Ore Geology Reviews 161 (2023) 105651

Contents lists available at ScienceDirect

Ore Geology Reviews

journal homepage: www.elsevier.com/locate/oregeorev

Mineral prospectivity mapping integrated with geological map knowledge

graph and geochemical data: A case study of gold deposits at Raofeng area,
Shaanxi Province
Qun Yan a, Linfu Xue a, *, Yongsheng Li b, Rui Wang a, Bo Wu a, Ke Ding a, Jianbang Wang a
a
College of Earth Sciences, Jilin University, Changchun 130061, China
b
Development and Research Center of China Geological Survey, Ministry of Natural Resources, Beijing 100037, China

A R T I C L E I N F O A B S T R A C T

Keywords: The knowledge graph in geological applications is highly regarded, and its usefulness requires the support of
Geological map knowledge graph geological survey data. Therefore, this paper proposes a mineral prospectivity mapping (MPM) method that
Knowledge embedding integrates geological map knowledge graphs with geochemical data. The geological map knowledge graph is
Spatial alignment of knowledge and space
used in the vectorized form as knowledge features in the training of the mineral prediction model. It is also used
Mineral Prospectivity Mapping of gold deposits
for knowledge-based ore-forming interpretation of prospective areas, achieving bidirectional application of
knowledge in prediction and interpretation of prediction results. Unlike most knowledge graphs, the geological
map knowledge graph represents spatial correlations and geological features in the form of “entity-relation-
entity” triplets, with geological map units as entities and assigned geological attributes. Based on this, an MPM
process driven by knowledge graph and data is proposed: (1) Convert the geological map knowledge graph into
low-dimensional vectors using node2VEC. (2) Align geological knowledge with geochemical data spatially
through knowledge embedding gridization, achieving a forward fusion of knowledge and data. (3) Collect the
integrated data using window sampling as the training dataset for DenseNet to train the mineral prospection
model. A case study was conducted to demonstrate the advantages of the method for producing a potential map
linked to gold mineralization in the Rao Feng area, Shaanxi Province, China. The results show that the pro­
spective areas with knowledge graph embedding and geochemical data are better and delineate six unminer­
alized prospective areas. Ore-forming interpretations based on knowledge graph information were conducted on
the prospective areas, demonstrating strong consistency between the ore-forming geological features of the
prediction areas and known mineral deposits.

1. Introduction prospective areas (Brown et al., 2000; Yang et al., 2022;). The deep
learning model represented by the convolutional neural network (CNN)
In recent years, data-driven quantitative geosciences research has and its variant model has been widely used in MPM (Li et al., 2023b), for
made significant progress (Chen and Wu, 2018; Chen et al., 2021a; Yin example, the Siamese Network (Ding et al., 2022), the Deconvolution
and Li, 2022; Anand et al., 2016; Zhang et al., 2023a; Xu and Zuo, 2023), Network (Wang et al., 2021a), the Generative Adversarial Network (Li
there are many models for MPM, which are mainly data-driven (Ghe­ et al., 2022), and CNN combined a direct sampling algorithm (Yang
zelbash et al., 2022; Zuo, 2020; Zuo et al., 2022). Deep learning as a et al., 2022).
powerful tool for data analysis has been widely applied in the field of However, the data-driven geological applications are mostly focused
geochemistry analysis (Chen and Sui, 2022; Fan et al., 2019; Li et al., on image recognition (Polat et al., 2021; Izadi et al., 2017; Xu and Zuo,
2023a; Zhao et al., 2021; Mueller and Grunsky, 2016; Zhang et al., 2023; Shi et al., 2023), and geophysical and geochemical anomalies
2023b) and MPM (Chen and Sui, 2022; Li et al., 2023b; Yang et al., (Wang et al., 2021b), whereas integration with geological knowledge is
2021; Yin and Li, 2022; Zuo and Xu, 2022). MPM is a multi-process, lacking. Knowledge-driven geological research development has
multi-principle, and computer-assisted approach used to delineate received much attention (Qiu et al., 2018; Qiu et al., 2020; Wang et al.,

* Corresponding author.
E-mail address: [email protected] (L. Xue).

https://doi.org/10.1016/j.oregeorev.2023.105651
Received 5 June 2023; Received in revised form 27 August 2023; Accepted 31 August 2023
Available online 9 September 2023
0169-1368/© 2023 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
Q. Yan et al. Ore Geology Reviews 161 (2023) 105651

Fig. 1. Schematic map of the research area (①-⑪ represent brittle and ductile faults; 1 Jinpengou gold deposit; 2 Zhanfang gold deposit; 3 Zhoujiawan gold deposit;
4 Yangpingwan gold deposit; 5 Lengjinggou gold deposit; 6 Xiangzizhai gold deposit).

2021b; Enkhsaikhan et al., 2021). Knowledge graph (Paulheim and knowledge graphs is mostly used for organizing and discovering text
Cimiano, 2016; Ji et al., 2022), as an important form of knowledge knowledge (Qiu et al., 2020; Wang et al., 2018a; Zhu et al., 2017; Chen
representation, can express the geological evolution by establishing the et al., 2021b). This includes constructing ore deposit knowledge graphs
geological entities and their relationships. This provides an effective (Qun et al., 2021; Zhou et al., 2021a), extracting geological knowledge
approach to organize geological data by graph structures and stan­ graph entities and their relationships (Qiu et al., 2019; Wang et al.,
dardized semantic links. Geological knowledge represented by 2018b), and analyzing metallogenic mechanism with knowledge graphs

2
Q. Yan et al. Ore Geology Reviews 161 (2023) 105651

Fig. 2. Gridding of geochemical data.

3
Q. Yan et al. Ore Geology Reviews 161 (2023) 105651

(Holden et al., 2019; Enkhsaikhan et al., 2021; Medjadba et al., 2020). ore host, whose distribution is controlled by the NW-trending brittle and
Text-based geological knowledge graph analysis lacks spatial attri­ ductile shear zones. The strata are intruded by many small intermediate-
bute expressions (Zhang et al., 2023c; Zhou et al., 2021b). Geological felsic dykes (Li et al., 2009).
maps use three types of elements (spot, line, and polygon). Combined Currently, a total of six gold deposits/prospects are discovered in the
with textual attributes such as legends, they describe the spatial corre­ study area, among which the Yangpingwan is the largest. The ore-
lations and mineralization characteristics of regional geological struc­ bearing Meiziya Formation is dominated by the carbonaceous two-
tures, volcanic activities, and mineral occurrences. Converting this into mica(-quartz) schist in the thrust nappe, with widespread develop­
a more comprehensible knowledge graph not only breaks the limitations ment of cleavages and fractures (Li et al., 2009). Fine biotite-
of conventional geological maps that use geometric shapes to represent orthoclase/-albite veins intruded along the schistose zone, and have
geological phenomena, but also equips geological knowledge graphs close spatial relation with the gold mineralization (Liu et al., 2016).
with the capability to express spatial properties. Previous studies have shown that the mineralization occurred in the
In geological analysis, knowledge-driven geological analyses only Late Indosinian (Triassic)-Early Yanshanian (Jurassic-Cretaceous)
aims to summarize regularity. Hence, the results lack practicality since period, and the metallogeny is interpreted as sedimentary-modified type
there is little support from geophysical and geochemical data (Zuo et al., (Li et al., 2009; Zheng and Yan, 2016). From the Paleozoic to early
2023). Data-driven geological analysis aims to discover regularity, and Mesozoic, the region may have undergone subsidence and sedimenta­
its results are poor in interpretability since the knowledge is not inte­ tion, forming the ore-bearing sequence. The multiphase tectonic activ­
grated (Shi et al., 2023). Pure Deep learning algorithms can make full ities during the Indosinian-Yanshanian orogeny may have resulted in the
use of available data and have consistency with the input data, but gold enrichment and mineralization. During the Yanshanian orogeny,
predictions may be geologically inconsistent or implausible owing to intense regional tectonic movements and thrusting shearing activities
extrapolation or observational biases (Zuo et al., 2023; Reichstein et al., may have controlled and formed gold deposits (Li et al., 2009; Liu et al.,
2019). The hybrid method of knowledge-driven and data-driven is 2016). Therefore, the gold ore formation is mainly controlled by ductile
crucial to the MPM (Godefroy et al., 2019; Zuo et al., 2023). For and brittle shear. The ore-forming materials may have mainly sourced
example, word embeddings is applied to classify and map lithological from the strata, and the hydrothermal alteration zone is controlled by
descriptions in a 3-D space (Fuentes et al., 2020). Satellite data, aero­ the NW-trending brittle and ductile shear zone.
magnetic and airborne radiometric data were integrated using fuzzy
logic and index overlay methods in a combination of knowledge and 2.2. Geochemical data collection and pre-processing
data-driven way (Riahi et al., 2021).
In this study, we focused on the prospection method to integrate According to the knowledge graph of the geological map of the study
geological map knowledge graphs with geochemical data. The area (Fig. 4b), the main chemically anomalous marker elements of the
geochemical are closely related to the formation of ore deposits (Li et al., gold mines in the area are Au, As, Mo, Cd, Cu, and Zn. Therefore, these
2021), their upper and lower limits play an important role in evaluating six elements data were forward integrated with the knowledge graph of
the scope of the metallogenic province (Wang and Hua, 2020). The the geological map.
spatial characteristics and semantic correlations of faults, strata, in­ As the locations of the geochemical samples are unevenly distrib­
trusions, and alteration zones were expressed in the form of geological uted, the Kriging interpolation method was used to grid the data (Li
map knowledge graphs. Node2vec (Grover and Leskovec, 2016) was et al., 2021; Ding et al., 2022). The spatially dispersed values were
used to embed the knowledge graph nodes. Metallogenic knowledge, converted into regularly distributed values to minimize the local noise,
such as ore deposits, strata, faults, and intrusions were transformed into and to better reflect the spatial patterns of element anomaly variables in
low-dimensional vectors, which correlate spatially with geochemical a unified spatial structure. To ensure the response of the unit grid to the
data to achieve the forward integration of metallogenic knowledge with entire geochemical anomalies, the grid was set as to 100 m (100 m ×
geochemical data. The well integrated-knowledge and geochemical data 100 m). Six types of geochemical interpolation grid data were obtained,
were used as the training data for Dense Convolutional Network (Den­ with each type having 384 grids (x-direction) and 477 grids (y-direc­
seNet) (Huang et al., 2017) to train the prospection model. tion), totaling of 183,168 grids.
The geochemical field of all elements is distributed in a NW-trend,
2. Local geology and data parallel to the local fault orientation. The anomaly of all elements rep­
resents the bead-like distribution, and becomes more prominent toward
2.1. Geological and metallogenic characteristics the southeast (Fig. 2). The small but high-amplitude Au anomaly shows
independent coherence and has five concentration centers. It is poorly
The study area is structurally located in the southern Qinling tectonic integrated with other elements. Mo, Cd, Cu, and Zn elements are well
belt (middle of the Qinling orogen), belonging to the Shiquan-Xunyang integrated, with large, high-amplitude, and distinct combination cen­
gold ore belt. Local strata have undergone multi-phase deformation and ters. Their morphology is consistent with the distribution range. The Au
intense folding, forming widespread schistosity (Zhang et al., 2013). The and As anomalies are prominent near the No. 5 and No. 6 gold deposit.
exposed strata are the Ordovician-Silurian Meiziya Formation meta- The Cu and Zn anomalies are large, and typical gold deposits in the area
clastic rocks, dominated by carbonaceous two-mica quartz schist, sili­ are mostly distributed within these high Cu and Zn anomalies.
ceous slate, and thick siliceous beds (Li, 1997).
Folding structures are widespread locally, and shear structures are 3. Geological map knowledge graph construction
dominated by ductile ones (with some brittle ones). The large-scale
ductile shear extends across the entire area in a NW- to near E-W 3.1. Entity construction of geological maps
trend. The ductile shear zone is a gold ore-controlling structure (Yang
et al., 2016). The gold orebodies and occurrences are mostly concen­ As nodes of the knowledge graph, entities represent the specific
trated in the brittle and ductile shear zones and the nearby schistose geological information in geological maps, and can be classified into
zones (Wang et al., 2018b). The superposition of multistage structures different types using labels. Polygons in geological maps indicate the
has locally formed wide synclines and anticlines (Liu et al., 2016; Xu stratigraphic unit, intrusive body (incl. dike), alteration zone, and
et al. 2018). geochemical anomaly. Lines in geological maps indicate faults. Spots in
Magmatic rocks are mainly outcropped in the southwestern part of geological maps indicate mineral occurrences. Geological map entities
the study area, and has no obvious gold metallogenic link (Fig. 1). The are divided into seven types of labels, including stratigraphic unit, fault,
Upper Ordovician-lower Middle Silurian Meiziya Formation is the main intrusion (incl. dike), alteration zone, geochemical anomaly, minerals,

4
Q. Yan et al. Ore Geology Reviews 161 (2023) 105651

Fig. 3. Entity relationship structure diagram of geological map knowledge graph.

and ore deposit (Fig. 3). Jinpengou gold deposit, are connected to the stratigraphic unit of the
To better distinguish the geologic unit in geological maps, non- Silurian Meiziya Formation through their relationship with the “ore-
repeating numerical labels were added to the entities of stratigraphic bearing strata”. The ore-bearing strata at Yangpingwan (ID = 4) are
unit and fault (FID: fault ID number; ID: stratigraphic ID number). Ac­ dominated by “carbonaceous sericite-quartz schist, garnet-bearing two-
cording to the geological information represented by the data polygon, mica quartz schist, metasandstone”. The ore-bearing strata at Xiang­
line, and spot in the map, “lithology” and “chronological stratigraphic zizhai (ID = 137) is “garnet-bearing quartz two-mica schist, garnet-
unit” attributes were assigned to the entity of stratigraphic unit and bearing two-mica quartz schist, and two-mica quartz schist”. The
intrusion (incl. dike). The fault “length, width, orientation ” in the “north alteration zone 1” entity and the Yangpingwan gold deposit are
geological map were mapped as the three attributes of the fault entity. connected with the “Minette veinlets” and “biotitealbite veins” in the
The “alteration type” attribute was added to the alteration zone entity, relationship of “occurrence”. The alteration zone is characterized by
whilst the “geochemical anomaly type” attribute was added to the “silicification”, “pyritization” and “biotitization” (Fig. 4a). Except for
geochemical anomaly entity. A total of 480 geological map entities were the Jinpan gold deposit, all the other three deposit entities occur in the
obtained from the local geological map, including 304 stratigraphic No. 5 brittle-ductile fault zone (FID = 5), with NW-trike and dip angle of
unit, 66 intrusion (incl. dike), 58 fault, 6 gold deposit, 11 geochemical 22◦ to the northeast (Fig. 4a). The Jinpengou gold deposit occurs within
anomaly, 5 alteration zone, and 30 mineral entities. the No. 22 NW-trending fault (Fig. 4b). The relative angle between the
No. 22 (FID = 22) and No. 14 (FID = 14) fault is zero (R angle: relative
azimuth between faults), indicating that the two faults are parallel and
3.2. Relationship between different entities
1.1 km apart. In terms of geochemical anomalies, the four deposits have
Au-Mo-Cd-Zn (Jinpengou), Au-As-Cu (Yangpingwan), Au-As-Mo-Cu
The relationship between regional strata, faults, and gold deposits is
(Xiangzizhai), and Au-Mo-Cd (Zhanfang) anomalies (Fig. 4b).
expressed as a triple form of “entity 1-relationship-entity 2”. A total of
817 relationships were connected based on 480 geological map entities.
The distance relationship between fault entities within 5 km was 4. Method
constructed based on the distance on the map (Fig. 3). Two attributes of
“relative azimuth (R Angle) and shortest distance (dis)” were added to The mineral prospection method integrated geological map knowl­
the relationship. The relative azimuth indicates whether the two faults edge graph and geochemical data proposed here is both knowledge and
intersect or separate. The distance indicates the spatial correlation be­ data-driven: The knowledge comprises the embedded knowledge from
tween the two adjacent faults. A “controlled” or “occurring” relationship the geological map knowledge graph, whilst the data comprise the
was constructed between the fractured entity and the deposit entity. A gridding data of Au, As, Mo, Cd, Cu, and Zn, which are sediment mea­
“occurring” or “ore-bearing strata” relationship was constructed be­ surements in aquatic systems. The geological map knowledge graph was
tween stratigraphic units and ore deposit entities, which reflected the embedded with metallogenic knowledge based on node2vec. The
location and ore-bearing strata of the ore deposit, respectively. A “cut­ knowledge was transformed into low dimensional vectors, and spatially
ting through” relationship was constructed between fault, stratigraphic aligned with geochemical gridding data to perform the forward inte­
unit, and intrusion or dike entities, reflecting the sequence of tectonic gration of knowledge and geochemical data (Fig. 5a). Subsequently, the
evolution. A “is” or “occurring” relationship was constructed between well-integrated data were sampled by window sliding (Fig. 5b). Finally,
the alteration zone and geochemical anomaly marker entities to connect the prospection model was trained and predicted by training a
the ore deposit, mineral, and fault entities and reflect the alteration and DenseNet.
geochemical anomaly characteristics of the connected entity (Fig. 3).
The knowledge graph can express the similarity and differences of 4.1. Embedding of geological map knowledge graph
metallogenic information in an intuitive form (Fig. 4). Four entities of
gold deposit, including the Yangpingwan, Xiangzizhai, Lingjinggou, and The method of node2vec (Grover and Leskovec, 2016) based on node

5
Q. Yan et al. Ore Geology Reviews 161 (2023) 105651

Fig. 4. Metallogenic characteristics of knowledge graph.

6
Q. Yan et al. Ore Geology Reviews 161 (2023) 105651

Fig. 5. Knowledge and geochemical data prospection and prediction framework based on DenseNet（In the figure, GU stands for geologic unit, F stands for fault.).

Fig. 6. Schematic diagram of knowledge and data alignment.

labels was adopted here. The vertex sequence was extracted through a (Formula 1). The strategies were sampled through the smooth transition
biased second-order random walk. To explore the global features con­ field, and the adjacent points were sampled through probability, which
tained in the knowledge graph network structure, the destination vertex reflected the sampling characteristics of depth and breadth-first. This
was mapped into low-dimensional vectors based on the co-occurrence method has resolved the topological separation of homogeneity and
relationship of adjacent nodes. structural equivalence (Ribeiro et al., 2017) in knowledge graph net­
In this method, the subgraph data labeled as ore deposit (Ore) and works, and best preserved the correlation of adjacent points. Skip-gram
stratigraphic unit (GU), ore deposit (Ore) and fault (F), stratigraphic unit (Laurens and Hinton 2008) was used to learn vertex embedding. The
(GU) and fault (F) were first loaded (Fig. 5a). Then, the walk sequence walk paths were trained, and each graph node was converted into a
was generated biased second-order random walk (Perozzi et al.,2014) vector.

7
Q. Yan et al. Ore Geology Reviews 161 (2023) 105651

Fig. 7. Visualization of geological map knowledge graph embedding features.

8
Q. Yan et al. Ore Geology Reviews 161 (2023) 105651

Fig. 8. Gold mineral potential map.

Table 1 Table 2
Impact of window size and positive/negative ratio on prediction results. Impact of model parameters on prediction results.
Window Practical range Sample Positive/ Test accuracy Batch Learning Attenuation Test accuracy
size size negative （%) size rate rate （%)

8*8 800 m*800 m 177,190 1:3 85.72 DenseNet (k = 32 0.001 0.03 81.39
8*8 800 m*800 m 177,190 1:4 72.43 12)
8*8 800 m*800 m 177,190 1:5 70.61 DenseNet (k = 32 0.001 0.03 88.98
16*16 1600 m*1600 170,478 1:3 88.54 24)
m DenseNet (k = 32 0.001 0.03 88.38
16*16 1600 m*1600 170,478 1:4 90.72 32)
m DenseNet (k = 32 0.001 0.07 86.81
16*16 1600 m*1600 170,478 1:5 92.12 48)
m DenseNet (k = 64 0.001 0.01 90.03
24*24 2400 m*2400 163,894 1:3 93.98 12)
m DenseNet (k = 64 0.001 0.01 95.39
24*24 2400 m*2400 163,894 1:4 91.10 24)
m DenseNet (k ¼ 64 0.001 0.01 97.78
24*24 2400 m*2400 163,894 1:5 92.39 32)
m DenseNet (k = 64 0.001 0.03 92.52
48)

9
Q. Yan et al. Ore Geology Reviews 161 (2023) 105651

Fig. 9. Loss rate and accuracy of prediction model training.

Fig. 10. Comparison of the prospecting potential area obtained from different knowledge embedding.

10
Q. Yan et al. Ore Geology Reviews 161 (2023) 105651

Fig. 11. Maps showing (a-b) MPM of CNN using only geochemical data (c) MPM of DenseNet using only geochemical data (d-e) MPM of CNN using integrated data
(f) MPM of DenseNet using integrated data.

{
Πvx /Z if(v, x) ∈ E Consequently, the knowledge embedding gridding was completed in the
P(ci = x|ci− = v) = (1)
1
0otherwise geological graphic units.

In formula (1), v is the current node, P is the probability of accessing I = ae− bx

(3)
the next node ×, V represents the current node, x represents the next
In formula (3), a represents the influenced amplitude value, a = 1.0,
node to be accessed, Πvx is the non-normalized transition probability
and b represents the attenuation coefficient, the larger the value, the
between nodes v and ×, and Z is the normalization constant.
faster the attenuation rate. As the distance between geological elements
⎧
⎨ 1/p ifdtx = 0 increases, the impact on the metallogenic geology within the grid
αpq (t, x) = 1if dtx = 1 (2) exponentially decreases with distance. X represents the distance be­
⎩
1/qif dtx = 2 tween geological elements. I represent the correlation between geolog­
ical graphics and grids.
In formula (2), Πvx = αpq (t, x) • wvx , wvx is the edge weight between
Normally, only the spatial distance relationship between geological
node v and ×, dtx is the distance between the previous node and the node
graphics in the grid units is considered when calculating the relative
to be accessed, dtx equal to 0 indicates returning the previously accessed
distance, whereas the geological graphic units were ignored. The
node from the current node, T is previously accessed node. Two
embedding of geological map knowledge has assigned knowledge
hyperparameters, p, and q, control the probability of walking to
feature value to the geological graphics (Fig. 6). Each geologic unit has
different nodes. If q > 1, it tends to visit the nodes closed to t, and if p <
independent ore-related knowledge embedding, and hence the geolog­
1, it tends to visit the nodes far from t. Due to the metallogenic char­
ical map could represent the metallogenic characteristics (both spatial
acteristics in the geological map being within a certain spatial range,
and knowledge dimensions), and align with geochemical gridding with a
two connected entities express the spatial information of two neigh­
unified grid size.
boring geological units on a geological map. Therefore, q = 3 is set to
access and embed the nodes adjacent to the deposit, which avoids the
effects of remote entities. 4.3. DenseNet-based model training of gold MPM

Different types of spatial structures represent different geological

4.2. Spatial alignment of knowledge and geochemical gridding data environments. As anomalous spatial structures have a certain size, the
data can be sampled with the sliding window method (Li et al., 2021).
Geological map of the study area was divided into grids of 100 m × Here, the window size was set as 24 × 24, and the sliding step of the
100 m. By calculating the shortest distance between the grid center and window was set to 1. The ore-bearing samples were set as positive
geologic unit (Formula (3)), the graphic units of the geological map, samples and the non-ore-bearing samples were set as negative samples.
such as polygon and line were decomposed (Ding et al., 2022). This step The ratio of positive to negative samples was set as 1:3, and randomly
gives the geologic map the same grid cells as the chemically gridding. selected the non-ore-bearing samples as negative samples. The features
Subsequently, the embedded knowledge is arranged in the order of the of the positive samples were extracted by training DenseNet, and the
geologic unit graphic IDs in the map, which correlates the embedded grid units outside the known mineral occurrence were predicted
knowledge to the geological graphic units. Sequential embedding vec­ (Fig. 5b).
tors were then spliced with the geologic unit gridding vectors, and the DenseNet (Zhou et al., 2021c) is a variant of the CNN model, which
knowledge attributes were superimposed for each geologic unit. can achieve feature reuse through feature connection in the channel

11
Q. Yan et al. Ore Geology Reviews 161 (2023) 105651

dimension. This approach can strengthen feature propagation, and altered ductile and brittle structures. P7 and P8 coincide with the
greatly reduce the number of parameters. DenseNet divides the network elevated Mo-Cd-Cu-Zn anomalies. The exposed strata at P10 comprise
into four densely-connected blocks for convolution and pooling. Each the Ordovician Banjiuguan Formation (ID = 167) and Silurian Meiziya
block has the same number of layers, i.e., convolutional layer (Conv), BN Formation (ID = 167). The lithologic assemblage includes the biotite-
Layer, and ReLU layer. Before the 3 × 3 convolution, the input feature bearing porphyroblastic schist, two-mica schist, metasandstone with
map is compressed to reduce dimension by 1 × 1 convolution. The minor dolomite interbeds. The alteration is dominated by silicification,
convolution layer was used to extract the input information, and acti­ pyritization, biotitization and chloritization. The exposed strata at P12
vate output features by executing ReLU. Therefore, the nonlinear com­ comprise the Silurian Meiziya Formation (ID = 277), and ductile and
bination function in each block became BN + ReLU + 1x1 Conv + BN + brittle shear zones (FID = 11) are widespread. This area has good pro­
ReLU + 3x3 Conv. Adjacent dense blocks were connected using an specting potential, as supported by the Au-As anomalies.
average pooling, with a step of 2 and a 1x1 convolution transition. The
last dense block performed a global average pooling to extract abnormal 6. Discussion
features, and a Softmax classifier was added for classification prediction.
6.1. Impact of sample parameters
5. Results
In the study, the sampling window size is set to an integer multiple of
5.1. Embedding features of knowledge graph the grid cell, i.e., 8, 16, and 24 times. The tests are conducted for positive
and negative sample ratios of 1:3, 1:4, and 1:5. The samples are divided
The three-dimensional embedding of ore-bearing strata and ore de­ into training sets and test sets at the ratio of 7:3. For DenseNet, K is set to
posits, faults and ore deposits, strata, faults and ore deposits, and 12 to perform the prediction test. The results show that test accuracy of
alteration zones and ore deposits is mapped onto the geologic units. The 93.98%, when the window size is 24 × 24 and the positive and negative
same color represents the consistency of knowledge. Different knowl­ sample ratio is 1:3 (Table 1).
edge embedding assigns different knowledge features to the geologic
unit (Fig. 7). 6.2. Impact of the training parameters of DenseNet model
The ore-bearing strata of No. 4–6 gold deposits have similar
knowledge embedding (Fig. 7a), i.e., Silurian metamorphic strata. The In the dense block of DenseNet, if the function of each layer generates
faults of No. 1, 2, and 4 gold deposits have similar knowledge embed­ k feature maps, then the nth layer has k0 + (n-1) k feature maps, and
ding (Fig. 7b), i.e., NW-trending brittle and ductile shear zones. No. 5 each layer can access all previous feature map in the block. k0 represents
and 6 gold deposits are located in the altered, NW-trending brittle and the channel number of the input layer, and the channel number of the
ductile shear zones. The embedding of strata and faults reflects the de­ input layer is 11 in this study. The hyperparameter k is the growth rate of
gree of influence from faults on the strata: No. 1, 2, and 4 gold deposits the network, which determines the number of the learning character­
are hosted by the Ordovician-Silurian metamorphic rocks, and istics of each layer in DenseNet, thus affecting the training model
controlled by the NW-trending faults (Fig. 7c). Local mineralization is quality. The hyperparameter k is set as 12, 24, 32, and 48 in the
controlled by the NW-trending faults, and the staggered color distribu­ experiment. When the model is trained, the batch size (set to be 32 and
tion in the southwest shows that secondary faults are extensively 64 here) would affect the generalization ability of the model.
developed (Fig. 7c). No. 1 and 2 gold deposits have similar embedding The window size is set as 24 × 24. A total of 163,894 training data
alteration knowledge with No. 5 and 6 gold deposits. No. 4 gold deposit were collected, among which 70% is used for model training and 30%
exhibits independent embedding alteration knowledge (Fig. 7d). for testing. The model is optimized with Adam (Kingma and Ba, 2014),
using 0.001 empirical learning rate, 64 batch size, and 20 training
5.2. Results of gold MPM period. From the experiment, we found that the best performance was
achieved at k = 32 (Table 2). The predicted model is saturated when it is
The well-integrated geochemical and geological map knowledge was learned to the 7th period. The loss curve has a convergence value of 0.20
used to train the gold ore prediction model with DenseNet, and the (Fig. 9a), and the accuracy is 97.78% (Fig. 9b).
prediction was performed for the entire study area. The predicted gold
mineralization area accounts for 8% of the total area, encompassing all 6.3. Impact of different knowledge graph embeddings
six existing gold deposits (Fig. 8). P1-P6 represent the six predicted
mineralization areas, whilst P7- P12 represent the six new predicted Using the trained prediction model, the results obtained from
areas. In each new predicted area, the graphic units were used to different knowledge embedding and from geochemical data were
interpret the results, combining the geological information and compared. The prediction results were then evaluated, based on the
geochemical anomalies of knowledge graph nodes. number of known mineral occurrences fallen into the predicted area.
Previous studies have shown that the formation of gold deposits in The prospecting potential area obtained from the geochemical data
this area is controlled by nappe shear structures, gold bodies mainly and knowledge embedding of ore deposit and strata (Fig. 10a) tends to
distributed in the brittle-ductile shear fault zone and the strong lamellar distribute in the Ordovician-Silurian metamorphic formation, capturing
zone nearby. The deposit mainly occurs in the Meiziya Formation of Late 83.3% of known gold occurrence (but not No. 1 gold deposit). The
Ordovician and Early Middle Silurian. Mineralization is closely related predicted area from the geochemical data and knowledge embedding of
to carboniferous sericite-quartz schist and biotite porphyritic sericite- ore deposit and fault (Fig. 10b) tends to distribute in the NW-trending
quartz schist, especially the contact surface of these two lithologic faults. The predicted area from only geochemical data tends to show
layers, which may be a favorable position for the production of gold ore horizontal and vertical grid arrangement (Fig. 10c), consistent with the
bodies (Xu et al. 2018). The intrusive rock mass is not developed and is high As-Mo-Cu-Zn anomalies. The area captured 66.67% of known gold
produced in the form of dikes. Quartz dikes, granite dikes, and alkaline occurrence, although No. 1 and No. 3 gold deposits fall outside of it. The
oblique Yunhuang (porphyry) dikes are closely related to gold miner­ predicted area from the geochemical data and knowledge embedding of
alization (Li et al., 2009). ore deposit, faults, strata, and alteration zones (Fig. 10d) is consistent
The exposed strata at P7-P9 and P11 comprise the Silurian Meiziya with the high Au-As-Mo-Cd-Cu-Zn anomalies and geological map
Formation (ID = 172), which is dominated by carbonaceous two-mica knowledge graph embedding features. Therefore, the predicted area
quartz schist, siliceous slate, and thick siliceous rock. The structural integrated with the geological map knowledge graph is well-controlled,
features are strong schistosity and silicification, superimposed by the and consistent with the local ore geological features.

12
Q. Yan et al. Ore Geology Reviews 161 (2023) 105651

6.4. Effectiveness of integrated data on CNN Appendix A. Supplementary data

The geological map knowledge and geochemical data integration Supplementary data to this article can be found online at https://doi.
(using the same sampling parameters as the DenseNet model training org/10.1016/j.oregeorev.2023.105651.
dataset), was adopted to predict local gold mineralization and compare
with the results using only geochemical data. The predicted area (using References
only six types of geochemical data) accounts for 9.8% of the total area
(Fig. 11a), and the distribution of the predicted area does not align with Anand, R.R., Aspandiar, M.F., Noble, R.R.P., 2016. A review of metal transfer
mechanisms through transported cover with emphasis on the vadose zone within the
that of elevated anomalies. When geochemical data were integrated Australian regolith. Ore Geol. Rev. 73, 394–416. https://doi.org/10.1016/j.
with the embedded geological map knowledge, the predicted area ac­ oregeorev.2015.06.018.
counts for 10.1% of the total area (Fig. 11d), consistent with the strata Brown, W.M., Gedeon, T.D., Groves, D.I., Barnes, R.G., 2000. Artificial neural networks:
A new method for mineral prospectivity mapping. Aust. J. Earth Sci. 47 (4),
distribution and fault knowledge. This further validates that embedding 757–770. https://doi.org/10.1046/j.1440-0952.2000.00807.x.
geological map knowledge can effectively constrain the predicted area Chen, C., Cai, F., Hu, X., Zheng, J., Ling, Y., Chen, H., 2021a. An entity-graph based
distribution, generating a stable prediction that conforms to geological reasoning method for fact verification. Inf. Process. Manag. 58 (3), 102472.
Chen, Y., Sui, Y., 2022. Dictionary learning for integration of evidential layers for
characteristics(Fig e and f). mineral prospectivity modeling. Ore Geol. Rev. 141, 104649.
Comparing the prediction results based on the DenseNet model Chen, Y., Wu, W., 2018. Isolation forest as an alternative data-driven mineral
(Fig. 11e and f), the CNN-based predictions have a larger predicted area, prospectivity mapping method with a higher data-processing efficiency. Nat. Resour.
Res. 28 (1), 31–46. https://doi.org/10.1007/s11053-018-9375-6.
which is not conducive to accurate exploration targeting. Due to the
Chen, Y., Zheng, C., Sun, G., 2021b. Gold prospectivity modeling by combination of
enhanced feature in the DenseNet model (Huang et al., 2017), it can laplacian eigenmaps and least angle regression. Nat. Resour. Res. 31 (4), 2023–2040.
better extract the features of integrated data, yielding a more precise https://doi.org/10.1007/s11053-021-09942-1.
target delineation in the exploration district. Ding, K.e., Xue, L., Ran, X., Wang, J., Yan, Q., 2022. Siamese network based prospecting
prediction method: A case study from the Au deposit in the Chongli mineral
concentrate area in Zhangjiakou, Hebei Province. China. Ore Geology Reviews 148,
7. Conclusions 105024.
Enkhsaikhan, M., Holden, E.-J., Duuring, P., Liu, W., 2021. Understanding ore-forming
conditions using machine reading of text. Ore Geol. Rev. 135, 104200.
In the study, we propose a method that integrates geological map Fan, R., Wang, L., Yan, J., Song, W., Zhu, Y., Chen, X., 2019. Deep learning-based named
knowledge graph with geochemical data and incorporates it into MPM. entity recognition and knowledge graph construction for geological hazards. ISPRS
The method includes three major steps: (1) Embed geological map Int. J. Geo Inf. 9 (1) https://doi.org/10.3390/ijgi9010015.
Fuentes, I., Padarian, J., Iwanaga, T., Willem Vervoort, R., 2020. 3D lithological mapping
knowledge graph with node2vec; (2) Integrate geologic map knowledge of borehole descriptions using word embeddings. Comput. Geosci. 141, 104516.
with geochemical data via knowledge embedding gridding; (3) Train Ghezelbash, R., Maghsoudi, A., Shamekhi, M., Pradhan, B., Daviran, M., 2022. Genetic
gold ore prediction model on DenseNet. algorithm to optimize the SVM and K-means algorithms for mapping of mineral
prospectivity. Neural Comput. & Applic. 35 (1), 719–733. https://doi.org/10.1007/
We constructed a MPM model for the Raofeng area, and compared s00521-022-07766-5.
the prediction results of knowledge embedding and those using only Godefroy, G., Caumon, G., Laurent, G., Bonneau, F., 2019. Structural interpretation of
chemical exploration data, which yielded the following conclusions: sparse fault data using graph theory and geological rules. Math. Geosci. 51 (8),
1091–1107. https://doi.org/10.1007/s11004-019-09800-0.
Grover, A., Leskovec, J., 2016. node2vec: scalable feature learning for networks. KDD
(1) The predicted area obtained from only geochemical data ac­ 2016, 855–864. https://doi.org/10.1145/2939672.2939754.
counts for 6% of the total area, encompassing 66.67% of known Holden, E.-J., Liu, W., Horrocks, T., Wang, R., Wedge, D., Duuring, P., Beardsmore, T.,
gold deposits. The predicted area with integrated knowledge 2019. GeoDocA – fast analysis of geological content in mineral exploration reports: a
text mining approach. Ore Geol. Rev. 111, 102919.
embedding accounts for 8% of the total area, encompassing all Huang, G., Liu, Z., Van Der Maaten, L., and Weinberger, K. Q. (2017). Densely Connected
the known gold deposits (i.e., higher prediction accuracy). The Convolutional Networks 2017 IEEE Conference on Computer Vision and Pattern
embedding of geological map knowledge effectively controls the Recognition (CVPR),111. https://CVPR/10.1109/j.oregeorev.2017.243.
Izadi, H., Sadri, J., Bayati, M., 2017. An intelligent system for mineral identification in
predicted area distribution, which makes exploration targeting thin sections based on a cascade approach. Comput. Geosci. 99, 37–49. https://doi.
more consistent with ore-forming geological features, except that org/10.1016/j.cageo.2016.10.010.
the predicted area is slightly larger than using only geochemical Ji, S., Pan, S., Cambria, E., Marttinen, P., Yu, P.S., 2022. A survey on knowledge graphs:
representation, acquisition, and applications. IEEE Trans Neural Netw Learn Syst 33
data. (2), 494–514. https://doi.org/10.1109/TNNLS.2021.3070843.
(2) For the metallogenic interpretation based on attributes of Kingma, D., and Ba, J. (2014). Adam: A Method for Stochastic Optimization. Computer
knowledge entities of the geological units (in which the predicted Science. abs/1412.6980. https://doi.org/10.48550/arXiv.1412.6980.
Laurens, V.D.M., Hinton, G., 2008. Visualizing data using t-SNE. J. Mach. Learn. Res. 9
area is located), P12 has good potential for gold formation. P12 is (2605), 2579–2605. https://doi.org/10.48550/arXiv.2108.01301.
located in the ductile–brittle shear zone. It has favorable gold Li, H., 1997. Geological features of gold deposits in the Shiquan─Hanyin northern gold
mineralization conditions, as supported by the elevated Au-As belt. Geology of Shaanxi 02, 48–57 in Chinese.
Li, Q., Chen, G., Luo, L., 2023a. Mineral prospectivity mapping using attention-based
anomalies and indicators for silicic, pyrite, and biotite alteration.
convolutional neural network. Ore Geol. Rev. 156, 105381.
Li, F., Hou, J., Yang, H., Wang, Q., 2009. Geological and geochemical characteristics and
Declaration of Competing Interest gold enrichment regularity of Yangpingwan gold deposit in Shiquan. Shaanxi
Province. Geology in China 36 (01), 174–183 in Chinese.
Li, Y.-S., Peng, C., Ran, X.-J., Xue, L.-F., Chai, S.-L., 2021. Soil geochemical prospecting
The authors declare that they have no known competing financial prediction method based on deep convolutional neural networks-Taking Daqiao
interests or personal relationships that could have appeared to influence Gold Mine in Gansu Province. China as an example. China Geology 4 (3), 1–14. https
the work reported in this paper. ://doi.org/10.31035/cg2021044.
Li, X.-M., Zhang, Y.-X., Li, Z.-K., Zhao, X.-F., Zuo, R.-G., Xiao, F., Zheng, Y.i., 2023b.
Discrimination of Pb-Zn deposit types using sphalerite geochemistry: New insights
Data availability from machine learning algorithm. Geosci. Front. 14 (4), 101580.
Li, T., Zuo, R., Zhao, X., Zhao, K., 2022. Mapping prospectivity for regolith-hosted REE
deposits via convolutional neural network with generative adversarial network
I have shared the code at the attach files step. augmented data. Ore Geol. Rev. 142, 104693.
Liu, J.D., Wang, Z., Li, X., 2016. Research on the geological characteristics and
Acknowledgment prospecting sign of yangpingwan gold deposit in Shiquan. Shaanxi. West-China
Exploration Engineering 28 (03), 150–153 in Chinese.
Medjadba, Y., Hu, D., Liu, W., Yu, X., 2020. Combining graph clustering and quantitative
This research was funded by the Geological Survey Project of China association rules for knowledge discovery in geochemical data problem. IEEE Access
Geological Survey (DD20160050) and was allowed to be published as its 8, 40453–40473. https://doi.org/10.1109/access.2019.2948800.
phased result.

13
Q. Yan et al. Ore Geology Reviews 161 (2023) 105651

Mueller, U.A., Grunsky, E.C., 2016. Multivariate spatial analysis of lake sediment Journal of Petrology 37(12): 3880-3892(in Chinese with English abstract).https:
geochemical data; Melville Peninsula, Nunavut, Canada. Appl. Geochem. 75, //doi.org/ 10.18654/1000-0569/2021.12.16.
247–262. https://doi.org/10.1016/j.apgeochem.2016.02.007. Yang, X., Han, K., Wu, X., Wang, B., Wang, X., Yang, H., He, H., Chao, H., Yu, H., 2016.
Paulheim, H., Cimiano, P., 2016. Knowledge graph refinement: A survey of approaches The structural deformation and tectonic evolution of intra-continental orogeny in
and evaluation methods. Semantic Web 8 (3), 489–508. https://doi.org/10.3233/ south Qinling orogen: Structural deformation analysis of the northern part of
sw-160218. Shiquan-Hanyin belt in the late indosinian-Yanshanian period. in Chinese with
Perozzi, B., Al-Rfou, R., & Skiena, S. (2014). DeepWalk: online learning of social English abstract Earth Sci. Front. 23 (04), 72–80. https://doi.org/10.13745/j.
representations Proceedings of the 20th ACM SIGKDD international conference on esf.2016.04.006.
Knowledge discovery and data mining, New York, New York, USA. https://doi.org/ Yang, F., Wang, Z., Zuo, R., Sun, S., Zhou, B., 2022. Quantification of uncertainty
10.1145/2623330.2623732. associated with evidence layers in mineral prospectivity mapping using direct
Polat, Ö., Polat, A., Ekici, T., 2021. Automatic classification of volcanic rocks from thin sampling and convolutional neural network. Nat. Resour. Res. 32 (1), 79–98.
section images using transfer learning networks. Neural Comput. & Applic. 33 (18), https://doi.org/10.1007/s11053-022-10144-6.
11531–11540. https://doi.org/10.1007/s00521-021-05849-3. Yin, J., Li, N., 2022. Ensemble learning models with a Bayesian optimization algorithm
Qiu, Q., Xie, Z., Wu, L., Li, W., 2018. DGeoSegmenter: A dictionary-based Chinese word for mineral prospectivity mapping. Ore Geol. Rev. 145, 104916.
segmenter for the geoscience domain. Comput. Geosci. 121, 1–11. https://doi.org/ Zhang, J., Chen, G., Wang, J., Zhang, Q., Kong, W., Ren, J., 2013. Geology of deposits in
10.1016/j.cageo.2018.08.006. northern Ankang-Shiquan. Shaanxi Province. Geology of Shaanxi 31 (02), 47–52 in
Qiu, Q., Xie, Z., Wu, L., Tao, L., Li, W., 2019. BiLSTM-CRF for geological named entity Chinese with English abstract.
recognition from the geoscience literature. Earth Sci. Inf. 12 (4), 565–579. https:// Zhang, L., Hou, M., Chen, A., Zhong, H., Ogg, J.G., Zheng, D., 2023a. Construction of a
doi.org/10.1007/s12145-019-00390-3. fluvial facies knowledge graph and its application in sedimentary facies
Qiu, Q., Xie, Z., Wu, L., Tao, L., 2020. Automatic spatiotemporal and semantic identification. Geosci. Front. 14 (2), 101521.
information extraction from unstructured geoscience reports using text mining Zhang, Z., Li, Y., Wang, G., Carranza, E.J.M., Yang, S., Sha, D., Fan, J., Zhang, X.,
techniques. Earth Sci. Inf. 13 (4), 1393–1410. https://doi.org/10.1007/s12145-020- Dong, Y., 2023b. Supervised mineral prospectivity mapping via class-balanced focal
00527-9. loss function on imbalanced geoscience datasets. Math. Geosci. https://doi.org/
Qun, Y., Linfu, X., Zeyu, L., Xin, G., Rui, W., Junhao, D., 2021. Construction of deposit 10.1007/s11004-023-10065-x.
model-oriented knowledge graph. IOP Conference Series: Earth and Environmental Zhang, Z.-J., Zhou, Y., Zhang, P., 2023c. Crucial geochemical signal identification for Cu-
Science. 671 (1), 012034. fertile magmas in paleo-tethyan arc based on machine learning. Math. Geosci. 55 (6),
Reichstein, M., Camps-Valls, G., Stevens, B., Jung, M., Denzler, J., Carvalhais, N., 799–828.
Prabhat., 2019. Deep learning and process understanding for data-driven Earth Zhao, J., Mo, X., Zhao, P., Chen, Y., 2021. Using the graph-cut method to segment the
system science. Nature 566 (7743), 195–204. https://doi.org/10.1038/s41586-019- mineralization area in the gejiu region of Yunnan Province. China. Mathematical
0912-1. Geosciences 53 (7), 1617–1642. https://doi.org/10.1007/s11004-021-09933-1.
Riahi, S., Bahroudi, A., Abedi, M., Aslani, S., Lentz, D.R., 2021. Evidential data Zheng, C., Yan, Q., Li, H.g and Yang, M. (2016). Soil Geochemical Feature and
integration to produce porphyry Cu prospectivity map, using a combination of prospecting Indicator of Silurian System Gold Deposit in Xiangzizhai-Tongqianxia
knowledge and data-driven methods. Geophys. Prospect. 70 (2), 421–437. https:// Region of Shaanxi. Gansu Metallurgy, 38(03), 66-69+72 (in Chinese with English
doi.org/10.1111/1365-2478.13169. abstract). https://doi.org/10.16042/j.cnki.cn62-1053/tf.2016.03.019.
Ribeiro, L. F. R., Saverese, P. H. P., and Figueiredo, D. R. (2017). struc2vec Proceedings Zhou, C., W, H., Wang, C., Hou, Z., Zheng, Z., Shen, S., Cheng, Q., F, Z,, Wang, X., Lu,H.,
of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Fan, J., Hu ,X,, Hou, M. and Zhu, Y. (2021). Research on geological knowledge graph
Data Mining, https://dl.acm.org/doi/abs/10.1145/3097983.3098061. in the era of big data. Scientia Sinica(Terrae) 51(07): 1070-1079(in Chinese).
Shi, Z., Zuo, R., Zhou, B., 2023. Deep reinforcement learning for mineral prospectivity Zhou, Y., Zhang, Q., Huang, Y., Yang, W., Xiao, F., Ji, J., Han, F., Tang, L., Ouyang, C.,
mapping. Math. Geosci. 55 (6), 773–797. Shen, W., 2021b. Construction knowledge graph for porphyry copper desposit in the
Wang, Z., Hua, S., 2020. Geochemical anomaly evaluation and mineralization prediction Qingzhou-Hangzhou Bay area: Insight into knowledge graph based mineral resource
of gold deposits. China Metal Bulletin 01, 47–48 in Chinese. prediction and evalution. in Chinese with English abstract Earth Sci. Front. 28 (03),
Wang, Z., Li, L., Zeng, D., Wu, X., 2021a. Incorporating prior knowledge from 67–75. https://doi.org/10.13745/j.esf.sf.2021.1.2.
counterfactuals into knowledge graph reasoning. Knowl.-Based Syst. 223, 107035. Zhou, Y., Zhou, T., Zhou, T., Fu, H., Liu, J., Shao, L., 2021c. Contrast-attentive thoracic
Wang, P., Liu, J., Li, L., 2018b. Geology, ORE-controlling factors and prospecting disease recognition with dual-weighting graph reasoning. IEEE Trans. Med. Imaging
direction of gold deposits in black rock series in YangaPingauan of Shiquan county 40 (4), 1196–1206. https://doi.org/10.1109/TMI.2021.3049498.
and Lanmugou of XunYang county. Shaanxi province. Geology of Shaanxi 36 (02), Zhu, Y., Zhou, W., Xu, Y., Liu, J., Tan, Y., 2017. Intelligent learning for knowledge graph
19–26 in Chinese with English adstract. towards geological data. Sci. Program. 2017, 1–13. https://doi.org/10.1155/2017/
Wang, C., Ma, X., Chen, J., Chen, J., 2018a. Information extraction and knowledge graph 5072427.
construction from geoscience literature. Comput. Geosci. 112, 112–120. https://doi. Zuo, R., 2020. Geodata science-based mineral prospectivity mapping: a review. Nat.
org/10.1016/j.cageo.2017.12.007. Resour. Res. 29 (6), 3415–3424. https://doi.org/10.1007/s11053-020-09700-9.
Wang, Z., Zuo, R., Liu, H., 2021b. Lithological mapping based on fully convolutional Zuo, R., Xu, Y., 2022. Graph deep learning model for mapping mineral prospectivity.
network and multi-source geological data. Remote Sens. (Basel) 13 (23). https://doi. Math. Geosci. 55 (1), 1–21. https://doi.org/10.1007/s11004-022-10015-z.
org/10.3390/rs13234860. Zuo, R., Luo, Z., Xiong, Y., Yin, B., 2022. A geologically constrained variational
Xu, F., Li, W., Song, G., Wang, M., Kang, Q., Cao, Y., Zhu, W., 2018. Ore-controlling autoencoder for mineral prospectivity mapping. Nat. Resour. Res. 31 (3),
factors and Metallogenic mode of gold deposit in Shiquan-Xunyang area. Shaanxi. 1121–1133. https://doi.org/10.1007/s11053-022-10050-x.
Mineral Exploration 01, 70–78 in Chinese. Zuo, R., Xiong, Y., Wang, Z., Wang, J., Kreuzer, O.P., 2023. A new generation of artificial
Xu, Y., Zuo, R., 2023. An interpretable graph attention network for mineral prospectivity intelligence algorithms for mineral prospectivity mapping. Nat. Resour. Res. https://
mapping. Math. Geosci. https://doi.org/10.1007/s11004-023-10076-8. doi.org/10.1007/s11053-023-10237-w.
Yang, M., Xue ,L., Ran, X-J., Sang, X-J., Yan,Q. and Dai, J-H. (2021). An intelligent
mineral geological survey method - an example of the Bridge-Yawan area in Gansu.

14