An intelligent mineral prediction method based on multi-element information fusion

Abstract

The application discloses an intelligent mineral prediction method based on multi-element information fusion, and relates to the technical field of mineral exploration. The method comprises the following steps: obtaining a real data set by preprocessing geochemical data and remote sensing data; obtaining a simulation data set by data enhancement on the processed data; dividing the data set obtained by combining the real data set and the simulation data set into a training set and a test set; extracting detailed features from the data after feature extraction to obtain a first output feature map; extracting global features from the data after feature extraction to obtain a second output feature map; fusing the first output feature map and the second output feature map by a weighted average method to obtain an image fusion result; inputting the image fusion result into a feature decoder module to obtain a final fusion result; and inputting the final fusion result into a mineral prediction model based on a convolutional neural network to obtain a prediction result map. Thus, the fused image is used for mineral prediction, and the accuracy of mineral prediction is improved.

Description

Technical Field

[0001] This invention relates to the field of geological exploration technology, and in particular to an intelligent mineral prediction method based on multi-source information fusion. Background Technology

[0002] Due to limitations imposed by natural conditions and economic costs, large-scale geochemical exploration data is often difficult to obtain in practice, thus affecting the accuracy, reliability, and sustainability of geochemical data research. The development of remote sensing technology has enabled remote sensing satellites to rapidly acquire spectral information of ground features across large areas and at multiple spatial, temporal, and spectral scales. This vast amount of Earth observation data has been widely applied in geoscientific research. Generally speaking, compared to remote sensing data, geochemical data has lower spatial resolution and contains less spatial detail. However, fusing geochemical data with remote sensing data can effectively improve the spatial resolution of geochemical data and integrate multi-source information, thereby providing a richer and more accurate description for geoscientific analysis.

[0003] How to better integrate geological data from different dimensions for mineral resource prediction and improve prediction accuracy has become an urgent problem to be solved. Summary of the Invention

[0004] This invention provides an intelligent mineral prediction method based on multi-source information fusion, which can effectively integrate geochemical data and remote sensing data, improve the spatial resolution of geochemical data, and thus provide reference data information for areas lacking large-scale geochemical data, thereby improving the accuracy of mineral prediction.

[0005] According to one aspect of this disclosure, an intelligent mineral resource prediction method based on multi-source information fusion is provided, the method comprising:

[0006] Acquire geochemical and remote sensing data of the area to be predicted;

[0007] Preprocess geochemical data;

[0008] Preprocessing remote sensing data;

[0009] The preprocessed data forms the real dataset;

[0010] Data augmentation was performed on the processed geochemical and remote sensing data to obtain a simulated dataset.

[0011] The dataset, which combines real and simulated datasets, is divided into training and testing sets.

[0012] Feature extraction is performed on the training set data;

[0013] The first output feature map is obtained by extracting detailed features from the data after feature extraction using a residual multi-scale feature encoder.

[0014] A second global feature map is extracted from the feature-extracted data using a two-stage sampling feature encoder.

[0015] The first output feature map and the second output feature map are fused by weighted averaging to obtain the image fusion result;

[0016] The image fusion result is input into the feature decoder module to obtain the final fusion result;

[0017] The final fusion result is input into a mineral exploration prediction model based on a convolutional neural network to obtain the prediction result map.

[0018] In one possible implementation, geochemical data and remote sensing data of the area to be predicted are acquired, including:

[0019] Acquire Sentinel-2 satellite data as remote sensing data;

[0020] Obtain water-washed sediment measurement data for the area to be predicted at at least two scales as geochemical data.

[0021] In one possible implementation, geochemical data is point data obtained by testing the elemental content of field samples;

[0022] Preprocessing of geochemical data includes: converting geochemical point data into raster images, performing spatial geometric correction, interpolation transformation, and resampling;

[0023] Resampling is the process of extracting low-resolution data from high-resolution data.

[0024] In one possible implementation, the remote sensing data is preprocessed, including atmospheric correction, image resampling, fusion mosaicking, and cropping of the remote sensing image data in raster data format.

[0025] Atmospheric correction converts the total radiance in remote sensing data into the true reflectance of the Earth's surface.

[0026] Image resampling is the process of resampling a digital image composed of discrete data after sampling, based on the required pixel positions or pixel spacing, to form a new image after geometric transformation.

[0027] In one possible implementation, data augmentation is performed on the processed geochemical data and remote sensing data to obtain a simulated dataset, including:

[0028] The scale of the processed geochemical data is augmented: the 1:200,000 geochemical data is downsampled, and the points in the 4×4 neighborhood of any geochemical point are averaged to form new pixels. Then, the new pixels are combined to form a new geochemical dataset.

[0029] Scale-up of the processed remote sensing data: Downsample the preprocessed remote sensing data to a resolution of 40 meters.

[0030] In one possible implementation, a first output feature map is obtained by extracting detailed features from the feature-extracted data using a residual multi-scale feature encoder, including:

[0031] After feature extraction, the geochemical data is upsampled and then stitched together with the remote sensing data to form 13-channel input data. The input data is then sequentially fed into a convolutional layer, a batch regularization layer, a ReLU activation function, and a hollow spatial pyramid pooling module to obtain the first output feature map, which has 256 channels.

[0032] In one possible implementation, a second global feature map is extracted from the feature-extracted data using a two-stage sampling feature encoder, including:

[0033] After downsampling the remote sensing data after feature extraction, it is concatenated with the geochemical data after feature extraction to form 13-channel input data, resulting in 13-channel input feature maps. The input feature maps are then sequentially input into the first-level upsampling module and the second-level upsampling module to obtain the second output feature map with 256 channels. The two upsampling modules include convolutional layers, batch regularization layers, and ReLU activation functions.

[0034] Compared with the prior art, the beneficial effects of the present invention are:

[0035] This disclosure presents an intelligent mineral prediction method based on multi-source information fusion. It proposes a dual-encoder network structure for extracting features at different scales and establishes a remote sensing geochemical data fusion model using a hybrid convolutional neural network. Based on the analysis of existing deep learning fusion algorithms, a novel multi-source data fusion method based on convolutional neural networks is constructed. This method can preserve important feature information and texture details in remote sensing images and geochemical data. By combining ASPP, residual blocks, and self-supervised modules, this method can better combine low-frequency and high-frequency features extracted by the network for feature fusion, achieving a nonlinear mapping relationship between low-resolution and high-resolution data.

[0036] By adding residual neural networks, hollow spatial pyramid pooling, and deep supervision modules to the basic structure of convolutional neural networks, a multi-source data fusion model based on hybrid convolutional neural networks is constructed, combining feature maps from multiple sources at different scales. This model can integrate multi-scale features from multiple sources, improving feature capture performance. This section also considers the significant resolution difference between remote sensing and geochemical data (geochemical data has a sampling point interval of 2000m, while Sentinel-2 data has a resolution of 10m). A feature selection method is designed to establish the relationship between remote sensing and geochemical data, enabling their fusion. The fused image is then used for mineral prediction, improving the accuracy of mineral prediction. Attached Figure Description

[0037] Figure 1 A block diagram of an intelligent mineral prediction method based on multi-source information fusion according to an embodiment of the present disclosure is shown. Detailed Implementation

[0038] Various exemplary embodiments, features, and aspects of this disclosure will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0039] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments.

[0040] Furthermore, to better illustrate this disclosure, numerous specific details are set forth in the following detailed description. Those skilled in the art will understand that this disclosure can be practiced without certain specific details. In some instances, methods, means, components, and circuits well known to those skilled in the art have not been described in detail in order to highlight the main points of this disclosure.

[0041] According to one aspect of this disclosure, an intelligent mineral resource prediction method based on multi-source information fusion is provided, the method comprising:

[0042] S01, Acquire geochemical and remote sensing data of the area to be predicted; preprocess the geochemical data; preprocess the remote sensing data;

[0043] S02, The preprocessed data forms a real dataset; Data augmentation is performed on the processed geochemical data and remote sensing data to obtain a simulated dataset;

[0044] S03, the dataset resulting from the combination of the real dataset and the simulated dataset is divided into a training set and a test set;

[0045] S04, extract features from the training set data;

[0046] S05, the first output feature map is obtained by extracting detailed features from the data after feature extraction through the residual multi-scale feature encoder;

[0047] S06, the second output feature map of global features is extracted from the data after feature extraction through a two-stage sampling feature encoder; the first output feature map and the second output feature map are fused by weighted averaging to obtain the image fusion result;

[0048] S07, The image fusion result is input into the feature decoder module to obtain the final fusion result;

[0049] S08, the final fusion result is input into the mineral exploration prediction model based on convolutional neural networks to obtain the prediction result map.

[0050] The first and second output feature maps are fused using a weighted average to obtain the image fusion result. For example, the image fusion result can be obtained using equation (1): (1)

[0051] Where i covers all spatial locations of the feature map, c is the channel index, M(x) represents the feature map output by the residual multi-scale encoder module, and S(x) represents the feature map generated by the two-stage sampling encoder. express Convolution operation, The channel dimension connection is represented by F(x), which represents the feature map generated after effective fusion by the fusion module.

[0052] First, the outputs of the two feature encoders are concatenated to obtain a feature map with 512 channels. Since the feature fusion operation used in this module causes a rapid increase in channel dimensionality, a 1×1 convolution is used to map all feature maps to the same dimensional space, reducing computational cost and yielding the final feature map. Compared to fully connected layers, this fusion layer reduces parameter computation while performing dimensionality reduction on the feature maps, improving algorithm efficiency. The feature fusion layer does not change the length and width of the input feature maps; it only reduces the number of channels, resulting in an output feature map with 1 channel, ultimately yielding the predicted image fusion result.

[0053] The multi-source data fusion model adds a fusion module between the encoder and decoder, which can effectively integrate data from different sources, thereby enhancing the network's representational capabilities. This fusion module can be directly embedded into existing network architectures, employing an end-to-end training approach to improve the overall model's efficiency and performance.

[0054] The feature decoder module takes as input a feature map with one channel output from the feature fusion layer. This feature map contains fused information from remote sensing and geochemical data. The function of the feature decoder module is to restore the feature map to the original image resolution and further improve the fusion effect. The residual module brings two significant advantages to the hybrid convolutional neural network model. First, the residual module can effectively prevent gradient vanishing and degradation problems in deep networks, accelerating the feature fusion process. Second, the residual module can retain low-level feature information from the original input, enhancing the feature encoder's ability to extract spectral and spatial features from Sentinel-2 remote sensing images and geochemical data, avoiding the loss of feature information after consecutive convolutional layers. This paper constructs a hybrid convolutional neural network by incorporating a residual connection module of the residual network, combining it with convolutional layers to achieve feature decoding.

[0055] The feature decoder module has the following structure: it consists of four residual connection modules, each comprising two convolutional layers and one skip connection. The convolutional layers have the following parameters: 3×3 kernel size, stride of 1, ReLU activation function, followed by a batch normalization (BN) layer. The skip connection concatenates features from different levels, achieving the fusion of high-level and low-level features. The first three residual modules have 128 input and output channels, while the last residual module has 128 input channels and 64 output channels. Finally, a 1×1 convolutional layer reduces the number of channels in the feature map to 1, outputting the final fusion result. This module effectively preserves and transmits complete feature information, mitigating the accuracy degradation caused by network degradation. Through multiple residual connections, this module can fuse feature information at different scales as much as possible, further improving the fusion effect.

[0056] In one possible implementation, geochemical data and remote sensing data of the area to be predicted are acquired, including:

[0057] Acquire Sentinel-2 satellite data as remote sensing data;

[0058] Obtain water-washed sediment measurement data for the area to be predicted at at least two scales as geochemical data.

[0059] In one possible implementation, geochemical data is point data obtained by testing the elemental content of field samples;

[0060] Preprocessing of geochemical data includes: converting geochemical point data into raster images, performing spatial geometric correction, interpolation transformation, and resampling;

[0061] Resampling is the process of extracting low-resolution data from high-resolution data.

[0062] In one possible implementation, the remote sensing data is preprocessed, including atmospheric correction, image resampling, fusion mosaicking, and cropping of the remote sensing image data in raster data format.

[0063] Atmospheric correction converts the total radiance in remote sensing data into the true reflectance of the Earth's surface.

[0064] Image resampling is the process of resampling a digital image composed of discrete data after sampling, based on the required pixel positions or pixel spacing, to form a new image after geometric transformation.

[0065] In one possible implementation, data augmentation is performed on the processed geochemical data and remote sensing data to obtain a simulated dataset, including:

[0066] The scale of the processed geochemical data is augmented: the 1:200,000 geochemical data is downsampled, and the points in the 4×4 neighborhood of any geochemical point are averaged to form new pixels. Then, the new pixels are combined to form a new geochemical dataset.

[0067] Scale-up of the processed remote sensing data: Downsample the preprocessed remote sensing data to a resolution of 40 meters.

[0068] In one possible implementation, a first output feature map is obtained by extracting detailed features from the feature-extracted data using a residual multi-scale feature encoder, including:

[0069] After feature extraction, the geochemical data is upsampled and then stitched together with the remote sensing data to form 13-channel input data. The input data is then sequentially fed into a convolutional layer, a batch regularization layer, a ReLU activation function, and a hollow spatial pyramid pooling module to obtain the first output feature map, which has 256 channels.

[0070] To improve feature extraction performance, the residual multi-scale feature enhancement module encodes features at multiple levels. Unlike traditional deep neural networks, this module does not use pooling or other downsampling operations, making it easier to capture target features without losing resolution. Simultaneously, this module borrows ideas from the hollow spatial pyramid pooling structure and residual structure, and adds skip connections, enabling it to effectively learn spatial detail information from remote sensing images. The input data for this module includes remote sensing data and geochemical data. Since the resolution of geochemical data is lower than that of remote sensing data, it needs to be upsampled to 16 times the original image size. The upsampled geochemical data is then concatenated with the remote sensing data to form 13 channels of input data (12 channels from remote sensing data + 1 channel from geochemical element concentration data). This data is first fed into a convolutional layer with a 7×7 kernel size, a stride of 3, and 64 output channels. This convolutional layer is followed by a batch regularization layer and a ReLU activation function, transforming the linear function, which can only express a very few cases in convolution operations, into data capable of approximating any function through nonlinear operations. To maximize the utilization of feature information at different scales, residual structured blocks are used to improve the efficiency of feature extraction while reducing computational complexity and suppressing gradient vanishing. In the feature encoder branch, two 3×3 convolutional layers are employed, with skip connections added between them. These two convolutional layers have 64 and 128 input channels, respectively, and 128 output channels for each. After processing by the residual structured blocks, the feature map enters the next layer of the feature encoder.

[0071] To address the challenge of extracting feature information at different scales during the fusion of remote sensing and geochemical data, a dilated spatial pyramid pooling structure is incorporated into the feature encoder 1. This encoder utilizes multi-scale dilated convolutions to increase the receptive field without reducing the feature map resolution, thereby capturing feature information at different scales and facilitating the fusion of multi-source data. Furthermore, unlike traditional convolutional neural networks with fixed single-size kernels, the ASPP module in this encoder can pool feature maps at different resolutions, effectively addressing discontinuities in the feature map background and contributing to the acquisition of complete remote sensing and geochemical boundary texture information. The ASPP module is specifically configured with four independent branches, each sampling the feature map using dilated convolutions with different dilation rates. The dilation rates of the four branches are (1, 1, 1), (3, 3, 12), (3, 3, 24), and (3, 3, 36), respectively. The dilation rate represents the spacing between each value in the convolution kernel. For example, (3, 3, 12) indicates a kernel size of 3×3 with a dilation rate of 12. Employing four dilated convolutional kernels with different sizes and dilation rates allows for the acquisition of more comprehensive detailed information from the input image while maintaining spatial resolution. The features output from each dilated convolution are generated through multiple layers of extraction from shallow to deep using the spatial pyramid module, with each layer containing 128 channels. The original dilated spatial pyramid pooling module does not perform direct pooling operations; instead, it uses dilated convolutions to change the receptive field and capture multi-scale target feature information. To enhance the representation of global information features, adaptive average pooling is added after the ASPP module, reducing the length and width of the feature maps to 1. To increase the network width and capture local details, a 1×1 convolutional dimensionality reduction strategy, referencing GoogleNet, is used. This layer is further enhanced by adding a 1×1 convolutional layer within the spatial pyramid pooling module to measure the importance of each channel's feature map and preserve the original information from the previous layer. This strengthens feature exchange between channels, reduces the dimensionality of the feature maps, and yields 128 more effective channel feature maps. Finally, filter center points are used to extract small-scale details from geochemical data and remote sensing images. Finally, the feature maps from each branch are concatenated to obtain the output feature map of the spatial pyramid pooling module with 256 channels.

[0072] In one possible implementation, a second global feature map is extracted from the feature-extracted data using a two-stage sampling feature encoder, including:

[0073] After downsampling the remote sensing data after feature extraction, it is concatenated with the geochemical data after feature extraction to form 13-channel input data, resulting in 13-channel input feature maps. The input feature maps are then sequentially input into the first-level upsampling module and the second-level upsampling module to obtain the second output feature map with 256 channels. The two upsampling modules include convolutional layers, batch regularization layers, and ReLU activation functions.

[0074] A two-stage sampling feature encoder gradually reduces the resolution of the feature map through convolution and pooling operations to extract higher-level feature information. To effectively preserve the detailed information of the geochemical data, lower-level network information is used to restore the resolution of the feature map.

[0075] First, the remote sensing data is downsampled to reduce its spatial size to 1 / 16 of the original image size. Then, the remote sensing and geochemical data are concatenated along the channel dimension to obtain 13 channels of input features. To enhance the input features, a two-stage upsampling module (two-stage sampling feature encoder) is designed, consisting of two upsampling modules. Each upsampling module uses the UpsamplingBilinear2d module in PyTorch, specifically designed for 2D data, to perform bilinear interpolation (bilinear interpolation is used in many classic semantic segmentation networks, such as the FCN network that laid the foundation for the semantic segmentation encoder-decoder framework). This involves calculating the weighted average of four pixels in the neighborhood of the new pixel to generate a new pixel value. In deep networks, convolutional layers are typically placed before and after bilinear interpolation. In this model, each upsampling module also includes two convolutional layers to adjust the resolution of the feature map, thus specifying that the target output size is twice the input size. The parameters of the convolutional layers are set as follows: kernel size 3×3, stride 1, using BN layers and the ReLU activation function, and the calculated values ​​do not overwrite the original input. The first-stage upsampling module upsamples the spatial size of the input feature map to 64×64 and increases the number of channels to 128. The second-stage upsampling module further upsamples the output of the first-stage module to 128×128 and increases the number of channels to 256. In this way, the output of the two-stage sampling feature enhancement module has the same spatial size and number of channels as the output of the residual multi-scale encoder (residual multi-scale feature encoder), which facilitates subsequent feature fusion.

[0076] In summary, the two-stage sampling encoder is designed to establish the relationship between remote sensing data and geochemical data on a large scale, focusing more on global feature correspondences. The remote sensing data input to this module is downsampled to ignore texture details in the image, thereby reducing the learning difficulty of the model, and then the extracted texture detail features are restored through the upsampling module.

[0077] The final fusion result is input into a mineral exploration prediction model based on convolutional neural networks to obtain the prediction result map. A multi-source data fusion model using hybrid convolutional neural networks establishes the relationship between remote sensing and geochemical data, achieving their fusion. Low-frequency and high-frequency features are combined for feature fusion, realizing a nonlinear mapping relationship between low-resolution and high-resolution data. The fused data is used for mineral prediction, significantly improving prediction accuracy.

[0078] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A smart mineral resource prediction method based on multi-source information fusion, characterized in that, The method includes: Acquire geochemical and remote sensing data of the area to be predicted; preprocess the geochemical data; preprocess the remote sensing data; The preprocessed data forms a real dataset; data augmentation is performed on the processed geochemical data and remote sensing data to obtain a simulated dataset. The dataset, which combines real and simulated datasets, is divided into training and testing sets. Feature extraction is performed on the training set data; The residual multi-scale feature encoder extracts detailed features from the extracted data to obtain the first output feature map; the two-stage sampling feature encoder extracts global features from the extracted data to obtain the second output feature map. The first output feature map and the second output feature map are fused by weighted averaging to obtain the image fusion result; The image fusion result is input into the feature decoder module to obtain the final fusion result; The final fusion result is input into a mineral exploration prediction model based on a convolutional neural network to obtain the prediction result map; Acquire geochemical and remote sensing data of the area to be predicted, including: Acquire Sentinel-2 satellite data as remote sensing data; Obtain water-washed sediment measurement data for the area to be predicted at at least two scales as geochemical data; Geochemical data are point data obtained by testing the elemental content of field samples. Preprocessing of geochemical data includes: converting geochemical point data into raster images, performing spatial geometric correction, interpolation transformation, and resampling; Resampling is the process of extracting low-resolution data from high-resolution data; Preprocessing of remote sensing data includes: atmospheric correction, image resampling, fusion mosaicking, and cropping of remote sensing image data in raster data format; Atmospheric correction converts the total radiance in remote sensing data into the true reflectance of the Earth's surface. Image resampling is the process of resampling a digital image composed of discrete data after sampling, based on the required pixel positions or pixel spacing, thereby forming a new image after geometric transformation. Data augmentation was performed on the processed geochemical and remote sensing data to obtain a simulated dataset, including: The scale of the processed geochemical data is augmented: the 1:200,000 geochemical data is downsampled, and the points in the 4×4 neighborhood of any geochemical point are averaged to form new pixels. Then, the new pixels are combined to form a new geochemical dataset. Scale-up of the processed remote sensing data: Downsample the preprocessed remote sensing data to a resolution of 40 meters; The first output feature map is obtained by extracting detailed features from the data after feature extraction using a residual multi-scale feature encoder, including: After feature extraction, the geochemical data is upsampled and then concatenated with the remote sensing data after feature extraction to form 13-channel input data. The input data is then sequentially fed into a convolutional layer, a batch regularization layer, a ReLU activation function, and a hollow spatial pyramid pooling module to obtain the first output feature map, which has 256 channels. A second global feature map is extracted from the feature-extracted data using a two-stage sampling feature encoder, including: After downsampling the remote sensing data after feature extraction, it is concatenated with the geochemical data after feature extraction to form 13-channel input data, resulting in 13-channel input feature maps. The input feature maps are then sequentially input into the first-level upsampling module and the second-level upsampling module to obtain the second output feature map with 256 channels. The two upsampling modules include convolutional layers, batch regularization layers, and ReLU activation functions.

Images