Extraction of NW-Trending, Ore-Conducting Basement Hidden Faults in Manganese Ore Concentration Area Based on Multi-Source Data in Northeastern Guizhou, China

Abstract

The Datangpo-type Mn ore deposits in northeastern Guizhou (southern China) are a relatively newly discovered type of sedimentary exhalative manganese ore deposit. Previous three-dimensional geological modeling has revealed an NW-trending trough-like depression that obliquely intersects the ENE-trending Nanhua Rift within the Nanhua System in this area. This depression likely represents a paleorift that was present before the metallogenetic period; its intersection with the Nanhua Rift corresponds precisely with the area in which a series of super-large and large new-type Mn ore deposits are located. Here, we used remote sensing image processing techniques, along with hierarchical spatial data fusion and mining methods adopted for exploration, to investigate this paleorift. Specifically, Bouguer gravity data were used to obtain middle–lower-crust structural information; aeromagnetic ΔT data were used to obtain middle–upper-crust structural information; and remote sensing and outcrop data coupled with regional geological survey, mineral exploration, and geochemical exploration data were used to obtain near-surface structural information. Combining these data, we determined the control that different deep tectonic frameworks exert on the formation and distribution of Mn ore deposits within the study area. This study proposes a new conceptual method and technical protocol permitting an improved understanding of the material source and mineralization pattern of Mn ore deposits within the study area, while verifying the existence of the NW-trending Tongren Paleorift.

1. Introduction

An area of concentrated manganese (Mn) ore deposits within the Nanhua System in northeastern Guizhou (southern China) is currently considered to have particularly high resource potential [1,2,3]. This area encompasses four hidden super-large (Daotuo, Gaodi, Pujue, and Taoziping) and multiple medium–large rhodochrosite ore deposits (Datangpo, Lijiawan, Yanglizhang, and Yangjiawan). The mineralization pattern of this ore field involves sedimentary exhalative mineralization of submarine Mn-bearing fluids [4], as determined by investigations of its ore-forming conditions and ore-controlling factors. Recently, three-dimensional (3-D) visual geological modeling of this ore field was undertaken during an intensive survey of key mineral resources in Guizhou Province [5]. A group of ENE-trending large-scale uplift and depression structures, corresponding to the Nanhua Rift, was found within the Nanhua System. Additionally, a group of NW-trending trough-like structures, which are approximately 40 km wide and traverse the entire Mn ore field, developed within the Banxi (Pt₃bx) and Fanjingshan (Pt₃f) groups in the Nanhua Rift basement; these obliquely intersect the graben–horst system of the ENE-trending Nanhua Rift [5]. A map of the tectonic framework of the Nanhua Rift basement (Figure 1) reveals that the ENE-trending uplift and depression structures ①–⑥ correspond to various level III graben and horst structures within the northwestern Nanhua Rift. In the NW-trending trough-like depression zone, two parallel secondary trough-like structures ((1) and (3) in Figure 1) were developed, with a ridge-like secondary uplift ((2) in Figure 1) sandwiched in between. The intersection between the NW-trending trough-like tectonic zone and ENE-trending Songtao–Guzhang level III graben (③ in Figure 1) corresponds precisely to the area in which the Songtao Mn ore field is located. The super-large, large, and central Mn ore deposits are mostly found in basement rocks at the intersection of these two trough-like structures and within level IV grabens of the Songtao–Guzhang level III graben in the overlying Nanhua Rift. Various lines of evidence indicate that the NW-trending trough-like depression zone likely represents a paleorift formed before the metallogenic period, and that its intersection with the Nanhua Rift controlled the diapiric upwelling of ore-forming materials [4,6,7,8,9,10,11].

The present study was undertaken to identify, assess, and verify both the nature and scale of the NW-trending basement depression zone; determine the level of its control on the formation of the Mn ore deposits; and understand its structural characteristics and spatial relationship with the graben–horst structures in the Nanhua Rift. Taking into account the ways in which different geological structures reflect at different depths using different geophysical exploration methods, appropriate spatial data fusion and mining techniques were chosen [12,13,14,15,16,17,18,19,20]; these included Gaussian filtering, optimal threshold processing, line feature expansion, and convolutional neural network (CNN) technology [21,22,23,24,25,26,27], which have proven successful in elucidating structural information from various depth levels within the Earth’s crust [28,29]. Gravity data reflect the superposition effects caused by geological bodies with different densities, sizes, and depths below the surface; such data can reflect the specific occurrence state of a target body via the method of continuation inversion. Specifically, the upward continuation method effectively suppresses high-frequency anomalies generated by shallow geological structures while enhancing low-frequency anomalies associated with large-scale, deeply buried geological bodies. It is among the most widely employed and effective techniques for separating regional from local components in gravity data, as well as for smoothing and denoising anomalous signals. Therefore, on the basis of various spatial continuation processing methods, image processing techniques were introduced to integrate and mine gravity and aeromagnetic data [30]. Lower–middle-crust structural information was obtained using Bouguer gravity data, middle–upper-crust structural information was obtained using aeromagnetic ΔT data, and near-surface structural information was obtained using mineral exploration and multispectral remote sensing data. The data were used to identify diapiric upwelling channels of deep Mn-bearing ore fluids and reveal ore-conducting, ore-matching, and ore-hosting structures within the lower, middle, and upper crust, as well as at/near the surface. We then interpreted the mechanisms underlying the accumulation and mineralization of the large amount of Mn-containing minerals within the study area. This research improves the prospecting prediction model used for this type of hidden Mn ore deposit, providing guidance for future prospecting prediction work.

2. Data Sources

The types of data used in this study are listed in

Table 1. Data used for extracting basement structures.

Type	Scale	Data Type	Objective	Data Preprocessing
Gravity data	1:200,000	TXT	Lower-crust fault information	Extension, residual field analysis
Aeromagnetic data	1:200,000	JPG	Middle-crust fault information	Polarization, extension
Remote sensing data	1:200,000	TIFF	Surface fault information	Radiation correction, atmospheric correction

. Remote sensing data were retrieved from Landsat 8 satellites, including data from three visible near-infrared bands (0.52–0.86 µm), six shortwave infrared bands (1.60–2.43 µm), and five thermal infrared bands (8.125–11.65 µm) (Figure 2). Aeromagnetic data included 1:200,000 aeromagnetic ΔT and aeromagnetic ΔT polarized data. Gravity data comprised 1:200,000 Bouguer gravity data (Figure 3), which were subjected to 20 km upward continuation, first-order derivation, and subtraction from the original gravity data to obtain the residual gravity anomaly field.

3. Methods

Our investigation of ore-controlling structures using geophysical data employed a typical spatial data mining protocol, involving the introduction of image processing techniques and adoption of various geophysical exploration approaches to identify regional basement structures at different depths (Figure 4). Using image processing techniques such as enhancement, segmentation, grayscale gradient analysis, Gaussian filtering [31], optimal threshold processing [32,33], and line feature expansion [34,35], gravity anomaly boundaries in Bouguer gravity data can be identified and extracted quickly and accurately, and the effects of noise and small patches can be overcome; this facilitates the identification and extraction of deep hidden faults in basement paleorift structures. The essential details of the method are as follows. First, during the preprocessing stage, upward continuation was applied to the gravity data; polarization and upward continuation were applied to the aeromagnetic data; radiometric and atmospheric corrections were performed on the remote sensing data; and borehole-constrained three-dimensional (3-D) modeling was conducted on the geological data. Then, during the extraction phase, based on upward continuation and polarization processing, U-Net CNN and linear segment detector (LSD) lineament extraction techniques were used for edge detection and extraction of linear body clusters from geophysical fields of different scales, thereby delineating linear structures within the lower and middle crust. Concurrently, linear structures at the surface were identified from the pre-processed remote sensing data using the LSD method. Furthermore, linear structures in the upper crust were obtained from the 3-D geological model. Finally, a comprehensive analysis was conducted on the extracted linear structures from various crustal depths with respect to upwelling channels of deep manganese-bearing minerals. This analysis aimed to elucidate the relationship between the Tongren Paleorift and the distribution of manganese deposits.

Meanwhile, this paper introduces seismic exploration profiles to further verify the effectiveness of the proposed method. The parameters of the seismic exploration profiles are as follows: Seismic exploration was performed using a 428XL multichannel telemetry digital seismometer (SERCEL, Carquefou, France). For the survey line, two lines and one shot (line distance: 20 m) were used. The geophone model was a 20DX-10 with a six-in-series combination. The observation system used 6000-20-20-20-6000 with 120 stacks. An integrated processing–interpretation model was adopted to obtain the exploration results. Processing flow design and parameter testing were conducted by focusing on fidelity and amplitude preservation, i.e., specifically protecting low-frequency signals and improving the accuracy of deep seismic imaging precision.

3.1. Linear Feature Extraction Based on U-Net CNN

Bouguer gravity anomalies are the sum of the gravitational effects produced by all the inhomogeneous density bodies underground, and they are important basic data for gravity exploration [36]. Bouguer gravity anomaly residual data generally refer to the remaining signals obtained after calculating the Bouguer gravity anomaly and applying further processing (such as trend removal or filtering), which are used to highlight local geological features. It can reflect the undulations of the Moho surface undulation, changes in crustal thickness, and variations in the distribution of inhomogeneous density bodies (such as ore bodies, structures). Currently, hidden fault information is interpreted and extracted from Bouguer gravity anomaly images using visual identification methods following specific guidelines. The success of this approach requires professional training, knowledge, and experience. However, the quality of visual interpretation is restricted by multiple factors, including the familiarity of the interpreter with the target area and their experience in interpretation, which can result in a high level of subjectivity. Additionally, owing to the low efficiency of the visual identification approach, it is often difficult to meet the needs of studies focusing on complex structures over large areas.

In this study, we introduced image processing techniques and proposed a hidden fault extraction method based on Bouguer gravity anomaly images. Following preprocessing with denoising, the Bouguer gravity anomaly residual images were trained using U-Net CNN [37,38,39], which is based on a full CNN with improved convergence speed and prediction accuracy. Its core feature is the symmetric U-shaped network structure comprising an encoder and a decoder. The encoder progressively extracts features and reduces spatial resolution through convolution and downsampling, while the decoder restores resolution and generates segmentation results through upsampling and skip connections. The skip connections in U-Net CNN concatenate the high-resolution, detail-rich feature maps from the encoder with the upsampled, semantically rich feature maps from the corresponding layers in the decoder, which helps preserve fine details and improve segmentation accuracy. Meanwhile, the U-Net CNN extracts linear structural orientations via foreground and background annotation of Bouguer gravity anomaly residual images and automatically generates an interpretation of the linear structures [22,40,41,42,43,44]. This technique can considerably improve identification efficiency and facilitates the detection of deep-seated, hidden, ore-controlling structures. The workflow of the procedure, shown in Figure 5, was as follows:

(1)

Bouguer gravity anomaly residual images were preprocessed via upward continuation for denoising.

(2)

The U-Net CNN architecture designed in this study (Figure 6) can be divided into down-sampling and up-sampling stages. The network consisted of only convolutional and pooling layers, with no fully connected layer. Here, the down-sampling stage is referred to as the encoder, and the up-sampling stage is referred to as the decoder. The encoder comprised four submodules, each of which had two 3 × 3 convolutional layers. Each convolutional layer was followed by a rectified linear unit. Each submodule was followed by a down-sampling layer, implemented through 2 × 2 max pooling with a step size of 2. The decoder, which had a layer number corresponding to that of the encoder, also consisted of four submodules. Each module first used 2 × 2 up-sampling transposed convolution to reduce the number of feature channels by half. Then, the channels were serially connected to the corresponding features of the encoder. This was followed by two 3 × 3 convolutional layers, each of which was followed by a rectified linear unit. The final layer used 1 × 1 convolution to map the feature vectors to the required number of categories.

(3)

A cross-entropy loss function was established and then used to represent the similarity between the actual output probability and the target output probability of the U-Net CNN (Loss). The calculation formula can be expressed as Formula (1):

$$Loss=-\left(ylog\right(\hat{y})+(1-y\left)log\right(1-\hat{y}\left)\right)$$

(1)

where $\hat{y}$ is the probability of a positive sample being predicted by the model and y is the sample label (positive samples were assigned a value of 1 and negative samples were assigned a value of 0); the foreground and background of the Bouguer gravity anomaly residual images were annotated as 0 and 1, respectively.

(4)

Data enhancement was conducted on single samples in the Bouguer gravity anomaly residual images; this included image distortion, rotation with 50% probability, horizontal flipping with 80% probability, vertical flipping with 30% probability, cropping of 50% of the region, random generation of images with specified resolution, and generation of a specified number of images from the pipeline.

(5)

On the basis of the fast parallel refinement algorithm, the skeletons of features expressing the spatial distribution of hidden faults were extracted from the Bouguer gravity anomaly residual images, i.e., those pixels that met specific conditions were corroded such that only the skeletons were displayed.

(6)

The skeletons in the Bouguer gravity anomaly residual images were linearly fitted using the least squares method. The skeleton extraction process was as follows. First, the binary image was preprocessed to remove noise and outliers. Second, through a pixel-by-pixel refinement operation, pixels were deleted according to a predefined refinement template. Third, the refinement operation was repeated until no further refinement was possible. Finally, the refined skeleton was post-processed to remove unreasonable branches and isolated points.

3.2. Linear Feature Extraction Based on Line Segment Detection

Gravity data converted into an image have the following two characteristics: (1) a frequency spectrum characteristic, i.e., the line segment is a high-frequency component with medium–low-frequency components on either side, and (2) a wave spectrum characteristic, i.e., the difference between the line segment and features on either side in terms of brightness or grayscale values. Existing linear structure extraction methods mostly realize automatic or semiautomatic extraction based on either the frequency spectrum characteristics or the wave spectrum characteristics. Such methods can be broadly divided into two processes: linear feature enhancement and linear feature extraction.

Here, taking into account the two linear expression characteristics, a method for the local extraction of linear structures was designed based on the LSD [45,46], which is a linear feature extraction and matching algorithm capable of self-controlling the number of false detections on line segments and obtaining detection results with subpixel-level accuracy [47,48,49]. The method involved the following stages (Figure 7).

(1)

The Bouguer gravity data of the study area were subjected to 20 km upward continuation, and the extended data were converted into 8-bit raster grayscale images. Then, Otsu’s method was employed to determine the optimal threshold for segmentation of the grayscale images that permitted grayscale images of high and low gravity anomalies to be obtained [50].

(2)

The gradient of each pixel was calculated using a 2 × 2 template, and the gradients of an entire image were quickly sorted using a non-recursive method. Subsequently, the gradient threshold was determined using an adaptive algorithm, and small gradient values in smooth areas or areas with slow changes in gradient were omitted. The pixel gradient was calculated as follows:

Let the grayscale value of the pixel point with coordinates $(x,y)$ be $i(x,y)$. Then, the gradient value $G(x,y)$ of this pixel point can be calculated as follows:

$$g_x(x,y)={\displaystyle \frac{i(x+1,y)+i(x+1,y+1)-i(x,y)-i(x,y+1)}{2}}$$

(2)

$$g_y(x,y)={\displaystyle \frac{i(x,y+1)+i(x+1,y+1)-i(x,y)-i(x+1,y)}{2}}$$

(3)

The gradient amplitude was calculated as follows:

$$G(x,y)=\sqrt{g_x^2(x,y)+g_y^2(x,y)}$$

(4)

Preferably, the formula for calculating the gradient threshold is expressed in the following form:

$$\rho ={\displaystyle \frac{q}{\sin \tau }}$$

(5)

where τ is the tolerance of region growing, $\rho$ is the gradient threshold, and $q$ is the possible error caused by the quantization effect.

(3)

The pixel point not currently used in the gradient sorting and having the largest amplitude was taken as the seed point, and its horizontal line angle was considered as the initial region angle. Eight neighboring regions were searched to determine the point at which the region angle deviated less than the tolerance value. This point was then merged into the region, and the region angle was updated.

(4)

The minimum circumscribed rectangle containing all pixel points in the linear region, constructed using the rotating calipers method, was then used for straight line extraction.

(5)

The extracted line segments were processed using the line feature expansion method and sorted in descending order of length. The left and right endpoints of the current longest and unused line segment were taken as the base points, and the endpoints of other unused line segments were used as candidate points to traverse line segments. Then, the endpoint closest to the base points was selected from the candidate list to generate a new long line segment representing the graben and horst boundary of the hidden basement.

From an optimization perspective, Gaussian filtering should be used to smooth the segmented images to eliminate noise and small polygons before calculating the pixel gradient.

4. Results and Discussion

Using our newly proposed method, i.e., through spatial data mining of aeromagnetic ΔT anomalies, we verified crystalline basement developed beneath the Nanhua Rift in northeastern Guizhou, as previously revealed via 3-D modeling [51,52]. Specifically, in the Neoproterozoic Banxi and Fanjingshan groups, NW-trending trough-like negative structures were present, which intersected the Nanhua Rift and controlled the sedimentary exhalative process of superabundant Mn-bearing fluids. Traces of these structures in mid-deep structural layers were discovered using mining Bouguer gravity anomaly residual data; this confirmed that these structures penetrated mid-deep layers. Ultimately, exploration of seismic reflection profiles revealed that these structures likely reflect a Qingbaikouan Paleorift that existed before the Nanhua Rift [4].

4.1. Reflection of NW-Trending Early Depression Zone in the Crystalline Basement

Only through geophysical data mining can it be verified whether an NW-trending trough-like depression zone within the Neoproterozoic Banxi and Fanjingshan groups exists in the basement of the Nanhua Rift in northeastern Guizhou, as previously suggested via 3-D geological modeling; this zone (Tongren Paleorift) is inferred to intersect the Nanhua Rift and control the sedimentary exhalative mineralization of superabundant Mn-bearing fluids.

Neoproterozoic and underlying Mesoproterozoic, Paleoproterozoic, and even Archaean strata in this region comprise metamorphic rocks with strong ferromagnetism, commonly known as the magnetic crystalline basement; their differences in structural components are often well reflected in magnetic anomalies [53,54,55]. Therefore, collecting magnetic anomaly data offers the possibility of understanding the characteristics of graben and horst structures in middle- and upper-crustal regions now represented by the structural layers of the basement.

Analysis of available regional aeromagnetic ΔT data from northeastern Guizhou revealed remarkable aeromagnetic anomalies in this area [56]. The aeromagnetic ΔT anomaly contour map data were extended by a 10 km upward continuation, and trough-like negative anomalies representing the existing sediments, i.e., linear structures, were extracted using the LSD algorithm. A total of four groups of relatively prominent linear body clusters trending in the NW, ENE, NE, and NNE directions were identified (Figure 8). Among those four sets of linear structural clusters extracted, the NW- and ENE-trending alignments exhibit the highest negative anomalies and are consequently identified as the strongest clusters, (Figure 8), indicating that these two groups are major structures controlling crystalline basement structural differentiation and bearing weakly magnetic sediments. The negative anomaly belt of ENE-trending linear body clusters extends in the same direction as that of the Songtao–Guzhang level III graben in the Nanhua Rift and is also consistent with the direction of extension of the long axis of all large or super-large Mn ore deposits; this renders them credible as a reflection of graben and horst structures in the Nanhua Rift. The negative magnetic anomaly belt of NW-trending linear body clusters represents a pre-Nanhuanian Paleorift [4], which will be discussed later. The NW-trending linear body clusters are consistent with the direction of a beaded arrangement of large and super-large ore deposits. For example, the NW-trending linear body clusters through Tongren–Gaodi and Xixibao–Xiaochayuan, which both contain multiple large and super-large Mn ore deposits. This further illustrates that the intersection between the two groups of strong linear body clusters served as a channel that controlled the diapiric upwelling of deep Mn-bearing fluids. The ENE-trending level IV grabens in the Nanhua Rift are ore-hosting structures for the sedimentary exhalative mineralization of rhodochrosite ores. The NE-trending linear body clusters are likely Caledonian–Hercynian structures, whereas the NNE-trending linear body clusters could relate to Yanshanian tectonism.

4.2. Reflection of NW-Trending Early Rift in the Middle–Lower Crust

If the NW-trending structures identified above do indeed represent a rift of the Banxi–Fanjingshan period that controlled the diapiric upwelling of deep Mn-bearing fluids, this would inevitably involve structural differentiation in the middle crust, leaving imprints in middle–lower crustal layers. To confirm this hypothesis, a method for potential field extension in the frequency domain was chosen for data preprocessing in this study. First, the Bouguer gravity anomaly data were extended using 6 km upward continuation. Then, the first-order derivative was solved, and the original gravity was subtracted to obtain a gravity anomaly residual planar graph (as shown in Figure 3).

Following data preprocessing, the gravity anomaly residual planar data were subjected to manual annotation of both the foreground and background, and edge cutting. Subsequently, the gravity anomaly residual planar data were extended to images with a resolution of 160 × 160 pixels (including 120 training data images and 120 label data images) using the Augmentor tool.

The specific steps were as follows: (1) random distortion of images, (2) rotation of images by 90° with 50% probability, (3) rotation of images by 270° with 50% probability, (4) flipping of images horizontally with 80% probability, (5) flipping of images vertically with 30% probability, (6) random cropping of 50% of the region, (7) random generation of images with specified resolution, and (8) generation of a specified number of images from the pipeline. Figure 9 and Figure 10 show the status of both the source images and annotated images after data enhancement processing.

The U-Net CNN architecture based on the deep learning framework TensorFlow 2.0 is provided in

Table 2. U-Net CNN architecture in the study area.

Module	Network Layer	Nuclear Size	Output Size (Fill)
	Input image	-	160 × 160 × 1
Decoder	Convolution layer 1 (ReLu)	3 × 3	160 × 160 × 64
	Convolution layer 2 (ReLu)	3 × 3	160 × 168 × 64
	Maximum pooling layer	2 × 2	80 × 80 × 64
	Convolution layer 3 (ReLu)	3 × 3	80 × 80 × 128
	Convolution layer 4 (ReLu)	3 × 3	80 × 80 × 128
	Maximum pooling layer	2 × 2	40 × 40 × 128
	Convolution layer 5 (ReLu)	3 × 3	40 × 40 × 256
	Convolution layer 6 (ReLu)	3 × 3	40 × 40 × 256
	Maximum pooling layer	2 × 2	20 × 20 × 256
	Convolution layer 7 (ReLu)	3 × 3	20 × 20 × 512
	Convolution layer 8 (ReLu)	3 × 3	20 × 20 × 512
	Maximum pooling layer	2 × 2	10 × 10 × 512
	Convolution layer 9 (ReLu)	3 × 3	10 × 10 × 1024
Encoder	Convolution layer 10 (ReLu)	3 × 3	10 × 10 × 1024
	Upward convolution + feature concatenation	2 × 2	20 × 20 × 1024
	Convolution layer 11 (ReLu)	3 × 3	20 × 20 × 512
	Convolution layer 12 (ReLu)	3 × 3	20 × 20 × 512
	Upward convolution + feature concatenation	2 × 2	40 × 40 × 512
	Convolution layer 13 (ReLu)	3 × 3	40 × 0 × 56
	Convolution layer 14 (ReLu)	3 × 3	40 × 0 × 56
	Upward convolution + feature concatenation	2 × 2	80 × 80 × 256
	Convolution layer 15 (ReLu)	3 × 3	80 × 80 × 128
	Convolution layer 16 (ReLu)	3 × 3	80 × 80 × 128
	Upward convolution + feature concatenation	2 × 2	160 × 160 × 128
	Convolution layer 17 (ReLu)	3 × 3	160 × 160 × 64
	Convolution layer 18 (ReLu)	3 × 3	160 × 160 × 64
Classification	Convolution layer 19 (ReLu)	1 × 1	160 × 160 × 2

, and its model training parameters are listed in

Table 3. Training parameters of Bouguer gravity anomaly data in the study area.

Epochs	Cross-Loss	Binary Classification Accuracy	Crossover Ratio	Dice Coefficient	Accuracy Rate
1	0.9866	0.4643	0.3275	0.4990	0.5010
5	0.6933	0.4691	0.3289	0.5002	0.6931
10	0.4586	0.8249	0.5298	0.7030	0.8701
15	0.2908	0.8756	0.6810	0.8186	0.9501
20	0.2692	0.8812	0.7001	0.8314	0.9566

.

Figure 11 shows the training flowchart of the test samples for the residual planar graph of the Bouguer gravity anomaly trend in the study area. Specifically, Figure 11a shows the test samples of the residual planar graph, Figure 11b displays the manual labels corresponding to the test samples, and Figure 11c presents the prediction results of the U-Net CNN.

Using the trained U-Net CNN model and the LSD extraction method, the Bouguer gravity anomaly data of the study area were processed, and four groups of linear body clusters—the ENE-, NW-, NE-, and NNE-trending groups that precisely matched the aeromagnetic ΔT anomaly contour data—were extracted from the Bouguer gravity anomaly data (Figure 12). As shown in Figure 12, Figure 12a is the residual planar graph of Bouguer gravity anomalies; Figure 12b is the linear structures extracted from Figure 12a using U-Net CNN and LSD; and Figure 12c presents the superimposed results of aeromagnetic anomalies and Bouguer gravity anomalies interpreted manually. This result provides further evidence that the development of the Mn ore field in northeastern Guizhou relates to the intersection of NW-trending structures with ENE-oriented basement and its trough-like depression zone (Figure 12). Our data, combined with analysis of Mn ore exploration data in the study area and a field outcrop study in the Fanjingshan area, affirm that the uplift and depression zones in the ENE-trending basement are reflections of the Nanhua Rift graben and horst structures in the Neoproterozoic Banxi and Fanjingshan groups, and that the trough-like depression zone in the NW-trending basement is a reflection of an early rift structure developed in the Neoproterozoic Banxi and Fanjingshan groups, i.e., the Tongren Paleorift [4].

We infer that the NW-trending early paleorift developed in the Banxi Group in the Nanhua Rift basement. The NW-trending linear body clusters through Tongren and Xixibao reflect two secondary grabens in this paleorift, whereas the intersection between the NW-trending early rift and ENE-oriented late rift spatially relates to the Mn ore concentration area.

4.3. Surface Reflection of Basement Structure

Landform orientation trends and linear features can reveal whether basement fault structures have been active over a long period and have had an impact on the surface of the Earth. Remote sensing and digital elevation model (DEM) data contain abundant landform orientation trends and linear features; thus, they can be used as resources for data mining. Landsat 8 satellite operational land imager (OLI) and DEM data from the study area were selected to extract landform orientation trends and linear features. Before extracting targets from remote sensing imagery, it is necessary to perform radiometric calibration and atmospheric correction on the original remote sensing image. The radiometric calibration of Landsat 8 OLI data includes laboratory calibration, onboard/star calibration, and site calibration. The first two procedures are conducted by the US Geological Survey. The third procedure facilitates radiation calibration of Landsat 8 OLI data using the ENVI 5.3 radiation calibration tool. Atmospheric correction is conducted to correct errors caused by the equipment itself and obtain the real physical model parameters of the observed object. In this study, the FLAASH (Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes) method with high correction accuracy was selected for Landsat 8 OLI data correction.

Using preprocessed remote sensing imagery as described above, continuous linear body features were extracted from the Landsat 8 OLI data using the LSD algorithm (Figure 13). Specifically, Figure 13a illustrates the feature outlines segmented through U-Net CNN from the DEM (white for background, black for foreground). Figure 13b shows the skeletons of the feature outlines. Figure 13c displays straight lines fitted according to the skeletons of the feature outlines. Figure 13d presents the linear structure extracted from the Landsat 8/OLI image using the original LSD algorithm. Figure 13e depicts the linear structure extracted from the Landsat 8/OLI image using the improved LSD algorithm. Finally, Figure 13f represents the linear structures obtained by fusing the results from Figure 13c,e. The orientations of the relatively prominent linear body clusters are evidently concentrated in three directions: NNE, NE, and ENE. However, in contrast to the features automatically extracted from the residual planar graph of gravity anomalies, the linear bodies extracted from the remote sensing images and DEM data are short and scattered. Thus, it is impossible to determine the structural information of the linear bodies in the study area directly; statistical analysis is required.

According to the sum of the lengths of the linear bodies in a given direction, the lengths of linear structures in the study area extracted using the LSD algorithm and U-Net CNN can be divided into four categories: short, intermediate, long, and extremely long. The short category accounts for most structures, indicating that linear structural information expressed at the surface is mostly scattered and inconspicuous (

Table 4. Classification of line segment length extracted based on Landsat 8/OLI image and DEM data.

ID	Level	Length (km)	LSD		U-Net		Fusion Result
			Quantity	Percent (%)	Quantity	Percent (%)	Quantity	Percent (%)
1	Short	<30	10,420	96.9	31	51.7	10,451	96.67
2	Intermediate	30–60	332	3.1	9	15	341	3.15
3	Long	60–90	0	0	8	13.3	8	0.07
4	Extremely long	>90	0	0	12	20	12	0.11

).

A density map of linear body clusters in the study area was extracted based on Landsat 8/OLI imagery and DEM data (Figure 14). The primary orientation of the linear structures was NNE, with angles of 18–25° from the north; the secondary orientation was NE, with angles of 45–50° from the north; and the tertiary orientation was ENE, with angles of 60–70° from the north. The orientations of these three groups of linear body clusters agree with the results obtained from mining aeromagnetic ΔT anomalies and Bouguer gravity anomalies. The primary, secondary, and tertiary orientations of the three sets of structural features and their relatively late activity in the strata of the study area imply Yanshanian, Caledonian–Hercynian, and Nanhuanian tectonic processes, respectively, bringing notable impacts on regional landforms. The NNE-trending linear body clusters have the largest total length, reflecting strong activity in the late Yanshanian period. Notably, the NW-trending linear body clusters, which were not mined, most likely reflect deep burial and late and weak activity. To further investigate the NW-trending basement structures, data mining was repeated using the U-Net CNN.

On the basis of in-depth observation of remote sensing and DEM images, we noted that all three major groups of linear body clusters were truncated in their orientation and showed intermittent extension. The lines connecting the truncated sites presented the remarkable characteristic of being linear body clusters with NW extension. This raises the question of whether such a feature reflects the existence of an NW-trending ancient hidden basement structure. Using the preprocessed Landsat 8 satellite OLI and DEM data from the study area, we repeated the manual process of annotation of image foreground, image background, and edge cutting for the sites where the three groups of linear body clusters were truncated in orientation. Then, the data were extended using the Augmentor tool to obtain 120 training data images and 120 label data images with a resolution of 160 × 160 pixels. Finally, the U-Net CNN was applied to enable calculation and prediction (Figure 15). Specifically, Figure 15a shows the DEM test samples, Figure 15b displays the corresponding manual labels, and Figure 15c depicts the mining results of the U-Net CNN. A set of NW-trending belts (depicted in black alternating with white in Figure 15c) is evident in the center of the study area, equivalent to the location of the NW-oriented paleorift identified using aeromagnetic ΔT and Bouguer gravity anomalies. This is inferred to reflect the Tongren Paleorift superposed on the Nanhua Rift basement.

The above data demonstrate the existence of the Qingbaikouan NW-oriented paleorift (Tongren Paleorift), which developed on the Nanhua Rift basement and is closely tied to the formation of the Songtao Mn ore field.

4.4. Verification of Deep Geological Structure via Seismic Reflection Exploration

On the basis of results obtained from data fusion and mining, an ENE-oriented short seismic survey line (No. DZ001) with a length of 18 km was arranged through the site of origin of the Mn ore deposits. This survey line was parallel to the direction of extension of the Nanhua Rift and nearly perpendicular to the NW-oriented trough. High-quality images were acquired from the metamorphic rock series in the Fanjingshan group underlying the Nanhua System. Finally, the geological structure of the Nanhua System and its crystalline basement was interpreted using the stacked profile of seismic reflection waves (Figure 16). The Neoproterozoic Nanhua System (Pt₃nh), Banxi Group (Pt₃bx), Tongchang Formation (Pt₃tc), and Huixiangping Formation (Pt₃hx) of the Fanjingshan group, as well as the Sinian System (Z), were identified. Additionally, NW-trending graben and horst structures composed of multiple normal faults trending NE and SW (e.g., F_D0–F_D3) were observed.

Specifically, F_D0 strikes NE, with considerable changes in stratigraphic structure on either side, which likely equates to the SW boundary of the NW-oriented Tongren Rift. Meanwhile, F_D1 strikes SW, and together with F_D0, forms a y-shaped structure that jointly controls the secondary graben on the SW side of the Tongren Rift (Figure 16 ①). F_D1 is also a boundary fault between the secondary graben on the SW side of the Tongren Rift and the central secondary horst; on its right side lies the central secondary horst in the NW-oriented Tongren Rift (Figure 16 ②). Moreover, F_D1-1 is a secondary fault of F_D1, F_D2 is a secondary fault in the SW-oriented secondary graben of the Tongren Rift, and F_D3 is an early nappe fault developed in the Huixiangping formation of the Fanjingshan group in the Tongren Rift basement. The F_D0, F_D1, and F_D2 faults are nearly orthogonal to the Nanhua Rift and control its lateral secondary graben and horst structures. These results corroborate the existence of the NW-oriented Tongren Paleorift, as revealed via 3-D geological modeling and the mining of aeromagnetic anomaly, Bouguer gravity anomaly, and remote sensing and DEM data. A funnel-shaped chaotic reflection zone, which extends vertically upwards from the top of the upper mantle and reaches 45 km in height [4,9,10], was also detected by our seismic reflection exploration. The upper side of this zone faces the rhodochrosite mineral ore concentration area and likely served as the diapiric exhalation channel for Mn-bearing fluids in this area. Furthermore, field surveys and drill-hole verification confirmed that the area contains concealed Mn ore bodies with a thickness of 5.47 m and an average grade of 18.66%, including rich Mn ore bodies with a thickness of 1.67 m and an average grade of 27.25%. The newly identified Mn ore resource in the target area exceeded 20.00 million tons of Mn ore, classifying the Mn ore deposit as large-scale [29,57].

5. Conclusions

Based on 3-D geological modeling of a Mn ore field in northeastern Guizhou, China, this study processed aeromagnetic ΔT, Bouguer gravity, and remote sensing and DEM data via upward continuation to verify the structure of the ore field. Image processing and analysis techniques were also introduced, and hierarchical spatial data fusion and mining were implemented based on the U-Net CNN. Four groups of linear body clusters with NW, ENE, NE, and NNE trends were extracted from the crystalline basement of the Nanhua Rift and middle–lower crust; these four groups of linear body clusters correspond to the main structural lines of the Proterozoic Qingbaikouan Tongren Paleorift, Neoproterozoic Nanhua Rift, Caledonian–Hercynian structure, and Yanshanian structure, respectively.

As discovered by 3-D geological modeling and confirmed by geological, geophysical, geochemical, and remote sensing spatial data mining, the NW-trending trough-like depression zone developed in the basement of the Qingbaikouan System formed part of the Neoproterozoic Qingbaikouan Tongren Paleorift before the Mn ore metallogenic period and intersects with the ENE-trending Nanhua Rift that formed in the Mn ore metallogenic period. The intersection of the two major rifts corresponds precisely to the location of a series of new-type Mn ore deposits; this implies that the two rifts jointly controlled the diapiric exhalative sedimentary mineralization of deep Mn-bearing fluids and the regional distribution of Mn ore deposits.

There are two main limitations of this study: (1) the proposed method is suitable only for endogenous external minerals, where mineral genesis is closely related to the structure, such as submarine Mn-containing fluid exhalative sedimentary Mn ore deposits; (2) the results must be verified through interpretation of seismic reflection data. Despite these limitations, structural information can be extracted from different depths in different regions using the proposed method, which could guide future prospecting for hidden solid ore deposits.

6. Patents

This work is related to a patent [ZL 2022 1 0601998.5] entitled [A method and system for extracting hidden basal ancient rifts based on image processing]. The patent was filed on 30 May 2022, and granted on 9 April 2024. The inventors are K.X., C.K., C.W., Y.T, and others.