Integration of β-VAE, DWT, and GMM for Sentinel-2-based mapping of hydrothermal alteration and iron speciation in the Zafarghand area, Iran

Abstract

Remote sensing (RS) technologies have emerged as efficient and cost-effective tools for the rapid identification of hydrothermal alteration zones, thereby supporting mineral exploration. The Zafarghand area in northeastern Isfahan Province, situated within the Central Iran structural zone of the Urmia–Dokhtar magmatic arc, hosts a porphyry system with multiple alteration types, including phyllic, propylitic, argillic, potassic, and silicic zones. This study aimed to develop an unsupervised and data-driven framework for mapping hydrothermal alteration using high-resolution satellite imagery. To achieve this, Sentinel-2 data were processed through a β-Variational Autoencoder (β-VAE) Deep Learning (DL) model, enriched by Discrete Wavelet Transform (DWT), and subsequently clustered with a Gaussian Mixture Model (GMM). The proposed approach successfully identified alteration zones and iron-bearing minerals (Fe²⁺, Fe³⁺, and iron oxides) at a 20-m spatial resolution. Validation using geochemical data showed classification accuracies of 94.74% and 95.74% for the phyllic and propylitic zones, respectively. In addition, buffered pixel-based validation using confusion matrices yielded overall accuracies of 94.5% and 86.9% for the phyllic and propylitic maps, confirming the statistical robustness of the proposed framework. Due to the limited distribution of argillic and silicic alteration, the depth of the potassic zone, and the lack of sufficient geochemical data for their validation, this study mainly focused on identifying phyllic and propylitic alteration along with iron oxides.These findings demonstrate that integrating DL, wavelet analysis, and probabilistic clustering with Sentinel-2 imagery provides a robust framework for detecting hydrothermal alteration zones. While the current study primarily focused on phyllic, propylitic, and iron-oxide zones due to limitations in data availability, the framework is transferable to other porphyry systems, provided that suitable reference data exist for validation.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-36349-w.

Introduction

Over the past decades, the landscape of mineral exploration has shifted fundamentally. As near-surface reserves have been extensively depleted, exploration efforts increasingly target deeper and less accessible deposits^1–3. Remote sensing (RS) has emerged as a cost-effective reconnaissance tool, enabling large-scale evaluation of mineral potential at the early stages of exploration. By reducing initial costs and guiding field investigations, RS plays a crucial role in defining exploration targets. Nevertheless, its effectiveness depends strongly on integration with field and laboratory data, since RS alone cannot provide fully reliable results^3–7.

Porphyry copper–molybdenum (Cu–Mo) deposits are among the most important sources of critical industrial metals. These systems are genetically and spatially associated with hydrothermal alteration zones, which serve as key indicators in exploration^8–10. Alteration zones typically include phyllic, argillic, and propylitic types, characterized by distinct clay minerals and diagnostic spectral signatures in the short-wave infrared (SWIR) region^9,11–17. Such signatures are detectable through RS techniques and can serve as indirect proxies for areas enriched in Cu and by-product critical metals such as cobalt (Co), tellurium (Te), rhenium (Re), and rare earth elements (REEs)^{14– 15}.

Recent advances in satellite missions, particularly Sentinel-2 with its high spatial resolution and multispectral coverage, have opened new opportunities for mapping hydrothermal alteration zones. At the same time, artificial intelligence (AI) and Deep Learning (DL) have become powerful tools for analyzing multispectral and hyperspectral data. These approaches automatically extract hidden and complex patterns, and have shown success in mineral exploration, lithological interpretation, and geological mapping^17–20. However, the implementation of DL methods often requires complex architectures, substantial computational resources, and large volumes of labeled data. Despite notable advancements, the use of advanced generative models—such as the β-Variational Autoencoder (β-VAE)—for analyzing multispectral data remains limited, with most studies focusing instead on hyperspectral datasets. Several recent works have attempted to apply generative models in related settings. For example, a geologically constrained VAE approach has been developed for mineral prospectivity mapping by incorporating geological constraints, while another study demonstrated that synthetic multispectral imagery generated from Sentinel-2 data can serve as effective augmentation where ground-truth labels are scarce^{21– 22}. These examples highlight the potential of generative approaches, although their application to multispectral alteration mapping remains rare. This gap highlights the opportunity to integrate DL with multiscale analysis techniques, such as the Discrete Wavelet Transform (DWT), to enhance feature extraction and clustering performance.

Despite these advances, important gaps remain. The application of generative models like β-VAE to multispectral RS imagery has been limited, with most studies focusing instead on hyperspectral datasets. Similarly, spatial–textural features are often underutilized in mineral exploration, even though their integration with spectral data can improve detection accuracy. In porphyry systems, alteration halos typically extend over scales of hundreds of meters to several kilometers. Therefore, even at the 20 m resolution of Sentinel-2, meaningful spatial–textural information can be extracted, particularly in the transitional areas between the phyllic and propylitic zones. Integrating these spatial features with spectral data provides a more robust basis for reducing classification errors and improving mapping accuracy. Given the 20 m resolution of Sentinel-2, the potential to derive meaningful textural attributes and incorporate them into DL frameworks warrants further investigation. Moreover, despite the adequate spectral capabilities of the Sentinel-2 sensor for detecting iron-bearing minerals and alteration zones, few studies have addressed the accurate and independent discrimination of Fe²⁺, Fe³⁺, iron oxides, as well as phyllic and propylitic zones^11,23–25. Addressing these gaps requires a comprehensive approach that combines spectral and spatial information within an unsupervised learning framework.

The Urmia–Dokhtar Volcanic Arc (UDMA) in Iran is a major metallogenic belt hosting numerous porphyry deposits. Within this arc, the Zafarghand area—located in the northeastern part of Isfahan Province (Fig. 1)—represents a significant porphyry Cu system that has been investigated through RS and geochemical studies since 2006²⁶. Previous research in this area has applied conventional RS techniques using ASTER and Landsat-8 imagery, including band ratios, least squares fitting, Principal Component Analysis (PCA), and matched filtering, to delineate alteration zones²⁷. More recent work has explored advanced statistical approaches such as singularity analysis, fuzzy logic integration, and DWT applied to ASTER data, which improved mapping accuracy^{28– 29}. However, these methods largely rely on classical or statistical algorithms, leaving the potential of modern DL-based frameworks underexplored. Classical approaches such as PCA, Independent Component Analysis (ICA), K-means Clustering, Spectral Angel Mapper (SAM), Matched Filttering (MF), and Suport Vector Machine (SVM) remain widely used and effective in many alteration mapping studies. However, like any method, they may face limitations under conditions of strong spectral overlap or limited reference data.

Fig. 1.

(a) Location of the Zafarghand granitoid massif and the studied area in UDMA, (b) Regional-scale satellite image showing the Zafarghand granitoid massif and surrounding areas, obtained from the Copernicus Open Access Hub (https://www.copernicus.eu/en) and used for spatial reference purposes, (c) Geological map Zafarghand granitoid massif, (d) The location of the study area is shown on a Sentinel-2 Level-2 A image obtained from the Copernicus Open Access Hub (https://www.copernicus.eu/en) image using a RGB = 11-12-8, generated using ENVI software (version 5.6; https://www.nv5geospatialsoftware.com/), where pinkish tones correspond to phyllic alteration zones and greenish tones indicate propylitic alteration zones, and (e) Alteration map of the studied area, showing the location of geochemical sampling points (Green points represent samples from propylitic alteration zones, red points correspond to phyllic alteration, and black points indicate other alteration types such as argillic, potassic, and siliceous); map generation and cartographic visualization were performed using ArcMap software (version 10.8.2; https://www.esri.com).

This study aims to accurately identify and differentiate hydrothermal alteration zones within the Zafarghand porphyry Cu system using Sentinel-2 imagery integrated with DL and statistical clustering techniques. Specifically, we propose a novel unsupervised framework that combines β-VAE for latent feature extraction, DWT for multiscale spectral–spatial enrichment, and GMM for probabilistic clustering. Unlike conventional approaches, this integration enables the precise delineation of phyllic and propylitic zones as well as iron-bearing phases (Fe²⁺, Fe³⁺, and iron oxides). The proposed framework addresses existing limitations and offers a scalable, cost-effective tool for early-stage mineral exploration in complex geological environments.

Regional setting

The Ardestan region is located in the northeast (NE) of Isfahan Province, within the central part of the Urmia–Dokhtar magmatic arc, while its southwestern margin is associated with the Sanandaj–Sirjan zone. Rock outcrops in the area range from the Paleozoic to the Quaternary, but the dominant exposures consist of Tertiary volcanic and pyroclastic rocks of intermediate to mafic composition, which are attributed to Middle to Late Eocene magmatic activity. In addition, several intrusive bodies of semi-acidic to acidic composition have been identified in the region: a semi-deep intrusive phase assigned to the Oligocene–Early Miocene, and a shallow intrusive phase to the Pliocene. These intrusions have played a key role in hydrothermal alteration and mineralization, leading to the development of porphyry- and vein-type deposits containing copper, molybdenum, gold, lead, zinc, and iron. The rock units of the region can generally be classified into three main groups. The first group consists of sedimentary rocks such as limestone, dolomitic limestone, conglomerate, sandstone, shale, and marl. The second group is composed of igneous rocks, including volcanic types such as rhyolite, andesite, and basalt, together with intrusive bodies of granite, diorite, monzodiorite, and gabbro. The third group comprises pyroclastic rocks, represented by tuff, ignimbrite, rhyolitic tuff, sandy tuff, and sheared tuff^30–33.

Structurally, the region is influenced by several major faults with northwest–southeast and east–west orientations, including the Milajerd–Zefreh, Kachumesqal–Genian, and Kuh-e-Doshakh faults. These structures have played a significant role in channeling hydrothermal fluids and in the localization of alteration and mineralization zones. Consequently, different types of mineralization have been reported in the area, including copper at Zefreh, gold at Kuh-e-Dorojin and Kuh-e-Boroni, lead at Kachumesqal, zinc at Shurghestan, and iron deposits in southwestern Ardestan.

Study area

The Zafarghand massif is located within the UDMA, positioned northeast of Isfahan and to the south of Ardestan, as delineated by Nabavi³⁴. Figure 1a illustrates its geographic location and tectonic context within the arc, whereas Fig. 1b displays a satellite image of the Zafarghand granitoid massif together with the surrounding villages^35–38.

The lithological units of the Zafarghand region can be broadly divided into volcanic, volcaniclastic, intrusive, and Quaternary deposits. Volcanic activity began in the Early Eocene and peaked during the Middle to Late Eocene, producing an extensive sequence of pyroclastic and effusive rocks such as brecciated tuffs, sandy tuffs, andesitic to basaltic-andesitic lavas, dacitic and gray rhyodacitic units, as well as rhyolite, rhyodacite, and rhyolitic tuffs. These volcanic formations are widely exposed throughout the area, with outcrops extending for several kilometers. The granitoid massif is cut by numerous dykes, and the volcanic strata are intersected by multiple intrusive bodies of differing ages^{35– 36}.

The intrusive rocks of the study area vary considerably in size and geometry, ranging from small-scale bodies to large plutonic masses. Field observations indicate a wide lithological spectrum that includes granite and alkali feldspar granite, granodiorite, diorite, quartz diorite, gabbro, and gabbrodiorite, as well as aplites, microgranular mafic enclaves, and andesitic dykes. Based on cross-cutting relationships, these dykes can be classified into two generations: the first generation (D1), mainly andesitic in composition, intrudes the older volcanic host rocks but is cross-cut by granitoid intrusions, while the second generation (D2), also andesitic, postdates the formation of the Zafarghand granitoid massif and intrudes directly into it^{35– 36}.

Younger Quaternary deposits consist of reworked clastic sediments derived from older lithologies, along with fluvial channel sediments and deposits associated with alluvial fans and debris flows. Overall, the study area covers approximately 8 km² in the northern part of the Zafarghand granitoid massif, west of Zafarghand village, between 52°23′55″ to 52°26′30″ E and 33°10′30″ to 33°11′52″ N (Fig. 1d).

Hydrothermal alteration

Hydrothermal alteration in the Zafarghand porphyry system is well developed and exhibits a distinct zonal pattern from the center toward the periphery. The potassic alteration zone, characterized by secondary biotite, alkali feldspar, and magnetite, occurs adjacent to the intrusive bodies. Although it is economically significant, this zone is relatively limited in extent. In contrast, the phyllic zone, distinguished by quartz, sericite, chlorite, and iron oxides, is the most widespread type of alteration in the region and represents the principal zone associated with Cu and Au mineralization. The argillic zone, composed of kaolinite and illite with hematite and limonite, commonly overlaps the phyllic alteration. While it is of minor economic importance, it plays a complementary role in understanding the evolution of hydrothermal fluids. The propylitic zone, resulting from the alteration of ferromagnesian and plagioclase minerals in dacitic rocks, is the most extensive alteration type and forms the outer halo of the system, delineating the boundary of the hydrothermal system rather than marking the main locus of mineralization. In addition, silicic alteration occurs locally, overprinting or replacing other zones and is typically represented by quartz veins and veinlets^{36– 37}.

In general, the significance of these alteration zones lies in the fact that each represents a specific physicochemical condition of the mineralizing fluids, and their spatial relationship to the intrusive body can serve as a reliable guide for delineating areas with high mineralization potential. Because the mineral assemblages characteristic of each alteration zone display distinct spectral signatures, they can be effectively detected and mapped using remote sensing techniques. This approach is particularly valuable in mineral exploration, as it enables the identification of prospective areas prior to costly field operations, thereby making the exploration process more focused and efficient^39–41.

Structural control on alteration and mineralization

Structural geology in the Zafarghand porphyry deposit has played a crucial role in the emplacement of intrusive bodies and the localization of mineralization. The intrusion of Miocene granodiorite stocks into Eocene volcanic units (andesite to rhyodacite) not only induced initial thermal alteration, but also, in combination with magmatic mixing processes, generated a network of pervasive fractures and numerous microfractures within the host rocks. These fractures were progressively widened by the increasing pressure of hydrothermal fluids and subsequently reactivated by sudden pressure drops from lithostatic to hydrostatic conditions, becoming the primary conduits for the ascent of mineralizing solutions. Such structural and physicochemical conditions provided a favorable environment for the concentration of phyllic and potassic alterations in the vicinity of the intrusive bodies, whereas the propylitic zone developed mainly at the outer margin of the hydrothermal system^{33,41– 42}.

In addition to microscale fractures, the presence of regional faults with a dominant northwest–southeast orientation (e.g., the South Ardestan fault) and subsidiary east–west structures has played a critical role in controlling fluid flow and focusing Cu and Fe mineralization. These faults not only guided the emplacement of granodiorite intrusions, but also controlled the principal migration pathways of hydrothermal fluids and the localization of alteration zones. The concentration of phyllic and potassic alteration along major fractures and in the vicinity of northwest–southeast faults highlights the decisive role of these structures in creating the physicochemical conditions required for ore deposition. In contrast, argillic and propylitic alteration zones predominantly occur in distal settings, away from the main fluid conduits, and reflect the progressive evolution of the hydrothermal system on a regional scale³³.

As illustrated in the alteration map (Fig. 1e), a clear spatial relationship exists between faults, fractures, and alteration zones. The major faults in the region coincide with the concentration of phyllic and potassic alteration, while propylitic zones occur as outer halos surrounding them. This evidence suggests that the network of faults and fractures, in conjunction with the emplacement of granodiorite intrusions, played a fundamental role in controlling hydrothermal alteration and localizing porphyry mineralization in the Zafarghand area.

Mineralization

Mineralization in the Zafarghand porphyry deposit formed as a result of the intrusion of Miocene granodiorite bodies and associated hydrothermal processes, and is of significant geological and economic importance. Two principal stages of mineralization are recognized: hypogene and supergene. The hypogene stage is represented by pyrite, chalcopyrite, molybdenite, and bornite, occurring as veins, veinlets, and disseminations within both volcanic and intrusive units closely associated with the porphyry intrusions. In the supergene zone, secondary minerals such as chalcocite, covellite, malachite, and azurite were produced through the interaction of meteoric waters with primary sulfides. In addition, oxide and hydroxide minerals including magnetite, hematite, and goethite have been identified at surface outcrops. Mineralogical and geochemical evidence indicates that the Zafarghand intrusion belongs to the I-type and magnetite series. The abundance of hematite and magnetite, coupled with the absence of pyrrhotite, reflects oxidizing magmatic conditions that facilitated the transport of substantial amounts of Cu into the upper crust. These features, together with processes such as the replacement of magmatic biotite by bornite and the development of covellite along the margins of chalcopyrite crystals, played a key role in the formation of sulfide mineralization and subsequent supergene enrichment^{42– 43}.

Mineralization in this deposit is closely associated with alteration zones. The potassic and phyllic zones show the strongest relationship with sulfide mineralization, where pyrite, chalcopyrite, and bornite are most abundant. In contrast, argillic and propylitic zones are mainly developed at the periphery of the system and are associated with weaker mineralization. At the surface, the development of oxidized and supergene zones has resulted in the formation of secondary Cu minerals such as malachite and azurite. Collectively, this evidence suggests that the Zafarghand deposit represents a typical example of an oxide-type porphyry Cu deposit (OPCD) within the Urmia–Dokhtar magmatic belt^{44– 45}.

Materials and methods

Material

This study employed Sentinel-2 multispectral imagery, which provides 13 spectral bands spanning wavelengths from 440 to 2200 nm. The Sentinel-2 constellation, comprising Sentinel-2 A and Sentinel-2B, captures imagery in the visible, near-infrared (NIR), and shortwave infrared (SWIR) ranges, with four bands at 10 m, six at 20 m, and three at 60 m resolution (Fig. 2). The MSI sensor design—including Band 8 A (865 nm)—enables effective detection of iron oxide minerals and atmospherically corrected surface reflectance (Level-2 A products) suitable for spectral analysis^46–50.

Fig. 2.

The spectral and spatial attributes of the Sentinel-2 MSI bands at 10, 20 and 60 m resolutions.

The overall methodological workflow is illustrated in Fig. 3. It outlines the sequential stages, beginning with Sentinel-2 data preprocessing, followed by latent feature extraction using β-VAE, multiscale feature analysis via 2D-DWT, and spectral–spatial clustering with GMM. The final step involves validation of alteration zones using field samples and band ratios.

Fig. 3.

Schematic diagram of the proposed workflow, illustrating the data preprocessing steps, latent feature extraction using β-VAE, multiscale analysis via 2D-DWT, spectral–spatial clustering with GMM, and validation of alteration zones.

Owing to their high spatial resolution and appropriate spectral range, Sentinel-2 images offer strong capabilities for detecting alteration zones and iron-bearing minerals (Fe²⁺, Fe³⁺, and iron oxides). To focus on hydrothermal alteration mapping, six 20 m bands (Bands 5, 6, 7, 8 A, 11, and 12) were selected. These bands are widely applied in alteration studies due to their diagnostic sensitivity to vegetation, phyllic and propylitic alteration minerals, and Fe²⁺/Fe³⁺ absorption features in the VNIR–SWIR ranges^33–35. Certain hydroxyl-bearing minerals, which exhibit diagnostic absorption features in the 2.10–2.28 μm range, lead to a notable decrease in reflectance in Band 12 of the Sentinel-2 A MSI sensor, while showing high reflectance in Band 11^23,51.

In this study, six Sentinel-2 bands (5, 6, 7, 8 A, 11, and 12) were selected for feature extraction and modeling due to their high sensitivity to hydrothermal alteration minerals and iron-bearing phases. It should be noted that Bands 3 (Green, 10 m) and 4 (Red, 10 m) were not used in the modeling process; however, they were employed only at the validation stage for computing widely used spectral indices of Fe²⁺ and Fe³⁺. This step allowed the model results to be cross-checked against standard band-ratio criteria reported in the literature. Thus, Bands 3 and 4 served exclusively as external references for validation and were not part of the model input.

The Sentinel-2 Level-2 A imagery used in this study was acquired on 8 July 2021 by the Sentinel-2 A satellite. The Level-2 A product provides atmospherically corrected bottom-of-atmosphere (BOA) reflectance generated by the Sen2Cor processor; therefore, no additional atmospheric correction was required. The image was clipped to the spatial extent of the study area, covering approximately 263 × 172 pixels. Since the Level-2 A data are distributed as scaled integers, digital numbers (DNs) were converted to surface reflectance using the scaling factor of 0.0001 (i.e., Reflectance = DN × 0.0001) recommended by European Space Agency (ESA)^52–54. The selected scene was cloud-free, eliminating the need for additional cloud or shadow masking.

To validate the suitability of the data for statistical modeling, distributional analysis was performed prior to applying β-VAE, DWT, and GMM techniques. Histogram plots and Q–Q diagrams confirmed that the reflectance values exhibited approximately normal distribution (Fig. 4). This verification supported the use of variance-based and probabilistic methods in the subsequent modeling steps. All data processing and modeling procedures were implemented using the Python programming language.

Fig. 4.

Examination of surface reflectance distributions for selected Sentinel-2 bands (Bands 5, 6, 7, 8 A, 11, and 12): (a) Histograms of pixel values for each band, showing an approximately normal distribution of reflectance data. (b) Q–Q plots for each band, used to assess distribution normality, confirming a close alignment with the theoretical normal line (red).

Field observations

In this study, a total of 251 samples were collected from the area, consisting of 67 soil samples and 184 rock samples. Prior to geochemical analyses, macroscopic examinations of the rock samples were conducted to determine their alteration characteristics. Based on these field-scale observations, each sample was assigned to a specific alteration type, and this preliminary classification provided the framework for subsequent petrographic and geochemical investigations. Among the rock samples, 114 are related to phyllic alteration and 47 to propylitic alteration. Based on field observations and petrographic analyses, the rock samples are classified into two main groups: intrusive rocks, including diorite, quartz diorite, and microdiorite, and volcanic rocks, including dacite, rhyodacite, and rhyolite. The distribution of the collected lithologies is shown in Fig. 5. In addition, several composite or altered lithologies, such as dacite–rhyodacite and diorite–dacite, were also recognized.

Fig. 5.

Distribution of collected soil and rock samples in the study area. Lithological types are indicated for the rock samples, and their locations are shown in relation to the main hydrothermal alteration zones.

To determine the chemical composition of these samples, the assays were carried out at Amdel Mineral Laboratories (Adelaide, South Australia), where 43 elements were analyzed using ICP-MS after four-acid digestion, while gold content was measured by the Fire Assay method. The detection limits of the ICP-MS instrument for the 43 analyzed elements are presented in Supplementary Table S1. Geological evidence indicates that the igneous rocks of the region have been affected by hydrothermal fluids, and the predominant alteration types include potassic, phyllic, argillic, and propylitic. For phyllic alteration, particular attention was given to the elements S, Fe, and Al (Supplementary Table S2), as they are considered diagnostic for the formation of sericite, quartz, and pyrite, which characterize this alteration type. For propylitic alteration, the elements Fe, Mg, and Ca were measured (Supplementary Table S3), since they are key indicators of chlorite, epidote, and calcite, which are typical minerals of this alteration zone⁸. Alteration classification was performed based on macroscopic field observations and confirmed by petrographic analyses of diagnostic minerals. Geochemical data were subsequently used only to document and interpret the consistency between field observations and elemental behavior, not for labeling or modeling purposes. The spatial distribution of these alterations is shown in Fig. 1e.

Due to the absence of field spectral measurements from the region, the USGS spectral library was employed as a reliable source to obtain reference spectra of alteration minerals and iron oxides. These spectra were resampled to match the Sentinel-2 MSI band responses, enabling direct comparison with Sentinel-2 surface reflectance data (Fig. 6).

Fig. 6.

Comparison of original spectral signatures from the USGS spectral library (black) with their resampled counterparts to Sentinel-2 band positions (red) for key hydrothermal minerals (chlorite, goethite, hematite, jarosite, kaolinite, magnetite, and muscovite).

For partial validation of the results, real spectra were extracted from the Sentinel-2 imagery within alteration indicator zones. The selection of these zones was feasible owing to the widespread and superficial exposure of indicator minerals in the phyllic (muscovite) and propylitic (chlorite) zones, where pixels with high relative purity were chosen for analysis (Fig. 7). A comparison between the real spectra of these minerals and the resampled library spectra demonstrated that their spectral behavior is well captured within the Sentinel-2 band range, and their diagnostic absorption features can be identified.

Fig. 7.

Comparison between Sentinel-2 image spectra (blue) and resampled USGS library spectra (red) for muscovite (left) and chlorite (right), illustrating the consistency and deviations between measured and resampled spectral responses.

In contrast, extracting reliable spectra for iron-bearing minerals such as goethite, hematite, and magnetite was constrained. These minerals typically occur as fine-scale speckles, thin surface coatings, or are intermingled with other surface components such as soil, bedrock, vegetation, and shadow. Moreover, the relatively broad bandwidths of Sentinel-2 bands across key iron absorption regions reduce the capacity to resolve subtle spectral features. Consequently, obtaining clean and representative pixels for these minerals was not possible. Nevertheless, the resampled library spectra for iron oxides were adopted as a dependable reference for qualitative interpretation and for identifying potential spatial patterns.

For quantitative evaluation, comparisons between Sentinel-2 image spectra and resampled library spectra were conducted using the root mean square error (RMSE) index. The results indicated that the agreement for muscovite (RMSE = 0.155) was substantially better than for chlorite (RMSE = 0.204). This outcome highlights that Sentinel-2 data are effective in representing the spectral characteristics of phyllic zone indicator minerals and provide a robust basis for hydrothermal alteration analyses.

Petrographic analyses

During the field surveys, a total of 184 rock samples were collected along trenches designed to cross major alteration zones and lithological boundaries. For each sample, geographic coordinates (GPS) and detailed specimen descriptions were recorded. Out of these, 37 representative samples were selected for the preparation of thin sections. Section preparation followed standard procedures, including cutting, mounting, and grinding to a thickness of ~ 30 μm for transmitted-light studies. Petrographic analyses were carried out on dacite, rhyodacite, and diorite samples, which correspond to the main lithologies analyzed in the geochemical dataset⁵⁵ (Fig. 8). In dacite thin sections, clay minerals occur as fine, highly birefringent shreds (presumably kaolinite or alunite) intimately associated with quartz and sericite (Fig. 8b). In diorite, alteration and ore minerals are concentrated among tremolite–actinolite and plagioclase crystals, with Cu-bearing phases mainly appearing as fine inclusions in tremolite–actinolite, sericite, clay minerals, and epidote (Fig. 8a). These observations confirm the mineralogical expression of both phyllic (quartz, sericite) and propylitic (chlorite, epidote, calcite) alteration, in agreement with the geochemical classification of alteration zones. The optical identification of alteration minerals was based on standard petrographic criteria: sericite, recognized by feldspar replacement and fine-grained aggregates; chlorite, showing green pleochroism and low interference colors; epidote, characterized by high relief and anomalous interference colors; calcite, with strong birefringence and distinct cleavage; and quartz, with no cleavage and wavy extinction. The mineral assemblages, together with their textural relationships (vein, replacement, and disseminated forms), were used to delineate the phyllic and propylitic alteration zones.

Fig. 8.

(a) Ore minerals concentrated among tremolite–actinolite and plagioclase in diorite (PPL (transmitted light)). (b) Clay minerals (kaolinite/alunite) with quartz and sericite in dacite, filling intracrystal voids (XPL).

It should be noted that these petrographic thin sections were selected specifically to match the lithologies of the geochemical dataset. Additional thin sections from other lithologies (e.g., andesite, granodiorite), used to validate the alteration zoning in the field, are presented separately in the Field Verification section.

Main processing

β-VAE: β-Variational autoencoder modeling

Autoencoders are commonly used for dimensionality reduction and compact feature extraction. Building on the latent variable modeling framework, the Variational Autoencoder (VAE) has been introduced as a deep generative model that learns latent variables and enables both efficient representation learning and generative sampling^56–58. However, a key limitation of standard VAE is the entanglement of latent features, where multiple underlying factors are compressed into a single, less interpretable representation.

To overcome this, the β-VAE introduces a control coefficient β into the loss function, balancing the reconstruction error against the Kullback–Leibler (KL) divergence term. By increasing β, the model emphasizes disentangling the latent space, producing more interpretable and independent features. In this study, the β-VAE framework was applied to Sentinel-2 multispectral imagery to extract fully unsupervised latent representations relevant to hydrothermal alterations.

The encoder architecture consisted of three fully connected layers with 128, 64, and 32 units, each using the ReLU activation function. The encoder produced two outputs, the mean vector (µ) and the logarithm of the variance (logσ²), which parameterized an 8-dimensional latent space. Latent vectors were sampled using the reparameterization trick to ensure differentiability during training. The decoder mirrored the encoder structure in reverse, with ReLU activations and a final linear output layer to reconstruct the original 10-dimensional input vector. To train the network, the Adam optimizer was employed with mean squared error (MSE) as the reconstruction loss, combined with the KL divergence term. Early stopping based on validation loss was applied to prevent overfitting. In this study, the value of β was set to 2 to balance reconstruction quality and disentanglement.

The complete objective function of the β-VAE is expressed as follows⁵⁸:

1

where q_ϕ(z∣x) is the approximate posterior distribution with parameters µ(x) and logσ²(x), and the prior distribution p(z)= (0,I) is assumed to be isotropic^{59– 60}.

When β = 1, the model reduces to the standard VAE, while larger β values encourage disentangled latent representations⁶¹. Latent vectors z are sampled from a Gaussian distribution parameterized by the encoder outputs µ(x) and σ(x).

After training the β-VAE model—by selecting an optimal value for β based on resolution criteria—only the mean vector µ(x) of each satellite tile is extracted. These vectors are used to construct a latent matrix Z ∈ R^N×M, which serves as the input for the subsequent DWT-based analysis.

In the context of this study, β-VAE was not merely a theoretical generative model but a practical tool for extracting compact, disentangled, and unsupervised features from Sentinel-2 data. This ensured that the latent space directly captured the spectral–spatial variability associated with hydrothermal alterations.

Discrete wavelet transform (DWT) for multiscale feature extraction

Unlike the Fourier transform, which only works in the frequency domain, the wavelet transform preserves both frequency and spatial context. This property allows a signal or image to be decomposed into multiple resolution levels, enabling feature extraction across different scales^{62– 63}. Wavelets are derived from a mother function and scaled/shifted for multiresolution analysis.

In practice, DWT is preferred over CWT because of its computational efficiency and ease of implementation^{64– 65}. Because this research employs satellite imagery, the wavelet transform is extended into 2D-DWT so that spatial variations within the image can be accurately captured and analyzed. The two-dimensional discrete wavelet transform is defined as follows:

2

Within this formulation, variable a functions as the scale factor, dictating how much the wavelet is expanded or contracted across both spatial dimensions. Meanwhile, b and c serve as translation parameters along the x- and y-axes, respectively, pinpointing where in the image the wavelet analysis takes place^{66– 67}.

As a result of the 2D wavelet decomposition, the image or signal is divided into four frequency subbands. The LL subband contains low-frequency information in both dimensions and serves as a coarse approximation of the original image. The LH and HL subbands capture horizontal and vertical edge details, respectively, whereas the HH subband accentuates diagonal features and other high-frequency content^29,64–67.

In practice, applying 2D-DWT to latent features may produce positive and negative values, depending on the orientation of the filters and the statistical properties of the β-VAE outputs. This behavior reflects representational differences in the latent space and does not indicate a weakness or defect in the model.

These subbands form the foundation for a wide range of applications, including feature extraction, data compression, and image content analysis. Because wavelet analysis is performed locally, the DWT is particularly effective at capturing features across multiple spatial scales.

In this study, the role of the DWT is to enrich the latent features obtained from β-VAE by embedding multiscale spatial detail. This ensures that both coarse- and fine-scale hydrothermal alteration patterns are preserved, overcoming the limitations of conventional spectral methods such as PCA and simple band ratios.

Gaussian mixture model (GMM) for probabilistic clustering

The Gaussian Mixture Model (GMM) is a statistical method that represents data as a mixture of multiple Gaussian distributions. Unlike K-means clustering, which assigns each sample to a single cluster, GMM applies a soft assignment strategy, allowing each point to belong to multiple clusters with associated probabilities. This probabilistic nature makes GMM more effective in separating clusters with overlapping or ambiguous boundaries^{68– 69}.

Each Gaussian component is defined by its mean, covariance, and mixing coefficient, which together determine its contribution to the overall mixture. To compute the overall likelihood of an observation, GMM combines all Gaussian components through weighted summation as follows:

3

In this context, P(x_n) represents the overall probability of observing the sample x_n under the mixture model, calculated by accounting for the contributions of all Gaussian components. This flexibility enables GMM to capture complex and overlapping structures in data, making it well suited for applications in geological analysis, satellite image classification, and latent feature modeling^{63– 64}.

In this study, the input to the GMM consisted of the LL components derived from the 2D-DWT applied to the β-VAE latent features. These low-frequency subbands preserve the dominant spectral–spatial structures of the image while reducing noise and redundant detail. A hybrid GMM was then employed to probabilistically cluster pixels with similar feature signatures, providing a flexible framework for alteration zone detection.

Validation methodology

To evaluate the accuracy and reliability of the proposed algorithm, two complementary approaches were applied: Tagged geochemical points, which independently confirmed the phyllic and propylitic alteration zones, were used as reference data to validate the corresponding clusters. The locations of these data are shown in Fig. 1e. The overlap between these points and the GMM clustering results was used to calculate the classification accuracy of each zone.

Due to the lack of field data for clusters associated with iron minerals and iron oxides (such as Fe²⁺ and ferric phases), standard spectral indices based on Sentinel-2 band ratios were used (Table 1). These indices, by combining visible, NIR, and SWIR bands—particularly bands 3 and 4 for Fe²⁺ and Fe³⁺—can reveal the characteristic absorption features of iron minerals. It should be emphasized that bands 3 and 4 were not used in the main modeling process (β-VAE → DWT → GMM), as the focus of the model was on the 20 m bands sensitive to hydrothermal alteration (5, 6, 7, 8 A, 11, and 12). However, these two bands were purposefully used in the validation phase to verify the model results against the established spectral indices for Fe²⁺ and Fe³⁺. This approach allowed the model to focus on the optimal bands for feature extraction while, at the same time, ensuring that its outputs were consistent with standard spectral criteria.

Table 1.

Validation results for iron-related alteration zones using Sentinel-2 band ratio expressions.

Iron Component	Band Ratio Expression
Ferric iron (Fe³⁺)
Ferric Oxides
Ferrous iron (Fe²⁺)

Results and discussion

To extract compact and meaningful representations from the 10-dimensional feature vectors—comprising six Sentinel-2 bands and four band ratios related to Fe²⁺, Fe³⁺, and ferric oxides—the β-VAE architecture was employed. This unsupervised learning model enables probabilistic encoding of input data and supports both reconstruction and the discovery of latent structures.

As shown in Fig. 9a, both training and validation errors decrease rapidly during the initial epochs and then plateau, indicating effective model convergence without signs of overfitting. The final validation error stabilizes near zero, demonstrating the model’s high accuracy in reconstructing the input data. To qualitatively evaluate the model’s performance in reconstructing input data, randomly selected samples of original and reconstructed feature vectors are presented in Fig. 9b. The strong agreement between the blue (original) and orange (reconstructed) curves suggests that the model effectively preserved the spectral structure of the input data with acceptable accuracy. Minor deviations at specific points may indicate a reduced contribution of certain latent dimensions in representing the input information, which is consistent with the regularizing effect imposed by the β parameter. Upon completion of training, the latent vectors (Z) for the entire image dataset were extracted and reshaped into spatial maps with dimensions height×width×8, which were subsequently used in the following stages of analysis. The reconstruction error distribution shown in Fig. 9c indicates that the majority of pixels exhibit very low error values. The distribution is strongly right-skewed, with most values concentrated near zero, suggesting that the model consistently and accurately reconstructs the input data across the entire image. This distribution also provides a solid foundation for potential anomaly detection, as pixels with higher reconstruction errors may signify atypical spectral behavior.

Fig. 9.

(a) Trend of decreasing training and validation errors or MSE across epochs, indicating rapid convergence and absence of overfitting. (b) Comparison between original and reconstructed feature vectors for a representative pixel, showing strong agreement between real and reconstructed signals and (c) Distribution of reconstruction error across the entire image, demonstrating stable model performance, with the majority of pixel errors concentrated near zero.

The eight dimensions extracted from the latent space—representing compressed encodings of the region’s spectral and structural patterns—are visualized as spatial maps in Fig. 10.

Fig. 10.

Visualization of the eight latent components extracted using the β-VAE model (β = 2), representing compressed spectral–spatial features derived from the six-band Sentinel-2 image stack and associated band ratio inputs, generated in the Python programming environment.

To evaluate the discriminative resolution of the β-VAE model’s latent dimensions with respect to alteration types, a stepwise analysis was performed using labeled samples. Based on visual interpretation of the Sentinel-2 band combination (RGB = 11–12–8), phyllic and propylitic alteration areas were preliminarily delineated (Fig. 1d). These regions were extracted using an ROI file and comprised 951 pixels for the phyllic class and 2,152 pixels for the propylitic class, from an image of size 171 × 263 pixels. The corresponding latent values for both classes were extracted, and the results are presented as box plots in Fig. 11. According to these plots, several latent dimensions—specifically Latents 2, 4, 6, and 7—show significant statistical differences between the two classes, effectively capturing their distinct spectral–structural patterns. For instance, in Latent Dimension 4, the phyllic class predominantly exhibits negative values, while the propylitic class lies in the positive range, indicating strong class separability. In contrast, dimensions such as Latents 5 and 8 do not display meaningful differentiation. These findings suggest that, despite being trained in an unsupervised manner, the β-VAE model successfully encodes interpretable latent features that reflect underlying geological variability. Such representations can be effectively leveraged in downstream tasks, including clustering, classification, and alteration zone mapping.

Fig. 11.

Boxplots of latent dimension values extracted from the β-VAE model, categorized by alteration type labels (phyllic and propylitic). The plots illustrate the separability of specific latent dimensions with respect to alteration classes.

Final interpretation of these layers is carried out in conjunction with wavelet transformation and GMM clustering. Importantly, the final result is not derived from any single latent layer, but rather from the integration of information across multiple dimensions.

After reconstructing the images using the β-VAE model, the DWT was applied to more accurately analyze the spectral–spatial structures present in the model outputs. By performing DWT on the reconstructed images, frequency subcomponents—LL, LH, HL, and HH—were extracted (Fig. 12a-h). The LL component, which captures low-frequency and general structural information, successfully preserved the primary alteration patterns.

Fig. 12.

Low-frequency (LL) components of the eight latent features, obtained by applying the 2D-DWT to the β-VAE outputs (a-h). These LL components capture dominant spatial structures that are relevant to hydrothermal alteration mapping.

In certain latent features, such as images Figs. 12b, d, e, and g, the target areas appeared with negative values. This may be attributed to the wavelet filtering process, which can represent specific spectral–spatial variations of alteration zones in the LL subband using negative intensities. Additionally, the orientation of the learned parameters in the β-VAE model may lead to some features being encoded with reversed contrast relative to others. This behavior does not indicate a flaw in the model but rather reflects a form of representational differentiation within the latent space.

Before the geological interpretation of the clusters, a statistical validation of the GMM results was conducted. For this purpose, the model was tested over a range of cluster numbers (k = 2 to 10), and the Bayesian Information Criterion (BIC) and Akaike Information Criterion (AIC) were calculated (Fig. 13). Both indices, which evaluate the balance between model accuracy and complexity, reached their minimum values at k = 5, indicating this as the optimal solution. To further ensure robustness, additional indices were examined: the Integrated Completed Likelihood (ICL), which penalizes cluster overlap; the Silhouette index, which measures cluster separation and compactness; the Davies–Bouldin Index (DBI), which evaluates the ratio of intra-cluster dispersion to inter-cluster separation; and a bootstrap stability test based on the Adjusted Rand Index (ARI), which assesses the consistency of the cluster structure under resampling. All of these measures converged on the same result, confirming k = 5 as optimal: Silhouette and ARI were maximized, while DBI was minimized. This convergence demonstrates that the choice of five clusters is not only statistically valid but also geologically meaningful and reliable.

Fig. 13.

Structural and stability indices versus number of clusters (k). The convergence of Silhouette, DBI, and Bootstrap Stability ARI suggests k = 5 as the optimal solution.

The results of GMM clustering revealed that one of the alteration zones exhibits phyllic characteristics accompanied by a strong trivalent iron (Fe³⁺) signal, interpreted as a phyllic zone enriched in Fe³⁺. A comparative analysis of the spectral clustering output further indicated that portions of the Fe³⁺-rich regions overlap with propylitic zones. This spatial overlap may be attributed to the presence of Fe³⁺-bearing epidote within the propylitic zone or to secondary oxidation processes affecting iron-bearing phases. The yellow areas highlighted in the clustering results (Fig. 14a), marked by the prominent presence of trivalent iron oxides such as hematite, goethite, or limonite, generally do not show direct spatial overlap with the phyllic and propylitic zones identified on the overlaping map. This phenomenon is interpreted as the result of surface-level secondary oxidation or supergene processes. Under such conditions, pyrite or other iron-bearing minerals are oxidized upon exposure to oxidizing fluids, leading to the formation of secondary ferric phases such as limonite or hematite. In Fig. 14b, phyllic alteration clusters are highlighted in red, whereas propylitic alteration clusters are displayed in green.

Fig. 14.

(a) Spatial distribution of iron-related alteration zones classified by GMM, including Fe²⁺, Fe³⁺, ferric oxides, and a mixed phyllic–iron oxide zone and (b) Spectral–spatial mapping of phyllic and propylitic alteration zones derived from Sentinel-2 latent features. Each pixel is assigned to the class with the highest membership probability (Hard classification). Classification was performed in the Python programming environment, and map generation and visualization were carried out using ArcMap software (version 10.8.2; https://www.esri.com).

The Fe²⁺ cluster generally signifies the presence of ferrous minerals such as chlorite, amphibole, or pyroxene. In this study, parts of the Fe²⁺-rich areas do not overlap directly with the mapped propylitic zone and instead appear as separate domains. These ferrous zones may represent regions of weak alteration, reduced environments, or even unaltered parent rock. As such, they may serve as useful baselines for comparison with secondary alteration patterns.

The core of this study relies on Sentinel-2 sensor data with a spatial resolution of 20 m. While this resolution is suitable for mapping surface alteration zones, the limited spectral coverage—particularly in the SWIR bands—restricts the ability to accurately and independently distinguish between phyllic and propylitic alterations at the spectral level.

The validation of phyllic and propylitic alteration was carried out at two complementary levels. In the first step, the direct overlap between the GMM clusters and the geochemical/petrographic labels showed that the algorithm performed very well in distinguishing these two alteration zones: 94.74% of phyllic samples (108 out of 114 points) and 95.74% of propylitic samples (45 out of 47 points) fell within their corresponding alteration zones on the model map (Fig. 15; Table 2). These point-based results confirm the strong agreement between the modeled spectral responses and the field evidence and provide a solid initial basis for confidence in the phyllic–propylitic separation.

Fig. 15.

Validation of phyllic and propylitic alteration zones using labeled geochemical data (114 phyllic and 47 propylitic samples). Geochemical reference points are overlaid on the final spectral–spatial clustering map, demonstrating strong spatial agreement with the predicted alteration zones. The map represents the hard classification output (maximum membership probability rule), and the transitional boundaries are consistent with the geological characteristics of hydrothermal alteration zones. Clustering results were obtained in the Python programming environment, and map generation and visualization were performed using ArcMap software (version 10.8.2; https://www.esri.com).

Table 2.

Sample-based and buffered pixel-based validation results for phyllic and propylitic alteration zones. Coverage accuracy is computed from the overlap between labeled geochemical samples and the alteration map, while confusion matrices, overall accuracy, and 95% CI are derived from buffered pixel masks around the same samples.

Alteration Zone	Total Geochemical Points	Points Inside Alteration Map	Coverage Accuracy (%)	Buffered Confusion (TP / FN / FP / TN, pixels)	Overall Accuracy (%)	95% CI
Phyllic	114	108	94.74	898 / 213 / 427 / 10,033	94.5	94.1–94.9
Propylitic	47	45	95.74	299 / 491 / 1026 / 9755	86.9	86.3–87.5

In the second step, buffered reference masks were employed to represent the spatial extent of the alteration halos around each sample, and pixel-scale validation was performed. For each of the phyllic (114 samples) and propylitic (47 samples) labels, a 5 × 5-pixel kernel was defined around the sample location to capture the spatial footprint of the alteration halo. Comparison of these buffered masks with the GMM-derived alteration maps enabled the construction of confusion matrices and the computation of quantitative metrics (Overall accuracy) across the entire scene comprising 11,571 pixels. The results of this analysis show that, in the phyllic zone, the model exhibits very stable and reliable performance. The phyllic confusion matrix consists of TP = 898 (reference phyllic pixels correctly classified as phyllic), FN = 213 (reference phyllic pixels detected as non-phyllic), FP = 427 (non-phyllic pixels labeled as phyllic), and TN = 10,033 (reference non-phyllic pixels correctly detected as non-phyllic). This matrix yields an overall accuracy of 94.5%, and the 95% confidence interval (CI) of the overall accuracy lies in the range of 94.1–94.9%, indicating high statistical stability of the model. For the phyllic class, the recall is approximately 0.81, the F1-score is about 0.74, and the intersection-over-union (IoU) coefficient is about 0.58. These values indicate that the majority of the reference phyllic halos are recovered by the model and that there is substantial spatial overlap between the GMM-derived map and the buffered phyllic masks, although some peripheral pixels are also labeled as phyllic.

For the propylitic class, the buffered confusion matrix (TP/FN/FP/TN = 299/491/1026/9755) yields an overall accuracy of 86.9% (95% CI: 86.3–87.5%). In areas where reference data are available, the model exhibits a conservative behavior and tends to extract the core propylitic alteration zones with relatively high confidence, without generating extensive false positive halos over the background. It should be noted that parts of the propylitic zone—particularly in the southern and southwestern sectors of the study area—lack sampling and field labels. In these regions, the model has proposed prospective areas purely based on spectral signatures, spatial patterns, and proximity to geological structures, and direct validation against existing labels has not been possible. Nevertheless, the pixels from these regions were included in the computation of the global validation metrics, which naturally leads to slightly lower pixel-based performance indices for the propylitic class compared with the point-based results and the areas covered by reference data. The values reported in Table 2 therefore represent a conservative estimate of the overall performance of the model across the entire scene.

Overall, the combination of point-based results (coverage accuracy) and buffered pixel-level evaluation demonstrates that the GMM model performs strongly and reliably in identifying and discriminating between phyllic and propylitic zones, both in terms of agreement with field data and statistical stability. However, in the propylitic zone, the model tends to deliberately emphasize more reliable cores, leaving some of the marginal areas with higher uncertainty open as potential targets for further exploration.

In certain parts of the study area, regions were identified as propylitic alteration zones for which direct validation was not feasible due to the absence of reference data or ground sampling. These areas were therefore analyzed solely based on their spectral signatures, spatial patterns, and proximity to geological structures.

The outputs of the band ratio calculations, based on the expressions provided in Table 3, are presented in Fig. 16. This method employed predefined thresholds capable of identifying pixels with high reflectance in iron absorption bands, which are typically indicative of surface alteration and secondary oxidation processes. To quantitatively assess the agreement between the GMM classification and these spectral indicators, confusion-matrix–based metrics were computed for three iron alteration types (ferric iron, ferric oxides and ferrous iron), and are summarized in Table 3.

Fig. 16.

Band ratio outputs for (a) Ferric iron (Fe³⁺), (b) Ferric oxides, and (c) Ferrous iron (Fe²⁺). Band ratio processing was performed using ENVI software (version 5.6; https://www.nv5geospatialsoftware.com/), and the resulting raster outputs were processed for grid generation and map visualization using ArcMap software (version 10.8.2; https://www.esri.com).

For ferric iron, defined by the band ratio , the GMM-derived alteration mask shows very strong agreement with the band-ratio mask, yielding a precision of 97.9%, recall of 96.3%, an IoU (Intersection-over-Union) of 94.3% and an F1 score of 97.1%. The overall accuracy for this class reaches 98.5% (95% CI: 98.23–98.68%), indicating that both false positives and false negatives are rare. Ferric oxides, mapped using , also exhibit high but slightly lower performance, with a precision of 91.5%, recall of 90.3%, IoU of 83.3% and F1 of 90.9%; the overall accuracy is 97.0% (95% CI: 96.72–97.34%). Ferrous iron (Fe²⁺), defined by the composite ratio , shows the lowest—but still robust—agreement, with precision and recall of 90.2% and 89.4%, respectively, an IoU of 81.5% and an F1 score of 89.8%; the corresponding overall accuracy is 97.5% (95% CI: 97.25–97.82%).

Table 3.

Validation results for iron-related alteration zones using Sentinel-2 band ratio expressions.

Alteration type	Band ratio expr.	Threshold	GMM pixels	Band-mask pixels	TP (Overlap)	FP	FN	TN (N = 11571)	Precision (%)	Recall (%)	IoU (%)	F1 (%)	Overall Accuracy (%)	95% CI
Ferric iron (Fe³⁺)	4 / 3	1.15	3051	3100	2986	65	114	8406	97.87	96.32	94.34	97.09	98.45	[98.23–98.68]
Ferric Oxides	11 / 8	1.13	1876	1900	1716	160	184	9511	91.47	90.32	83.30	90.89	97.03	[96.72–97.34]
Ferrous iron (Fe²⁺)	12/8 + 3/4	1.07	1387	1400	1251	136	149	10,035	90.19	89.36	81.45	89.77	97.54	[97.25–97.82]

Taken together, these results confirm that the clusters extracted by the GMM are broadly consistent with the spatial distribution of the iron-sensitive band-ratio anomalies. The agreement is particularly strong for ferric iron alteration, whereas ferric oxides and especially ferrous iron show slightly higher uncertainty, which is consistent with the more complex and less distinctive spectral response of Fe²⁺-bearing phases. Nonetheless, the consistently high precision, recall and overall accuracy across all three iron alteration types demonstrate that the proposed GMM-based approach provides a reliable representation of iron-related surface alteration patterns in the study area.

The results of this study demonstrate that the combination of β-VAE, DWT, and GMM provides an efficient and unsupervised framework for hydrothermal alteration zoning using Sentinel-2 imagery. The high classification accuracy in phyllic (94.74%) and propylitic (95.74%) zones highlights the framework’s ability to extract hidden features at the resolution of subtle yet geologically important spectral–spatial variations that are often difficult to detect using purely spectral methods. Notably, the successful discrimination of iron species, including Fe²⁺, Fe³⁺, and iron oxides (with an accuracy exceeding 90%), confirms the framework’s capability to analyze iron speciation—an aspect critical for understanding oxidation and supergene enrichment processes in porphyry systems.

Comparison with previous studies in the Zafarghand region further underscores the advantages of the proposed approach. Earlier research using ASTER and Landsat-8 imagery combined with methods such as PCA, LS-fit, and Matched Filtering^27,28 partially delineated alteration halos but was limited in distinguishing overlapping zones or identifying iron species. More recent studies employing MAD and DWT on ASTER data²⁹ achieved higher accuracy in anomaly detection, yet remained constrained by conventional statistical modeling. By contrast, the present framework leverages the generative capacity of β-VAE to extract latent, compact features from multispectral data and integrates GMM probabilistic clustering, which avoids the rigid boundaries imposed by algorithms such as K-means. These strengths allow for more accurate separation of phyllic and propylitic zones, along with the independent identification of iron-bearing species.

From a geological perspective, distinguishing between phyllic and propylitic zones is particularly important for porphyry exploration. Phyllic alteration, typically characterized by sericitization and Fe³⁺ enrichment, commonly develops near the mineralization center and therefore represents a high exploration priority. In contrast, propylitic alteration forms peripheral halos that delineate the spatial framework of the hydrothermal system. Independent identification of iron oxides and Fe²⁺-rich zones is also significant, as these features may signal supergene oxidation processes or the preservation of relatively unaltered host rocks. Collectively, these findings provide a clearer understanding of the hydrothermal evolution of the Zafarghand system and support the optimization of exploration strategies.

Nevertheless, several limitations must be acknowledged. First, the 20 m spatial resolution and limited SWIR spectral coverage of Sentinel-2 reduce the ability to resolve fine-scale alteration minerals, particularly argillic and silicic zones. It should be noted that argillic, potassic, and silicic zones are present in the Zafarghand system, but due to the limited areal extent and 20-meter resolution of Sentinel-2, their spectral signal is combined with the background and their independent separation is not possible within the present framework. Second, the absence of geochemical data restricted the quantitative evaluation of these less extensive zones. Finally, although the proposed approach is entirely data-driven and requires no labeled training data, its accuracy remains dependent on the quality of the input imagery and the availability of at least minimal reference data for validation.

Despite these constraints, the framework exhibits strong scalability. Because the method is unsupervised, it can be applied to large, unexplored areas lacking prior information on alteration zones. In such cases, the outputs can serve as a valuable preliminary tool for identifying targets for field investigation. In known areas, the method can refine existing maps and improve the delineation of alteration zones. Future research could further enhance mapping accuracy and geological interpretation by integrating this framework with hyperspectral datasets (e.g., PRISMA and EnMAP) as well as geophysical and structural data. In this study, only the low-frequency sub-band (LL) was used as the basis for clustering to capture the main signal of regional-scale alteration. In future research, however, a combination of high-frequency sub-bands (LH, HL, HH) could be incorporated to provide a more detailed examination of alteration boundaries and fluid-related processes. Future research could also employ automatic feature selection or weighting methods (e.g., feature importance or mutual information) to better quantify the contribution of individual latent dimensions and further optimize the efficiency of the framework.

Field verification

In contrast to the petrographic thin sections described in Sect. 3.3, which were directly tied to the geochemical dataset, the images presented here illustrate alteration as observed in the field context. Thin sections from granodiorite, andesite, and rhyodacite, together with field photographs, confirm the presence and distribution of phyllic and propylitic zones^36,70 (Figs. 17 and 18). These results provide independent validation of the alteration patterns delineated by remote sensing.

Fig. 17.

(a) Field view of the phyllic alteration zone developed within Eocene volcanic rocks. (b) Photomicrograph (XPL) showing quartz (Qz) and sericitized plagioclase with chlorite (Chl) in granodiorite. (c) Photomicrograph (PPL) of andesite displaying sericitized plagioclase and disseminated pyrite (Py) with chlorite alteration. (d) Field photograph of hydrothermal breccia hosted by altered granodiorite.

Fig. 18.

(a) Field view of the propylitic alteration zone, (b) Field photograph showing epidote mineralization in altered host rocks, (c) Photomicrograph (PPL) illustrating epidotization and chloritization of plagioclase (Plg) with quartz (Qz) in rhyodacite, (d) Photomicrograph (XPL) showing development of epidote (Epi) and calcite (Cal) associated with propylitic alteration in granodiorite. (Qz: quartz; Chl: chlorite; Epi: epidote; Cal: calcite)

Field observations and petrographic analyses conducted to support RS interpretations confirm the presence of rhyolite–rhyodacite subvolcanic units in the central portion of the study area. These rocks display pronounced phyllic alteration (Fig. 17a). The identification of sericite, quartz, iron oxides (from sulfide oxidation), and manganese-bearing phases in thin sections further substantiates the intensity and character of the alteration. Phyllic alteration corresponds to hydrolytic reactions affecting mafic and intermediate minerals. In granodioritic rocks, this alteration manifests as the transformation of plagioclase and mafic minerals into sericite (Fig. 17b), while in andesitic rocks, plagioclase alters to sericite and mafic components are replaced by chlorite (Fig. 17c).

Hydrothermal fluid activity has led to the widespread development of silicic stockworks, pervasive silicification, and accumulations of iron oxides observable in surface outcrops. Additionally, hydrothermal breccia is notably well-developed within the altered granodioritic groundmass (Fig. 17d).

Propylitic alteration is distributed around the margins of the phyllic alteration zone (Fig. 18a). In several areas of the study site, this type of alteration is particularly extensive—likely the result of interactions between meteoric fluids and various intrusive and volcanic host rocks, including granodiorite, dacite, and rhyodacite. This process is evidenced by the presence of veinlets and stringers composed of epidote, quartz, albite, chlorite, and carbonate minerals that cross-cut the altered zones (Fig. 18b). Plagioclase phenocrysts, which constitute approximately 25–33% of the rock volume, are frequently altered to secondary minerals such as chlorite, epidote, kaolinite, and sericite (Figs. 18c and d).

Conclusion

This study aimed to accurately identify hydrothermal alteration zones in the Zafarghand region by introducing and implementing an integrated framework combining a β-Variational Autoencoder (β-VAE), discrete wavelet transform (DWT), and Gaussian Mixture Model (GMM) clustering. Leveraging multispectral Sentinel-2 data and operating in a fully unsupervised manner, the framework successfully generated latent, compact, and structured representations of the region’s spectral–spatial characteristics. The analysis results demonstrated that the β-VAE model effectively extracts complex, multidimensional latent features from Sentinel-2 imagery, allowing clear spectral differentiation between phyllic and propylitic zones in the GMM clustering output. At the sample level, the classification accuracies for these zones, validated using field-based geochemical data, were 94.74% for the phyllic zone and 95.74% for the propylitic zone—highlighting the high effectiveness of the proposed framework in distinguishing hydrothermally altered areas. This point-based validation was further supported by buffered, pixel-level confusion-matrix analyses, which yielded overall accuracies of 94.5% for phyllic and 86.9% for propylitic alteration, with narrow 95% confidence intervals, confirming the statistical robustness of the mapping results.

Furthermore, the final output enabled the independent separation of iron-bearing compounds, including Fe²⁺, Fe³⁺, and iron oxides. The identification accuracies for these classes were 90.19%, 97.87%, and 91.47%, respectively, and were accompanied by high F1-scores and IoU values, particularly for ferric iron, demonstrating the framework’s strong capability in capturing geochemical variations associated with hydrothermal alteration and redox processes—capabilities that are often underachieved in conventional RS approaches.

The main advantage of this framework compared to classical methods such as PCA, band ratios, and supervised classification lies in three aspects: (i) the extraction of latent and compact features by β-VAE, (ii) the multiscale enrichment of spectral–spatial patterns by DWT, and (iii) the probabilistic clustering by GMM that reduces classification errors in transition regions. In addition, the buffered validation revealed a conservative behaviour of the model in mapping propylitic alteration: while the cores of well-constrained propylitic zones are reliably identified where reference data are available, more diffuse halos and southern/southwestern targets without ground control are mapped as prospective areas rather than fully validated classes.

Overall, the results of this study demonstrate that combining DL, multiscale analysis, and statistical clustering using Sentinel-2 imagery offers an effective, economical, and scalable solution for identifying alteration zones in geologically complex environments. The implementation of these methods in the Zafarghand area—despite the limited dimensions of the Sentinel-2 scene—showed that even in the absence of dense field data, promising areas can be identified in the early stages, while known alteration zones can be mapped with high confidence. Consequently, this framework can be used in known regions to refine existing maps and in unknown regions for initial discovery. Future enhancement of this framework through the integration of hyperspectral, geophysical, and structural data could represent a significant step toward more comprehensive modeling of hydrothermal systems.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Conceptualization, S.S.Gh.; Methodology, S.S.Gh. and S.E.K.; Software, S.E.K.; Validation, S.S.Gh. and S.E.K.; Formal analysis, S.E.K.; Investigation, S.E.K.; Resources, S.S.Gh.; Data curation, S.E.K.; Writing—original draft preparation, S.E.K.; Writing—review and editing, S.S.Gh.; Visualization, S.E.K.; Supervision, S.S.Gh.; Project administration, S.S.Gh.; Funding acquisition, not applicable. All authors have read and agreed to the published version of the manuscript.

Data availability

The Sentinel-2 satellite data used in this study are publicly available and can be freely accessed from the Copernicus Open Access Hub ( [https://www.copernicus.eu](https:/www.copernicus.eu) ).

Declarations

Competing interests

The authors declare no competing interests.