Artificial Intelligence-Based Hyperspectral Classification of Rare Earth Element-Related Heavy Mineral Sand

Abstract

Heavy minerals, such as Rutile, Ilmenite and Zircon, and other essential trace elements are important in modern technology development. The integration of hyperspectral imaging and artificial intelligence presents a promising approach for the accurate identification of heavy minerals, especially Rare Earth Element (REE)–bearing phases such as Monazite. This study evaluates three AI classifiers, Support Vector Machine (SVM), Neural Networks (NNs) and Convolutional Neural Networks (CNNs), for their performance in classifying ten different minerals distributed across six grain size groups ranging from 125 μm to over 300 μm. The analysis focuses on how grain size affects spectral reflectance and classification accuracy. Among the tested models, SVM consistently outperformed NN and CNN, achieving the highest precision, recall and spectral similarity, particularly within the 150–300 μm grain size range. CNN showed the lowest performance and frequently misclassified spectrally similar minerals, such as Zircon and Rutile, likely due to its 1D architecture and limited spatial representation. Monazite, notable for its strong Nd³⁺ absorption features, was accurately identified across applicable grain sizes, highlighting its reliability for REE detection. Spectral Angle Mapper (SAM) analysis confirmed that SVM and NN maintained better spectral similarity than CNN. In general, the results highlight the significant influence of grain size, spectral similarity and dataset size on classification accuracy and the overall effectiveness of AI models in hyperspectral mineral analysis.

1. Introduction

Heavy mineral sands are an important resource that sustains the infrastructure of modern-day technology. These sands often contain high concentrations of economically valuable minerals, such as Rutile, Ilmenite and Zircon, which currently are considered key commodities for industries such as aerospace, electronics, energy and ceramics, as stated by [1,2]. In addition to these, heavy mineral sands may host Rare Earth Element (REE)-bearing minerals such as Monazite that particularly gained global interest due to their importance in clean energy technologies, high-performance magnets and defense systems, according to [3,4]. The heavy mineral sands used in this experiment closely resemble those found in Mozambique, which contain high grade Ilmenite, Rutile, Zircon and Monazite. Among known global deposits, the heavy mineral sands found in Mozambique are particularly notable for their exceptionally high concentrations of total heavy minerals, making them one of the most economically significant deposits worldwide [5].

Accurate and efficient identification and quantification of these resources are essential for effective resource evaluation, geological exploration and mining operations. Conventional techniques such as chemical composition and spectral analysis have been employed to accurately identify these minerals. Refs. [6,7,8] describe laboratory methods such as X-ray Diffraction (XRD), Scanning Electron Microscopy (SEM) and Inductively Coupled Plasma Mass Spectrometry or Optical Emission Spectroscopy (ICP-MS/OES) that are often time consuming, expensive and poorly suited for rapid or large-scale assessments. Such limitations have prompted the integration of Hyperspectral Imaging (HSI) in geosciences, which is a rapid, non-destructive technique and simultaneously captures spectral and spatial information [9].

HSI systems capture reflectance data over hundreds of contiguous spectral bands, enabling detection of subtle mineralogical differences. However, the high dimensionality and spectral redundancy inherent in such data pose significant challenges for accurate interpretation and classification, often requiring the use of advanced feature selection or dimensionality reduction techniques, as highlighted in [9]. This has led to increased adoption of Artificial Intelligence (AI), particularly Machine Learning (ML) and Deep Learning (DL) algorithms, for automated mineral identification. AI has been successfully applied in areas that cover mineral processing, mineral prediction, mineral exploration mapping, chemical exploration anomaly mapping, geological mapping, core drilling mapping and mineral phase segmentation of X-ray microcomputer tomography data [10]. On the other hand, machine learning techniques have reportedly been applied successfully in lithological classification [11], mineral mapping [12,13,14] and core logging automation [15]. These applications demonstrate the potential of ML to improve both the efficiency and accuracy of mineral exploration processes.

Ref. [16] categorized machine learning into five distinct paradigms, each with specific implications for hyperspectral mineral classification:

i.

Supervised Learning: Relies on labeled datasets for training and has been widely used in hyperspectral image classification. Algorithms such as Support Vector Machine (SVM), Artificial Neural Networks (ANNs), random forests and decision trees fall into this group. These models use pre-defined input–output mapping to learn classifications effectively. In recent mineral identification studies, SVM has widely been recognized for its robustness in high-dimensionality feature spaces [17], which demonstrated superior classification performances [18,19].

ii.

Unsupervised Learning: Works with unlabeled data and identifies clusters or structures based on inherent features. This method is helpful when ground truth labels are not available, though it may struggle with mineral classification due to spectral similarities.

iii.

Semi-Supervised Learning: Combines a small labeled dataset with a larger unlabeled one to improve learning efficiency. Ref. [18] highlights that deep networks trained with semi-supervised learning can significantly enhance feature extraction and classification when labels are sparse.

iv.

Active Learning: Iteratively queries the most informative unlabeled samples to be manually labeled, thereby reducing the overall labeling effort. This approach is useful in geological datasets where expert labeling is expensive or time consuming.

v.

Transfer Learning: A machine learning technique that uses knowledge learned from one domain and applies it to another, often requiring minimal additional training. This is particularly relevant in geoscience, where labeled hyperspectral datasets are often scarce, but related domains can offer transferable spectral patterns.

Each paradigm offers unique benefits and trade-offs depending on the dataset structure, label availability and computational constraints.

Deep leaning approaches, especially Convolutional Neural Networks (CNNs), have gained attention due to their capacity to process spatial and spectral information simultaneously. CNNs learn hierarchical feature representations, making them suitable for classifying hyperspectral images with complex textures. However, CNNs also face challenges such as the risk of overfitting on small datasets and the lack of spectral generalizability when spatial context is underutilized, as described by [20,21]. This necessitates careful model selection, preprocessing and validation when applying deep leaning models to hyperspectral tasks.

A major challenge in hyperspectral analysis lies in its dimensionality, where numerous redundant or irrelevant spectral bands can reduce classification accuracy. This problem is compounded by the scarcity of high-quality labeled samples, which increases the risk of overfitting and limits the generalizability of models. These challenges are particularly critical in the identification of REE-bearing minerals such as Monazite, where subtle spectral features must be distinguished from spectrally similar phases.

Feature selection methods and band reduction techniques attempt to address this issue but may introduce new biases. Deep learning partially mitigated these problems by integrating automatic feature learning and dimensionality reduction within the model architecture itself, according to [22]. To improve efficiency and scalability, recent studies have incorporated parallel computing (CPU/TPU acceleration), ensemble learning and hybrid machine learning, deep learning frameworks for hyperspectral classification [23]. CNN, in particular, can exploit spatial–spectral correlations better than traditional ML models but only when appropriately trained with sufficient and balanced datasets [24]. As refs. [10,25] emphasized, small and uneven datasets can lead to overfitting, reducing the generalizability of deep learning models.

This study investigates the classification of heavy minerals, particularly REE-bearing Monazite, from hyperspectral images using AI-based classifiers. It evaluates the performance of SVM, NN and CNN in distinguishing ten minerals across six particle size groups. The main objective is to assess how grain size affects spectral behavior and classification accuracy. It is stated that particle size has a significant impact on reflectance [26,27], where finer particles generally exhibit higher reflectance due to increased surface scattering, while larger particles reflect less, likely due to stronger internal absorption. By examining comparative performance to traditional machine learning and deep learning models, this research contributes to advanced non-destructive mineral identification methods, with potential applications in exploration, environmental monitoring and remote sensing.

2. Related Work

In mineral exploration, accurate mineral identification remains essential. Conventional techniques typically involve manual identification based on physical attributes, such as shape and color, or laboratory-based chemical analysis using X-ray Diffraction (XRD), Scanning Electron Microscopy (SEM) or spectroscopy. Notwithstanding, these methods are normally labor intensive, time consuming and costly, especially when applied on large datasets or remote environments. Hyperspectral Imaging (HSI) has emerged as a transformative tool in geoscience, enabling detailed mineralogical mapping by capturing spatial and spectral information simultaneously [8,28,29]. Because of the high dimensionality and complexity of hyperspectral data, classical image processing and shallow machine learning techniques face limitations in scalability and performance. The growing accessibility of HSI data has sparked considerable interest in using Artificial Intelligence (AI) for automated mineral classification.

Feed forward-neural networks, such as Convolutional Neural Networks (CNNs), have demonstrated high efficiency in pixel-based classification, anomaly and target detection and lithological discrimination in hyperspectral images [30,31]. Recent developments in AI, particularly in machine and deep learning, have enabled new opportunities in hyperspectral mineral classification. These architectures can automatically extract hierarchical features from spectral data and have proven especially robust in scenarios with overlapping spectral responses. Deep Neural Networks (DNNs), with fewer handcrafted features and deeper layers, learn directly from raw data and scale more effectively to complex geological environments, reducing dependence on manual feature engineering [32,33].

According to [10], AI models for mineral identification can be grouped into three main categories: (i) Artificial Neural Networks (ANNs), which often depend on expert curated datasets; (ii) machine learning models, which include statistical and rule-based algorithms; and (iii) deep learning models, which use hierarchical feature learning to approximate human like decision making. Deep learning approaches, such as CNN and ResNet, are increasingly being favored due to their ability to learn directly from unstructured data without domain specific heuristics. For instance, ref. [10] demonstrated that Resnet-based classifiers achieve over 91% accuracy in classifying minerals from lower resolution imagery, offering a reliable alternative where high quality data is not available.

Further development of these algorithms has resulted in specialized models such as Mineral-ResNet and CNN-LSTM hybrids. Ref. [31] compared these models against traditional ML methods, such kNN, SVM, LR, RF, XGB and CatBoost, reporting classification accuracy exceeding 90% using raw hyperspectral data. Their approach minimized preprocessing steps like feature selection or augmentation, simplifying the pipeline and reducing bias. Moreover, Ref. [34] applied CNNs to classify both mineral types and grain size distributions in hyperspectral images, optimizing training using Adam and stochastic gradient descent techniques. Ref. [35] introduced a modular CNN framework tailored for thin section classification, indicating the flexibility of deep learning in imaging resolutions.

To enhance classification performance for REE-bearing heavy mineral sands, recent studies have incorporated transfer learning, adaptive feature selection and hybrid model stacking. Ref. [36] introduced a Deep Residual Network (D-ResNet) for beach sand mineral exploration, highlighting its ability to identify Ilmenite, Monazite and Zircon with high spatial fidelity. On the other hand, Ref. [37] explored transfer learning for its effectiveness in processes involving limited labeled data. They highlighted the need for robust domain adaptation, especially in remote or poorly mapped geological environments. Equally, ref. [38] proposed a segmentation classification pipeline that integrated hybrid spectral spatial features for grain level classification in placer deposits. These innovations address key challenges such as spectral confusion, dimensionality reduction and spatial heterogeneity.

Despite these advancements, several limitations remain. Spectral mixing limited annotated datasets and lack of benchmarking datasets that restrict the generalizability of models [29]. Traditional machine learning algorithms, such as decision trees or SVM, often lack the capacity to model the high intraclass variability found in natural mineral samples, as described by [6]. Rule-based methods, including PCA and PLS, are prone to overfitting and often require extensive preprocessing to remove redundant bands. Deep learning models, though powerful, are constrained by the availability of large-labeled datasets and computational resources. Ref. [39] emphasized the need for robust frameworks capable of domain transfers and uncertainty quantification to scale DL applications across diverse geological terrains.

To address these gaps, this study builds upon state-of-the-art AI frameworks to develop a deep learning-based classification system for hyperspectral identification of heavy mineral sands enriched with rare earth elements. The models emphasize scalable classification with minimal preprocessing, aiming to replace large intensive laboratory workflows with AI-driven alternatives that are both cost effective and geologically available.

3. Materials and Methods

To address the mineral identification challenges posed by high-dimensional hyperspectral data, this study integrates machine learning and deep learning approaches capable of learning discriminative spectral–spatial features for REE-bearing minerals. It proposes an artificial intelligence-driven framework for classifying hyperspectral images of heavy mineral sand, with a particular emphasis on identifying REE-bearing minerals such as Monazite. AI, in this context, refers to both machine learning techniques, such as SVM and NN, and deep learning, which in this case is CNN. Their selection was due to their ability to extract and learn complex spectral patterns from high-dimensional data [40,41], which might enable precise classification of REE-related mineral explorations.

Traditional cameras use three wavelength spectrum bands in the visible range (RGB) to capture data, which limits the acquisition of more detailed spectral information for mineral discrimination. In contrast, hyperspectral imaging offers detailed spectral signatures, making it possible to distinguish minerals with overlapping color profiles and distinct absorption features, as stated by [42]. The development in recent years of hyperspectral cameras with more spectral information has gained space in various fields of geosciences. The hyperspectral data was acquired using the SPECIM IQ camera, which captures reflectance across the Visible and Near-Infrared (VNIR) range of 400–1000 nm wavelength with a spectral resolution of 7 nm, yielding 204 spectral bands and 512 spatial pixels per band. To minimize edge noise commonly present in hyperspectral sensors, the experiment only focused on the 450–950 nm range, assuming that this range balances information content and spectral stability for mineral classification.

The dataset includes ten common minerals found in heavy mineral sands, including REE-related minerals like Monazite and industrially significant minerals such as Ilmenite, Rutile, Zircon and Staurolite as well as other minerals including Epidote, Kyanite, Pleonaste (ferroan spinel), Tourmaline and quartz. To ensure imaging consistency and reduce intraclass variability as a result of grain size or surface texture, the samples were collected and processed in a controlled workflow, following three main steps: crushing raw materials and photographing, hyperspectral data capture and supervised image classification using AI.

The hyperspectral images captured were processed using a MATLAB-based graphical user interface application called APiS (AI-powered intelligence spectrum analyzer) [43] developed from MATLAB version R2023a. Initial steps involved spectral visualization of each mineral to understand their distinct reflectance behavior. The application enables the user to select a working spectral range and apply either machine- or deep learning-based classification workflows. All models were trained using supervised classification, and the data was divided into 90% for training and 10% for testing. The mineral classes were annotated manually.

Three models were evaluated:

i.

Convolutional neural network (CNN): one layer, followed by batch normalization, ReLU activation, global average pooling, a fully connected layer and SoftMax output, was implemented to learn the spectral relationships and optimized for computational efficiency. It was selected for its proven ability to model spatial hierarchies and extra high-level spectral features from raw data [43] (Li, L., Iskander, M. et al., 2023).

ii.

Support vector machine (SVM) was implemented using the Radial Basis Function (RBF) Kernel. This algorithm is particularly effective for high-dimensional feature spaces and has demonstrated strong performance with limited training data, making it robust for small sample hyperspectral applications [42].

iii.

Neural Network (NN): a fully connected feed forward neural network was trained as the baseline deep learning method, requiring fewer computational resources and having the capability to recognize patterns in all spectral dimensions [33].

The model performance was evaluated using accuracy, precision and recall metrics for each class.

Although the camera covers the full 400 to 1000 nm range, the experiment only focuses on the 450–950 nm range to avoid spectral noise and ensure robustness in the classification results while preserving critical absorption features, which, in this case, can be relevant for REE-bearing mineral identification. The use of this range was based on preliminary tests that showed signal degradation at both spectral extremes. The complete workflow ranging from sample preparation to image classification can be visualized in Figure 1 and represents a structured AI-based approach developed to support the identification of heavy mineral sand and the REE-targeted exploration.

3.1. Data Collection

To conduct hyperspectral classification of heavy mineral samples, nine mineral types, Epidote, Ilmenite, Kyanite, Monazite, Pleonaste (Ferroan Spinel), Rutile, Staurolite, Tourmaline and Zircon, were mechanically pulverized into six standardized grain size groups: 125 μm, 150 μm, 180 μm, 250 μm, 300 μm and above 300 μm (Figure 2). The crushing was performed using the Pulverisette 6 mill at 400 rpm, with at least two cycles to ensure adequate fragmentation, particularly for reducing oversized particles.

Quartz was included in the dataset as a reference material but was not subjected to crushing and placed only in the group above 300 μm. Following pulverization, Ilmenite, Kyanite, Pleonaste, Rutile, Staurolite, Tourmaline and Zircon were successfully fractionated into all six target sizes. Meanwhile, Epidote yielded only fractions of 125 μm and greater than 300 μm, while Monazite provided 125 μm and 150 μm grain sizes. All the samples were then prepared for hyperspectral data acquisition and classification analysis.

3.2. Hyperspectral Data Acquisition and Feature Extraction

The grain size fractions were imaged using the hyperspectral camera under controlled conditions at the Resources Management Laboratory in the School of Engineering at Hokkaido University. Each group was photographed within a dark room setup using halogen lights at a 45° angle of incidence to minimize ambient interference and optimize reflectance capture (Figure 3). The camera was positioned 30 cm above the sample to ensure image consistency. Each acquisition involved an integration time of 30 ms and recording time of 45 s. Image quality was validated in real time using the built-in system interface (Figure 4), which displayed histograms to guide capture decisions. In this process, white peaks indicated optimal exposure, whereas blue and red peaks signaled underexposed or oversaturated conditions, respectively. Additionally, the position of the histogram relative to the threshold was used to confirm whether the illumination adequately covered the spectral range. For instance, if the values fell on the left side of the threshold, it indicated insufficient illumination across the wavelength range.

Figure 5 represents reflectance profiles for all ten minerals (including Quartz) in their uncrushed form, establishing baseline spectra. In contrast, Figure 6 displays spectral profiles on the crushed minerals grouped by grain sizes, demonstrating how particle size influences spectral reflectance and absorption, especially in the near-infrared (700–950 nm) region. Spectral similarity between certain mineral groups posed a challenge throughout the classification process, which is believed to be attributable not just to algorithmic limitation but to the inherent spectral overlap between mineral classes. While SVM managed to reduce some of these errors, complete separation remained difficult, reinforcing the importance of high-resolution reference spectral data and potential use of extraction techniques also supported by [44,45,46].

3.3. Classification Dataset Preparation

Each sample group was labeled and organized into twelve distinct classes for supervised classification. The mineral classes included Epidote, Ilmenite, Kyanite, Monazite, Pleonaste (Ferroan Spinel), Quartz, Rutile, Staurolite, Tourmaline and Zircon. Adding to that, two non-mineral classes, background (BG) and case (i.e., the sample container), were also included. The inclusion of BG and case was made in an attempt to reduce the risk of misclassification, particularly due to spectral similarities between these surfaces and certain minerals, where background regions or container surfaces were sometimes incorrectly labeled as mineral classes.

3.4. Model Training and Algorithm Selection

Two different approaches were used for classification.

3.4.1. Classical Machine Learning

The uncrushed mineral hyperspectral signatures were used to train and test 36 machine learning algorithms, including the following:

i.

Decision trees: fine, medium and coarse.

ii.

Discriminant: linear and quadratic.

iii.

Logistic Regression: binary GLM and efficient.

iv.

Naive Bayes: Gaussian and Kernel based.

v.

Support Vector Machine: linear, quadratic, cubic, Kernel-based and Gaussian.

vi.

K-Nearest Neighbors: weighted, fine, medium, coarse, cosine and cubic.

vii.

Neural Network: narrow, medium, wide, bilayered and trilayered.

viii.

Ensemble methods: RUS Boosted, Boosted Trees, Bagged Trees, Subspaces Discriminant and others.

Each model was trained with 90% of the data, reserving 10% for testing using the APiS MATLAB interface. On the two best-performing models, we computed accuracy, precision, recall and F1-score metrics to compare their performance.

3.4.2. Deep Learning (CNN)

Simultaneously, a 1D Convolutional Neural Network (CNN) model was trained separately using the default configuration provided by the software suite (APiS) developed from MATLAB (R2023a) and published in 2024. The performance of CNN was compared with the top classical models to assess suitability for mineral classification under varying grain sizes. It is suggested that 1D-CNN, though suitable for spectral information, may be inadequate in complex mineralogical contexts, particularly when spatial cues are essential for discrimination. As ref. [47] argued, 1D-CNNs do not incorporate spatial information and may struggle with intraclass variability, a pattern observed in this study. Incorporating 2D- or 3D-CNNs or hybrid models could improve performance in future applications.

3.5. Evaluation Metrics and Spectral Angle Mapper Analysis

To assess the rendering of these models, metrics such as precision (the number of predicted classes that were labeled correctly), recall (the number of instances in which a class was classified correctly) and F1-score (the comparison between the two, in other words, the overall classification performance) were computed on the confusion matrices generated from the classification [48].

To further evaluate how each model performed in particle size group classification, the Spectral Angle Mapper (SAM) was computed in each classified image using python. The SAM algorithm finds similarities between the reference spectrum from the ground truth data and the classified image [49]. It provides a better understanding of the quality and accuracy of the classification results based on the classified images. The statistics presented in Table 3 were generated from the classified images in group I. The mean SAM illustrates the spectral similarities between the reference data and the output, and a lower value may represent better classification accuracy. Even though it was computed post-classification, SAM may be used to validate and improve the classification by determining the accuracy and consistency of the classification, including, in this case, the evaluation of model performance.

4. Results and Discussion

4.1. Particle Size Effects on Reflectance and Classification

The primary challenge in hyperspectral data analysis lies in the effective processing and classification of spectral signatures. The reflectance profiles (Figure 6) illustrate the clear relationship between particle sizes and spectral response of different minerals. In general, smaller grain sizes (125 μm) exhibit higher reflectance due to increased surface scattering; meanwhile, coarser grains (>300 μm) reflect less light as a result of stronger internal absorption.

This behavior varies by mineral, where Ilmenite, Tourmaline and Pleonaste consistently demonstrate low reflectance across all sizes. Notably, Tourmaline showed higher reflectance in its uncrushed state, which decreased significantly after pulverization. Kyanite, which exhibited average reflectance prior to grinding, revealed increased reflectance in the crushed form. Epidote, presented in both the 125 μm and above 300 μm groups, had lower reflectance than Staurolite in its ungrounded state but surpassed Staurolite at 125 μm. Above 300 μm, the two minerals showed comparable reflectance. Zircon out-reflected Rutile in the uncrushed form as well as in the crushed ranges between 125 μm and 180 μm.

Figure 6. Reflectance profiles of nine crushed samples across gain size groups (125 μm to above 300 μm). Spectral reflectance is shown for Monazite, Kyanite, Epidote, Rutile, Staurolite, Zircon, Tourmaline, Pleonaste and Ilmenite. Smaller particles (125–150 μm) generally exhibit higher reflectance due to enhanced surface scattering, while particles above 300 μm display increased absorption and reduced reflectance. These trends are specifically pronounced in the Near-Infrared (NIR) region (700–950 nm), where spectral differences impact classification accuracy in AI models. Monazite retained a stable spectral profile across all sizes, with distinctive absorption features at 580 nm, 740 nm, 800 nm and 870 nm related to the presence of Nd³⁺ phases. These features match previous reports [49,50,51] and confirm Monazite’s potential as a reliable REE indicator mineral. Importantly, while many minerals showed altered profiles after crushing (Figure 6), Monazite’s spectral stability persisted, enhancing its classification reliability.

Figure 6. Reflectance profiles of nine crushed samples across gain size groups (125 μm to above 300 μm). Spectral reflectance is shown for Monazite, Kyanite, Epidote, Rutile, Staurolite, Zircon, Tourmaline, Pleonaste and Ilmenite. Smaller particles (125–150 μm) generally exhibit higher reflectance due to enhanced surface scattering, while particles above 300 μm display increased absorption and reduced reflectance. These trends are specifically pronounced in the Near-Infrared (NIR) region (700–950 nm), where spectral differences impact classification accuracy in AI models. Monazite retained a stable spectral profile across all sizes, with distinctive absorption features at 580 nm, 740 nm, 800 nm and 870 nm related to the presence of Nd³⁺ phases. These features match previous reports [49,50,51] and confirm Monazite’s potential as a reliable REE indicator mineral. Importantly, while many minerals showed altered profiles after crushing (Figure 6), Monazite’s spectral stability persisted, enhancing its classification reliability.

These results confirm the inverse relationship between particle sizes and reflectance and highlight that size-dependent spectral shifts can significantly influence classification outcomes, particularly in minerals with overlapping absorption features.

4.2. Model Performance Across Particle Sizes

Quadratic SVM and medium NN consistently outperformed other tested algorithms, achieving accuracies above 95% (

Table 1. Validation and test performance metrics for AI models across particle sizes.

Particle Size (μm)	Model Type	Accuracy % (Validation)	Total Cost (Validation)	Accuracy % (Test)	Total Cost (Test)
125	SVM	95.03	303	96.45	24
	NN	97.39	159	98.82	8
	CNN	74.37
150	SVM	97.06	509	97.29	52
	NN	97.43	445	98.28	33
	CNN	86.16
180	SVM	96.85	510	96.55	62
	NN	96.74	527	97.61	43
	CNN	82.53
250	SVM	96.85	510	96.55	62
	NN	96.74	527	97.61	43
	CNN	82.53
300	SVM	96.85	510	96.55	62
	NN	96.74	527	97.61	43
	CNN	82.53
>300	SVM	96.04	283	96.22	30
	NN	96.77	231	95.47	36
	CNN	59.89

). 1D-CNN models showed variable results, ranging from 59% to 86% accuracy, with performance decreasing for larger particle sizes (>300).

SVM and NN correctly identified most minerals across all sizes (Figure 7, Figure 8, Figure 9, Figure 10, Figure 11 and Figure 12b), with the highest robustness in the 150–300 μm range. Misclassifications were most common among spectrally similar minerals, such as Pleonaste and Ilmenite as Tourmaline or Kyanite as Zircon. CNN (Figure 7, Figure 8, Figure 9, Figure 10, Figure 11 and Figure 12c), on the other hand, identified Monazite, Rutile, Ilmenite and Tourmaline but showed reduced precision, especially for Pleonaste, Ilmenite and Staurolite. Larger particles amplified classification errors due to loss of fine spectral detail and increased shadowing effects.

4.3. Classification Metrics and Spectral Fidelity

F1-score analysis showed SVM consistently exceeding 85% for most classes, including for background and case. CNN exhibited greater variability, with sharp drops for spectrally similar or size-sensitive minerals, such as Pleonaste, Ilmenite and Staurolite. Monazite, Kyanite and Zircon achieved high F1-scores in both models, though SVM remained more robust, as displayed in Table 2. A similar pattern was also reported in general (Table 3), where SVM attained the highest accuracy (97.07%) and balanced F1-score (86.85%). CNN, however, recorded the lowest performance metrics, with reduced precision (73.57%) and recall (70.06%), particularly in the fine-grain classification, such as 125 μm.

Spectral Angle Mapper (SAM) results, displayed in Table 4 (generated from Group I), indicated stronger spectral fidelity for SVM and NN, especially in the range of 150 to 300 μm, where SAM mean values were lowest. CNN models systematically produced higher SAM values, reflecting greater spectral distortion and misclassification. Additional quality metrics such as entropy, PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index Measure) confirmed superior spatial and spectral integrity in SVM and NN results.

Table 2. Comparative classification performance of AI models across grain size groups, reported using precision, recall and F1-score.

125 μm	SVM			NN			CNN
Class	Precision	Recall	F1-Score	Precision	Recall	F1-Score	Precision	Recall	F1-Score
BG	97.71	99.71	98.71	96.87	99.56	98.2	83.87	97.52	90.13
Case	85.06	83.93	84.49	90.63	98.34	94.33	46.88	10.81	17.45
Epidote	81.94	88.06	84.87	88.65	94.57	91.51	57.58	80.85	67.23
Ilmenite	87.85	82.82	85.26	90.05	88.37	89.2	48.84	44.21	46.41
Kyanite	96.74	98.47	97.6	97.09	99.17	98.12	84.27	96.02	89.73
Monazite	93.02	100	96.37	93.06	100	96.39	41.51	100	58.67
Pleonaste	89.6	91.83	90.7	92.84	95.17	93.99	39.68	47.62	43.29
Rutile	94.23	97.34	95.76	94.27	98.99	96.58	62.07	72.02	66.65
Staurolite	84.67	75.76	79.97	91.21	83.63	87.24	23.33	7.61	11.42
Tourmaline	91.68	93.31	92.48	95.21	96.45	95.83	63.64	71.43	67.31
Zircon	97.41	99.36	98.38	98.49	99.82	99.15	74.91	88.49	81.11
150 μm	SVM			NN			CNN
Class	Precision	Recall	F1-Score	Precision	Recall	F1-Score	Precision	Recall	F1-Score
BG	99.57	99.95	99.76	99.56	99.98	99.77	93.11	99.8	96.34
Case	92.74	96.09	94.39	95.75	99.25	97.47	86.83	84.13	85.46
Ilmenite	88	87.11	87.55	87.36	81.08	84.11	56.29	53.37	54.8
Kyanite	97.37	98.95	98.15	97.97	98.8	98.38	94.26	95.27	94.76
Monazite	95.45	100	97.66	93.61	100	96.7	87.19	100	93.2
Pleonaste	89.37	89.71	89.54	90.38	89.78	90.08	24.12	29.09	26.35
Rutile	95.94	94.69	95.31	96.49	97.38	96.93	75.78	73.78	74.77
Staurolite	88.66	83.42	85.96	86.55	77.33	81.67	60.58	46.67	52.76
Tourmaline	91.2	95.61	93.35	91.77	96.25	93.95	48.44	77.02	59.61
Zircon	98.89	99.71	99.3	97.59	99.71	98.64	81.47	89.29	85.2
180/250/300 μm	SVM			NN			CNN
Class	Precision	Recall	F1-Score	Precision	Recall	F1-Score	Precision	Recall	F1-Score
BG	99.41	99.91	99.66	99.54	99.98	99.76	94.61	98.59	96.56
Case	91.37	93.95	92.64	95.73	98.28	96.99	83.88	84.53	84.2
Ilmenite	90	89.03	89.51	87.83	85.18	86.49	63.35	46.15	53.41
Kyanite	97.57	98.8	98.18	97.93	99.05	98.49	88.46	87.5	87.98
Pleonaste	89.85	90.6	90.22	88.83	88.39	88.61	42.86	51.56	46.79
Rutile	95.18	96.59	95.88	96.62	97.81	97.21	86.73	55.38	67.46
Staurolite	88.26	84.46	86.32	88.7	83.48	86.01	41.84	48.13	44.78
Tourmaline	94.46	96.68	95.56	91.13	94.65	92.85	53.28	54.17	53.72
Zircon	98.33	99.36	98.84	98.02	99.78	98.89	76.35	90.21	82.68
>300 μm	SVM			NN			CNN
Class	Precision	Recall	F1-Score	Precision	Recall	F1-Score	Precision	Recall	F1-Score
BG	98.66	99.87	99.26	98.98	99.94	99.46	80.22	99.86	88.91
Case	93.97	94.83	94.4	97.77	99.28	98.52	76.4	43.88	55.79
Epidote	80.56	88.52	84.36	87.37	88.04	87.7	54.75	59.38	56.96
Ilmenite	86.78	86.84	86.81	88.22	85.56	86.87	43.65	39.57	41.52
Kyanite	95.5	98.69	97.07	96.84	97.87	97.35	90.05	95.78	92.83
Pleonaste	85.47	88.44	86.94	92.02	86.1	88.96	28	10.45	15.15
Quartz	99.29	100	99.64	98.13	100	99.06	100	100	100
Rutile	95	100	97.43	91.99	100	95.81	38.82	37.08	37.93
Staurolite	84.42	78.34	81.26	80.26	77.78	79	4.35	1.96	2.69
Tourmaline	90.53	92.4	91.46	90.79	93.08	91.92	78.18	84.04	80.99
Zircon	95.96	98.46	97.19	95.6	99.28	97.41	76.44	73.04	74.71

Table 2. Comparative classification performance of AI models across grain size groups, reported using precision, recall and F1-score.

125 μm	SVM			NN			CNN
Class	Precision	Recall	F1-Score	Precision	Recall	F1-Score	Precision	Recall	F1-Score
BG	97.71	99.71	98.71	96.87	99.56	98.2	83.87	97.52	90.13
Case	85.06	83.93	84.49	90.63	98.34	94.33	46.88	10.81	17.45
Epidote	81.94	88.06	84.87	88.65	94.57	91.51	57.58	80.85	67.23
Ilmenite	87.85	82.82	85.26	90.05	88.37	89.2	48.84	44.21	46.41
Kyanite	96.74	98.47	97.6	97.09	99.17	98.12	84.27	96.02	89.73
Monazite	93.02	100	96.37	93.06	100	96.39	41.51	100	58.67
Pleonaste	89.6	91.83	90.7	92.84	95.17	93.99	39.68	47.62	43.29
Rutile	94.23	97.34	95.76	94.27	98.99	96.58	62.07	72.02	66.65
Staurolite	84.67	75.76	79.97	91.21	83.63	87.24	23.33	7.61	11.42
Tourmaline	91.68	93.31	92.48	95.21	96.45	95.83	63.64	71.43	67.31
Zircon	97.41	99.36	98.38	98.49	99.82	99.15	74.91	88.49	81.11
150 μm	SVM			NN			CNN
Class	Precision	Recall	F1-Score	Precision	Recall	F1-Score	Precision	Recall	F1-Score
BG	99.57	99.95	99.76	99.56	99.98	99.77	93.11	99.8	96.34
Case	92.74	96.09	94.39	95.75	99.25	97.47	86.83	84.13	85.46
Ilmenite	88	87.11	87.55	87.36	81.08	84.11	56.29	53.37	54.8
Kyanite	97.37	98.95	98.15	97.97	98.8	98.38	94.26	95.27	94.76
Monazite	95.45	100	97.66	93.61	100	96.7	87.19	100	93.2
Pleonaste	89.37	89.71	89.54	90.38	89.78	90.08	24.12	29.09	26.35
Rutile	95.94	94.69	95.31	96.49	97.38	96.93	75.78	73.78	74.77
Staurolite	88.66	83.42	85.96	86.55	77.33	81.67	60.58	46.67	52.76
Tourmaline	91.2	95.61	93.35	91.77	96.25	93.95	48.44	77.02	59.61
Zircon	98.89	99.71	99.3	97.59	99.71	98.64	81.47	89.29	85.2
180/250/300 μm	SVM			NN			CNN
Class	Precision	Recall	F1-Score	Precision	Recall	F1-Score	Precision	Recall	F1-Score
BG	99.41	99.91	99.66	99.54	99.98	99.76	94.61	98.59	96.56
Case	91.37	93.95	92.64	95.73	98.28	96.99	83.88	84.53	84.2
Ilmenite	90	89.03	89.51	87.83	85.18	86.49	63.35	46.15	53.41
Kyanite	97.57	98.8	98.18	97.93	99.05	98.49	88.46	87.5	87.98
Pleonaste	89.85	90.6	90.22	88.83	88.39	88.61	42.86	51.56	46.79
Rutile	95.18	96.59	95.88	96.62	97.81	97.21	86.73	55.38	67.46
Staurolite	88.26	84.46	86.32	88.7	83.48	86.01	41.84	48.13	44.78
Tourmaline	94.46	96.68	95.56	91.13	94.65	92.85	53.28	54.17	53.72
Zircon	98.33	99.36	98.84	98.02	99.78	98.89	76.35	90.21	82.68
>300 μm	SVM			NN			CNN
Class	Precision	Recall	F1-Score	Precision	Recall	F1-Score	Precision	Recall	F1-Score
BG	98.66	99.87	99.26	98.98	99.94	99.46	80.22	99.86	88.91
Case	93.97	94.83	94.4	97.77	99.28	98.52	76.4	43.88	55.79
Epidote	80.56	88.52	84.36	87.37	88.04	87.7	54.75	59.38	56.96
Ilmenite	86.78	86.84	86.81	88.22	85.56	86.87	43.65	39.57	41.52
Kyanite	95.5	98.69	97.07	96.84	97.87	97.35	90.05	95.78	92.83
Pleonaste	85.47	88.44	86.94	92.02	86.1	88.96	28	10.45	15.15
Quartz	99.29	100	99.64	98.13	100	99.06	100	100	100
Rutile	95	100	97.43	91.99	100	95.81	38.82	37.08	37.93
Staurolite	84.42	78.34	81.26	80.26	77.78	79	4.35	1.96	2.69
Tourmaline	90.53	92.4	91.46	90.79	93.08	91.92	78.18	84.04	80.99
Zircon	95.96	98.46	97.19	95.6	99.28	97.41	76.44	73.04	74.71

Table 3. Overall classification performance of AI models (SVM, NN and CNN).

Model	Accuracy	Precision	Recall	F1-Score
SVM	97.07	86.92	87.04	86.85
CNN	87.86	73.57	70.06	70.1
NN	96.40	88.63	88.47	88.54

Table 3. Overall classification performance of AI models (SVM, NN and CNN).

Model	Accuracy	Precision	Recall	F1-Score
SVM	97.07	86.92	87.04	86.85
CNN	87.86	73.57	70.06	70.1
NN	96.40	88.63	88.47	88.54

Table 4. Spectral Angle Mapper (SAM) and reflectance statistics across particle sizes for AI models.

Particle Size (μm)	Model	Mean SAM	Mean Intensity	Std Dev	Median Intensity	Skewness	Kurtosis	Entropy	Edge Density	SSIM	PSNR
125	CNN	0.659	0.530	0.477	0.667	−0.116	−1.903	37,159.137	0.306	0.854	19.659
	NN	0.659	0.530	0.477	0.667	−0.116	−1.903	37,159.137	0.306	0.854	19.659
	SVM	0.625	0.553	0.465	0.667	−0.203	−1.837	52,027.082	0.307	0.807	20.020
150	CNN	0.570	0.584	0.427	0.667	−0.281	−1.639	155,009.620	0.316	0.862	20.144
	NN	0.560	0.594	0.424	0.667	−0.334	−1.599	152,966.480	0.315	0.834	20.120
	SVM	0.560	0.594	0.424	0.667	−0.334	−1.599	152,966.480	0.315	0.834	20.120
180	CNN	0.594	0.561	0.430	0.333	−0.130	−1.718	167,020.220	0.318	0.910	22.010
	NN	0.563	0.595	0.427	0.667	−0.324	−1.623	149,673.390	0.316	0.865	21.686
	SVM	0.563	0.595	0.427	0.667	−0.324	−1.623	149,673.390	0.316	0.865	21.686
250	CNN	0.591	0.565	0.430	0.333	−0.155	−1.714	163,420.980	0.318	0.879	20.578
	NN	0.550	0.607	0.425	0.667	−0.387	−1.578	144,747.860	0.313	0.845	20.379
	SVM	0.550	0.607	0.425	0.667	−0.387	−1.578	144,747.860	0.313	0.845	20.379
300	CNN	0.596	0.559	0.430	0.333	−0.121	−1.721	167,090.450	0.319	0.907	22.100
	NN	0.553	0.610	0.430	1.000	−0.404	−1.587	134,489.800	0.316	0.837	21.084
	SVM	0.553	0.610	0.430	1.000	−0.404	−1.587	134,489.800	0.316	0.837	21.084
+300	CNN	0.629	0.546	0.462	0.667	−0.161	−1.837	60,438.332	0.310	0.850	20.328
	NN	0.673	0.520	0.480	0.667	−0.077	−1.922	32,796.434	0.307	0.836	19.492
	SVM	0.673	0.520	0.480	0.667	−0.077	−1.922	32,796.434	0.307	0.836	19.492

Table 4. Spectral Angle Mapper (SAM) and reflectance statistics across particle sizes for AI models.

Particle Size (μm)	Model	Mean SAM	Mean Intensity	Std Dev	Median Intensity	Skewness	Kurtosis	Entropy	Edge Density	SSIM	PSNR
125	CNN	0.659	0.530	0.477	0.667	−0.116	−1.903	37,159.137	0.306	0.854	19.659
	NN	0.659	0.530	0.477	0.667	−0.116	−1.903	37,159.137	0.306	0.854	19.659
	SVM	0.625	0.553	0.465	0.667	−0.203	−1.837	52,027.082	0.307	0.807	20.020
150	CNN	0.570	0.584	0.427	0.667	−0.281	−1.639	155,009.620	0.316	0.862	20.144
	NN	0.560	0.594	0.424	0.667	−0.334	−1.599	152,966.480	0.315	0.834	20.120
	SVM	0.560	0.594	0.424	0.667	−0.334	−1.599	152,966.480	0.315	0.834	20.120
180	CNN	0.594	0.561	0.430	0.333	−0.130	−1.718	167,020.220	0.318	0.910	22.010
	NN	0.563	0.595	0.427	0.667	−0.324	−1.623	149,673.390	0.316	0.865	21.686
	SVM	0.563	0.595	0.427	0.667	−0.324	−1.623	149,673.390	0.316	0.865	21.686
250	CNN	0.591	0.565	0.430	0.333	−0.155	−1.714	163,420.980	0.318	0.879	20.578
	NN	0.550	0.607	0.425	0.667	−0.387	−1.578	144,747.860	0.313	0.845	20.379
	SVM	0.550	0.607	0.425	0.667	−0.387	−1.578	144,747.860	0.313	0.845	20.379
300	CNN	0.596	0.559	0.430	0.333	−0.121	−1.721	167,090.450	0.319	0.907	22.100
	NN	0.553	0.610	0.430	1.000	−0.404	−1.587	134,489.800	0.316	0.837	21.084
	SVM	0.553	0.610	0.430	1.000	−0.404	−1.587	134,489.800	0.316	0.837	21.084
+300	CNN	0.629	0.546	0.462	0.667	−0.161	−1.837	60,438.332	0.310	0.850	20.328
	NN	0.673	0.520	0.480	0.667	−0.077	−1.922	32,796.434	0.307	0.836	19.492
	SVM	0.673	0.520	0.480	0.667	−0.077	−1.922	32,796.434	0.307	0.836	19.492

4.4. Impact of Training Dataset Structure

Two training strategies were tested: Group I—using only the size-specific reference spectra for each particle size group—and Group II—included all ten reference spectra used at once for every size group. Group I produced more spectrally constant and visually accurate maps. Conversely, Group II yielded slightly higher numerical metrics but introduced more visible misclassification, suggesting inflated accuracy from dataset mismatch. This supports the need for representative, size-specific reference data to ensure reliable classification in AI-based hyperspectral workflows.

4.5. Misclassification Patterns and Spectral Overlap

Persistent misclassification errors were linked to spectral similarity between certain minerals. Common confusion included Pleonaste and Ilmenite as Tourmaline, Kyanite as Zircon and Staurolite as Zircon or Rutile. These errors occurred in both SVM and CNN results but were more pronounced in CNN.

Such misclassifications arise from both algorithmic limitations and the inherent overlap in mineral spectra. Even with high-performing models, complete separations remain difficult without enhanced preprocessing. Advanced spectra feature extraction methods, such as continuum removal or derivative spectroscopy, could help reduce overlap and improve class separability.

4.6. Optimal Conditions for REE-Bearing HMS Mapping

Across all analyses, particle size between 150 and 300 μm provided the best balance between spectral sharpness and surface scattering, resulting in lower SAM mean values, higher classification accuracy and improved visual quality. Sizes outside this range showed increased misclassification, likely due to either spectral distortion in finer grains or reduced reflectance and shadowing in coarser grains.

Monazite’s spectral stability and distinct Nd³⁺ absorption features make it an especially reliable target for AI-driven hyperspectral classification in REE exploration. However, mineral provenance must also be considered, as compositional variability from different source regions can influence spectral behavior and model performance.

5. Conclusions

This study evaluated the performance of artificial intelligence models, such as Support Vector Machine (SVM), Neural Network (NN) and Convolutional Neural Network (CNN), in the classification of rare earth elements bearing heavy mineral sands from hyperspectral images. The results demonstrated that SVM consistently outperformed both NN and CNN (Table 4), particularly in terms of precision, recall and spectral fidelity (Table 2). While NN showed comparable accuracy, its lower precision suggests difficulty in distinguishing spectrally similar minerals. CNN implemented as 1D architecture underperformed across all metrics, likely due to its reliance on large datasets and absence of spatial context in this configuration.

Out of ten tested minerals, Monazite, Ilmenite, Tourmaline, quartz, Zircon, Rutile, Kyanite and Epidote were at least fully or partially identified. Monazite, in particular, showed robust and consistent classification across particle sizes due to its unique Nd³⁺ absorption features, making it a reliable spectral indicator for hyperspectral REE detection. However, Ilmenite, Pleonaste, Rutile and Staurolite often presented overlapping spectral signatures, leading to persistent misclassification, especially in finer grain sizes. This study further confirms that particle size significantly influences reflectance and classification accuracy, with the size 150 to 300 μm range yielding optimal results. Smaller particles tended to absorb more and obscure spectral signatures. Additionally, there is a need for more robust architectures and larger training sets in future works.

In conclusion, this research highlights the potential of machine learning, particularly SVM, for hyperspectral mineral classification in heavy mineral sands. While high classification accuracy was achieved, spectral overlap, grain-size effects and model limitations indicate the need for further exploration into hybrid model architectures, advanced feature extraction and expanded datasets to enhance classification reliability in complex mineralogical environments. To our knowledge, this is one of the first studies to systematically examine the combined effect of particle size and AI model design on hyperspectral classification of REE-related heavy mineral sands. These insights could support future airborne and laboratory hyperspectral exploration workflows, especially in optimizing model selection, sample preparation and data interpretation in REE-targeted surveys.