Mineral Prospectivity Mapping Based on Remote Sensing and Machine Learning in the Hatu Area, China

Abstract

The Hatu region in the Western Junggar, Xinjiang, is one of the most significant gold metallogenic concentration areas in China. Gold mineralization is primarily controlled by several parallel NE-trending strike-slip faults and Late Paleozoic granitic plutons, accompanied by multiple stages of hydrothermal activity. To enhance the objectivity and accuracy of mineral prospecting prediction, this study develops an integrated forecasting framework that combines multi-source remote sensing datasets with machine learning techniques. Alteration anomalies related to iron staining and hydroxyl-bearing minerals are extracted from ASTER data, alteration mineral mapping is performed using GF-5 hyperspectral imagery, and Landsat-9 data is used for structural interpretation to refine the regional metallogenic framework. On this basis, these multi-source remote sensing products are then integrated to delineate five prospective metallogenic areas (T1–T5). Subsequently, a Random Forest (RF) model optimized by the Grey Wolf Optimizer (GWO) algorithm is employed to quantitatively integrate key evidence layers, including alteration, structure, and geochemistry, for estimating mineralization probability. The results show that the GWO-RF model effectively concentrates anomalous areas and identifies two high-confidence targets, Y1 and Y2, both with mineralization probabilities exceeding 0.8. Among them, the Y1 target is associated with the Bieluagaxi pluton and exhibits strong montmorillonitization, chloritization, and iron-staining alteration, typical for magmatic–hydrothermal controlled mineralization. In contrast, the Y2 target is strictly controlled by the Anqi Fault and its subsidiary faults, primarily characterized by linear chloritization and iron-staining anomalies indicative of structure–hydrothermal mineralization. Field verification confirms the significant metallogenic potential of both Y1 and Y2, demonstrating the effectiveness of integrating multi-source remote sensing and machine learning for predicting orogenic gold systems. This approach not only deepens the understanding of the diverse gold mineralization processes in the Western Junggar but also provides a transferable methodology and case study for improving regional mineral exploration accuracy.

1. Introduction

Accretionary orogens are major hosts of global gold deposits, where mineralization is typically governed by long-lived strike-slip fault systems, syn- to post-collisional magmatism, and extensive fluid–rock interactions. The Central Asian Orogenic Belt (CAOB), representing the world’s largest Phanerozoic accretionary orogen, is a product formed through multiple episodes of subduction, accretion, and collision during the evolution of the Paleo-Asian Ocean [1,2,3,4]. Owing to its complexity and well-preserved geological record, the CAOB is often regarded as a natural field site for investigating the metallogenic principles and predictive methodologies of orogenic gold systems [1,5,6]. Situated in the eastern segment of the CAOB, the Western Junggar region of northwestern China represents a complex product of the convergence, subduction, and accretion of the Paleo-Asian Ocean [7]. It is a well-known metallogenic province that experienced widespread gold mineralization during the Late Carboniferous to Early Permian. Within this region, the Hatu area in southern Western Junggar hosts a belt-like distribution of gold deposits aligned along multiple northeast-trending strike-slip fault zones, such as the Hatu, Baobei, Huilvshan, Gezigou, and Mentougou deposits, making it a key area for extracting ore-forming information and developing prediction models for orogenic gold deposits [8,9].

The extraction of alteration mineral distributions using remote sensing data is considered a crucial approach for identifying mineralization and delineating prospective areas [10,11,12,13,14]. ASTER data possesses diagnostic advantages in the Short-Wave Infrared (SWIR) bands and is widely applied to extracting iron-staining and hydroxyl alterations [15,16,17,18]. Such applications include identifying alteration mineral assemblages in porphyry copper systems [19,20,21,22], delineating alteration zones related to epithermal deposits [11,23,24,25], and assessing the mineralization potential of gold-bearing alteration areas [11,26,27,28]. The Advanced Hyperspectral Imager (AHSI) onboard GF-5 covers 400–2500 nm with 330 narrow bands, offers high spectral resolution, and is particularly capable of finely identifying alteration mineral assemblages in the SWIR range [29,30,31,32,33]. In studies related to orogenic gold deposits, the data can effectively detect key alteration minerals associated with gold mineralization, such as muscovite, pyrophyllite, chlorite, and kaolinite. By analyzing the spectral responses of these minerals, it is possible to map alteration zones and examine their spatial zoning, thereby providing direct evidence for delineating prospective areas [34,35,36]. Therefore, the integrated application of ASTER and GF-5 data can enhance diagnostic identification capabilities while simultaneously improving alteration mapping accuracy [37,38]. Furthermore, Landsat-9 data provides the up-to-date high-resolution imagery for structural interpretation, facilitating the identification of fault trends, circular structures, and their spatial distribution [13,39].

Traditional exploration methods mainly rely on geological mapping, geochemical anomalies, and single-source remote sensing interpretation, showing clear limitations under the complex tectonic setting and superimposed mineralization of the Western Junggar [40]. Previous research indicates that, in the Hatu area, hydroxyl and iron-staining alteration anomalies detected by remote sensing exhibit strong spatial correlation with geochemical Au-As-Sb anomalies [41]; however, effectively focusing on anomalous areas and improving prospecting accuracy requires the integration of multi-source datasets. Additionally, early remote sensing interpretation relied heavily on manual visual analysis or empirical overlay methods, making it difficult to quantify results and scale them for regional applications [42]. With the rapid development of machine learning algorithms in mineral prospectivity mapping [43,44,45,46,47], algorithms such as Random Forest (RF), Support Vector Machine (SVM), and XG-Boost have been widely applied in mineral prediction [48,49]. Furthermore, emerging techniques that integrate signal processing and deep learning, such as fractal-wavelet analysis and image fusion based on deep learning (FDL), are demonstrating powerful capabilities in enhancing anomaly detection and integrating heterogeneous exploration datasets (e.g., remote sensing, geophysics, geochemistry) for mineral prospectivity mapping [50,51]. Among these, RF demonstrates strong performance in handling non-linear relationships and high-dimensional geoscientific data. Furthermore, swarm intelligence optimization algorithms, such as Grey Wolf Optimizer (GWO), can enhance model generalization capability and help prevent overfitting [52].

Currently, mineral prospectivity mapping of the Hatu gold deposit concentration area in the Western Junggar still lacks systematic research that integrates multi-source remote sensing data with optimized machine learning models. To address this gap, this study aims to establish an integrated framework for gold prospectivity mapping in the Hatu area of the Western Junggar by synthesizing multi-source remote sensing datasets and applying machine learning approaches, specifically including the following:

  (i)

Integrating ASTER, GF-5, and Landsat-9 data to extract iron-staining anomalies, hydroxyl anomalies, and hyperspectral alteration mineral assemblages, and combining these with geological, geochemical, and deposit (occurrence) information;

 (ii)

Building an RF model based on fishnet grid and using the GWO algorithm to tune hyperparameters, thereby achieving quantitative integration of multi-source evidence layers;

(iii)

Comparing qualitative results from multi-source overlay with quantitative results from machine-learning-based probability models, progressing from broad perspective areas to specific exploration targets and verifying prediction accuracy through field reconnaissance.

In summary, this paper seeks to establish a comprehensive gold exploration model through machine learning analysis of remote sensing information, metallogenic geological characteristics, and geochemical anomalies, providing methodological support for mineral prospectivity mapping of orogenic gold deposits in the CAOB.

2. Regional Geological Background

The Western Junggar region is located at the convergence of the Kazakhstan, Tarim, and Siberian plates [53,54,55,56] and represents an important component of CAOB (Figure 1a). It borders the Altay orogen to the north, connects with the Tianshan orogen to the south, and adjoins the Junggar Basin to the east (Figure 1b). The region preserves the tectonic framework of Paleozoic multiple oceanic-island accretion and juvenile crustal growth [55], displays multi-stage tectonic and magmatic activities, and forms the Au-Cu-Mo-W-Sb-Be-U metallogenic series under the Late Carboniferous-Permian post-collisional extensional setting in the south of Western Junggar. This tectonic environment has produced diverse deposits, including high-temperature hydrothermal tungsten deposits, porphyry copper–molybdenum systems, volcanic–hydrothermal beryllium–polymetallic deposits, and orogenic gold deposits [48,49,50,51,52,53,54,55,56,57,58,59,60]. The region is characterized by widespread Devonian-Carboniferous marine volcanic–sedimentary formations and ophiolitic mélanges [61]. During the Permian period, the area evolved into a continental volcanic–molasse sedimentary environment. Collectively, this sequence records the tectonic evolution of the Paleo-Asian Ocean, including its subduction, closure, and post-collisional extension [62], thereby providing a favorable dynamic setting for deep magmatic activity, fluid migration, and metal mineralization.

The Hatu area mainly exposes Devonian-Carboniferous volcanic–sedimentary formations (Figure 1c), including the following:

• Tuffaceous siltstone and tuff of the Baogutu Formation (C₁b);
• Tuffaceous sandstone and tuffaceous siltstone of the Xibeikulasi Formation (C₁x);
• Welded tuff, andesite, and rhyolite of the Tailegula Formation (C₁t);
• Shallow-marine intermediate tuff and tuffaceous sandstone of the Kelumudi Formation (D₂k);
• Argillaceous siltstone and siltstone of the Baerleike Formation (D₂b).

Among these, the Xibeikulasi and Baogutu Formations constitute the most important ore-hosting strata in this region. The magmatic rocks in the area are dominated by Late Paleozoic A-type granites, which commonly appear as large, triangular-shaped batholiths such as the Tiechanggou pluton. The interior of these batholiths exhibits weak gold mineralization [8,61], and only a few gold deposits or occurrences are found within them. However, several deposits occur south of the pluton [8,63], indicating it exerts a controlling effect on gold mineralization. In addition, Late Carboniferous I-type granites, represented by smaller plutons such as the Bieluagaxi and Baobei plutons, show strong gold metallogenic potential and serve as the principal ore-bearing plutons. Gold mineralization is most developed near its contact with surrounding country rocks, and most of the discovered gold deposits (occurrences) in the area display a close genetic relationship with these intrusions [8,64]. Structurally, the NE-trending Hatu and Anqi faults are the primary ore-controlling structures, while NW-, EW-, and NE-trending secondary faults act as important ore-hosting structures. Multi-stage fault activity has facilitated gold enrichment along fracture zones and their intersections, accompanied by intense hydrothermal alteration dominated by silicification, sericitization, pyritization, chloritization, kaolinization, and illitization [65].

The Hatu gold deposit, the most well-known super-large deposit in this area, has reported cumulative production and resources of approximately 62 tons at an average grade of about 4.99 g/t Au [66]. Individual orebodies within the deposit commonly show grades ranging from 4.3 to 16.6 g/t Au, with localized zones reaching up to 300 g/t Au [67]. It is hosted within a sequence of felsic volcaniclastic rocks and mafic volcanic rocks belonging to the Tailegula Formation. Its formation is primarily controlled by the Anqi Fault and associated intersecting structural zones. The economic orebodies are mainly divided into auriferous quartz vein-type and disseminated altered rock-type, characterized by strong silicification, sericitization, pyritization, and chlorite–epidote alteration. The K26 gold deposit occurs within volcaniclastic rocks of the Baerleike formation and is influenced by both the Bieluagaxi pluton and nearby fault intersection zones. Its ore types include quartz vein-type and altered rock-type, characterized by silicification, sericitization, and chloritization, among other components. The Gezigou East deposit lies in the hanging wall of the Anqi Fault and is closely associated with acidic magmatic activity. It is accompanied by pronounced Au-As-Sb geochemical anomalies and displays alteration assemblages including chloritization, carbonatization, and sericitization [68,69]. The Huilvshan gold deposit is located in altered basalt of the Baogutu Formation, showing well-developed chloritization, silicification, and sideritization, and is clearly controlled by the NE-trending Anqi Fault along with secondary NE-, NEE-, and near-EW-trending fault zones. In summary, the gold mineralization in this area is mainly controlled by NE-trending faults and subsidiary structural systems, which commonly develop notable alteration signatures such as silicification, sericitization, and chloritization. In addition, some mineralized zones are closely related to granitic plutons, demonstrating excellent innate advantages for remote sensing-based information extraction.

Figure 1. Tectonic locations of West Junggar and the Hatu gold district. (a) Simplified tectonic map of the Central Asian Orogenic Belt (CAOB), showing its position between the East European, Siberian, Tarim, and North China cratons [70]; (b) regional geological framework of North Xinjiang, illustrating the West Junggar terrane and adjacent tectonic units [71]; (c) geological map of the Hatu gold district, showing major lithostratigraphic units, granitoid intrusions, and fault systems.

Figure 1. Tectonic locations of West Junggar and the Hatu gold district. (a) Simplified tectonic map of the Central Asian Orogenic Belt (CAOB), showing its position between the East European, Siberian, Tarim, and North China cratons [70]; (b) regional geological framework of North Xinjiang, illustrating the West Junggar terrane and adjacent tectonic units [71]; (c) geological map of the Hatu gold district, showing major lithostratigraphic units, granitoid intrusions, and fault systems.

3. Data and Methods

3.1. Data Sources

To construct a comprehensive evidence system of “structure-alteration-mineral-geochemistry-mineral occurrences”, this study integrated three types of remote sensing data with regional geological, geochemical exploration, and mineral deposit datasets. For the remote sensing component, ASTER multispectral data, Landsat-9 satellite data, and GF-5 AHSI hyperspectral data were selected.

ASTER is a sensor onboard NASA’s Terra satellite, jointly developed by the National Aeronautics and Space Administration (NASA, Washington, DC, USA) and the Ministry of Economy, Trade and Industry of Japan (METI, Tokyo, Japan), and launched on 18 December 1999. It contains 14 spectral bands, including three Visible and Near-Infrared bands (VNIR, 0.52–0.86 μm, spatial resolution 15 m), six Short-Wave Infrared bands (SWIR, 1.60–2.43 μm, spatial resolution 30 m), and five Thermal Infrared bands (TIR, 8.125–11.65 μm, spatial resolution 90 m) [72,73]. The two ASTER scenes used in this study were downloaded from Earthdata, with IDs AST_L1T_00309212002052642_20150425054420_7401 and AST_L1T_00309212002052651_20150425054420_7405. Both scenes contain minimal cloud or snow cover and were applied to extract iron oxide and hydroxide alteration minerals.

Landsat-9 satellite was successfully launched on 27 September 2021. As an upgraded version of Landsat-8, it incorporates several technical improvements while inheriting the excellent performance of the predecessor. Landsat-9 provides higher-resolution and higher-quality images, and its Earth-observation capability surpasses that of Landsat-8. The satellite is equipped with the Operational Land Imager 2 (OLI-2) and the Thermal Infrared Sensor 2 (TIRS-2), both developed by the National Aeronautics and Space Administration (NASA, Washington, DC, USA) Landsat-9 contains nine multispectral bands with 30-m spatial resolution, one 15-m panchromatic band, and 100-m thermal infrared bands. The dataset used in this study was downloaded from the United States Geological Survey (USGS), with the identification number LC09_L1TP_145028_20241104_20241104_02_T1. This data was applied for structural interpretation and enhancement of lithological boundaries.

China’s Gaofen-5 satellite was launched on 9 May 2018. The Advanced Visible and Shortwave Infrared Hyperspectral Camera (AHSI) onboard the satellite contains 330 spectral bands, including 150 Visible and Near-Infrared bands (VNIR, 0.39–1.03 μm; spectral resolution 5 nm) and 180 Shortwave Infrared bands (SWIR, 1.0–2.5 μm; spectral resolution 10 nm). Each AHSI scene covers an area of approximately 60 × 60 km² with a spatial resolution of 30 m [74]. The GF-5 AHSI image used in this study was obtained from the China Centre for Resources Satellite Data and Application, with the identification number GF5_AHSI_E84.43_N46.08_20190731_006532_L10000052492. This cloud-free dataset was employed to extract weak information of alteration minerals such as muscovite, limonite, montmorillonite, goethite, amphibole, kaolinite, chlorite, clinochlore, and plagioclase in the Hatu gold deposit area.

In this study, all remote sensing datasets were selected from cloud-free scenes acquired during comparable dry-season periods to minimize the influence of seasonal vegetation and surface condition variability. GF-5 hyperspectral imagery was used for SAM-based alteration mineral identification, whereas ASTER data was employed to derive iron-staining and hydroxyl alteration indicators using PCA-based methods.

In addition to remote sensing data, this study also incorporates multiple types of auxiliary materials. The first is the 1:50,000 regional geological survey dataset, including stratum-lithology distribution, locations of granitic intrusive bodies, and characteristics of their contact zones. This data was used to constrain remote sensing interpretation results and support subsequent modeling analyses. The second source is a 12.5 m resolution digital elevation model (DEM), from which slope maps were generated to analyze the terrain morphology and its spatial correlation with geological structures and alteration zones for subsequent geomorphological interpretation. The third source is geochemical data, primarily consisting of geochemical exploration anomalies for elements such as Au, As, Sb, Ag, Pb, Zn, and Mo. After determining anomaly thresholds, this data provides quantitative geochemical evidence layers for the model. The fourth source is the mineral–geological database, covering the spatial distribution of 86 gold deposits (occurrences) and used for constructing labels for both the training and validation sets.

3.2. Data Preprocessing

In mineral prospectivity mapping studies, the consistency and comparability of multi-source data directly influence the reliability of subsequent modeling results. Therefore, before data utilization, this study performed systematic preprocessing and standardization on the remote sensing images, geological information, and geochemical datasets, converting them into quantifiable evidence layers under a unified spatial reference framework.

For the remote sensing data, the ASTER and Landsat-9 datasets used in this study are Level 1T products that have already undergone precise terrain correction. Therefore, only radiometric calibration, atmospheric correction, image fusion, and regional cropping were required to meet the research objectives. However, the high-resolution five-band hyperspectral data requires radiometric calibration, band combination, atmospheric correction, orthorectification, band removal, and additional post-processing steps before being used for refined extraction of alteration information in this study. Through band removal, water vapor absorption bands (no-data bands), low Signal-to-Noise Ratio (SNR) bands, and detector overlap bands were removed, resulting in final hyperspectral data containing 285 bands (

Table 1. List of bands removed from GF-5 hyperspectral data.

Band Number	Wavelength Position/nm	Reason for Exclusion	Spectral Range	Number of Excluded Bands
Band1~band2	390~394	low signal-to-noise ratio	VNIR	2
Band191~band207	1342~1477	Water vapor absorption band, no data	SWIR	43
Band247~band261	1813~1931	bad track/dead band
Band261~band266	1805~1973	Water vapor absorption band, no data
Band325~band330	2471~2513	low signal-to-noise ratio

).

Despite these preprocessing measures, GF-5 remote sensing images may still contain noise, leading to abrupt changes or interruptions in pixel spectra. To mitigate this issue, the data was smoothed using the Savitzky-Golay (S-G) filtering method, thereby producing more reliable characteristic reflectance curves, as illustrated in Figure 2.

Atmospheric correction was performed using the FLAASH module in ENVI 5.3 based on an atmospheric radiative transfer model, producing reflectance data scaled between 0 and 1.0, which provides the foundation for subsequent remote sensing interpretation.

In addition to remote sensing images, the geological and geochemical datasets were also subjected to standardized processing. Structural vector data was calculated spatially to determine the length proportion and intersection densities within grid cells, enabling quantitative assessment of structural control on mineralization. The geochemical dataset covers Au, As, Sb, Ag, Pb, Zn, Mo, Bi, Sn, and other ore-sensitive elements. After anomaly threshold determination and spatial interpolation, the data was converted into gridded formats. Through this processing workflow, all datasets were converted into spatially consistent and quantitatively comparable evidence layers.

3.3. Main Research Methods

3.3.1. Principal Component Analysis

Principal Component Analysis (PCA) is a commonly used remote sensing data processing method that transforms multiple correlated bands into several new, independent principal components through linear transformation, thereby maximizing variance information and reducing data redundancy [75]. The core concept is to project the original band matrix X into a new feature space to maximize variance and emphasize abnormal information. This is expressed as follows:

Y = XW,

(1)

where W is the eigenvector matrix and Y is the principal component matrix. By examining the loadings of diagnostic bands on the eigenvectors, iron-stained and hydroxyl alterations can be identified and distinguished from background information [15,16]. In this study, diagnostic bands of the ASTER dataset were selected based on the spectral characteristics of various alteration minerals, and PCA was applied to extract iron-stained and hydroxyl alteration anomalies. This approach effectively improves the accuracy of alteration zones and reveals the spatial distribution patterns of hydrothermal alteration closely associated with mineralization systems.

3.3.2. Spectral Hourglass Method

The Spectral Hourglass Method (SHM) is a commonly used processing workflow for hyperspectral data that involves endmember extraction and mineral identification [76,77]. Its core steps include the following:

-

Applying the Minimum Noise Fraction (MNF) transformation for dimensionality reduction to suppress noise and retain principal feature components;

-

Employing the Pixel Purity Index (PPI) and n-Dimensional Visualization (n-D) methods to identify potential endmembers;

-

Utilizing the Spectral Angle Mapper (SAM) algorithm within the ENVI 5.3 software (L3Harris Geospatial, Broomfield, CO, USA) for spectral similarity discrimination against the USGS spectral library, thereby extracting spatial distribution information of target alteration minerals [76,78,79].

This method has been widely applied in mineral mapping using GF-5 AHSI hyperspectral data and can effectively enhance the accuracy and reliability of alteration identification [80]. Given the 30 m spatial resolution of the GF-5 hyperspectral data, SAM-derived results are interpreted as indicators of alteration mineral presence and spatial distribution patterns, rather than as pixel-level quantitative mineral abundance estimates.

3.3.3. Optimum Index Factor (OIF)

The Optimum Index Factor (OIF) was applied to identify the most informative band combinations of Landsat-9 data. OIF integrates standard deviation and inter-band correlation to select combinations with high information content and minimal redundancy [81]. OIF is mathematically expressed as follows:

$$OIF=\sum _{i=1}^3{S}_i/\sum _{r=1}^3\left|R_{ij}\right|$$

(2)

where S_i is the standard deviation of the i-th band, and R_ij represents the correlation coefficient between bands i and j. A larger OIF value indicates that the band combination contains more non-repetitive information, making it suitable for remote sensing interpretation for structures, lithologies, and alterations. Moreover, comparative evaluation indicated that the 6-5-2 false-color composite is capable of providing the clearest structural and lithological discrimination; therefore, it was adopted for regional fault interpretation.

3.3.4. Random Forest (RF)

RF is an ensemble machine learning algorithm that integrates multiple decision trees for classification and regression [82]. It is particularly effective for processing high-dimensional and nonlinear geoscientific datasets. In this study, RF was employed to quantify mineralization potential by combining multi-source evidence layers, including alteration information, structural attributes, and geochemical anomalies. Known gold deposits were labeled as “1”, while non-mineralized grid cells were labeled as “0”. The model output represents the probability of mineralization within each grid cell, providing a quantitative basis for delineating prospectivity zones [83].

3.3.5. Grey Wolf Optimizer (GWO)

The GWO is a swarm intelligence optimization algorithm inspired by the social hierarchy and hunting behavior of grey wolf packs [52]. Its fundamental principle involves simulating the encircling and chasing of prey by wolves at different hierarchical levels (α, β, δ, and ω), thereby achieving a dynamic balance between global search and local exploitation. The combined effect of its hierarchical information-transfer mechanism and adaptive search strategy enables effective balance between global exploration and local exploitation capabilities, ensuring both algorithmic stability and simplicity in parameter configuration [52,84,85] (Figure 3). In this study, GWO was employed to optimize key hyperparameters of the RF model, such as the number of decision trees and the maximum number of features, thus enhancing the model’s generalization ability and prediction accuracy.

4. Remote Sensing Mapping Results and Prospective Zone Delineation

4.1. Geological Structural Interpretation

To highlight the structural characteristics of the study area, Gram–Schmidt fusion was applied to Landsat-9 multispectral imagery and the 15-m panchromatic band. Based on the Optimum Index Factor (OIF) analysis, the 6-5-2 band combination was selected for false-color composite to enhance the recognition of faults and lithological boundaries. Geological structures are often expressed in remote sensing images as linear features aligned along specific directions and marked by water systems, micro-landforms, lithological boundaries, and other indicators [87]. Remote sensing imagery can clearly reflect various geological structural elements, particularly fault structures and circular features that are closely associated with mineralization and ore control [88]. Building upon existing geological data, detailed remote sensing interpretation of linear and circular structures within the study area was carried out using Landsat-9 imagery (Figure 4).

The interpretation results indicate that linear structures are highly developed within the study area (Figure 4a), with dominant NE-trending features, followed by NNE, NEE, NW, WE, and NWW directions (Figure 4b). Among these, the NE-trending Hatu Fault and Anqi Fault constitute the two principal regional ore-controlling structures. They not only govern the spatial distribution of strata and plutons but also serve as vertical conduits for hydrothermal fluid migration. The associated feather fractures and horsetail-like secondary faults provide favorable conditions for ore-fluid accumulation at their intersections and within dilation zones. The younger NW-trending faults typically cut across the NE-trending structures, and their intersections commonly form favorable mineralization nodes.

Furthermore, circular structures can be locally identified within the study area (Figure 4c). Their morphology generally appears circular, semicircular, or polygonal, and their spatial distribution corresponds closely with granitic intrusions. These features reflect the superimposition and modification of overlying strata by deep magmatic activity. When they overlap with linear fractures, such zones often display concentrated alteration and well-developed mineralization anomalies. Overall, the structural interpretation results indicate that the main gold mineralization in the study area is strongly controlled by the intersections of NE-trending fractures with secondary NW- and NE-trending faults. This structural framework aligns closely with the spatial distribution of known gold deposits (including the Qi-I, Qi-II, Qi-III, Qi-IV, K26, Gezigou, and Huilvshan gold deposits), highlighting its obvious fracture structure ore-control characteristic.

4.2. PCA Results Based on ASTER Data

PCA was applied to the ASTER dataset to extract iron-stained and hydroxyl alteration anomalies associated with hydrothermal activity. Diagnostic band combinations were selected based on the spectral characteristics of alteration minerals (Figure 5): Bands 1-2-3-4 for iron-stained (

Table 2. Changes in principal components of bands 1, 2, 3, and 4 of ASTER data.

Eigenvectors	Band 1	Band 2	Band 3	Band 4
PC1	0.34959	0.430536	0.529391	0.642005
PC2	0.5327	0.492822	0.076608	−0.683732
PC3	0.301063	0.281731	−0.843976	0.343064
PC4	0.709492	−0.701707	0.039762	0.051447

), Bands 3-4-6-7 for Al-OH (

Table 3. Changes in principal components of bands 3, 4, 6, and 7 of ASTER data.

Eigenvectors	Band 3	Band 4	Band 6	Band 7
PC1	−0.407229	−0.573331	−0.492264	−0.512964
PC2	−0.899892	0.120127	0.272808	0.318338
PC3	−0.153123	0.767357	−0.133227	−0.60825
PC4	−0.030196	0.260816	−0.815784	0.515326

), and Bands 1-3-4-8 for Mg-OH alterations (

Table 4. Changes in principal components of bands 1, 3, 4, and 8 of ASTER data.

Eigenvectors	Band 1	Band 3	Band 4	Band 8
PC1	0.276339	0.479883	0.642564	0.529585
PC2	0.624727	0.567997	−0.434822	−0.31309
PC3	0.606791	−0.576741	−0.23665	0.493124
PC4	−0.406396	0.338321	−0.58484	0.615096

). The corresponding principal components (PC4, PC4, and PC3) were used to emphasize each alteration type. To translate the continuous pixel values in these component images into discrete alteration intensity levels, we applied a standardized thresholding technique based on statistical analysis of the anomaly distribution. Following established methodologies for alteration mapping [16,23], the thresholds were defined as the mean value plus multiples of the standard deviation (μ + nσ). Specifically, pixels with values exceeding μ + 3σ, μ + 2.5σ, and μ + 2σ were classified as First-Level, Second-Level, and Third-Level anomalies, respectively. In addition, the Crosta technique with a 5 × 5 median filter was adopted to suppress isolated noise, thereby enabling effective extraction of alteration information from the ASTER data (Figure 6 and Figure 7).

The iron-staining alteration anomalies in the study area were successfully extracted through PCA of the ASTER data and were classified into First-Level iron anomaly, Second-Level iron anomaly, and Third-Level iron anomaly according to their intensity (Figure 6a). The First-Level iron anomalies are mainly concentrated along the margins of the Tiechanggou pluton in the northeastern part and in the shallow-coverage zone in the northern part. They also appear sporadically in the central area and exhibit a linear distribution in the southwestern corner, which is related to the NW-trending deep fault. The Second-Level iron anomalies occur scattered around the First-Level anomalies, whereas the Third-Level anomalies show the most dispersed distribution, primarily occupying the highly eroded eastern part of the study area. Fishnet-based analysis of the three-dimensional and two-dimensional distributions of alteration anomalies (Figure 6b,c) indicates that iron-staining alteration is significantly enriched in the northern and central-eastern parts of the target area, where anomalies often occur in dense banded or planar forms. It is worth noting that, although the iron-staining alteration anomalies in the western area have a limited extent, they intensively concentrate First-Level and Second-Level iron anomalies. This distribution suggests the possible presence of structurally controlled hydrothermal activity.

The Al-OH anomalies in the study area exhibit clear spatial zoning. First-Level, high-intensity anomalies occur predominantly in the northern part of the Tiechanggou pluton, with smaller occurrences in the central and eastern sections of the region. Second-Level, medium-intensity anomalies are generally distributed around the margins of the First-Level zones, while Third-Level, weak-intensity anomalies are more concentrated in the western and southern parts of the study area. Mg-OH alteration anomalies are primarily developed along the periphery of the Tiechanggou pluton, with sporadic distribution still observed within the central part of the pluton. Medium- and low-intensity anomalies are widely dispersed across the central and eastern areas of the regions (Figure 7a). The comprehensive hydroxyl alteration anomalies, shown in both three-dimensional and two-dimensional spatial representations (Figure 7b,c), reveal significant enrichment zones in the northern and central parts of the study area, typically forming planar or banded patterns. Moreover, the anomaly intensity in the northern part of the Tiechanggou pluton displays a decreasing gradient trend that radiates outward from the pluton’s center.

Overall, the iron-staining and hydroxyl alteration anomalies extracted from the ASTER data correspond closely to NE-trending strike-slip faults, NW- and NE-trending intersection structures, and contact zones around plutons. This spatial consistency indicates that the magmatic–hydrothermal activity in the Hatu area is intense, and its fluid migration along fault-controlled pathways is clearly expressed.

4.3. Spectral Hourglass Method Analysis Results Based on GF-5 Data

To further extract the distribution characteristics of alteration minerals in the study area, this study employed GF-5 hyperspectral data for analysis using the Spectral Hourglass Method.

First, the Minimum Noise Fraction (MNF) transformation was applied to the GF-5 hyperspectral data for dimensionality reduction and noise suppression, effectively reducing inter-band correlation and enhancing diagnostic spectral features. The MNF transformation improves the SNR and concentrates most of the meaningful spectral information within the first few components [89]. In hyperspectral imagery, mixed pixels (those containing multiple surface materials) are common due to limited spatial resolution and surface heterogeneity, and their mixed spectra can obscure the diagnostic absorption features of individual minerals. Therefore, extracting spectrally pure pixels (endmembers) is a critical step in hyperspectral analysis [90]. Subsequently, the Pixel Purity Index (PPI) algorithm was employed to identify high-purity pixels, using 10,000 iterations and a threshold value of 2.5 [91,92]. With the n-Dimensional Visualizer tool in ENVI 5.3 software, alteration mineral endmembers were selected from n-dimensional scatter plots and exported as Regions of Interest (RoIs). The spectral data for these pure endmembers were then obtained, confirming the presence of alteration minerals such as muscovite, limonite, montmorillonite, goethite, hornblende, kaolinite, chlorite, clinochlore, and plagioclase in the study area. Using the built-in USGS spectral library in ENVI 5.3, a standard spectral library for the alteration of minerals in the region was established (Figure 8a) and resampled to match the spectral resolution of the GF-5 hyperspectral data (Figure 8b).

Finally, alteration mineral mapping was conducted using the standard spectral library. Through comparative analysis of the endmember spectral curves, nine mineral types were identified: muscovite, limonite, montmorillonite, goethite, hornblende, kaolinite, chlorite, clinochlore, and plagioclase. The entire study area was then classified using the Spectral Angle Mapper (SAM) method, completing the extraction of hyperspectral remote sensing-based alteration information [31] (Figure 9).

From the mineral mapping results of GF-5 hyperspectral data, it is evident that montmorillonitization is widely developed within the Middle Devonian–Early Carboniferous volcanic–sedimentary strata in the western part of the region, generally exhibiting areal distribution patterns. Montmorillonitization is also extensively developed within the Bieluagaxi pluton, indicating relatively strong intermediate argillic alteration that is clearly associated with magmatic–hydrothermal processes associated with the pluton. Similarly, a muscovitization anomaly zone has been identified in the northwestern part of the study area, reflecting pronounced intermediate-temperature hydrothermal activity. A NE-trending banded chlorite zone is interpreted in the central region, showing strong spatial consistency with the regional Anqi strike-slip fault distribution and demonstrating obvious fault-controlled characteristics. Furthermore, the Hatu gold mining area displays characteristic areal combinations of intense chloritization, clinochloritization, and montmorillonitization, with a spatial transition from clinochloritization in the interior to montmorillonitization toward the outer zones. This alteration zonation is generally regarded as a typical signature of orogenic gold deposits in the Western Junggar region [59]. Overall, the alteration mineral assemblages in the region present a distinct spatial distribution pattern: the northwestern part is dominated by muscovitization; the central and northeastern regions show large-area occurrences of chloritization and clinochloritization; and the western part is primarily characterized by montmorillonitization. The chloritization in the central part exhibits a NE-trending banded configuration with considerable lateral continuity, accompanied locally by limonite and goethite anomalies. These features indicate intense supergene oxidation and iron-staining alteration, as exemplified by the shallow oxidation zone in the northeastern piedmont of the Hatu gold mining area.

4.4. Multi-Source Remote Sensing Information Integration for Prediction

This study integrates stratigraphic, structural, and gold-occurrence data with multiple remote sensing products, including Landsat-9-based structural interpretation results, ASTER-derived hydroxyl and iron-staining alteration anomalies, and alteration minerals identified from GF-5 hyperspectral data (muscovite, limonite, montmorillonite, goethite, amphibole, kaolinite, chlorite, clinochlore, and plagioclase). These datasets were combined through spatial overlay analysis to identify zones of interest. The preliminary delineation of prospective areas was guided by the regional exploration model for orogenic gold deposits [93], resulting in the identification of five exploration targets in the Hatu area designated as T1, T2, T3, T4, and T5 (Figure 10). This outcome reflects the geological characteristics of multi-factor coupled ore control involving “fault structures—granitic plutons—mineralization alteration—gold occurrences”.

Among them, the T1 prospective area is located in the western part of the Bieluagaxi pluton, where intersections between NE- and NW-trending secondary faults form a fault-controlled ore-hosting network. Numerous gold occurrences are distributed along the margins of the pluton, with the main host horizons belonging to the Kelumudi and Baerleike Formations. Alteration assemblages such as muscovitization, chloritization, and montmorillonitization are well developed, reflecting a mineralization pattern controlled jointly by the magmatic–hydrothermal activity and fault structures centered on the small-sized pluton. The T2 prospective area encompasses the Hatu gold deposit (Qi-I–Qi-IV) and its surrounding zone. The strata are mainly composed of the Baogutu and Xibeikulasi Formations, exhibiting NE-trending zonal distributions of chloritization, clinochloritization, and goethitization. This area is confined between the Hatu Fault and the Anqi Fault, demonstrating clear geological characteristics of ore-control by major fault systems, with the NE-trending faults providing the most favorable ore-hosting structures [89]. The T3 prospective area is situated in the Gezigou gold deposit area in the central-eastern part of the study region. It features a NE-trending chloritization zone accompanied by Level 1–3 Mg-OH anomalies and Level 1–2 iron-staining anomalies. Furthermore, several gold deposits, including the Qi-V and Mandonggou gold deposits, occur at intersections of NW- and NE-trending secondary faults. The T4 prospective area lies in the eastern segment of the Hatu Fault in the northeastern part of the study region. Here, NE-trending faults intersect intensively with near-EW-trending secondary structures, forming numerous structural blocks. Extensive chloritization and clinochloritization occur in the area, together with localized intense iron-staining anomalies, demonstrating favorable metallogenic potential. The T5 prospective area is located in the footwall of the major Anqi Fault, where NE- and NW-trending secondary faults show good spatial correlation with known gold occurrences. The main ore-hosting strata are the Baogutu and Tailegula Formations, with widespread chloritization and clinochloritization, locally overprinted by iron-staining and sericitization, also suggesting notable mineralization potential.

Overall, the delineation of the T1-T5 exploration targets is based on the consistency of multi-source remote sensing information, highlighting the combined control of NE-trending strike-slip fault systems, small granitic plutons, and alteration zonation on gold mineralization. This framework provides target areas for subsequent quantitative machine learning validation.

5. Establishment and Optimization of the Prospecting Model

5.1. Evidence Layer Integration

In mineral prospectivity mapping, the introduction of standardized fishnet cells for spatial partitioning is a crucial step for achieving quantitative integration of multi-source geoscientific information. In this study, the Hatu area (1399 km²) was divided into 796 prediction units, each measuring 1.3 km × 1.3 km, to facilitate the integration of geological data, multi-source remote sensing information, and geochemical anomaly indicators.

Combining geological survey data with remote sensing interpretation results, the mineral prospectivity predictors were converted into quantifiable evidence layers within the GIS platform. These include indicators such as the area proportion of remote sensing anomalies, the length proportion of structures, intersection density, and the mean values of geochemical anomalies. After normalization, these parameters were defined as feature variables, forming the input factor set for the quantitative mineral prospectivity prediction model (

Table 5. Quantitative prospecting model of the Hatu gold mine area.

Data Source	No.	Evidence Layer	Eigenvalue
GF-5	1	Muscovite	Proportion of anomalous area within the fishnet
	2	Limonite
	3	Montmorillonite
	4	Goethite
	5	Hornblende
	6	Kaolinite
	7	Chlorite
	8	Clinochlore
	9	Plagioclase
ASTER	10	Iron staining	Proportion of anomalous area within the fishnet
	11	Al-OH
	12	Mg-OH
Landsat-9 and geological data	13	Structures	Proportion of structural length within the fishnet
	14	Structural intersections	Density of structural intersections
Geochemistry	15	Au, Pb, Ag, Zn, As, Bi, Mo, Sb, Sn	Anomaly mean values within the fishnet

).

Based on the comprehensive quantitative prediction model, a gridding approach was adopted to integrate multi-source information and establish a binary classification sample library with clear geological significance. The construction of this sample library mainly involves two components: the selection of positive and negative samples and the determination of predictive variables. The selection of positive and negative samples in the training dataset should ideally meet three conditions:

(1)

The number of positive and negative samples should be approximately balanced;

(2)

The spatial distance between positive and negative samples should not be excessively close;

(3)

Negative samples should be randomly and evenly distributed across the study area as much as possible [94].

Following these principles, this study used gridded known ore bodies as positive samples, while negative samples were randomly extracted from non-mineralized areas. In total, 86 positive samples and 86 negative samples were selected. For mineral prospectivity mapping, the final prediction results are divided into mineralized and non-mineralized classes. During training sample construction, mineralized units were assigned a label of 1, whereas non-mineralized units were labeled as 0.

5.2. Mineral Prospectivity Modeling and Accuracy Assessment

During model construction, this study adopted a stratified sampling strategy to divide the complete dataset into training and test sets at a 7:3 ratio. This approach ensured that the proportions of “1” and “0” labels in each subset remained consistent with those of the original dataset, thereby reducing potential evaluation bias caused by imbalanced sample distribution. Based on the partitioned datasets, three machine learning algorithms—Random Forest (RF), Support Vector Machine (SVM), and eXtreme Gradient Boosting (XG-Boost)—were systematically applied. Through comparative analysis of these algorithms, the model that best matched the mineralization patterns of the study area was ultimately identified.

Model accuracy analysis is essential for evaluating model performance and typically involves methods such as the confusion matrix, ROC curve and AUC value, cross-validation, learning curves, and feature importance analysis. Among these, the ROC curve is extensively used in machine learning and data mining research [95,96] and is frequently employed to compare the performance of different mineral prospectivity models. The Area Under the ROC Curve (AUC) is a threshold-related evaluation metric, and larger AUC values indicate better model classification performance. Together, the ROC curve and AUC provide a comprehensive assessment of classification accuracy. In addition, cross-validation and learning curves can further evaluate the model’s robustness and generalization ability [97,98,99]. In this study, a multi-dimensional evaluation framework was applied: model accuracy, precision, recall, and F1-score were calculated using the confusion matrix (

Table 6. Model evaluation metrics from the confusion matrix.

	ACC	Precision	Recall	F1	AUC
Random Forest	0.904	0.885	0.920	0.902	0.933
Support Vector Machine	0.692	0.698	0.698	0.688	0.808
XG-Boost	0.827	0.827	0.827	0.827	0.910

), while ROC curves and AUC values were generated (Figure 11) to systematically assess the classification effectiveness of the different models.

The results indicate that the RF model outperforms both SVM and XG-Boost in terms of accuracy, precision, recall, and AUC, demonstrating stronger generalization capability and better adaptability to high-dimensional geoscientific data. Therefore, RF was selected as the core predictive model for this study.

5.3. Grey Wolf Optimizer and Model Optimization Results

To further enhance model performance, GWO was introduced to conduct global optimization of the key hyperparameters in the RF model. This algorithm simulates the hierarchical search behavior of a grey wolf pack, enabling it to effectively avoid convergence toward local optima caused by inappropriate parameter settings. Additionally, it helps reduce overfitting, thereby enhancing the model’s generalization capability and ultimately improving the accuracy of mineral prospectivity prediction [55,100,101].

After optimization, the accuracy of the GWO-RF model increased from 0.885 to 0.942, and precision improved from 0.920 to 0.948, while recall and F1-score rose from 0.902 to 0.942, respectively. Furthermore, feature-importance analysis showed that the optimized model exhibited more focused feature selection, contributing to improved predictive accuracy. ROC curve analysis further confirmed these improvements, with the AUC value increasing from 0.933 to 0.95, indicating stronger capability in distinguishing between “non-prospective” and “prospective” categories (Figure 12).

The development of the SHapley Additive exPlanations (SHAP) value, proposed in prior research, has greatly enhanced the interpretability of machine learning models [102]. By attributing the contribution of each feature to the final prediction, SHAP provides a unified framework for interpreting model outputs. Owing to its solid theoretical foundation and broad applicability across different model types, this method has gained increasing attention in recent years [103]. Consequently, to improve interpretability, this study conducted a SHAP-based feature importance analysis using the PyCaret library (https://www.pycaret.org). This approach assigns importance scores to features according to their significance in predicting the target variable, thereby facilitating the role of each individual feature within the classification model. Feature importance analysis of the GWO-RF model using the SHAP algorithm produced the following ranking of feature contributions: Au > As > Structural density > Sb > Iron-stained alteration > Ag > Mo > Zn > Pb > Mg-OH. Among these, Au, As, structural density, and Sb exerted the strongest influence on the model’s prediction (Figure 13), greatly surpassing the contributions of other variables. This result underscores the close association between gold mineralization and zones of high fault-intersection density, emphasizing the dominant structural control on mineralization. Furthermore, the combined Au-As-Sb anomaly emerges as a key indicator for gold mineralization. These modeling results strongly align with the metallogenic characteristics and exploration criteria of orogenic gold deposits worldwide [104]. Additionally, Ag, Mo, Zn, Pb anomalies, and Mg-OH alteration patterns also show good mineralization correlation, suggesting that chloritization in this region carries notable indicative significance for gold mineralization.

Classification modeling and mineralization probability calculation were conducted based on the identified ore-controlling factors. Ultimately, the resulting spatial visualization of mineralization probability for each grid cell (Figure 14) enabled the quantitative delineation of gold prospective areas within the study region, providing a probabilistic foundation for subsequent target prioritization.

6. Discussion

6.1. Evaluation of Prediction Results

Based on the ore-controlling factors, classification modeling and mineralization probability were conducted, ultimately enabling secondary quantitative delineation of gold prospective areas through spatial visualization of mineralization probabilities for each grid cell (Figure 14). Comparative analysis showed that the five prospective areas preliminarily identified from the multi-source remote sensing overlay (Figure 10) exhibit strong spatial consistency with the quantitative prediction results generated by the RF-GWO optimized model (Figure 14). These areas largely coincide with the high-probability anomaly zones (exploration targets) derived from model calculations, validating the effectiveness of remote sensing anomalies in mineral prospectivity mapping.

To quantitatively validate the predictive performance of the model with respect to known mineralization, a spatial efficiency analysis was conducted. This analysis quantified the capture rate of known gold occurrences within the highest-probability zones identified by the model. Pixels with predicted probabilities exceeding 0.8 account for approximately 15% of the total study area, yet contain 84% of all known gold occurrences, yielding a high enrichment ratio of 5.6. This result provides robust quantitative support for a significant positive spatial correlation between the model-derived prospectivity and the distribution of known mineral deposits. Consequently, spatially contiguous areas with probability values consistently exceeding 0.8 were selected as final high-confidence exploration targets, designated as Y1 and Y2 based on their distinct metallogenic characteristics. The Y1 exploration target corresponds to the eastern portion of the T2 prospective area and its surrounding zone, while the Y2 target encompasses the T3 prospective area and its adjacent region. Other prospective areas with probability values below 0.8 were classified as high-risk zones and consequently excluded.

To further elucidate the geological significance of these targets, Figure 15 presents a final integrated map synthesizing multi-source remote sensing alteration signatures, the geological framework, major fault systems, and all known gold deposits and occurrences. This map provides a comprehensive geological context, clearly illustrating the close spatial association of the Y1 and Y2 targets with ore-controlling structures and well-developed hydrothermal alteration zones. The strong alignment between the model-derived targets and these mappable geological alteration patterns underscores the reliability and geological rationale of the prospectivity model.

Comprehensive Analysis of Target Areas

The Y1 exploration target area is mainly hosted within the Middle Devonian Kelumudi Formation and the Baerleike Formation, and it also exposes the Late Carboniferous Bieluagaxi pluton together with the southern part of the Tiechanggou pluton (Figure 16a,b). Gold occurrences are concentrated along regional NE-trending faults and associated secondary fault zones, with some distributed at the edge and within the interior of the Bieluagaxi pluton, as well as along the southern margin of the Tiechanggou pluton. The K26 gold deposit lies within this target area. Remote sensing results reveal widespread chloritization, montmorillonitization, limonitization, and kaolinization, accompanied by Level 1 iron-staining and hydroxyl anomalies, reflecting strong hydrothermal alteration. Machine learning further confirms that this area represents the high-probability mineralization core zone of the T1 prospective area. It contains multiple favorable mineralization conditions, including intersections of NE-, NW-, and near-EW-trending faults, as well as contact zones of Late Carboniferous granitic plutons. This demonstrates good prospects for discovering altered rock-type and quartz-vein-type gold orebodies, and its mineralization model aligns well with the typical characteristics of “fault structures + magmatic hydrothermal activity + alteration indicators”.

The Y2 mineralization target area is located in the northeastern peripheral region of the Hatu gold deposit (Figure 16c,d) and is strictly controlled by secondary fracture structures associated with the Anqi Fault and the Hatu Fault. Numerous gold occurrences display a banded distribution along NE-trending secondary fault zones, while the intersections of NE-trending and WE-trending fractures form favorable enrichment sites for gold deposits. Remote sensing interpretation reveals that this area contains well-developed NE-trending banded chlorite–iron-staining anomalies, locally associated with limonitization. Their spatial distribution shows strong spatial coupling with gold occurrences, demonstrating good mineralization alteration zone characteristics. Machine learning predictions further indicate that the mineralization probability in the Y2 target area reaches 0.8–0.9, consistent with the T2 prospective area, confirming that fault structures are the primary ore-controlling factors and that high-density segments constitute the core areas. Compared with the Y1 target area, Y2 more prominently exhibits gold mineralization controlled by NE-trending strike-slip faults and their secondary structures.

Overall, the Y1 target area reflects a magmatic–hydrothermal mineralization model associated with granite intrusion–fault intersections, while the Y2 target area more clearly presents a regional fault-channel-controlled medium- to low-temperature hydrothermal mineralization model. These findings indicate that qualitatively identifying favorable mineralization areas using remote sensing anomalies, followed by applying machine learning methods for quantitative constraints, can effectively narrow the target range and improve prediction accuracy, thereby providing optimal targets for the next step of exploration.

In addition, the geomorphological context derived from the 12.5 m resolution DEM provides further insight into the surface expression of these mineralization systems (Figure 16a,c). The spatial coupling between the high-probability targets (Y1 and Y2) and distinctive slope patterns underscores that geomorphological processes act as a critical interface. They modulate the exposure, preservation, and ultimately the remote sensing detectability of hydrothermal alteration in both intrusion-related and fault-controlled settings [105,106]. Although not used as an explicit predictive variable in the machine learning model, the consistency between terrain morphology and known geological/alteration patterns offers strong, independent support for the geological rationale behind the identified targets.

6.2. Field Verification

To verify the reliability of mineral prospectivity mapping using machine learning that integrates multi-source geological, remote sensing, and geochemical information, this study conducted field geological traverse surveys in the Y1 and Y2 exploration targets. Field verification revealed several newly identified NE-trending fault fracture zones near the eastern boundary of the Bieluagaxi pluton in the Y1 target area. Trench exposures show fracture widths ranging from 1 to 15 m, containing large-scale alteration fracture zones accompanied by intense limonitization, kaolinization, and montmorillonitization (Figure 17a). Limonitized chloritization zones (Figure 17b) and quartz veinlets (Figure 17c) are widely distributed within and adjacent to these fractures, with some veins extending along fault planes and showing strong correspondence with the remote-sensing-detection anomaly zones. The Au grade in the aforementioned zones reached up to 5.3 g/t, demonstrating excellent gold mineralization potential and confirming that the mineralization in this target area is closely linked to magmatic–tectonic hydrothermal processes associated with granitic pluton emplacement.

Within the volcanic strata of the Baogutu Formation in the central segment of the Anqi Fault in the Y2 exploration target, multiple parallel chlorite–epidote quartz veinlets (Figure 17d), limonitized quartz veins (Figure 17e), and other quartz veins (Figure 17f) are observed. The wall rocks display linear distributions of chloritization and epidotization, with extensive limonitization and minor malachitization developed in the upper part of the alteration zone. Gold grades in this alteration zone exceed 0.5 g/t, reflecting strong structural–hydrothermal controls on mineralization.

In summary, the montmorillonitization and chloritization zones in the Y1 target area, together with the iron-staining and chloritization anomalies in the Y2 target area (and their overlapping with Au-As-Sb anomalies, favorable geological units, and key structural features), have been verified to possess strong mineralization potential. Therefore, multi-source remote-sensing-based mineral prospectivity mapping under a machine learning framework is demonstrated to be effective. This approach not only enhances the scientific nature of exploration target delineation but also provides a robust basis for establishing regional metallogenic models and guiding subsequent deep drilling and exploration programs.

7. Conclusions

Based on the above, the following conclusions can be obtained:

(1)

This study comprehensively applied multi-source remote sensing data from ASTER, Landsat-9, and GF-5. The results show that iron-staining, hydroxyl anomalies, and alteration minerals, such as muscovitization, chloritization, and montmorillonitization in the Hatu area, are highly concentrated near granitic plutons and fault intersection zones. By integrating this multi-source remote sensing information, an exploration model was established, leading to the preliminary delineation of five prospective areas.

(2)

The RF model optimized with GWO was applied for quantitative integration of multi-source evidence layers. The machine learning results showed strong overall consistency with the remote sensing overlay predictions while significantly refining the potential mineralization zones through probability distribution constraints. Using a probability threshold of >0.8, two high-confidence exploration targets (Y1 and Y2) were ultimately selected.

(3)

The Y1 target area reflects mineralization dominated by magmatic hydrothermal processes, with granitic plutons serving as favorable geological units. It is primarily characterized by areal montmorillonitization and chloritization that are enriched along favorable fault structures. In contrast, the Y2 target area demonstrates mineralization controlled by structural hydrothermal activity concentrated along the Anqi Fault zone and its secondary faults, mainly showing linear chloritization and iron-staining anomalies enriched at intersections of NE-, NW-, and NE-trending structures. Field investigations have verified the rationality of delineating the Y1 and Y2 target areas, further confirming the scientific validity and reliability of the prediction methodology that integrates multi-source geological, remote sensing, and geochemical information through machine learning.

As for future work, further studies could focus on incorporating higher-quality remote sensing and geological datasets, explicitly integrating geomorphological indices as independent evidence layers, and testing additional machine learning algorithms to further improve prediction robustness. With improved spatial resolution and more robust endmember constraints, quantitative spectral unmixing approaches such as LSU or MTMF could also be explored to further investigate mineral abundance variations at finer spatial scales. Applying the proposed integrated framework to other gold districts in the Central Asian Orogenic Belt would also help to evaluate its transferability and practical value in regional mineral exploration.