Mineral resource intelligent prediction method based on AlphaEarth basic model

Abstract

The invention discloses a mineral resource intelligent prediction method based on an AlphaEarth basic model, and relates to the technical field of artificial intelligence, and the method comprises the steps: combining an enhanced feature set with a pixel space position code to obtain a position enhanced feature set, and training a Transform prediction model based on a global self-attention mechanism in combination with a mineralization supervision label set to obtain a position enhanced feature set; a Transform prediction model after training is completed is obtained; organizing the position enhancement feature set according to a tile mode to obtain a tiled position enhancement feature set, inputting the tiled position enhancement feature set into a trained Transform prediction model, and outputting a tiled mineralization grade probability set; and splicing the tile mineralization grade probability set into a mineralization grade probability graph according to spatial positions, and generating a mineralization potential index graph. According to the method, the spatial consistency and reproducibility of the input features are realized, and the intelligent drawing and target area delineation of the mineral product prospect in the research area are realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an intelligent prediction method for mineral resources based on the AlphaEarth basic model. Background Technology

[0002] Mineral resource prospecting technology has continuously evolved with the improvement of remote sensing observation capabilities, geophysical exploration capabilities, and geographic information processing capabilities. Early work was dominated by metallogenic geological theory, relying on the comprehensive discrimination of evidence layers such as lithology, structure, alteration, geochemical anomalies, and geophysical anomalies, and achieving spatial mapping through weight superposition, evidence theory, and multi-criteria evaluation. Subsequently, mineral prospecting frameworks represented by statistical learning emerged, using methods such as logistic regression, Naive Bayes, support vector machines, and random forests to map multi-source raster evidence layers into mineralization probabilities or prospect indices, realizing the transformation from experience-based interpretation to data-driven approaches. In recent years, deep learning methods have gradually entered the field of mineral prospecting. Convolutional neural networks and graph neural networks are used to extract multi-scale spatial textures and structural cues, and structures such as Transformers are used to model long-range dependencies. At the same time, basic models have developed rapidly in the field of remote sensing. The remote sensing embeddings output by basic models can express the comprehensive semantics of land cover, structure, and environmental elements in a unified vector space, providing a new feature base for cross-regional migration and weakly supervised scenarios.

[0003] Existing technologies still suffer from two key shortcomings. First, traditional mineral prospect prediction often relies on evidence layer construction and scale uniformity at the feature level. Resolution differences, projection differences, and missing pixels between evidence layers can easily cause spatial alignment errors. Furthermore, the statistical distribution of multi-source evidence layers is significantly affected by the regional geological background. When migrating to new study areas, evidence layer screening, normalization, and weight allocation need to be repeated, thus limiting generalization ability. Second, although deep learning methods can improve feature representation capabilities, in typical scenarios where mineral deposit supervision signals are sparse and class imbalanced, structures dominated by local receptive fields tend to weaken long-distance tectonic control relationships and the spatial continuity of metallogenic systems. Probability maps often exhibit local fragmentation, are highly dependent on post-processing, and struggle to simultaneously consider global spatial correlation and pixel-level scaling within a unified framework. Figure 1 To the point of being responsive. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a mineral resource intelligent prediction method based on the AlphaEarth basic model to solve the problem that the inconsistency of multi-source evidence layer feature expression across regions and the insufficient characterization of global spatial correlation in existing mineral prospect predictions lead to limited stability and transferability of mineralization probability mapping.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a method for intelligent prediction of mineral resources based on the AlphaEarth model, which includes: Remote sensing embedded raster data output by the AlphaEarth base model was obtained in the study area, and spatial benchmark unification and invalid pixel processing were performed to obtain a set of remote sensing embedded features. Based on the known spatial locations of mineral deposits in the study area, a multi-level buffer partition is constructed, and a set of mineralization monitoring labels is generated. The remote sensing embedded feature set is standardized and subjected to principal component transformation, and then fused into an enhanced feature set; The enhanced feature set is combined with the pixel spatial location encoding to obtain the location enhanced feature set, and then combined with the mineralization supervision label set to train a Transformer prediction model based on the global self-attention mechanism, thus obtaining the trained Transformer prediction model. The location enhancement feature set is organized into a tiled location enhancement feature set, which is then input into the trained Transformer prediction model to output a set of tiled mineralization level probabilities. The probability set of tiled mineralization grades is pieced together according to spatial location to form a mineralization grade probability map, and a mineralization potential index map is generated.

[0007] As a preferred embodiment of the intelligent mineral resource prediction method based on the AlphaEarth basic model described in this invention, the specific steps for acquiring remote sensing embedded raster data output by the AlphaEarth basic model in the study area are as follows: The study area boundary file is loaded into Google Earth Engine, rasterized to generate a study area range mask, and a study area coordinate system is established to generate a set of spatial references for the study area. Based on the spatial benchmark set of the study area, remote sensing embedded raster data output by the AlphaEarth base model was read in GoogleEarthEngine.

[0008] As a preferred embodiment of the intelligent mineral resource prediction method based on the AlphaEarth model described in this invention, the specific steps for obtaining the remote sensing embedded feature set are as follows: Perform study area range masking and clipping on the remote sensing embedded raster data, and then perform mosaicking to form a continuous coverage raster; The continuous coverage raster is projected to the coordinate system of the study area and resampled with resolution uniformity to obtain spatial reference unified remote sensing embedded raster data. The invalid pixel processing of the unified remote sensing embedded raster data of the spatial reference is performed by using the judgment rule of NaN pixel value to obtain invalid pixel processed remote sensing embedded raster data; The invalid pixel processing remote sensing embedded raster data is exported as a GeoTIFF file. The remote sensing embedding vector is read pixel by pixel according to the pixel index and associated with the pixel index to write it into the feature record, thus obtaining the remote sensing embedding feature set.

[0009] As a preferred embodiment of the intelligent mineral resource prediction method based on the AlphaEarth model described in this invention, the specific steps for generating the mineralization supervision label set are as follows: In ArcGIS Pro, the spatial locations of known mineral deposits in the study area are loaded and a coordinate reference check is performed. The spatial locations of known mineral deposits in the study area are then reprojected onto the coordinate system of the study area to generate a consistent set of mineral deposit features. Based on a consistent set of mineral point features, the Euclidean distance from the pixel center to the nearest mineral point is used as the pixel value within the study area. The pixel size of the mineral point distance grid is kept consistent with the pixel size of the remote sensing embedded grid data. The processing range of the mineral point distance grid is kept consistent with the mask of the study area range to obtain the mineral point distance grid. Based on the distance interval, the distance grid of the mining point is reclassified to obtain a multi-level buffer partition grid; Each distance interval of the multi-level buffered raster is assigned a corresponding mineralization level label, and the mineralization level label is associated with the pixel index and written into the label record to obtain a mineralization supervision label set.

[0010] As a preferred embodiment of the intelligent mineral resource prediction method based on the AlphaEarth model described in this invention, the specific steps for performing standardization and principal component transformation on the remote sensing embedded feature set are as follows: The remote sensing embedding feature set is expanded into a feature matrix according to the pixel index, and the mean and standard deviation of the effective pixels are calculated dimension by dimension. Zero mean unit variance transformation is performed to generate a standardized remote sensing embedding feature set. Based on the feature matrix corresponding to the standardized remote sensing embedded feature set, the covariance matrix is ​​calculated and the eigenvectors are solved to form the principal component projection matrix; The principal component projection matrices are sorted by variance contribution rate, and the principal component subsets whose cumulative variance explanation rate exceeds the explanation rate threshold are truncated. The standardized remote sensing embedded feature set is projected onto the principal component subspace to obtain the principal component feature set.

[0011] As a preferred embodiment of the intelligent mineral resource prediction method based on the AlphaEarth model described in this invention, the fusion is an enhanced feature set, and the specific steps are as follows: The pixel index is aligned with the standardized remote sensing embedded feature set and the principal component feature set by the same index, and the fused vector is obtained by vector concatenation. The fusion vectors corresponding to all pixel indices are aggregated and written into the enhanced feature set.

[0012] As a preferred embodiment of the intelligent mineral resource prediction method based on the AlphaEarth model described in this invention, the specific steps for obtaining the trained Transformer prediction model are as follows: Based on the coordinate system of the study area, the pixel index of the enhanced feature set is mapped to the two-dimensional coordinates of the pixels and encoded into a position embedding vector to form a position enhanced feature set; A Transformer prediction model is constructed using a stacked Transformer encoder structure. Based on the mineralization supervision label set and the location enhancement feature set, a training sample set is constructed, and the Transformer prediction model is trained to obtain the trained Transformer prediction model.

[0013] As a preferred embodiment of the intelligent prediction method for mineral resources based on the AlphaEarth basic model described in this invention, each layer of the Transformer encoder stack structure includes a multi-head self-attention structure, a feedforward fully connected structure, a residual connected structure, and a layer normalization structure. The encoder output of the Transformer prediction model is fed into the classification head, which sequentially performs global average pooling, Dropout, and fully connected mapping, and outputs the mineralization level prediction result through Softmax.

[0014] As a preferred embodiment of the intelligent mineral resource prediction method based on the AlphaEarth model described in this invention, the specific steps for outputting the probability set of tiled mineralization levels are as follows: According to the rasterization strategy, tile division rules are constructed within the raster range of the study area, and the pixel indices corresponding to the location enhancement feature set are grouped to generate a tile sequence covering the study area; Based on the tile sequence covering the study area, the location enhancement feature set is tiled to generate a tiled location enhancement feature set; The set of enhanced features at tiled locations is input into the trained Transformer prediction model in batches according to the tile sequence to perform forward inference, and the set of tiled mineralization level probabilities is output.

[0015] As a preferred embodiment of the intelligent mineral resource prediction method based on the AlphaEarth model described in this invention, the specific steps for generating the mineralization potential index map are as follows: The global pixel row and column index of the study area corresponding to each mineralization level probability vector is recovered by using the tile row and column index and the pixel row and column index within the tile, which are recorded in the probability set of tile mineralization level. Based on the global pixel row and column index of the study area, a framework for writing the mineralization level probability map of the raster range of the study area is established. Based on the mineralization grade probability map writing framework, the tiled mineralization grade probability set is backfilled tile by tile to generate the mineralization grade probability map. Pixel-by-pixel fusion is performed on the probabilities corresponding to non-zero mineralization levels in the mineralization level probability map to generate a mineralization potential index map.

[0016] The beneficial effects of this invention are as follows: by acquiring remote sensing embedded raster data output by the AlphaEarth basic model and performing spatial benchmark unification and invalid pixel processing to form a remote sensing embedded feature set, spatial consistency and reproducibility of input features are achieved; by expanding mineral point supervision to a mineralization supervision label set through multi-level buffer partitioning, supervision coverage is improved; by combining position encoding with global self-attention Transformer prediction model training and tiled inference stitching, mineralization level probability map and mineralization potential index map are output, realizing intelligent mapping of mineral prospect in the study area and target area delineation. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a smart prediction method for mineral resources based on the AlphaEarth basic model.

[0019] Figure 2 This is a flowchart illustrating the workings of a Transformer prediction model based on a global self-attention mechanism.

[0020] Figure 3 A flowchart for enhancing the feature set construction.

[0021] Figure 4 A flowchart for generating Transformer predictions and mineralization potential index maps. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for intelligent prediction of mineral resources based on the AlphaEarth basic model, including the following steps: S1: Obtain remote sensing embedded raster data output by the AlphaEarth base model in the study area, and perform spatial benchmark unification and invalid pixel processing to obtain a set of remote sensing embedded features. S1.1: Load the study area boundary file into Google Earth Engine, rasterize it to generate a study area range mask, establish a study area coordinate system, and generate a set of spatial references for the study area. Furthermore, the study area boundary file is loaded using Google Earth Engine. The study area boundary file is entered into the Google Earth Engine spatial object set in the form of vector features. The study area boundary file is converted into a binary raster consistent with the spatial extent of the study area through rasterization operations. The coverage position of the study area in the binary raster is assigned as a valid marker and forms a study area extent mask. The coordinate reference recorded in the study area boundary file is used to establish the study area coordinate system. The study area extent mask and the study area coordinate system are converged and written into the study area spatial reference set.

[0026] S1.2: Based on the spatial benchmark set of the study area, read the remote sensing embedded raster data output by the AlphaEarth basic model in GoogleEarthEngine; Furthermore, based on the spatial reference set of the study area, remote sensing embedded raster data output by the AlphaEarth basic model was read in GoogleEarthEngine. The reading action was completed by using the dataset identifier and band set of the remote sensing embedded raster data output by the AlphaEarth basic model as indexes. The loading result kept a set of remote sensing embedded components corresponding to each pixel position and formed remote sensing embedded raster data.

[0027] S1.3: Perform study area range masking and clipping on the remote sensing embedded raster data, and perform mosaic processing to form a continuous coverage raster; Furthermore, the remote sensing embedded raster data is subjected to study area range mask clipping. The study area range mask clipping uses the study area range mask as a pixel-level screening condition. Pixels marked as invalid by the study area range mask are removed from the remote sensing embedded raster data and the clipping result is output. The clipping result enters the mosaicking process. The mosaicking process controls the coverage priority according to the image set sorting rules and performs coverage merging on valid pixels at the same location. The mosaicking process outputs a continuous coverage raster.

[0028] It should be noted that the image set sorting rule is a rule for establishing the order of multiple images involved in the mosaicking process of remote sensing embedded raster data. The image set sorting rule constructs a sorting key based on the image attribute fields of the remote sensing embedded raster data and sorts the image set accordingly. The sorting key preferentially uses the time field to ensure that pixel values ​​with times closer to the target time period are selected at the same spatial location. When the time field cannot distinguish, the effective pixel ratio field is used to ensure that pixel values ​​with a lower invalid pixel ratio are selected at the same spatial location. The sorted image set is used in the mosaicking process of coverage merging to determine the writing order of pixel values ​​at the same spatial location.

[0029] S1.4: The continuous coverage raster is transformed to the coordinate system of the study area through projection, and resolution-consistent resampling is performed to obtain spatial reference unified remote sensing embedded raster data; Furthermore, the continuous coverage raster projection is transformed to the coordinate system of the study area. The projection transformation uses the coordinate system of the study area as the target coordinate reference and completes the pixel coordinate mapping. The projection transformation result enters the resolution-consistent resampling. The resolution-consistent resampling uses the target cell size recorded by the spatial reference set of the study area as the resampling scale and uses bilinear interpolation to calculate the resampled pixel value. The resolution-consistent resampling outputs the unified remote sensing embedded raster data of the spatial reference.

[0030] S1.5: The invalid pixel processing is performed on the unified remote sensing embedded raster data of the spatial reference using the judgment rule that the pixel value is NaN, and the invalid pixel processed remote sensing embedded raster data is obtained. Furthermore, a NaN-based determination rule is adopted to process the unified remote sensing embedded raster data of the spatial reference. The NaN-based determination rule performs validity judgment on each band component of the unified remote sensing embedded raster data of the spatial reference. Pixel positions of any component that meet the NaN condition are written into an invalid pixel mask. The invalid pixel mask is used to perform mask removal on the unified remote sensing embedded raster data of the spatial reference. The mask removal outputs invalid pixel processed remote sensing embedded raster data.

[0031] It should be noted that the NaN determination rule is a rule for performing validity judgment on pixel values ​​of spatial reference unified remote sensing embedded raster data. The NaN determination rule uses the non-numerical label of the numerical type as the judgment basis. For each pixel position of the spatial reference unified remote sensing embedded raster data, the pixel components are read band by band and NaN judgment is performed. When any band pixel component meets the NaN condition, the pixel position is written into the invalid pixel mask and the corresponding pixel position is removed from the spatial reference unified remote sensing embedded raster data in the subsequent invalid pixel processing. When the NaN determination rule is completed, the goal of reproducible labeling of invalid pixel positions is achieved.

[0032] S1.6: Export the invalid pixel processing remote sensing embedded raster data as a GeoTIFF file, read the remote sensing embedding vector pixel by pixel according to the pixel index and write it into the feature record to obtain the remote sensing embedding feature set; Furthermore, invalid pixel processing of remote sensing embedded raster data is exported as GeoTIFF files. The GeoTIFF file export uses the study area coordinate system to write the geographic reference and maintains the band order consistent with the remote sensing embedded component order. The GeoTIFF file enters a pixel-by-pixel reading process. The pixel-by-pixel reading process generates pixel indices based on row and column positions and uses the pixel indices to locate the pixel positions in the GeoTIFF file. The remote sensing embedded vectors corresponding to the pixel positions are read in band order and associated with the pixel indices to be written into feature records. The feature records are aggregated according to pixel indices to form a remote sensing embedded feature set.

[0033] S2: Based on the known spatial location of mineral deposits in the study area, a multi-level buffer partition is constructed, and a set of mineralization monitoring labels is generated; S2.1: Load the known spatial locations of mineral deposits in the study area into ArcGIS Pro and perform a coordinate reference check. Reproject the known spatial locations of mineral deposits in the study area to the coordinate system of the study area to generate a consistent mineral deposit feature set. Furthermore, ArcGIS Pro is used to load the spatial locations of known mineral deposits in the study area. ArcGIS Pro reads the coordinate reference information recorded in the layer properties of the known mineral deposit spatial locations in the study area and completes the coordinate reference check. The spatial locations of the known mineral deposits in the study area are transformed to the coordinate system of the study area through the reprojection tool. The reprojection result is written out as a consistent mineral deposit feature set.

[0034] S2.2: Based on a consistent set of mineral point features, the Euclidean distance from the pixel center to the nearest mineral point is used as the pixel value within the study area. The pixel size of the mineral point distance grid is kept consistent with the pixel size of the remote sensing embedded grid data. The processing range of the mineral point distance grid is kept consistent with the mask of the study area range to obtain the mineral point distance grid. Furthermore, a mineral point distance raster is generated based on a consistent mineral point feature set. ArcGIS Pro calls the Euclidean distance tool with the consistent mineral point feature set as the distance source. The Euclidean distance tool calculates the Euclidean distance from the center of each raster pixel to the nearest mineral point within the study area and writes it into the pixel value. The cell size of the Euclidean distance tool is set to the cell size of the remote sensing embedded raster data, and the processing range of the Euclidean distance tool is set to the mask coverage range of the study area. The Euclidean distance tool outputs the mineral point distance raster.

[0035] S2.3: Reclassify the distance grid of the mining point according to the distance interval to obtain a multi-level buffer partition grid; Furthermore, the distance raster to the mining point is reclassified to form a multi-level buffered partition raster. ArcGIS Pro calls the reclassification tool to read the distance raster to the mining point and load the distance range. The reclassification tool assigns the position of the pixel value of the distance raster to the corresponding partition code and outputs the multi-level buffered partition raster.

[0036] S2.4: Assign the corresponding mineralization level label to each distance interval of the multi-level buffered raster, and associate the mineralization level label with the pixel index and write it into the label record to obtain the mineralization supervision label set; Furthermore, ArcGIS Pro performs label assignment mapping on the multi-level buffer zone raster partition codes. The label assignment mapping adopts the correspondence between mineralization level labels and distance intervals, replacing the partition code of each pixel position of the multi-level buffer zone raster with a mineralization level label to form a mineralization level label raster. The mineralization level label raster enters the label record writing process. The label record writing process generates pixel indexes according to row and column positions and associates the pixel indexes with mineralization level labels to write them into the label records. The label records are aggregated to obtain a mineralization supervision label set.

[0037] S3: Perform standardization and principal component transformation on the remote sensing embedded feature set, and fuse them into an enhanced feature set; S3.1: Expand the remote sensing embedding feature set into a feature matrix according to the pixel index, and calculate the mean and standard deviation of the effective pixels dimension by dimension. Perform zero-mean unit variance transformation to generate a standardized remote sensing embedding feature set. Furthermore, the remote sensing embedding feature set is sorted by pixel index and written into the feature matrix. Each row of the feature matrix corresponds to a pixel index and each column corresponds to a component of the remote sensing embedding vector. The feature matrix is ​​combined with invalid pixel processing. After removing the invalid pixel mask of the remote sensing embedding raster data, the row corresponding to the invalid pixels is removed to form an effective pixel matrix. The mean vector and standard deviation vector of the effective pixel matrix are statistically calculated one dimension at a time. The zero mean unit variance transformation is performed on the effective pixel matrix one dimension at a time, and the result is written back according to the pixel index to form a standardized remote sensing embedding feature set.

[0038] S3.2: Based on the feature matrix corresponding to the standardized remote sensing embedded feature set, calculate the covariance matrix and solve the eigenvectors to form the principal component projection matrix; Furthermore, the feature matrix corresponding to the standardized remote sensing embedded feature set is read, the covariance matrix is ​​calculated and output, the covariance matrix enters the feature decomposition process and outputs the eigenvalue set and eigenvector set, and the eigenvector set is concatenated column by column and written into the principal component projection matrix.

[0039] S3.3: Sort the principal component projection matrices by variance contribution rate and extract the principal component subsets whose cumulative variance explanation rate exceeds the explanation rate threshold; Furthermore, the eigenvalue set is read and the variance contribution rate sequence is calculated. The variance contribution rate sequence is sorted and a cumulative variance explained rate sequence is generated. The cumulative variance explained rate sequence is compared with the explained rate threshold to obtain the principal component subspace dimension. The column vector slices of the principal component projection matrix corresponding to the principal component subspace dimension are written into the principal component subspace.

[0040] It should be noted that the explanatory rate threshold is obtained by constructing a validation set sample set from the training sample set and the mineralization supervision label set, evaluating the validation set classification index for each candidate explanatory rate threshold set, and selecting the optimal corresponding threshold for the validation set classification index and writing it into the explanatory rate threshold parameter record; the explanatory rate threshold value range is an interval greater than zero and not exceeding one, and the candidate explanatory rate threshold set is generated at fixed intervals within this interval.

[0041] S3.4: Project the standardized remote sensing embedded feature set onto the principal component subspace to obtain the principal component feature set; Furthermore, the standardized remote sensing embedding feature set and the principal component subspace are read. The standardized remote sensing embedding vectors are extracted row by row from the standardized remote sensing embedding feature set according to the pixel index. The principal component subspace performs linear projection on the standardized remote sensing embedding vectors and outputs the principal component vectors. The principal component vectors are written back according to the pixel index to form the principal component feature set.

[0042] S3.5: Perform same-index alignment on the pixel index of the standardized remote sensing embedded feature set and the principal component feature set, and obtain the fused vector through vector concatenation. The expression is: ; in, Pixel Index The corresponding fusion vector, Pixel Index The corresponding standardized remote sensing embedding vector, Pixel Index The corresponding principal component vectors, This is the transpose operator.

[0043] S3.6: Converge the fusion vectors corresponding to all pixel indices and write them into the enhanced feature set; Furthermore, the fusion vector is read and the corresponding pixel index is kept. The fusion vector is written into the enhanced feature record in pixel index order. The enhanced feature records are aggregated to form an enhanced feature set.

[0044] S4: Combine the enhanced feature set with the pixel spatial location encoding to obtain the location enhanced feature set, and combine it with the mineralization supervision label set to train a Transformer prediction model based on the global self-attention mechanism to obtain the trained Transformer prediction model. S4.1: Based on the coordinate system of the study area, the pixel index of the enhanced feature set is mapped to the two-dimensional coordinates of the pixels and encoded into a position embedding vector to form a position enhanced feature set; Furthermore, the enhanced feature set is read and the pixel index of the enhanced feature set is maintained. The pixel index is spatially mapped in the coordinate system of the study area to obtain the two-dimensional coordinates of the pixel. The two-dimensional coordinates of the pixel are obtained by corresponding conversion between the pixel row and column position and the raster geographic reference of the coordinate system of the study area. The two-dimensional coordinates of the pixel enter the location encoding process to generate the location embedding vector. The location embedding vector is combined with the enhanced feature set according to the pixel index and written into the location enhanced feature set.

[0045] S4.2: A Transformer prediction model is constructed using a stacked Transformer encoder structure; Each layer of the Transformer encoder stack structure includes a multi-head self-attention structure, a feedforward fully connected structure, a residual connected structure, and a layer normalization structure. The encoder output of the Transformer prediction model is fed into the classification head, which sequentially performs global average pooling, Dropout and fully connected mapping, and outputs the mineralization level prediction result through Softmax. Furthermore, starting with the location-enhanced feature set, the location-enhanced feature set is organized into input sequences in batches according to pixel index and written into the location-enhanced feature matrix X. The row vectors of the location-enhanced feature matrix X correspond to the order of pixel indices. The location-enhanced feature matrix X enters the first layer of the Transformer encoder stack structure and performs multi-head self-attention structure operation, feedforward fully connected structure operation, and residual connected structure synthesis layer normalization structure normalization in each layer in sequence. The output of the last layer of the Transformer encoder stack structure is written into the encoder mineralization level prediction T. The encoder mineralization level prediction T enters the classification head to obtain the mineralization level prediction result.

[0046] Using the location-enhanced feature matrix X as input, a linear mapping is performed on each attention head index to obtain the query matrix, key matrix, and value matrix. The query matrix and key matrix are multiplied, and scaling is performed using the key vector dimension parameter to obtain the attention scoring matrix. The attention scoring matrix is ​​then fed into a softmax function to obtain the attention weight matrix, expressed as: ; ; in, For the first Individual attention head mineralization level prediction For the first Attention weight matrix for each attention head. For the first The value matrix of each attention head, For the first A query matrix with attention heads For the first The key matrix of each attention head. The dimension parameter is the key vector.

[0047] All attention-based mineralization level predictions are concatenated along the feature dimension and then linearly mapped to the output to obtain multi-head self-attention structure mineralization level predictions.

[0048] It should be noted that the feedforward fully connected structure is used to perform nonlinear feature transformation on the row vector corresponding to each pixel index of the multi-head self-attention structure mineralization level prediction. The feedforward fully connected structure adopts a combination of two-level fully connected mapping and nonlinear activation to transform global dependency information into a higher-order representation that is more suitable for mineralization level discrimination and outputs the feedforward mineralization level prediction. The feedforward mineralization level prediction is one of the inputs to the residual connected structure.

[0049] The residual connection structure is used to maintain the continuous transmission of main path features and alleviate the training instability caused by deep stacking. The residual connection structure takes the multi-head self-attention structure position enhancement feature matrix and the feedforward mineralization level prediction as inputs of the same shape and performs same-position synthesis to obtain the residual synthesis matrix. The residual synthesis matrix enters the layer normalization structure to complete the scale stabilization process.

[0050] The layer normalization structure is used to normalize the row vectors of the residual synthesis matrix in the feature dimension and maintain numerical scale stability. The layer normalization structure calculates the mean statistic and variance statistic for each row vector and performs normalization. Then, it performs a linear transformation through the scaling parameter vector and the bias parameter vector to obtain the layer mineralization level prediction. The layer mineralization level prediction is used as the input of the next layer Transformer encoder stack structure and outputs the encoder mineralization level prediction T in the last layer.

[0051] The classification head is used to transform the encoder mineralization level prediction T into a mineralization level prediction result. The classification head performs global average pooling on the encoder mineralization level prediction T to obtain a global feature vector, and after Dropout, it enters a fully connected mapping to obtain a log-odds vector. The log-odds vector enters Softmax to obtain the mineralization level prediction result and forms a mineralization level probability vector.

[0052] S4.3: Based on the mineralization supervision label set and the location enhancement feature set, construct a training sample set and perform Transformer prediction model training to obtain the trained Transformer prediction model; Furthermore, a training sample set is constructed based on the mineralization supervision label set and the location enhancement feature set. During the construction of the training sample set, the pixel indices of the mineralization supervision label set and the location enhancement feature set are aligned to form sample pairs. After the sample pairs are written into the training sample set, the training set and validation set are divided. The division of the training set and validation set adopts a stratified sampling method of mineralization level labels to maintain the consistency of mineralization level label coverage. The Transformer prediction model training uses the training set sample pairs to perform forward computation to output the mineralization level prediction results and calculate the training loss in combination with the mineralization supervision label set. The training loss adopts the form of weighted Focal Loss and the class weight vector is calculated from the frequency of mineralization level labels and written into the training parameter record. During the training process, AdamW and cosine annealing learning rate scheduling are used to update the Transformer prediction model parameters. After each round of training, the validation set index is calculated and the parameter file corresponding to the best validation set index is recorded. The parameter file corresponding to the best validation set index is backfilled into the Transformer prediction model structure to obtain the trained Transformer prediction model.

[0053] S5: Organize the location enhancement feature set into a tiled location enhancement feature set, and input it into the trained Transformer prediction model to output a tiled mineralization level probability set; S5.1: Construct tile division rules within the grid range of the study area according to the grid tileization strategy, and group the pixel indices corresponding to the location enhancement feature set to generate a tile sequence covering the study area; Furthermore, the study area range mask and study area coordinate system are read and the study area grid range is generated. The study area grid range is represented by a set of row and column indices and is consistent with the pixel index. The grid tiling strategy generates tile division rules within the study area grid range. The tile division rules use a fixed window to slide blocks in the row and column directions. The fixed window generates tile row and column indices in the study area grid range according to the scanning order and assigns a set of pixel indexes to each tile row and column index. The set of pixel indexes is written into the tile record to form a tile sequence covering the study area.

[0054] S5.2: Based on the tile sequence covering the study area, perform tile-based organization on the location enhancement feature set to generate a tile-based location enhancement feature set; Furthermore, the tile sequence and location enhancement feature set covering the study area are read. The tile sequence covering the study area outputs a pixel index set for each tile. The pixel index set is used to extract the corresponding location enhancement feature vector from the location enhancement feature set and arrange them in the row and column order of the pixel index to form the feature record within the tile. The feature record within the tile is associated with the tile row and column index and written into the tiled location enhancement feature set. The mapping relationship between the tile row and column index and the pixel index set is maintained during the writing process of the tiled location enhancement feature set.

[0055] S5.3: Input the set of enhanced features at tiled locations into the trained Transformer prediction model in batches according to the tile sequence order to perform forward inference and output the set of tiled mineralization level probabilities; Furthermore, the tiled location enhancement feature set and the trained Transformer prediction model are read. The tiled location enhancement feature set is output as tile feature records in the order of the tile sequence and organized into inference input batches. The inference input batches are input into the trained Transformer prediction model to perform forward inference to obtain the mineralization level probability records within the tiles. The mineralization level probability records within the tiles maintain a one-to-one correspondence between the pixel index and the probability corresponding to the mineralization level label and are associated with the tile row and column indexes and written into the probability records. All probability records are aggregated to form the tiled mineralization level probability set.

[0056] S6: Piece together the probability set of tiled mineralization grades according to spatial location to form a mineralization grade probability map, and generate a mineralization potential index map; S6.1: Recover the global pixel row and column index of the study area corresponding to the probability vector of each mineralization level by using the tile row and column index and the pixel row and column index within the tile recorded in the tile mineralization level probability set; Furthermore, the tiled mineralization level probability set is read. The tiled mineralization level probability set records the tile row and column indices and the pixel row and column indices within the tiles. The tile row and column indices and the pixel row and column indices within the tiles are mapped to the global pixel row and column index of the study area based on the tile division rules. The global pixel column index of the study area is associated with the mineralization level probability vector in a one-to-one correspondence and written into the index mapping record.

[0057] S6.2: Based on the global pixel row and column index of the study area, establish a framework for writing the mineralization level probability map of the raster range of the study area; Furthermore, the global pixel row and column index of the study area is read and the raster range of the study area is established. The raster range of the study area is used to generate a raster writing structure according to the row and column range of the global pixel row and column index of the study area. The raster writing structure establishes a multi-band writing channel according to the mineralization level label dimension and forms a mineralization level probability map writing framework.

[0058] S6.3: Based on the mineralization level probability map writing framework, the tiled mineralization level probability set is backfilled tile by tile to generate a mineralization level probability map; Furthermore, the mineralization level probability map writing frame and the tiled mineralization level probability set are read. The tiled mineralization level probability set is output in the order of tile row and column indices, along with the pixel row and column indices and mineralization level probability vectors within the tiles. The pixel row and column indices within the tiles are converted into global pixel row and column indices for the study area through index mapping records. The global pixel row and column indices for the study area are used to locate the raster writing position of the mineralization level probability map writing frame. The mineralization level probability vectors are written into the mineralization level probability map writing frame in the order of mineralization level labels, and the entire area is backfilled. After the backfilling is completed, the mineralization level probability map is obtained.

[0059] S6.4: Perform pixel-by-pixel fusion on the probabilities corresponding to non-zero mineralization levels in the mineralization level probability map to generate a mineralization potential index map; ; in, Mineralization potential index map at pixel index The mineralization potential index value at the location. For mineralization grade label index, This is a set of labels for non-zero mineralization levels. Mineralization grade label The corresponding fusion weights, For the mineralization level probability map at pixel index Mineralization grade label The corresponding probability value; Furthermore, the mineralization level probability map and the mineralization supervision label set are read. The mineralization level labels in the mineralization supervision label set, excluding the background mineralization level labels, are aggregated to form a non-zero mineralization level label set. The mineralization supervision label set is statistically analyzed according to the mineralization level labels and normalized to obtain the fusion weight record. Within the grid range of the study area, the effective pixel positions are traversed by pixel index, and the probability values ​​corresponding to the non-zero mineralization level label set in the mineralization level probability map are read. The fusion weight record and the corresponding probability value are fused pixel by pixel to obtain the mineralization potential index value and written to the corresponding pixel position in the mineralization potential index map. After the full domain writing is completed, the mineralization potential index map is output.

[0060] In summary, this invention achieves spatial consistency and reproducibility of input features by: acquiring remote sensing embedded raster data output from the AlphaEarth base model and performing spatial benchmark unification and invalid pixel processing to form a remote sensing embedded feature set; expanding mineral point supervision into a mineralization supervision label set through multi-level buffer partitioning to improve supervision coverage; and outputting a mineralization level probability map and a mineralization potential index map by combining position encoding with global self-attention Transformer prediction model training and tiled inference stitching.

[0061] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent prediction of mineral resources based on the AlphaEarth basic model, characterized in that: include, Remote sensing embedded raster data output by the AlphaEarth base model was obtained in the study area, and spatial benchmark unification and invalid pixel processing were performed to obtain a set of remote sensing embedded features. Based on the known spatial locations of mineral deposits in the study area, a multi-level buffer partition is constructed, and a set of mineralization monitoring labels is generated. The remote sensing embedded feature set is standardized and subjected to principal component transformation, and then fused into an enhanced feature set; The enhanced feature set is combined with the pixel spatial location encoding to obtain the location enhanced feature set, and then combined with the mineralization supervision label set to train a Transformer prediction model based on the global self-attention mechanism, thus obtaining the trained Transformer prediction model. The location enhancement feature set is organized into a tiled location enhancement feature set, which is then input into the trained Transformer prediction model to output a set of tiled mineralization level probabilities. The probability set of tiled mineralization grades is pieced together according to spatial location to form a mineralization grade probability map, and a mineralization potential index map is generated.

2. The intelligent prediction method for mineral resources based on the AlphaEarth basic model as described in claim 1, characterized in that: The specific steps for obtaining the remote sensing embedded raster data output by the AlphaEarth base model in the study area are as follows: The study area boundary file is loaded into Google Earth Engine, rasterized to generate a study area range mask, and a study area coordinate system is established to generate a set of spatial references for the study area. Based on the spatial benchmark set of the study area, remote sensing embedded raster data output by the AlphaEarth base model was read in GoogleEarthEngine.

3. The intelligent prediction method for mineral resources based on the AlphaEarth basic model as described in claim 2, characterized in that: The specific steps for obtaining the remote sensing embedded feature set are as follows. Perform study area range masking and clipping on the remote sensing embedded raster data, and then perform mosaicking to form a continuous coverage raster; The continuous coverage raster is projected to the coordinate system of the study area and resampled with resolution uniformity to obtain spatial reference unified remote sensing embedded raster data. The invalid pixel processing of the unified remote sensing embedded raster data of the spatial reference is performed by using the judgment rule of NaN pixel value to obtain invalid pixel processed remote sensing embedded raster data; The invalid pixel processing remote sensing embedded raster data is exported as a GeoTIFF file. The remote sensing embedding vector is read pixel by pixel according to the pixel index and associated with the pixel index to write it into the feature record, thus obtaining the remote sensing embedding feature set.

4. The intelligent prediction method for mineral resources based on the AlphaEarth basic model as described in claim 3, characterized in that: The specific steps for generating the mineralization monitoring label set are as follows. In ArcGIS Pro, the spatial locations of known mineral deposits in the study area are loaded and a coordinate reference check is performed. The spatial locations of known mineral deposits in the study area are then reprojected onto the coordinate system of the study area to generate a consistent set of mineral deposit features. Based on a consistent set of mineral point features, the Euclidean distance from the pixel center to the nearest mineral point is used as the pixel value within the study area. The pixel size of the mineral point distance grid is kept consistent with the pixel size of the remote sensing embedded grid data. The processing range of the mineral point distance grid is kept consistent with the mask of the study area range to obtain the mineral point distance grid. Based on the distance interval, the distance grid of the mining point is reclassified to obtain a multi-level buffer partition grid; Each distance interval of the multi-level buffered raster is assigned a corresponding mineralization level label, and the mineralization level label is associated with the pixel index and written into the label record to obtain a mineralization supervision label set.

5. The intelligent prediction method for mineral resources based on the AlphaEarth basic model as described in claim 4, characterized in that: The specific steps for performing standardization and principal component transformation on the remote sensing embedded feature set are as follows: The remote sensing embedding feature set is expanded into a feature matrix according to the pixel index, and the mean and standard deviation of the effective pixels are calculated dimension by dimension. Zero mean unit variance transformation is performed to generate a standardized remote sensing embedding feature set. Based on the feature matrix corresponding to the standardized remote sensing embedded feature set, the covariance matrix is ​​calculated and the eigenvectors are solved to form the principal component projection matrix; The principal component projection matrices are sorted by variance contribution rate, and the principal component subsets whose cumulative variance explanation rate exceeds the explanation rate threshold are truncated. The standardized remote sensing embedded feature set is projected onto the principal component subspace to obtain the principal component feature set.

6. The intelligent prediction method for mineral resources based on the AlphaEarth basic model as described in claim 5, characterized in that: The fusion process results in an enhanced feature set, and the specific steps are as follows. The pixel index is aligned with the standardized remote sensing embedded feature set and the principal component feature set by the same index, and the fused vector is obtained by vector concatenation. The fusion vectors corresponding to all pixel indices are aggregated and written into the enhanced feature set.

7. The intelligent prediction method for mineral resources based on the AlphaEarth basic model as described in claim 6, characterized in that: The specific steps to obtain the trained Transformer prediction model are as follows. Based on the coordinate system of the study area, the pixel index of the enhanced feature set is mapped to the two-dimensional coordinates of the pixels and encoded into a position embedding vector to form a position enhanced feature set; A Transformer prediction model is constructed using a stacked Transformer encoder structure. Based on the mineralization supervision label set and the location enhancement feature set, a training sample set is constructed, and the Transformer prediction model is trained to obtain the trained Transformer prediction model.

8. The intelligent prediction method for mineral resources based on the AlphaEarth basic model as described in claim 7, characterized in that: Each layer of the Transformer encoder stack structure includes a multi-head self-attention structure, a feedforward fully connected structure, a residual connected structure, and a layer normalization structure. The encoder output of the Transformer prediction model is fed into the classification head, which sequentially performs global average pooling, Dropout, and fully connected mapping, and outputs the mineralization level prediction result through Softmax.

9. The intelligent prediction method for mineral resources based on the AlphaEarth basic model as described in claim 7, characterized in that: The specific steps for outputting the probability set of tile-like mineralization levels are as follows. According to the rasterization strategy, tile division rules are constructed within the raster range of the study area, and the pixel indices corresponding to the location enhancement feature set are grouped to generate a tile sequence covering the study area; Based on the tile sequence covering the study area, the location enhancement feature set is tiled to generate a tiled location enhancement feature set; The set of enhanced features at tiled locations is input into the trained Transformer prediction model in batches according to the tile sequence to perform forward inference, and the set of tiled mineralization level probabilities is output.

10. The intelligent prediction method for mineral resources based on the AlphaEarth basic model as described in claim 9, characterized in that: The specific steps for generating the mineralization potential index map are as follows. The global pixel row and column index of the study area corresponding to each mineralization level probability vector is recovered by using the tile row and column index and the pixel row and column index within the tile, which are recorded in the probability set of tile mineralization level. Based on the global pixel row and column index of the study area, a framework for writing the mineralization level probability map of the raster range of the study area is established. Based on the mineralization grade probability map writing framework, the tiled mineralization grade probability set is backfilled tile by tile to generate the mineralization grade probability map. Pixel-by-pixel fusion is performed on the probabilities corresponding to non-zero mineralization levels in the mineralization level probability map to generate a mineralization potential index map.

Images