Mafic-ultramafic intrusive complexes possess a considerable capacity for hosting Ni-Cu-platinum group element (PGE) sulfide deposits. However, the mapping of small outcrops over large areas by field surveys is time-consuming. In this study, WorldView-3 (WV-3) data with moderate spectral and very high spatial resolution were employed for mapping mafic-ultramafic units using spectral indices and the spatial-spectral transformer (SSTF) deep learning method in the Luotuoshan area of Beishan, Gansu Province, China. Based on representative reflectance signatures extracted from imagery of known locations, false-color composites of three-band ratios and a newly proposed short-wave infrared (SWIR) spectral index provided reasonable delineation of mafic-ultramafic rocks. The SSTF method facilitated mapping the occurrence of small mafic-ultramafic outcrops and defining much clearer boundaries, particularly for tiny units at meter scale. Moreover, the SSTF method is not sensitive to the occurrence of carbonate lenses that may affect the reflectance of outcrops. Field investigation and laboratory sample analyses confirmed the occurrence of mafic and ultramafic rocks with substantial metallic mineral potential in this area. Seven prospects were confirmed to be related to mafic-ultramafic intrusions during field validation, four of which contained metallic minerals such as chalcopyrite, pentlandite, pyrite, and chromite in the samples observed by scanning electron microscopy and energy dispersive spectrometry. This study proves that the spectral indices and SSTF deep learning method applied on WV-3 multispectral data are useful for discriminating small-sized mafic-ultramafic intrusive rocks (<100 m) for prospecting of local mineralization.

Magmatic nickel (Ni), copper (Cu), and platinum group element (PGE) sulfide deposits typically occur in mafic-ultramafic suites. These deposits contribute significantly to the global mining industry with approximately 56% of Ni, 5.5% of Cu, and 96% of PGEs (Naldrett, 1997; Mungall and Naldrett, 2008; Mudd and Jowitti, 2014; Barnes et al., 2017). The association between magmatic sulfide deposits and mafic or ultramafic rocks is well documented (Naldrett, 1997). The formation of magmatic Ni-Cu-PGE sulfide deposits is attributed to the segregation and concentration of liquid sulfide droplets from mafic or ultramafic magma and the partitioning of chalcophile elements into these from a silicate melt (Chai and Naldrett, 1992; Arndt et al., 2005; Lightfoot et al., 2012a). Therefore, mafic-ultramafic rocks are considered a key indicator for the possible presence of Ni-Cu-PGE sulfide deposits (Naldrett, 1997; Maier, 2005; Barnes et al., 2016; Graham et al., 2017).

The majority of ore-bearing mafic-ultramafic intrusions are characterized by areas that are less than a few square kilometers, with notable examples including the Norilsk-Talnakh giant Cu-Ni-PGE deposit (Russia) and Voisey’s Bay giant Ni-Cu-Co deposits (Canada), which are only 3 to 6 km2 in outcrop area (Amelin et al., 1999; Naldrett, 1999; Lightfoot et al., 2012b; Porter, 2016). Similarly, the outcrop areas of the Jinchuan and Kalatongke giant Ni-Cu deposits (China) are only 1.3 and 0.07 km2, respectively (Tang et al., 2006; Gao et al., 2012; Lu et al., 2019). The ore-bearing intrusions typically occur in clusters of veins, dikes, and sills, with lithology documented as peridotite, lherzolite, gabbro, and pyroxenite (Tang et al., 2006, 2012). It is worth noting that the small-sized ore-bearing mafic-ultramafic outcrops may be overlooked during field geologic surveys, as they may be much smaller than the mappable unit at a given scale during regional geologic mapping. Therefore, mapping the mafic-ultramafic rock outcrops of all sizes is essential for the exploration of Cu, Ni, and PGE mineralization (Barnes and Lightfoot, 2005; Liu et al., 2014; Rogge et al., 2014).

Multispectral and hyperspectral satellites provide feasible alternatives to traditional field surveys for lithologic mapping, providing significant advantages in terms of efficiency and cost-effectiveness (Rowan et al., 1977; Goetz et al., 1983; Kruse et al., 2003; Ninomiya et al., 2005; Nair and Mathew, 2012). Satellite images from the Thematic Mapper (TM), the Enhanced Thematic Mapper (ETM), and the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) have been widely used in lithological mapping and mineral delineation (Di Tommaso and Rubinstein, 2007; Pour and Hashim, 2012; Zadeh et al., 2014; Pour et al., 2018). Specifically, TM, ASTER, and simulated EnMAP data have been successfully used to discriminate ophiolite complexes and mafic-ultramafic rock units (Abrams et al., 1988; Rowan and Mars, 2003; Ninomiya et al., 2005; Rogge et al., 2014; Qasim et al., 2022). However, the low spatial resolution of these sensors (30 m) presents a limitation to the accurate mapping of small mafic-ultramafic rock outcrops.

WorldView-3 (WV-3) is the solitary commercially available imaging satellite that furnishes high spatial resolution multispectral images over both visible near-infrared (VNIR) and short-wave infrared (SWIR) spectral regions, which contain lithological and mineralogical information. WV-3 has eight VNIR bands with 1.24-m spatial resolution and eight SWIR bands with spatial resolution of 3.7 m (Mars, 2018). In addition, it also has a panchromatic band with 0.3-m resolution. This high spatial resolution is particularly valuable in the accurate identification of small rock units that might not be easily discernible with other satellite images. WV-3 has been used to accurately identify small rock units, ferric and ferrous minerals, hydroxyl group minerals, and carbonates (Mars, 2018; Ninomiya and Fu, 2019). Kruse and Perry (2013) effectively mapped various minerals including kaolinite, alunite, buddingtonite, muscovite, calcite, and hydrothermal silica at Cuprite, Nevada, using simulated WV-3 SWIR imagery. A subsequent study by Kruse et al. (2015) confirmed the mapping potential of WV-3 data using on-orbit WV-3 SWIR data at the same site. Mars (2018) successfully mapped different lithologic units at Mountain Pass, California, including muscovite-rich granitic gneiss, chlorite-epidote–rich migmatite, schist, gneiss, limestone, and dolostone using WV-3 data with band ratios, false-color composites, and logical operator algorithms. Sekandari et al. (2020) demonstrated that VNIR bands of WV-3 were useful in discriminating ferrous and ferric iron-bearing minerals in the Pb-Zn mineralization district of Kerman-Kashmar, Iran. Similarly, more recent studies have substantiated the potential of the VNIR and SWIR bands of WV-3 for lithologic and alteration mineral mapping (Sun et al., 2017; Pour et al., 2019; Son et al., 2021; Zhao et al., 2021). Nevertheless, no study has examined the use of WV-3 for mapping mafic-ultramafic rocks, which present a challenge due to their low reflectance and weak characteristic absorptions in the VNIR-SWIR region.

Previous studies have proposed several approaches to detect mafic-ultramafic rocks using hyperspectral data, such as band ratios, neural network classifiers, and linear spectral unmixing algorithms with different data sets (Rogge et al., 2010; Leverington and Moon, 2012; Rogge et al., 2014). The recent development of deep learning techniques has demonstrated significant potential in lithologic classification from multispectral remote sensing images. Owing to its robust modeling ability of combining spectral and spatial information, the deep learning method can deliver more accurate mapping results with continuous spatial patterns and well-defined edges. Accordingly, this study aims to investigate the transformer method and evaluate its performance in classifying rock types for mapping occurrence of small mafic-ultramafic outcrops. The main objectives are (1) to differentiate mafic-ultramafic rocks (e.g., gabbro, pyroxenite, and peridotite) from other bedrocks (e.g., carbonates, granite, and sandstone), which would assist in mineral exploration over a large area (100 km2) in Beishan, Northwest China, (2) to assess the ability of WV-3 data in mapping small-sized outcrops with low reflectance, and (3) to investigate the feasibility of the deep learning method in lithologic mapping with WV-3 high spatial resolution imagery.

Work area and geology

Beishan is located in northwest China and spans Xinjiang, Gansu, and Inner Mongolia (Fig. 1a). This semiarid area is suitable for the growth of sparse vegetation such as grasses and shrubs. The diverse geology, tectonics, and metallogenic environment make the Beishan area one of the largest mineral resource belts in China. More than 90 ore deposits (primarily gold, copper-nickel, and lead-zinc) have been discovered in this belt. More than 20 deposits are medium to large sized (Fig. 1b) (Yang et al., 2006). Mafic-ultramafic intrusions or complexes host Ni-Cu-Co deposits that are prevalent throughout the Beishan area (Yang et al., 2012). Over 70 such complexes have been mapped in the Beishan belt in the Gansu Province (Yang et al., 2006). A recent discovery is the Pobei superlarge Ni-Cu-Co deposit hosted by several cogenetic mafic-ultramafic intrusions in the Beishan belt (Yang et al., 2014; Ye et al., 2017). Some other well-known ore deposits in the belt include the Hongshishan chromium ore deposit and the Heishan Ni-Cu ore deposit (Yang et al., 2006). Therefore, the Beishan belt is recognized as an important mineral resource zone related to mafic-ultramafic intrusions. There exist some known mafic-ultramafic intrusions around the Luotuoshan area in southwestern Beishan, Gansu Province (Fig. 1b). However, limited field surveys have been conducted in this region to date. It is imperative to quickly map these mafic-ultramafic intrusions and assess their mineralization potential with the aid of high spatial resolution imagery.

The study area (Fig. 1c) predominantly comprises five Paleozoic units, namely the Permian Shuangbaotang Formation (P1s), the Carboniferous Shibanshan Formation (C2s), the Carboniferous Hongliuyuan Formation (C1h), the Devonian Dundunshan Group (D3dn), and the Ordovician Huaniushan Group (O1-2hnc). The Precambrian crystalline basement, which mainly consists of metamorphic rocks, is sporadically exposed in the study area. Unconformable overlying strata ranging from the Cambrian to Permian periods are commonly lower-grade metasedimentary rocks dominated by limestones, slates, phyllites, sandstones, and quartz schists (Ruan et al., 2021). Extensive tectonics and magmatic activity occurred during the Carboniferous and Permian periods (Xia et al., 2013), resulting in the emplacement of mafic-ultramafic complexes with similar rock types and early Permian ages (Su et al., 2012; Liu et al., 2017; Ruan et al., 2020). The complexes discovered so far are primarily distributed in the western part of the Beishan region and intrude the Proterozoic and Carboniferous strata (Jiang et al., 2006; Su et al., 2009, 2011). They are controlled by a complex fracture system with northeast, northwest, and nearly east-west trends. The orientation of the northeast and east-west fractures is nearly identical to tectonic lineations in the region, which are considered to be part of the regional faults.

Data and measurements

WorldView-3 multispectral data: Two WV-3 data sets (standard-level 2A product) with no cloud or snow and very little vegetation cover (<1%) were acquired for the study area on April 30, 2019. The data sets contain 16 bands ranging from VNIR to SWIR. The band positions are presented in Figure 2. The SWIR bands were provided with 7.5-m spatial resolution, whereas the VNIR bands were provided at 2.0-m spatial resolution. The SWIR data were spatially resampled to 2.0 m and coregistered to the VNIR data. The high spatial and moderate spectral resolution ensures the accurate extraction and identification of small outcrops at a scale of a few meters.

Samples and laboratory measurements: In the study area, a total of 37 hand samples were collected, including gabbro and pyroxenite. All samples had surface areas greater than 10 × 10 cm, with at least one weathered surface and one fresh surface for each sample.

The VNIR spectra of rock samples were acquired using the SR-3500 portable spectrometer from Spectral Evolution Inc., Lawrence, Massachusetts, USA. The spectrometer acquired high-resolution spectra (3 nm from 350 to 1,000 nm; 8 nm from 1,000 to 1,900 nm; and 6 nm from 1,900 to 2,500 nm) within the wavelength range of 350 to 2,500 nm. Spectral measurements were conducted on weathered surfaces and fresh surfaces separately, with 1 to 5 spot spectra collected per surface type. To ensure consistent illumination conditions, a contact probe (field of view of 25°) with an internal illumination source was used. Radiance spectra were calibrated and converted to reflectance using a white panel made of BaSO4. An average of 10 consecutive measurements was made for each saved spectrum.

In order to characterize the mineralogy of the rock samples, a total of 19 thin sections and six polished sections were prepared and investigated with a Nikon Labophot2 polarizing-light microscope. Additionally, the polished sections were sent for scanning electron microscopy (SEM) and energy dispersive spectrometry (EDS) analysis using an FEI Quanta 650 FEG SEM instrument at 20 kV to identify metallic minerals such as chalcopyrite, pentlandite, pyrite, and chromite.

Preprocessing of WV-3 images

The WV-3 radiance data were converted to surface reflectance using the Fast Line-of-Sight Atmospheric Analysis of Spectral Hypercubes (FLAASH) algorithm (Berk et al., 1998) with parameters of 617-km sensor altitude, 1,900-m ground elevation, subarctic summer atmospheric model, rural aerosol model, 40-km initial visibility, and no spectral polishing. The reflectance data were subsequently georeferenced to a local datum and coregistered with preexisting 1:10,000-scale digital topographic maps of the study area.

Spectral characteristics of rock units

Spectra from the Johns Hopkins University (JHU) spectral library were employed as preliminary reference (Fig. 2) to distinguish various types of mafic-ultramafic rocks from background. According to the local geologic settings, five mafic-ultramafic (diabase, basalt, gabbro, dunite, and picrite) and four nonmafic rock types (quartz monzonite, granite, granodiorite, and diorite) were selected for identification of potential spectral features. Mafic rocks, which contain iron oxide, pyroxene, and amphibole, exhibit absorption features from ferrous and ferric ions in the vicinity of 0.7 and 1.0 μm (Hunt, 1977; Gupta, 2003). Some mafic rocks, such as pyroxenite and gabbro, present Fe-OH and Mg-OH absorption features of biotite and hornblende at 2.32 and 2.38 μm (Rowan et al., 2005). Diorite has a broad absorption feature from 0.7 to 1.2 μm and another prominent feature near 2.33 μm. Rocks with hornblende and epidote show absorption features at 2.25, 2.32, and 2.39 μm. Additionally, the weathered surfaces of certain rock types (e.g., granite and quartz monzonite) exhibit diagnostic absorption features of Al-OH at 2.20 μm and Fe3+ features at 0.49 and 0.87 μm (Hunt and Ashley, 1979; Gaffey, 1986).

Even though WV-3 data only has 16 bands from VNIR to SWIR, it has been demonstrated to be effective in detecting minerals and mapping lithological units over areas at meter scale (Mars, 2018). WV-3 data have been utilized for the identification of typical alteration minerals including hydroxyl-bearing, iron oxide-bearing, and carbonate minerals (Kruse and Perry, 2013; Sun et al., 2017). The VNIR bands of WV-3 have proven to be useful for mapping lithological units by differentiating between ferrous and ferric iron-bearing minerals (Ye et al., 2017; Salehi and Tangestani, 2018; Sekandari et al., 2020). Specifically, prominent spectral absorption features in ferrous iron minerals such as epidote and chlorite can be identified by the obvious decrease in reflectance from VNIR band 3 to SWIR band 4. Ferric iron minerals show a decrease from VNIR band 2 to VNIR band 4 and from VNIR band 7 to SWIR band 1. The SWIR bands can be used to detect the absorption features of Al-OH (2.20 μm), Fe-OH and Mg-OH (2.30+ μm), and CO32 (2.16, 2.32–2.35 μm) found in clays, epidote, chlorite, and calcite/dolomite (Clark et al., 1993; Mars, 2018). This study aims to identify appropriate VNIR or SWIR bands and spectral indices for mapping mafic-ultramafic rocks in the work area.

Mapping mafic-ultramafic units

Spectral indices: Spectral indices that could enhance the spectral features of minerals (e.g., iron oxides and hydroxyl-bearing minerals) were used to highlight the spatial distribution of mafic and ultramafic rocks. The bands utilized for calculating spectral indices were based on previous studies (Rowan et al., 2005; Amer et al., 2010; Nair and Mathew, 2012), as well as on spectroscopic analysis of minerals available in the JHU spectral library.

Based on an existing geologic map (Fig. 1c), this study selected 11 dominant rock types from which representative spectra were extracted for image spectral analysis. The spectral signatures of the selected rock types displayed significant differences across the 16 bands of the WV-3 (Fig. 3). Based on the spectral features illustrated in Figure 2, three band ratios (SWIR bands 6/8, SWIR bands 5/3, and VNIR bands 3/2) were calculated and then merged into a false-color composite image to enhance the differences among rock types. The three band ratios were used to identify the three compositional indicators from local rock types: (1) Fe-OH and Mg-OH in mafic rocks such as pyroxenite and gabbro, (2) ferrous iron in diorite and peridotite, and (3) ferric iron in sandstone, schist, and granite.

To discriminate the mafic rocks (e.g., gabbro, diabase, and pyroxenite) from nonmafic rocks (e.g., diorite and granite), spectral index (SWIR band 4 + SWIR band 6)/(2 × SWIR band 5) was proposed. This index was expected to enhance the spectral difference between the mafic and nonmafic rocks, as the spectral slope from SWIR band 4 to SWIR band 6 is steeper for nonmafic rocks compared to the relatively flat slope observed in mafic rocks (Fig. 3).

Classification of rock units using deep learning: A spatial-spectral transformer (SSTF) was proposed for the final lithologic classification task (Fig. 4). The SSTF is a type of deep learning method called “transformer” that has demonstrated robust performance in various applications (Vaswani et al., 2017; Hong et al., 2022; Yu et al., 2022; Zhang et al., 2023). The transformer method originated from the natural language processing (NLP) and computer vision (CV) image analyses and has powerful modeling ability. The transformer network uses encoder and decoder architecture and relies solely on a self-attention mechanism that differs greatly from convolutional and recurrence neural networks. It treats the image as sequence data and builds encoder and decoder layers with the same architecture. The encoder transforms the input data to sequence representations with position information, and then the decoder converts the transformed representation to an output sequence. The encoder and decoder are connected by an attention mechanism that automatically adjusts the relationship between different locations and focuses on the areas that require attention. The essence of the attention mechanism is to adapt the weights for classifying pixels, through which the network can form a feature map for each class.

To represent the long-range dependence between input and output, an attention matrix was computed by a function of query (Q), keys (K), and values (V) in the form of vectors as below:

Attention(Q,K,V)=softmaxQKTdk V,

where softmax is a normalized function that transforms a vector into values in the range of 0 to 1, and dk represents the dimension of keys and is a scaling factor related to K. Furthermore, the transformer employs a multihead attention mechanism that allows parallel processing and speeds up the classification.

In this study, the WV-3 data were fed into the neural networks in two ways. One is a spatial transformation using 3-D Gabor filtering based on a 7 × 7 spatial window, and the other is per-pixel spectral embedding using a visual transformer (ViT) network modified by two modules of Groupwise Spectral Embedding (GSE) and Cross-Layer Adaptive Fusion (CAF) (which is suitable for multispectral data) (Fig. 4).

To make the network more adaptable to the high-resolution WV-3 data, we modified the transformer by adding a block mechanism that improves the extraction of spectral and spatial information (Fig. 4). The SSTF adopts two blocks of GSE and CAF proposed by Hong et al. (2022). While the ordinary transformer method often applies a bandwise classification to a spectral image, the SSTF uses the block of GSE that can learn pixelwise representation and address the spectral correlation between adjacent bands (Hong et al., 2022). Therefore, this strategy is more suitable for capturing spectral features such as spectral peaks and troughs of the WV-3 data. The block of CAF aims to enhance the connectivity across long-range network layers. Additionally, SSTF includes a feature extraction block of a three-dimensional Gabor filter (Shen and Jia, 2011) to detect both spectral and spatial variance of the data. In this Gabor block, varying frequencies (Rivard et al., 2008) and directions are defined. The Gabor filtering based on Gaussian kernel functions assists us in capturing more representative information (e.g., well-defined texture) from the image. The merit of Gabor features improves the classification performance that has been demonstrated in a series of studies (Hu et al., 2018; Huang et al., 2022). Given the task of this study, we used a framework for multiclass classification. The classifier was conducted using the reflectance data without any spectral transformation.

The number of pixels selected for each training class (rock type) is listed in Table 1. While the original geologic map featured 18 rock types (Fig. 1c), similar rock types were merged manually to address only compositional variations. Thus, a total of 11 classes were selected as regions of interest.

The classification accuracy of each rock type was calculated based on cross-validation, and a confusion matrix was constructed to describe the false alarms among different rock types. For the accuracy evaluation, pixels within each of the labeled regions of interest were randomly split into a training set (80%) and a testing set (20%).

Field validation

The accuracies of the two mapping methods were validated by a field campaign and laboratory analyses of samples. Spectral properties of these samples were assessed by measuring the reflectance of both fresh and weathered sample surfaces. A total of 37 rock samples, including gabbro, pyroxenite, peridotite, and olivine gabbro, were collected from 22 sites. Twenty-five (out of 37) samples containing sulfides were selected for further analysis based on rock types and visual assessment of hand samples. Thin sections or polished sections were prepared for petrographic analysis.

Detection of mafic-ultramafic outcrops from spectral indices

False-color composites of the three band ratios (red = SWIR bands 6/8 for Fe-OH, Mg-OH; green = SWIR bands 5/3 for Fe2+; blue = VNIR bands 3/2 for Fe3+) were constructed (Fig. 5b). Mafic-ultramafic units such as gabbro and pyroxenite are represented by red to dark-red colors owing to the presence of pyroxene, chlorite, and epidote. Diabase is indicated by light- to dark-green colors owing to the lack of a strong Fe2+ absorption that results in a relatively high SWIR band 5/3 value. Additionally, granite, which exhibits a strong Fe3+ absorption feature, is shown as blue owing to its high VNIR band 3/2 value. Compared to the true-color image (Fig. 5a), the false-color composite image of band ratios significantly enhanced rock types and highlighted the occurrence of mafic-ultramafic outcrops, particularly in the northeast part of the study area. Furthermore, small gabbro outcrops that were not visible on the natural color image and not mapped on the traditional geologic map were differentiated from large diabase sills in the false-color band ratio composite image (Fig. 5b).

Results of the SWIR spectral index (SWIR band 4 + SWIR band 6)/(2 × SWIR band 5) conveyed similar information, except the mafic-ultramafic rocks were represented as bright areas on the image (Fig. 5c), whereas the other rock types, including granite and diabase, demonstrated relatively low index values. This spectral index provides a convenient tool for quickly mapping the occurrence of mafic-ultramafic outcrops. Nonetheless, false positives occur when lenses of carbonate are present. These are mainly distributed in the southeast part of the study area (yellow arrows in Fig. 5c). In contrast, the false-color composite image (Fig. 5b) shows that the lenses are not mafic-ultramafic rocks.

Lithologic mapping using the SSTF classification method

Figure 6a shows the color-coded classification result of 11 rock types that was obtained using the SSTF deep learning method. Figure 6b shows only the distribution of mafic-ultramafic rocks (diabase and gabbro) overlaid on the true-color WV-3 image. The result shows more precise lithological boundary delineation and the detection of small outcrops at a scale of meters. Compared with the color composite of band ratios in Figure 5b, the deep learning classification result is able to map the mafic-ultramafic intrusions with clearly defined boundaries. Compared with the 1:50,000 geologic map in Figure 1c, the SSTF generated a very similar mapping of granite, sandstone, and diorite. For units that were labeled as gabbro, diabase, schist, siltstone, and limestone on the geologic map, the SSTF classification showed much more intraclass variability, thereby enabling the discrimination of compositional details. Large intrusions (D8 and D9 on Fig. 6), labeled as gabbro on the 1:50,000 geologic map, were adequately mapped in the classification result, providing additional support for the validity of the SSTF approach. More importantly, the SSTF classification revealed many small mafic-ultramafic intrusions that were not previously shown on the geologic map. For instance, large areas of diabase and schist (Fig. 1) located in the eastern part were mapped as distinct mafic units (D4 and D5 in Fig. 6a). Small intrusions of mafic-ultramafic rocks (D6 and D7 in Fig. 6) in the northeastern part were separated and accurately classified as gabbro, which represents a significant improvement over the prior geologic map wherein they were merged with adjacent units. This improved discrimination is important as it allows for the effective delineation of clustered, small outcrops. Furthermore, the SSTF mapped small outcrops of mafic rocks (D1, D2, and D3), all less than 100 m across, in the western part of the study area. These were misidentified as granite and schist in the 1:50,000 geologic map. Most importantly, the carbonate lenses (yellow arrows in Fig. 5) that affected the spectral index map had no influence on the SSTF classification results.

In general, results showed that SSTF outperformed the band ratio and spectral index by providing a more detailed and accurate map with clear lithologic boundaries. While all three methods were able to identify mafic-ultramafic rocks, the spectral index method had more limitations. First, it was not effective in defining clear boundaries for different rock types, particularly for small outcrops. Additionally, it incorrectly identified certain carbonate lenses as mafic-ultramafic in the southeast part of the study area (denoted by yellow arrows in Fig. 5c). Finally, the spectral index method lacked sensitivity in identifying all lithological units except mafic-ultramafic rocks, indicating its limited capability as a single band ratio technique.

The SSTF image and false-color composite image of band ratios both successfully delineated the locations of mafic-ultramafic intrusions with consistent results. Some lithological boundaries of more felsic rock units in the SSTF image were more distinct than those in the false-color composite image, indicating that the deep learning supervised classification method works better in defining boundaries between different rock types. The effectiveness of the SSTF in discriminating lithological units with subtle spectral differences can also be illustrated by its ability to distinguish granite porphyry from granite in the western part (inside the green line in Fig. 6a). The granitic rocks appear as similar green-blue colors on the false-color composite image of band ratios. The SSTF output image looks somewhat noisy when the size of outcrops is small. For example, the diorite outcrops (inside the white polygon in Fig. 6a) were identified as clusters of scattered pixels of diorite and diabase. The small clusters represent the widespread occurrence of diabase dikes in this area.

Cross-validation of the SSTF method with 20% training pixels produced an overall accuracy of 95.86% and a Kappa coefficient of 0.95 (Table 2). The accuracy for seven classes (including brick-red granite, dark-gray diorite, gabbro, sand and gravel, sandstone, siltstone and limestone, and flesh-red granite) was greater than 95%. The accuracy of diabase, schist, and granite porphyry ranged from 93.72 to 94.32%, whereas the accuracy of gray diorite was the lowest at 83.20%. The confusion between diabase and gray diorite was the highest because of similar mineralogy. Diabase and gray diorite show similar spectral features of Fe2+/Fe3+-bearing minerals (e.g., chlorite and epidote) at 0.60 to 1.21 μm (gray shadow in Fig. 3), whereas the absorption features of Al-OH, Fe-OH, and Mg-OH were not obvious at 2.16 to 2.33 μm (Fig. 3).

Field validation of mapping results

Field inspections were conducted to assess the accuracy of mapping results obtained by the false-color composite and SSTF methods. A total of seven small-sized mafic-ultramafic intrusions were examined (Fig. 7). None of them were mapped on traditional geologic maps. These intrusions, ranging in size from 1,000 to 70,000 m2, exhibit varying rock types including gabbro, pyroxenite, pyroxene peridotite, and olivine gabbro with strong weathering (Fig. 7). Four of the sites show significant mineralization with the obvious presence of chalcopyrite on fresh rock surfaces. Further laboratory analysis was conducted to determine the mineralogical, spectral, and alteration properties of these intrusions.

Three small mafic-ultramafic rocks (D1, D2, D3) were found in the southwest portion of the study area. D1 (80 m in diam) represented a newly identified mafic-ultramafic unit, which was confirmed by thin-section investigation to be strongly altered, serpentinized pyroxenite. Its surface is weathered to a dark-maroon color (Fig. 7a), whereas the fresh surface is black. D2 (60 m in diam) and D3 (30 m in diam) are both gabbroic units (Fig. 7b), with melanocratic minerals faded owing to weathering. It was observed that D1 and D3 were mapped as granite on the geologic map, whereas D2 was mapped as schist (Fig. 1c). All three intrusions were omitted by previous work done by traditional geologic surveys, most likely because of their small sizes.

Two mafic-ultramafic intrusions (D4 and D5) were discovered in the eastern part of the study area. The D4 intrusion was identified as olivine gabbro with 35% olivine, whereas D5 intrusion is gabbro with heterogeneous internal lithofacies and a gray weathered surface. As illustrated in Figure 8a, the fresh surface spectrum of D4 displays an absorption feature near 0.91 μm indicative of hypersthene (Clark et al., 1993; Yan et al., 2012), which was consistent with microscopic observation. In contrast, the weathered surface of D4 exhibits an absorption near 2.32 μm characteristic of serpentine. The serpentine formed along cracks and edges of olivine where the free iron formed magnetite. Absorption features near 2.35 μm indicative of epidote were observed in spectra of both fresh and weathered surfaces of D5. D4 and D5 intrusions were not shown in the geologic map likely because of their small sizes. The D4 intrusion was mapped as diabase covering a large area, and the D5 intrusion was mapped as schist (Fig. 1c).

Two newly discovered intrusions (D6 and D7) have been found in the northeast part of the study area. The D6 intrusion, which is less than 100 m in diameter (Fig. 7e), is identified as a pyroxene peridotite. The fresh surface of the D6 sample is dark green with widespread chalcopyrite. The D7 intrusion (about 55 m in diam) is a dark gabbro (Fig. 7f) with visible chalcopyrite. In thin section, it can be classified as olivine gabbro (Fig. 9b). As illustrated in Figure 9a, there is a narrow and shallow absorption feature near 2.23 μm due to Mg-OH vibrational overtones (Calvin and King, 1997; Ehlmann et al., 2010). This is indicative of chlorite. Also, there is a deep absorption near 2.35 μm characteristic of epidote and zoisite. Both of these two intrusions were mapped as part of a large area of diabase. Neither of them is found as separate intrusion on the 1:50,000 geologic map (Fig. 1c).

Effectiveness of WV-3 imagery for lithologic identification

Traditional geologic surveys are limited by their high cost, and they are not always effective in mapping small outcrops, as illustrated by the omission of some outcrops (D1-D7) in the 1:50,000-scale geologic map. Therefore, it is necessary to use high-resolution remote sensing imagery to guide further exploration. The high spatial resolution of VNIR and SWIR WV-3 data allows for the mapping of relatively small rock outcrops. Thermal infrared data (with a wavelength range of 7.55–12.5 μm, for example) is not effective in detecting small outcrops owing to the required lower spatial resolution.

Relationships between laboratory measured spectra and image spectra

Knowledge of rock mineralogy and spectroscopy is crucial in choosing the appropriate method for lithological mapping using satellite data (Karimzadeh and Tangestani, 2021). It is useful to collect spectra from field samples that can be compared to spectra from spaceborne sensors. Some rock types have low reflectance (<20%) (Figs. 8a, 9a) that can make it more difficult to match spectra. Although matching of the image spectra to laboratory spectra could be affected by atmospheric effects, instrument effects, spatial resolution, mixed pixels, and weathering (Yu et al., 2012), the shape of lab spectra for weathered gabbro for example is comparable to WV-3 gabbro spectra (Fig. 3). This demonstrates that the FLAASH algorithm can accurately convert WV-3 radiance data into reflectance (Berk et al., 1998; Mars, 2018).

Influencing factors of deep learning and traditional spectral indices

This study shows the efficiency of the spectral index approach to map large mafic-ultramafic intrusions, albeit with some limitations. Specifically, the presence of carbonate units was found to interfere with the spectral index results. The WV-3 spectra of the carbonate lenses display high reflectance at SWIR band 4 and low reflectance at SWIR band 8, which affected the spectral index (SWIR band 4 + SWIR band 6)/(2 × SWIR band 5). Further work is needed to minimize the impact of carbonates on the spectral index results.

The present study demonstrates the potential of using a false-color composite image of band ratios to detect mafic-ultramafic rocks and to map most of the lithologic boundaries. However, the SSTF approach is a more advanced, supervised method that relies on the accuracy of training samples for mapping. It is noteworthy that the carbonate lenses were not defined as mafic-ultramafic rocks either in the false-color composite image or in the SSTF results.

The high accuracy of the SSTF classification result shows the reliability of the deep learning method for lithological mapping. Compared to the other two methods employed in the study, the SSTF classification approach allowed more detailed identification of lithological units with greater intraclass variability. However, the classification accuracy could be affected by rock types of similar mineralogy. For example, the gray diorite in the northwest (inside the white polygon in Fig. 6a) was mostly confused with diabase owing to the similar spectral features of Fe2+/Fe3+ in VNIR wavelengths and weak spectral features of Fe-OH and Mg-OH. It was determined that the accuracy of the classification map generated by the deep learning algorithm is highly dependent on the quality of the samples used for training. In this study, training samples were collected based on the 1:50,000 geologic map and previous geologic knowledge. Therefore, there could be some level of uncertainty about the results.

Mineralization potential

This study has successfully identified several mafic-ultramafic occurrences that are absent from the geologic map. This has led to the discovery of four mafic-ultramafic intrusions with Cu-Ni-PGE mineralization. These are chalcopyrite-bearing intrusions of various rock types including gabbro, pyroxenite, and pyroxene peridotite. The presence of metallic minerals was confirmed by the analysis of SEM and EDS data.

Chalcopyrite, a copper iron sulfide mineral, was found in rock samples from all the D4, D5, D6, and D7 intrusions (Figs. 79). The metallic minerals in D4 are mainly pyrrhotite, chalcopyrite, and a small quantity of chromite, which developed along chalcopyrite fractures (Fig. 8b). D5 shows more evident mineralization, including pentlandite, pyrrhotite, pyrite, and chalcopyrite (Fig. 8c). Pyrite exhibits an idiomorphic crystalline-granular texture, whereas chalcopyrite developed along the edges of pyrite (Fig. 8d). Pentlandite, which has a xenomorphic granular texture, generally occurs in pyrite (Fig. 8c). The metallic minerals found in D6 are mainly pentlandite, pyrrhotite, pyrite, chalcopyrite, magnetite, and some chromite (Fig. 9c). Pentlandite and chalcopyrite exhibit an idiomorphic to subidiomorphic granular texture, while magnetite and pyrrhotite crystallized later. Spinel in D6 samples developed an obvious inverse zonal texture in the backscattered images (Fig. 9f). According to the EDS, the brighter region in the core is ferrochrome spinel, namely chromite, which is rich in Fe-Cr with a small amount of titanium (Fig. 9g). The darker zone on the edge is rich in Mg-Al, which is magnesium-aluminum spinel with less chromium and iron (Fig. 9h). In contrast, mineralization in D7 samples is less intense, with pyrite, chalcopyrite, and magnetite being the main metallic minerals present (Fig. 9d).

This study shows that the spectral indices and deep learning methods are useful for prospecting for Ni-Cu-PGE–bearing mafic-ultramafic intrusions with WV-3 data in this area of Beishan, Northwest China. Mafic-ultramafic intrusions that are associated with Cu-Ni mineralization in the area are typically smaller than 1 km2 (Tang et al., 2012). It is important to note that not all mafic-ultramafic intrusions in this region contain economic Ni-Cu-PGE mineralization (Guo et al., 2017).

This study demonstrates that relatively small mafic-ultramafic rock outcrops can be detected with WV-3 VNIR and SWIR multispectral data using spectral indices and the spatial-spectral transformer (SSTF) deep learning method. WV-3 data has sufficient spectral resolution for discriminating mafic-ultramafic rocks from other rock types in the study area, and its high spatial resolution helps to map smaller intrusions.

The composite of the three band ratios (SWIR band 6/band 8 in red for Fe-OH and Mg-OH; SWIR band 5/band 3 in green for Fe2+; VNIR band 3/band 2 in blue for Fe3+) and the newly proposed SWIR spectral index (SWIR band 4 + SWIR band 6)/(2 × SWIR band 5) provide simple but effective methods for detecting mafic-ultramafic rocks. However, the deep learning method proves to be more powerful and enables a more detailed discrimination of different rock types than the spectral indices-derived images. Most importantly, the SSTF classification method was not affected by the carbonate lenses that impacted the spectral index map. The contacts of rock types were also more clearly defined by the SSTF classification. This study shows that spectral indices and a deep learning method provide practical methods for detecting mafic-ultramafic outcrops. However, the attention-based SSTF method is more effective and results in higher classification accuracy.

This work was supported by the Key Laboratory of Strategic Mineral Resources of the Upper Yellow River, Ministry of Natural Resources (grant YSMRKF202203), the Natural Science Basic Research Program of Shaanxi Province (grant 2023-JC-ZD-18), the National Natural Science Foundation of China (grant 62001303), the Guangdong Basic and Applied Basic Research Foundation (grant 2023A1515012053), and the National Key Research and Development Program of China (2022YFB3903702).

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Chuntao Yin is a current geology Ph.D. candidate at the School of Earth Science and Resources, Chang’an University, China, with expectations to graduate in June 2024. She specializes in economic geology with a special interest in the application of remote sensing to the study of lithologic interpretation and mineral resource exploration. Her research interest is in the application of multispectral and hyperspectral imagery and spectral information to the extraction of mineralization information and inversion of quantitative elements, with research projects involving a number of image processing methods including false-color composite images, band ratios, logical operator algorithms, and deep learning techniques.

Lei Liu received B.S and M.S. degrees from Chang’an University and a Ph.D. degree from the Chinese Academy of Sciences in 2004, 2007, and 2011, respectively. Lei Liu is currently a professor in the School of Earth Science and Resources, Chang’an University, China. He specializes in lithologic and mineral mapping and mineral resource prospecting using multispectral and hyperspectral data.

Gold Open Access: This paper is published under the terms of the CC-BY-NC license.