and knowledge-based methods to address significant challenges in real-time ore grade control in mining operations using drill cuttings samples. Several of the methods presented in this paper have been reported in the literature, but the paper’s novelty lies in its pioneering use of hyperspectral scanning on fine-grinded drill cut samples, a departure from the typical focus on large rock samples presented in previous studies. Hyperspectral imaging and analysis of fine-grinded particle sizes of drill-cutting samples differ from those of large rock samples. Hyperspectral imaging of intact rock surfaces preserves their spatial context, while spectral imaging of grinded rock samples analyzes a homogenized sample. This can result in reduced spectral variability compared to larger, more heterogeneous rock samples, where distinct clusters of pixels represent different minerals. This innovative approach not only allows for a better understanding of ore and waste distribution in a small-scale mine bench but also paves the way to better control the ore loss and dilution in an operating mine. In addition, the paper presents a novel absorption-peak-based feature extraction approach to extract the most reliable and necessary spectral information for the classification of hyperspectral images. The frequency of these absorption peak features observed in ore and waste samples was used to develop the data-driven algorithm for sample classification. This novel data-driven approach is specifically designed for drill cutting samples, significantly improving the accuracy and efficiency of mineral resource management.

This study uses a gold open pit mine as a case study. Our dataset is an original hyperspectral image of the significant size taken from multiple drill cutting samples obtained from blast holes in the mine. The ore body is a Carlin-type gold deposit with extremely irregular and heterogeneous mineralization, highlighting the critical importance of the study. The results of the proposed hyperspectral methods are validated using the measured gold assay and XRD analysis along the blast hole intervals.

2. Study Area and Sampling

This study was conducted at the Gold Bar open pit mine in the southern Roberts Mountains in Eureka County, NV, USA. All the mineralization at the Gold Bar North (GBN) occurs as sediment-hosted, “Carlin-type” gold deposits. The deposits are hosted in well-bedded limestone of the Devonian Nevada Formation and are characterized by micron-sized gold and a distinct hydrothermal alteration characteristic [28,29]. Lower plate Paleozoic-aged carbonates comprise approximately 30% of the surface exposure and host all the known deposits in the GBN area. The Bartine Member of the McColley Canyon Formation is the main stratigraphic unit containing most of the significant mineralization. The bedded limestone has good primary porosity and lateral permeability that allows the movement of hydrothermal fluids [30]. Figure 1 shows the location of the Gold Bar mine and the local lithological map.

Mineralization in the GBN deposits is closely related to decalcification and, to a lesser extent, silicification along high-angle structures [31]. Carbon and calcite have commonly been remobilized. Calcite veins are typically found in the vicinity of mineralization. The decalcified rock can be either unoxidized (carbonaceous) or oxidized. Primary pyrite/marcasite and arsenopyrite generally replace iron-bearing minerals and form dissemination in unoxidized host rocks. The more intensely decalcified zones in the mineralized areas correlate well with higher gold grades [30].

Based on geological studies conducted at GBN, gold mineralization in oxidized ore is correlated with clay minerals (mostly illite), silicification, and iron oxides in limestone host rocks. The waste, however, mostly includes carbonaceous limestone, unmineralized host rock, and sulphide [30,32]. The mineralogical study of the ore deposit demonstrates that illite is the primary clay mineral and is identifiable in all the sedimentary units. In Carlin-type deposits, the amount and composition of illite are variable and can be used as a vector for the location of ore-bearing fluid conduits [33].

The mining operation at the Gold Bar mine employs a drilling and blasting process for rock fragmentation and extraction. Blast holes of 6.84-inch (17.37 cm) diameter are drilled by rotary drilling machines in a 15 ft × 19 ft (4.6 m × 5.8 m, burden × spacing). The blast holes have a 20 ft (6 m) depth, of which 2 ft (61 cm) are sub-drill (bench height of 18 ft—5.48 m). Figure 2a shows a plan view of the blast hole patterns for a blast location. The red polyline shows the boundary of the mineralized zone. The same figure shows the location of the five blast holes that were sampled for this study (BH156, BH167, BH177, BH186, and BH188). The drilling in BH188 was conducted to confirm that the area is entirely waste and outside the mineralization zone. This confirmation helps validate our classification model by providing a clear distinction between ore and waste zones. Figure 2b presents a cross-section of the five blast holes where the location of Yellowboy fault ore zone (thickness of 20–30 ft ≈ 6 m to 9 m) has been identified based on the information from the reserve block model. Figure 2c,d show a top view and an isometric view of blast holes drilled on a bench at the Gold Bar mine, respectively. The same figures show the pile of drill cuttings beside each blast hole that needs to be sampled by the ore control team for grade estimation and reconciliation. These holes are spaced about 15 feet (4.6 m) apart, except BH188, which was drilled in the unmineralized zone (See Figure 2a,b).

Throughout the drilling process of the five vertical blast holes, samples of drill cuttings were extracted at two-feet intervals along each hole. A total of 10 samples were collected for each hole, as depicted in Figure 3. No sample could be recovered from the first 2 ft (61 cm) of BH177. The collected 49 samples were analyzed at the mine using the Cyanide Assay Value (ppm) for gold assay analysis. In total, 49 assays were measured, representing the collected samples from the five blast holes. A statistical summary of the assay analysis has been provided in

Table 1. Statistical analysis of gold assay results for drill cutting samples.

Statistical Measures	BH 188	BH 156	BH 167	BH 177	BH 186
Mean (ppm)	0.11	1.79	1.18	1.08	0.13
Median (ppm)	0.09	1.72	1.01	0.99	0.05
Standard Deviation (ppm)	0.08	0.91	0.64	0.81	0.16
Range (ppm)	0.27	3.19	2.00	2.22	0.46

. The gold cut-off grade at the mine is 0.17 ppm. BH188 exhibited the lowest average gold grade at 0.11 (ppm) and a relatively small standard deviation of 0.077 ppm. Except for one sample, all the samples in this blast hole fell below the cut-off grade. Conversely, BH156 demonstrated the highest average gold grade of 1.79 ppm, with a large standard deviation of 0.911 ppm, indicating a significant variation in gold grades along the hole. Blast holes 167 and 177 had similar averages and standard deviations, with BH167 having a slightly higher mean (1.18 ppm) and BH177 having a slightly higher standard deviation (0.807 ppm). The average grade of BH186 (0.125 ppm) is above the cut-off grade and has a relatively low standard deviation (0.161 ppm). Based on the measured gold assays, only the samples collected from the uppermost two intervals of this hole were located within the mineralization zone.

Geological studies conducted at the mine indicate that the ore mineralogy of the GBN deposits consisted of an intermixing of oxidized and unoxidized refractory ores with variable leach recovery. Unoxidized mineralization occurred as pods within oxidized ore and as larger, deeper zones where supergene weathering processes had not yet reached [30]. Unoxidized refractory ore with black carbonaceous mineralization frequently carried high gold grades but had historically exhibited refractory and preg-robbing characteristics. The metallurgical evaluation concluded that this material should be mined as waste independent of gold grade as it could inhibit gold recovery from cyanide-amenable oxide ores [30].

3. Materials and Methods

3.1. Dataset

The dataset used in this research includes 49 samples collected from five blast holes. Each sample represents a two-foot depth of drill cuttings. All samples were scanned using Specim SisuROCK™ Oulu, Finland, a hyperspectral scanner with two spectral cameras capturing the reflectance range from the VNIR (400 to 1000 nm) to the SWIR (900 to 2500 nm). The camera was positioned at a height of 35 cm from the samples, and halogen lights covering the VNIR and SWIR range were used to ensure uniform illumination and high-quality spectral data. The scanner had an integration time of 2 ms with a spectral resolution of 6 nm and a spatial resolution of 200 spatial pixels in VNIR and 320 spatial pixels in the SWIR camera. Each pixel contains 88 VNIR and 256 SWIR bands. Figure 4 shows the true and false-color composite of hyperspectral VNIR and SWIR images of the scanned drill cuttings, respectively.

3.2. Hyperspectral Image Processing

This study investigated the feasibility of using hyperspectral imaging data to classify ore and waste samples and predict gold mineralization along the blast holes. The hyperspectral images of the drill cuttings were analyzed using ENVI-v6.0 (environment for visualizing images), a specialized multi-hyperspectral processing software and programming in Python-v3.12. The pre-processing of images involved several key steps. First, radiometric correction was applied to ensure accuracy in the captured data. Then, radiance data were converted to reflectance data by subtracting the dark current and normalizing the radiance measurement of the target to that of a Spectralon™ white panel. Finally, background removal was performed to isolate the relevant spectral information. These pre-processing steps were essential to prepare the data for subsequent analysis. Two approaches were implemented to detect the desired features along each blast hole. Figure 5 demonstrates the workflow of data-driven and knowledge-based approaches used in this study. The LSU was employed as a knowledge-based method to map the endmembers identified across the blast hole samples. In the data-driven approach, after pre-processing of the collected images, an absorption peak extraction method was applied to spectra, followed by a filtering analysis, which removed insignificant peaks, and the remaining peaks were utilized as inputs for developing the classifier model. A supervised classification method (GBC) was employed as a data-driven technique to discriminate the ore/waste classes based on training data.

3.2.1. Linear Spectral Unmixing

Unlike hyperspectral images of solid rock samples, which may consist of multiple clusters of pixels with similar spectra belonging to various constituent minerals, the drill cutting samples include fine particles of rock intercepted by the drilling bit. Therefore, each pixel of the hyperspectral image of drill cuttings may include a few endmembers. Based on the nature and characteristics of drill cuttings, using spectral unmixing methods helps to decompose the measured spectrum into a collection of constituent spectra. Spectral unmixing algorithms have been extensively implemented to decompose the mixed pixel spectra into their constituent endmember fractions. The LSU technique assumes that the measured spectrum of each pixel can be represented as a weighted average of the spectra of endmembers, as depicted in Figure 6 [34].

It is recognized that LSU can be considered both a knowledge-based and a data-driven approach, depending on the context and the implementation details. Asadzadeh et al. [35] mention that LSU is a data-driven approach when the endmembers are derived directly from the data without prior knowledge. In this case, various algorithms are used to extract endmember spectra from the hyperspectral image itself. Techniques like Principal Component Analysis (PCA), Independent Component Analysis (ICA), or algorithms specifically designed for endmember extraction (e.g., N-FINDR, Pixel Purity Index) are utilized to identify pure pixels or representative spectra from the dataset. However, in some cases, such as in this study, known spectral libraries or laboratory measurements of pure material spectra are used to perform the unmixing. The known endmember spectra guided the unmixing process by providing a reference against which the mixed pixel spectra are decomposed. Therefore, in our work, we used LSU as a knowledge-based technique.

Mateer [32] collected spectral endmembers from 57 chip trays, chipboards, and core samples from a single diamond drill hole at the Gold Bar mine site using the Terraspec Spectrometer. This was completed based on changes in lithology and alteration of the collected samples; most of the collected spectra also belonged to repeating endmembers. Representative spectra from the current study are shown in Figure 7. The gathered spectra were employed as candidate constituent endmembers within the framework of the LSU method. These endmembers represent distinct components present in the mining region and are crucial for decoding individual constituents from the complex blend of drill cuttings signals.

The spectroscopic investigations by Mateer [32] indicated that the increase in gold concentration in different samples is associated with organic carbon, quartz, illite, and kaolinite minerals. The most common mixed spectral signature is illite and calcite. This is not surprising as lithologies are primarily carbonates, siltstones, and mudstones. Wackstone, packstone, and grainstone are the most common carbonate classifications in the drill holes. An attempt has been made to categorize mixed illite-calcite spectral signatures based on these classifications. Mudstones and siltstones are a mix of clay and quartz with local carbonate. As a result, an unaltered mudstone spectrum should be illitic with possible secondary quartz and calcite. Preliminary results show a positive correlation between mudstones and their corresponding spectral signature. Illite is the primary clay mineral in the limestone and dolomite units. Illite was identified as normal, high aluminum, and ammoniated. Of the 22 samples analyzed, none contained ammonium, but the highest gold anomalies contained high aluminum illite. When illite is accompanied by silicification, it becomes higher in aluminum [32,33].

3.2.2. Gradient Boosting Classifier

Supervised classification is the dominant technique frequently employed for quantitative image data analysis. Basically, it translates the sensor measurement space into a field of classified labels [36]. In this study, a GBC, as a supervised classification algorithm, was employed to map the measured spectra of drill cuttings to ore/waste classes. This process encompasses three primary stages: peak extraction, training and classifying. Figure 8 illustrates the main stages of the classification procedure, beginning with image pre-processing. This includes general pre-processing steps such as radiometric correction and background removal. Following this, the procedure continues with peak extraction, filtering, and classification, which will be discussed in detail subsequently.

• The proposed absorption-based peak extraction and GBC

The study proposed a novel absorption-peak-based feature extraction approach to extract the most reliable and necessary spectral information. The extracted absorption features were used to develop the ML classifier model rather than using the results of linear spectral decomposition. In addition, previous studies mainly focused on pixel-based ML classification, which allocates a class of either ore or waste to a pixel. This approach is typically practical for satellite-based hyperspectral data in which the spatial resolution ranges from a few centimeters to meters. The pixel-based classification is inappropriate for the close-range hyperspectral data collected in this study, which has spatial resolutions ranging from a micrometer to millimeter scale.

To identify key absorption peaks in our dataset, we first used the first derivative analysis to identify all absorption peaks by detecting changes in slope from negative to positive. This method helps highlight the peaks accurately. Given the large size of our dataset, it is important to streamline it for efficiency. Therefore, we employed a second derivative analysis to reduce the number of absorption peaks, making the dataset more manageable. Additionally, we refined our analysis by applying further filters to the results of the second derivative analysis. This included setting a reflectance value threshold, typically at 0.95, to differentiate reliable absorption peaks. Peaks below this threshold were retained as primary features in each pixel’s spectrum, while those above were excluded from further analysis. These selected features served as the foundation for our subsequent analysis and interpretation. This methodological approach of peak extraction, data reduction, and refinement through filtering allowed us to focus on the most relevant spectral features, facilitating a deeper understanding of our dataset.

• Training

In machine learning, sufficient training samples are crucial for achieving high-performance and reliable results [37]. However, in this study, only 49 trays were available, which is insufficient for effective machine learning training. To address this limitation, instead of using the conventional upsampling method, each sample was divided into four parts, with each sub-sample considered an independent sample, resulting in 196 samples. Each tray (the original rock sample domain) represents two feet in length of the borehole data. The utilized upsampling approach assumes that each part belongs to one fourth of the mentioned length. Afterward, the proposed feature extraction was applied to each subsection and further used in the ML development process. One problem that could arise when using borehole data is split bias. This may occur when one datum point is allocated to the training set and its neighboring point to the testing set. To eliminate the split bias, instead of a random split, it was tried to keep all four sub-parts of one sample either in the training or testing dataset. This data augmentation approach helped increase the amount of training data. For this work, approximately 70% of the images were assigned to the training dataset, which comprised 33 trays representing 132 (33 × 4) tray parts. The remaining 30% of the images were designated as the test dataset, consisting of 16 trays and 64 (16 × 4) tray parts. This partitioning was applied to both ore and waste classes, ensuring that both sets contained representative samples. Figure 9 illustrates the distribution of the train and test datasets. The test dataset was selected from two specific boreholes: BH188 and BH156. BH188 was chosen because it is spatially distant from other boreholes, offering unique spatial perspectives within the dataset. Additionally, BH156 was selected because it contains trays with higher grades compared to other boreholes. This selection process considers the potential influence of both spatial relationships and grade variations within boreholes on the dataset. As a result, a non-random approach was used to ensure that the train and test datasets represent meaningful spatial and grade-related variations for effective model training and evaluation.

• Gradient Boosting algorithm

Gradient Boosting is a robust machine learning technique used for regression and classification tasks, employing an ensemble of decision trees to create a strong learner [38]. The process begins with an initial shallow decision tree labeled as the base model. Subsequently, new models are fitted to the residuals (errors) generated by the previous models, with each new model aiming to correct or reduce the errors. This iterative approach updates predictions step by step, utilizing gradient descent optimization to minimize a logistic loss function at each stage [39,40].

Gradient Boosting can effectively handle complex datasets with non-linear relationships and outliers, making it a proficient method for distinguishing ore from waste. Non-linear relationships between input features and target variables are common in geological datasets, and the discrimination of ore versus waste may involve complicated patterns and interactions between variables.

The application of the GBC in this study involved a meticulous process of cross-validation and hyperparameter tuning to achieve optimal model performance. Randomized Search Cross-Validation was employed to fine-tune the hyperparameters of the GBC, a method that efficiently explores a broad spectrum of parameter combinations within a manageable computational timeframe.

Table 2. Used ranges and the obtained optimal values from the fine-tuning process.

Parameter	Range	Optimal Value
Number of estimators	10–1000	845
Minimum samples required to split a node	2–20	3
Minimum samples required to be at a leaf node	2–20	14
Maximum depth of the individual trees	1–50	27
Fraction of features considered	0.1–0.9	0.20

presents the used ranges and the obtained optimal hyperparameter values from the fine-tuning process. The search space for these parameters was defined using random distributions. The Randomized Search Cross-Validation was executed with 100 iterations, utilizing a 5-fold, cross-validation to validate model performance. The evaluation metric chosen was the F1 Score. This rigorous tuning approach ensured that the GBC model was well calibrated for the training dataset, thereby enhancing its predictive accuracy and robustness.

3.2.3. Geochemical and Mineralogical Analysis

As mentioned earlier, all collected samples were analyzed at the mine using the Cyanide Assay Value (ppm) with a detection limit of 0.01 ppm gold for gold assay analysis. To highlight the reliability of our findings, 10 samples were sent to a commercial lab for XRD analysis. The analysis was conducted by SGS Canada Inc. using a BRUKER AXS D8 Advance Diffractometer. The test conditions included Co radiation at 35 kV and 40 mA, with a LYNXEYE_XE_T detector. Interpretations were made using the PDF2/PDF4 powder diffraction databases issued by the International Center for Diffraction Data (ICDD) and Diffrac Eva-v7.0 software. The detection limit ranged from 0.5% to 2%, strongly dependent on crystallinity.

4. Results

The 49 drill cutting samples were analyzed using knowledge-based and data-driven techniques to identify the gold mineralization along the five blast holes.

4.1. LSU Outcomes

The LSU method used knowledge and information about mineralization and endmembers obtained from the Gold Bar North deposit to classify ore and waste samples from the drill cuttings. Figure 10 presents the mapping of endmembers for each collected sample along the drill holes. This map provides valuable insights into the composition and distribution of materials within each tray. The same figure shows a bar chart of gold grade (ppm) for each sample interval. As can be seen, carbonaceous limestone (green) significantly emerged in the waste sample trays, and illite-calcite (red) appeared mostly in the ore trays.

The relationship between the LSU-mapped endmembers and gold mineralization is statistically discussed as follows. Figure 11 shows the relationship between the overall percentage of illite-calcite and quartz spectra observed in each tray sample and the gold assay results for the tray sample. The trendline illustrates an association between these spectra and the gold grade. The measured data exhibit a non-linear relationship and are not normally distributed, so Spearman correlation analysis is used to investigate the correlation between variables. The Spearman correlation coefficient ranges from −1 to 1. The results are shown in

Table 3. Spearman correlation coefficients between illite-Calcite + Quartz and gold assay.

Parameter	Value
Spearman Coefficient (Rs)	0.59
Number of samples	49
T	5.02
p-value	7.88 × 10−6

. Based on the results, it can be concluded that there is a statistically great moderate positive correlation (R_s = 0.5907) between the variables. A very small p-value (7.88332 × 10⁻⁶) suggests strong evidence against the null hypothesis of no correlation between the variables. This observation aligns consistently with well-established geological knowledge at the mine, as previously discussed in Section 3.2.1. Hence, the presence of illite-calcite and quartz spectra in the processed images can be used as reliable indicators for potential gold mineralization. Further investigation and analysis are warranted to confirm and expand upon these findings, which could potentially offer valuable insights into the underlying processes driving gold mineralization.

To bolster the reliability of our findings derived from LSU techniques, ten samples were selected for XRD analysis, as highlighted by the red rectangles in Figure 10.

Table 4. XRD analysis results for selected samples along with the fire assay gold analysis (Green: Ore, Red: Waste).

Minerals	Unit	BH 188	BH 156	BH 156	BH 167	BH 167	BH 177	BH 177	BH 177	BH 186	BH 186
		2–4	2–4	12–14	6–8	18–20	6–8	8–10	18–20	0–2	10–12
Calcite	%	56.9	6.9	74.5	38.0	65.0	49.2	71.5	33.5	48.6	75.3
Quartz	%	11.7	70.0	12.4	43.7	23.1	35.3	16.0	37.9	37.7	7.4
Dolomite	%	21.9	12.2	8.1	10.9	5.2	9.1	7.5	13.0	6.8	11.1
Muscovite	%	1.5	5.8	1.0	3.7	3.1	1.5	2.3	5.6	2.3	2.9
Epidote	%	2.4	-	1.3	-	-	0.9	0.7	1.4	1.3	-
Ankerite	%	-	1.1	0.3	1.0	0.7	0.9	-	0.9	0.5	1.9
Microcline	%	1.5	-	0.5	-	-	0.7	-	1.1	0.7	-
Albite	%	0.7	0.4	0.2	0.5	0.6	0.5	0.5	0.5	-	-
Goethite	%	-	0.6	0.2	0.3	0.3	0.2	0.3	0.5	0.1	-
Anatase	%	0.2	0.2	0.1	0.1	0.1	0.1	0.1	0.2	0.1	0.1
Barite	%	-	-	-	-	-	0.3	-	0.5	-	-
Biotite	%	-	0.7	-	-	-	-	-	-	-	-
Pyrite	%	-	-	-	-	0.2	-	-	0.2	0.2	-
Fluorapatite	%	-	0.1	-	-	-	-	-	0.2	-	-
Illite	%	2.8	1.6	1.4	1.8	1.3	1.0	1.2	4.1	1.5	1.3
Kaolinite	%	0.4	0.4	-	-	0.3	0.3	-	0.4	0.3	-
Gold assay	ppm	0.067	2.84	0.664	2.296	1.756	2.417	1.039	0.344	0.409	0.036

represents the results of the XRD analysis for the selected samples. Additionally, the last row of the table displays the results of the fire assay gold analysis for the samples. Two waste samples and eight ore samples were selected to cover various mineral compositions in drill cuttings.

It is important to note that XRD analysis identifies minerals in the bulk of the sample and may not directly correlate with surface measurements from hyperspectral mapping. The overall composition determined by XRD can differ from the surface-specific findings of hyperspectral mapping. Although XRD confirms the presence of minerals identified by hyperspectral mapping, the exact percentages of these minerals can differ due to the bulk vs. surface nature of the measurements. In all selected samples, the presence of quartz, calcite, and illite has been confirmed via XRD analysis in agreement with hyperspectral mapping results. Samples BH188-4 and BH186-12, situated within the waste zone, show a dominant percentage of calcite and dolomite in XRD analysis, consistent with mapped carbonaceous/unmineralized limestone using hyperspectral processing. In contrast to the waste zone samples, samples identified as ore exhibit a significant overall amount of calcite and quartz, with quartz being particularly abundant. This observation suggests a robust correlation between quartz content and gold assay values. Samples with higher quartz content tend to exhibit higher gold assay values. For instance, sample BH156-4, with the highest quartz content (70.0%), also has the highest gold assay value (2.84 ppm). This indicates that quartz could be a significant indicator mineral for gold mineralization in agreement with LSU outcomes.

Although the direct link between illite content and gold assays is not immediately apparent in the XRD dataset, geological investigations suggest a positive relationship between illite-calcite and gold mineralization. Considering the limestone host rock, a significantly higher proportion of calcite in the samples is expected to be found compared to illite. However, as XRD analysis does not quantify the percentage of illite-calcite, distinguishing between calcite originating from the unmineralized limestone host rock and calcite associated with illite-calcite linked to mineralization poses challenges. This also explains the mixed relationship between calcite content and gold assay values. For example, sample BH156-14 has a high calcite content (74.5%) but a relatively low gold assay value (0.664 ppm), while sample BH186-2 has a lower calcite content (48.6%) but a slightly lower gold assay value (0.409 ppm).

Statistical analysis was conducted on the LSU outcomes, revealing that the majority of the drill cuttings categorized as waste contains between 15–95% mapped carbonaceous limestone, with less than 19% illite-calcite or quartz. Thus, a 15% carbonaceous limestone spectra threshold can effectively characterize all the waste trays except one waste tray (BH 188, depth interval 8–10 m), as shown in

Table 5. Misclassified samples by LSU method.

Blast Hole	Depth	LSU Classification Status	Gold Assays	Carbonaceous Limestone (%)
	Interval		(ppm)
BH 167	0–2	misclassified as waste	0.293	26.92%
BH 177	12–14	misclassified as waste	0.379	17.02%
BH 156	12–14	misclassified as waste	0.664	15.2%
BH 188	10–12	misclassified as waste	0.327	26.78%
BH 188	8–10	misclassified as ore	0.085	6.93%

. In addition, using the same threshold, four ore trays were wrongly identified as waste (Table 5). The misclassified ore trays had a gold assay below the 30th percentile, indicating that the LSU method faces challenges in separating low-grade ore from waste. The overall accuracy of the LSU method was calculated to be 89.7% (see

Table 6. Confusion matrix and accuracy assessment of LSU method; Green: Ore, Orange: Waste, Purple: Misclassified.

Confusion Matrix				Accuracy Assessment
True Labels	Ore	28 TP	4 FN	Class	Overall Accuracy	Precision	Recall	F1-Score
	waste	1 FP	16 TN	Function	(TP+TN)/N	TP/(TP+FP)	TP/(TP+FN)	2TP/(2TP+FP+FN)
		Ore	waste	Ore	89.7	0.96	0.90	0.92
	Predicted Labels			Waste		0.80	0.94	0.86

for the confusion matrix and accuracy assessment). The accuracy rate was computed by dividing the number of correct predictions by the total number of predictions [41]. In this case, 44 out of 49 trays were predicted correctly.

The classification model demonstrates high overall accuracy in distinguishing between ore and waste, with particularly high precision and recall for the ore class. The metrics indicate that the model is effective in identifying ore, with fewer false positives and false negatives. For the waste class, while the recall is high, indicating that most actual waste samples are correctly identified, the precision is slightly lower, suggesting a higher number of false positives. The F1 scores for both classes show a good balance between precision and recall, highlighting the robustness of the model in practical applications.

4.2. GBC Outcomes

GBC was utilized as a classification method to differentiate between ore and waste samples, with a gold cut-off grade set at 0.17 ppm. The dataset comprised 132 training samples and 64 test samples, all categorized into ore and waste classes. The extracted absorption peaks are utilized as input features for classification. Figure 12 displays the classified output using the GBC on the testing dataset. Among the 64 test samples, the classifier misclassified three samples initially predicted as waste but were ore (highlighted in purple), which all belong to one tray (drill cutting sample obtained from BH188, 10–12 ft) with a grade of 0.33 ppm, considered as a low-grade tray (close to cut-off grade).

Additionally, the classifier accurately predicted the classes for 61 samples: 41 as ore (highlighted in green) and 20 as waste (highlighted in orange).

The performance of the classifier was evaluated through a comprehensive accuracy assessment using accuracy, precision, recall, and F1 score. In classification results, these metrics provide insights into different aspects of model performance and help understand how well the model is performing in terms of correct predictions and capturing relevant instances of each class. Further information on the evaluation indices can be found here [42,43,44].

Classification results demonstrate excellent performance in classifying ore with perfect precision (all predicted ore samples are correct) and high recall (correctly identifies most of the ore samples). For waste classification, while the precision is slightly lower (87%), the recall is good (100%), indicating that although there are some false positive predictions, all actual waste samples are correctly identified by the model. The F1 scores for both classes (0.96 for ore and 0.93 for waste) demonstrate a good overall balance between precision and recall, highlighting the model’s effectiveness in discriminating between ore and waste.

In the context of mining operations, the primary goal is to accurately identify ore to maximize recovery while minimizing the misclassification of valuable material as waste. Therefore, achieving perfect precision for ore classification is crucial, as it ensures that all predicted ore samples are indeed correct. High recall for ore is also essential, as it means most ore samples are correctly identified, reducing the risk of ore loss. Conversely, while high recall for waste classification ensures that all actual waste samples are identified, it is acceptable to have slightly lower precision for waste (87%), which indicates some false positives (waste samples mistakenly classified as ore/dilution). This balance reflects the operational priorities in mining, where avoiding misclassifying ore as waste takes precedence.

The spectral features of the three misclassified sub-samples from BH188, 10–12 ft, shown in Figure 12, were carefully inspected to determine possible reasons for the misclassification. First, the Feature Importance function was applied to the GBC model to identify the most significant spectral features. Figure 13 shows the important spectral features for distinguishing ore from waste. The five most important features are 2200 nm, 2213 nm, 1918 nm, 2207 nm, and 1150 nm.

The frequency of absorption peaks at these wavelengths was calculated for the three misclassified sub-samples of this blast hole interval. It was compared with the correctly classified sub-sample from the same tray and with the average frequency of absorption peaks for ore and waste samples. Figure 14 provides a graphical comparison of these key features. The green sub-sample is the one which has been correctly classified as ore. The three purple color sub-samples are the misclassified samples from the same tray. The analysis revealed that the frequency of absorption peaks for all of the four sub-samples at the key features are similar except for 1150 nm. For the correctly classified ore sample (green), the absorption peak frequency at 1150 nm was close to the average frequency of ore samples (red dash line), while the misclassified sub-samples had absorption peak frequencies at 1150 nm resembling those of waste samples (Blue dash line). This can be attributed to the inhomogeneous distribution of the features in this low-grade tray.

5. Discussion

Hyperspectral analysis techniques using the LSU and GBC methods were demonstrated to be valuable tools for discriminating between ore and waste of blast hole drill cuttings. The LSU approach requires extensive knowledge of mineralization and alteration and the examination of the spectral signatures of various minerals present in the study area. This method enables the identification of different mineralization components within the drill cuttings and provides valuable information about the abundance of various minerals along each blast hole. This information can be correlated with gold mineralization and can provide additional insight regarding the metallurgical characteristics of the mineralized zone (e.g., potential for preg-robbing). It should be mentioned that preg-robbing in gold leaching is a phenomenon where certain minerals (e.g., certain clay minerals, sulphide minerals) or organic materials (e.g., graphite or organic matter) present in the ore adsorbs gold cyanide complexes, thus reducing the amount of gold available for extraction by the leaching process and diminishing the overall efficiency of gold extraction.

For the Gold Bar North ore deposit, the coexistence of illite-calcite and quartz was found to be correlated with gold grade in the drill cutting samples. This correlation can be used as a proxy for rough estimation of gold grade along the blast holes, which provides more information about the quality and extent of the mineralized zone, which is essential for resource reconciliation.

In addition to the LSU method, the GBC method was employed to classify the drilling cuttings based on extracted spectral absorption peaks from hyperspectral data. This machine learning approach considers the probability distribution of each class, namely ore and waste, and assigns each sample to the class with the highest probability. The evaluation of accuracy metrics affirms the robust overall performance of the classification model. Compared to the LSU method, GBC represents a data-driven approach that considers the variations and uncertainties inherent in the data. This results in a more accurate classification, making it a more efficient method for distinguishing ore and waste. The GBC method achieved an accuracy rate of 97%, surpassing the accuracy rate of 89.8% obtained by the LSU method.

It is recognized that various surface properties of minerals, such as texture, roughness, moisture content, and weathering state, can significantly influence their spectral characteristics. Each of these factors could indeed be the focus of separate, detailed studies. However, in this research, the focus is on drill cuttings, which typically range in size from millimeters to centimeters. These samples retain surface properties such as texture and composition largely from their original drill cutting context. Notably, the spectra used in the study are local spectra derived directly from these samples rather than from a standard library. This consistency in mineral surface characteristics between field samples (drill cuttings) and laboratory analysis supports the validity of the findings in comparing spectral methods for ore and waste discrimination. The study is conducted in a controlled laboratory environment to understand the fundamental spectral responses of minerals under specific conditions. The algorithms and methodologies developed in this phase are applied to laboratory-scale data. For future work, the research will be extended to field conditions, where the effects of various factors in real-world mining scenarios will be investigated.

The proposed method offers a more detailed estimation of ore distribution and grade along the blast holes rather than relying on a single aggregate assessment. This approach enhances the delineation of the boundary between ore and waste along the hole and allows for the exclusion of sub-drill samples (the final part of each hole) from the average bench grade estimation. Integrating samples from the sub-drill portion, which pertains to materials in the lower bench, can potentially skew the average ore grade estimation for the blasting bench, either boosting or diluting it.

In numerous large-scale open pit operations, every blast site may encompass hundreds of blast holes. Manual sampling and labeling of drill cutting samples from these blast holes are both laborious and time-consuming. Additionally, sending these samples to a lab for assay analysis can incur considerable delays of days or weeks, along with significant costs. The proposed method holds promise in automating this process by providing comprehensive ore and waste characteristic information. The realization process for instantaneous analysis and real-time grade estimation involves using the developed classifier model directly while collecting drill cutting samples during blast hole drilling. This approach ensures that the classification results are promptly available, enabling immediate decision-making and optimizing the efficiency of mining operations. The practical implementation includes integrating the classifier with portable analytical tools to provide on-the-spot analysis, thereby facilitating rapid and accurate grade estimation. Nevertheless, the reliability of this approach must be assessed across various ore deposits to ensure its robustness.

6. Conclusions

In traditional manual drill-cut sampling techniques, only a small sample is manually collected, resulting in a substantial reduction of the overall drill cutting mass. However, leveraging hyperspectral imagery, the proposed method can markedly enhance the representativeness of sampling by processing a substantial portion of drill cuttings at frequent intervals along the blast hole. Furthermore, by implementing a real-time sampler/processor during drilling, results can be promptly accessible, allowing for a precise definition of the ore/waste boundary within the blast holes.

The Linear Spectral Unmixing method, fed by endmembers collected across the Gold Bar North deposit, was used to explore the correlation between the different endmembers and measured gold assays. The results showed that gold mineralization was significantly associated with the presence of illite-calcite and quartz in drill cuttings. Given the low-grade and heterogeneous nature of gold mineralization, a deep understanding of the important characteristics and different types of mineralization is critical for enhancing the performance of LSU as a knowledge-based method.

The use of the Gradient Boosting Classifier as a data-driven technique resulted in a more accurate and precise ore/waste classification than the LSU method. The F1 score for the GBC method was calculated to be 96% for the ore class and 93% for the waste class, demonstrating the strong performance of this model. With the exception of a single tray, all trays in the test dataset were precisely classified using this method. These results provide valuable insights into the relationship between pixel labels and gold grade measurements, showcasing the potential of image analysis techniques for predicting or assessing the presence of valuable minerals like gold in drilling cuttings.

Hyperspectral analysis, including knowledge-based (LSU) and data-driven (GBC) methods, demonstrates great promise for ore/waste discrimination and classification in blast hole drilling cuttings. Incorporating these techniques into mining operations can greatly enhance decision-making and production planning in the industry. Further investigations and analyses are required to enhance our understanding of these relationships and to develop more accurate predictive models for real-time ore grade estimation. Addressing challenges such as field conditions and data variability will be essential for the successful deployment and practical application of these methods in operational mining environments. For example, the accuracy of the Linear Spectral Unmixing (LSU) method depends heavily on the selection of appropriate endmembers. Any inaccuracies or biases in endmember selection can impact the reliability of the results. In future work, we also aim to develop regression models for predicting gold content, which typically requires a larger number of samples for effective training and testing. Although the use of hyperspectral imagery allows for a substantial portion of drill cuttings to be processed, the study was limited by the number of samples analyzed. A larger dataset would be beneficial to validate the findings and improve the robustness of the predictive models. Expanding our dataset will enable us to build and validate robust regression models to predict gold content with higher accuracy.