Geochemical anomaly map based on soliton fitting for evaluating mineralization potential

Abstract

The application discloses a geochemical anomaly map spot mineralization potential evaluation method based on soliton fitting, which comprises the following steps: obtaining a selected anomaly residual map and normalizing the selected anomaly residual map to obtain a selected anomaly map; performing high-fraction bright core map spot detection to obtain an integer label map; performing three-branch fast discriminant filtering based on the crust abundance Clark value to remove background noise map spots and output a Clark perception label map; performing image edge effect detection to exclude map spots with incomplete radial profiles caused by edge truncation; performing soliton radial fitting on the detected map spots to obtain soliton shape matching scores; performing cross-spot global normalization on the effective map spot scores to obtain intra-batch relative mineralization potential scores, and outputting a mineralization potential score map and a visualization file. The application embeds a soliton physical model in evaluation, realizes accurate quantification of anomaly spot potential under the constraint of a physical mechanism, and effectively improves the signal-to-noise ratio and reliability of the optimization of a prospecting target area.

Description

Technical Field

[0001] This invention relates to the field of quantitative interpretation technology of geochemical exploration data, specifically to a method for evaluating the mineralization potential of geochemical anomaly patches based on soliton fitting. Background Technology

[0002] Geochemical exploration requires separating mineralized anomalies from large-scale, multi-element data. A common practice is to manually delineate anomaly halos and then sort them using indicators such as coefficient of variation, concentration, or scale zoning. However, errors in actual processing do not only arise from concentration fluctuations; patch boundaries, geological background, and raster edges can also alter the evaluation results.

[0003] Manual delineation relies on the interpreter's experience; the boundaries and areas of the same anomalous map may differ when handled by different personnel. Comparing only elemental concentrations can also overlook spatial morphology: patches with central aggregation, radial decay, and different long-axis directions, even if their means are similar, may correspond to different geological processes.

[0004] Background undulations, enrichment through mechanical transport along waterways, tectonic fracture zone responses, and hydrothermal alteration enrichment may yield similar coefficients of variation. If only a single global statistic is used for scoring, these differently induced patches will be prioritized at the same level, and false positives will then be subject to further investigation.

[0005] Deep learning models are typically trained and inferred using a tile-based approach. Patches that fall at the edges of an image or at the edges of a stitched image may only have half a radial profile; directly comparing such truncated patches with complete patches will result in directional bias in the fitting scores.

[0006] Another easily overlooked aspect is the concentration unit. Clarke values ​​are given in ppm, ppb, or mass fraction, while network inputs are often normalized; if not restored to the same unit first, Clarke thresholds are not comparable to pixel values.

[0007] Abnormal halos often exhibit a localized pattern with a high center and a decaying outer edge. 2 The function exhibits similar unimodal decay characteristics and can serve as a morphological template for repeated calculations; existing methods rarely incorporate this morphological similarity, element abundance threshold, and border integrity into the same scoring process. Although the original fitting scores of each patch can be directly compared, the normalization range must still be specified when using them across batches to avoid mistaking interval scaling values ​​as occurrence probabilities defined by the samples. Summary of the Invention

[0008] To address the technical problems in existing technologies, such as the inconsistency between the Clarke threshold and the unit domain of network data, the continued participation of truncated patches in radial fitting, and the lack of batch caliber in the original fitting scores, this invention provides a method for evaluating the mineralization potential of geochemical anomaly patches based on soliton fitting. After standardizing concentration units, the method performs Clarke tri-branch discrimination, uses quadrilateral distance to exclude truncated patches, and then uses sech... 2 The template calculates the radial morphology score and generates a [0,1] relative mineralization potential map within the same evaluation batch for quantitative priority evaluation of mineral exploration target areas.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] A method for evaluating the mineralization potential of geochemical anomaly patches based on soliton fitting includes the following steps:

[0011] S10. Obtain a selected anomaly residual map of geochemical gridded data, perform Gaussian difference enhancement on the selected anomaly residual map and normalize it to a preset interval to obtain a selected anomaly map;

[0012] S20. Perform high-resolution bright kernel patch detection on the selected anomaly map, extract candidate patches through morphological processing and connected component filtering, and output an integer labeled map.

[0013] S30. Perform a three-branch fast discriminant filtering on the integer label map based on the crustal abundance Clarke value to remove background noise patches that are lower than the crustal background abundance level and output the Clarke-sensing label map.

[0014] S40. Perform image border effect detection on each patch in the Clark perception label map, and exclude patches whose bounding box edge is less than the image boundary distance from the preset border margin from the mineralization potential evaluation to obtain a set of valid patches.

[0015] S50. Perform soliton radial fitting on each of the effective soliton sets, extract the centroid of the soliton and calculate the radial distance and energy value, use a multi-scale soliton model to perform grid search fitting, and obtain the maximum goodness of fit of each soliton as the soliton morphology matching score Q.

[0016] S60. Perform cross-patch global normalization on the soliton morphology matching scores of effective patches with a fitting status of ok, and convert the original Q value into an intra-batch relative mineralization potential score. Among them, the patch pixels excluded by Clarke discrimination, border effect detection or energy validity check are assigned a value of zero, and the pixels outside the survey area are retained as no data.

[0017] S70. Save the relative mineralization potential score as a mineralization potential score map in a preset data format, and output a visualization file with an array of patch records for sorting of target areas in the same batch.

[0018] Specifically, the process of obtaining the selected anomaly residual map of the geochemical gridded data in step S10 includes:

[0019] Geochemical gridded data is input into a two-stage progressive autoencoder, which includes a first-stage weakly pre-trained network and a second-stage refined network.

[0020] The first-stage weak pre-trained network uses the Charbonnier loss function to learn the geochemical background laws; the second-stage refined network uses a multinomial loss function that includes soliton constraints and adaptive weight graphs to refine the reconstruction of mineral-induced anomaly regions.

[0021] The residual direction is determined based on the output semantics of the second-stage refined network: when the output is the background reconstruction result, the original input data is subtracted from the output; when the output is the anomaly enhancement result, the original input data is subtracted from the output; the resulting positive anomaly residual is used as the selected anomaly residual map.

[0022] Specifically, the process of performing high-resolution bright core patch detection on the selected anomaly map in step S20 includes:

[0023] Only a limited number of pixels located within the effective measurement area are counted. A preset high proportion quantile is used as the threshold, and pixels not lower than the threshold are recorded as bright kernel pixels.

[0024] The same four-connected structuring element is used to perform one morphological opening operation and one morphological closing operation on the brightness kernel binary map; after each operation, the invalid survey area pixels are reset to zero;

[0025] Connected components in the processed image are extracted using the four-connectivity rule. Connected components with an area smaller than a preset pixel threshold are filtered out. The remaining connected components are then reassigned with consecutive integer labels to generate an integer label map. When there are no finite pixels in the effective measurement area, the integer label map is an all-zero map.

[0026] Specifically, the process of performing a three-branch fast discriminant filter on the integer label map based on the crustal abundance Clark value in step S30 includes:

[0027] Obtain standard crustal abundance data corresponding to the target element; denormalize the original geochemical concentration field to the same concentration unit as the Clark value, use only the limited concentration values ​​within the patch for discrimination, and use the product of the Clark value and the preset factor as the discrimination threshold;

[0028] For each patch in the integer label map, perform the first discrimination: if there are no pixels in the patch below the discrimination threshold, then retain the patch and skip the mean calculation;

[0029] If all pixels within a patch are below the discrimination threshold, then perform the second discrimination and delete the patch;

[0030] If only some finite concentration values ​​within a patch are lower than the discrimination threshold, then the third discrimination is performed to calculate the mean of all finite concentration values ​​for that patch; if the mean is lower than the discrimination threshold, the patch is deleted; if the mean is not lower than the discrimination threshold, the patch is retained; if there are no finite concentration values ​​within a patch, the patch is deleted and a state of no data is recorded.

[0031] The labels of the deleted patches are reset to zero, generating a Clark perception label map.

[0032] Specifically, step S40, which involves performing image border effect detection on each patch in the Clarke-sensory labeling map, includes:

[0033] Traverse each patch in the Clark perception labeling map and calculate the edge coordinates of the patch's bounding box in the vertical and horizontal directions;

[0034] Determine whether the distance between the above edge coordinates and the corresponding boundary of the image is less than the border margin; if the distance of any edge is less than the border margin, then the patch is determined to be a border-truncated patch.

[0035] The border-truncated patches are excluded from the Clarke-sensory label map to obtain a set of valid patches; the excluded patches are not sent to step S50, and the radial samples in their missing directions are not included in R. 2 calculate.

[0036] Specifically, the process of performing soliton radial fitting on each patch in the effective patch set in step S50 includes:

[0037] Extract the centroid coordinates of the current patch and calculate the radial distance from each pixel in the patch to the centroid. After removing non-finite energy values, normalize the remaining energy values ​​to a preset range. If there are fewer than 3 finite energy values ​​or the range is not greater than the preset tolerance ε, set the soliton morphology matching score Q of the patch to zero.

[0038] A set of scale parameters based on the soliton solution of the KdV nonlinear partial differential equation is constructed; the amplitude A0 = 1 is set for the normalized energy value, and the soliton radial template is expressed as: pred(r;L) = sech 2 (r / L) = 1 / cosh 2 (r / L), where L is the scale parameter and r is the radial distance;

[0039] Calculate a predicted value for each L in the set of scale parameters, and normalize the sequence of predicted values;

[0040] Calculate the goodness of fit R between the normalized predicted value and the normalized energy value. 2= 1 - SSR / max(SST, ε), where SSR is the residual sum of squares, SST is the total sum of squares, and ε is a positive number to prevent the denominator from being zero;

[0041] Corresponding each scale parameter to R 2 The maximum value is compared with zero, and the larger of the two values ​​is taken as the soliton morphological matching score Q of the patch.

[0042] Specifically, the process of performing cross-patch global normalization in step S60 includes:

[0043] Collect the soliton morphological matching scores Q of the effective patches with the fitted state of ok, and calculate the maximum value Qmax and the minimum value Qmin.

[0044] When Qmax-Qmin is greater than the tolerance ε, the intra-batch relative mineralization potential score of the patch is calculated according to score = (Q-Qmin) / (Qmax-Qmin);

[0045] When the number of scalable patches is zero, the effective survey area is assigned a value of zero and no data is retained outside the survey area; when the number of scalable patches is one, or Qmax-Qmin is not greater than ε, the relative mineralization potential score of the scalable patches is assigned a preset median value; patches whose fitting state is not OK are kept as zero.

[0046] Specifically, step S70, which involves saving the relative mineralization potential score as a mineralization potential score map in a preset data format and outputting a visualization file with an array of patch records, includes:

[0047] The relative mineralization potential score is stored as a single-precision floating-point matrix; the patch record array contains the patch number, Q value, and in-batch score annotation; the matrix and record array are written to a configurable final results directory, and the heat map serves as a visualization copy;

[0048] The final results directory and the intermediate results directory generated during the evaluation process use different paths and are independent of each other; the reporting program only reads the final results directory.

[0049] Furthermore, when the two-stage energy map produced by the two-stage progressive autoencoder is unavailable, the selected anomaly map is directly input as an energy proxy field into step S50 to calculate the radial energy value, so as to ensure that the evaluation process operates normally in the absence of a pre-stored energy map.

[0050] Furthermore, the method for evaluating the mineralization potential of geochemical anomaly patches based on soliton fitting also includes applying the mineralization potential scoring map to the multi-element comprehensive mineral exploration target area ranking:

[0051] For each of the multiple target elements, perform steps S10 to S70 respectively to obtain the corresponding single-element mineralization potential rating map;

[0052] For spatially overlapping patches with consistent grids, the single-element mineralization potential score maps of each pixel are weighted and merged or the maximum value is taken to generate a multi-element comprehensive mineralization potential score map.

[0053] A descending list of target areas is generated based on the numerical values ​​of the multi-element comprehensive mineralization potential scoring map. Each record in the list must specify at least the element combination, patch number, grid range, fusion method, and score.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] (1) This invention introduces the KdV soliton physical model as the benchmark for mineralization morphology and performs multi-scale grid search fitting, thereby realizing the deep quantitative integration of the physical diffusion dynamics mechanism of ore-forming fluids and the evaluation of geochemical anomalies. This makes the evaluation indicators have clear geological and physical interpretation capabilities, and overcomes the evaluation bias caused by the lack of physical meaning of traditional pure statistical indicators.

[0056] (2) By introducing a three-branch fast discrimination and screening mechanism based on Clark value before soliton fitting, this invention achieves complete integration of mineralization potential evaluation and crustal abundance background constraint system, thereby effectively eliminating residual background patches with lower crustal background abundance and significantly improving the signal-to-noise ratio of mineral exploration target area selection.

[0057] (3) By introducing a border effect active elimination mechanism based on bounding box boundary margin, this invention achieves accurate identification and elimination of artifacts in the training and splicing edges of segmented tiles, thereby avoiding systematic distortion of the scoring system caused by incomplete radial profile of truncated patches and greatly improving the robustness of large-format data processing.

[0058] (4) By performing cross-plot global normalization on the maximum goodness of fit of all valid plots, this invention realizes the transformation of soliton morphology matching scores into intra-batch relative mineralization potential scores within a preset probability interval, thereby solving the problem of incomparability of evaluation indicators among plots of different elements and different sizes, and directly supporting the comprehensive priority ranking decision of multiple plots and multiple elements.

[0059] (5) The entire process of this invention does not require the participation of known mineral point marking. It is based entirely on the geometric-energy characteristics and physical model of the data itself for unsupervised evaluation. The output results can be seamlessly integrated with the geochemical exploration report, thus making it suitable for new areas without prior knowledge in the early stages of exploration. It has strong engineering universality. Attached Figure Description

[0060] Figure 1This is a schematic diagram of the mineralization potential evaluation process according to an embodiment of the present invention.

[0061] Figure 2 This is a comparison chart of the various processing stages and scoring results of an embodiment of the present invention. Detailed Implementation

[0062] The present invention will be further described below with reference to the accompanying drawings and embodiments. The embodiments of the present invention include, but are not limited to, the following embodiments.

[0063] Example

[0064] like Figures 1 to 2 As shown, this embodiment provides a method for evaluating the mineralization potential of geochemical anomaly patches based on soliton fitting. Using 1:50,000 soil geochemical measurement data from an important metallogenic belt in southwestern my country, four elements—Cu, Au, As, and Sb—are selected. The implementation process of this invention is described in detail. The computing platform is an Intel Core i7 processor with 32GB of memory, and the program is written in Python.

[0065] Step S10: Obtain the selected anomaly residual map of the geochemical gridded data, perform Gaussian difference enhancement on the selected anomaly residual map and normalize it to a preset interval to obtain the selected anomaly map.

[0066] A pre-trained two-stage progressive autoencoder is invoked to process geochemical grid data. This two-stage progressive autoencoder comprises a first-stage weakly pre-trained network and a second-stage refined network. The first-stage weakly pre-trained network uses the Charbonnier loss function to learn geochemical background patterns; the second-stage refined network uses a method containing sech... 2 Soliton constraints and a multinomial loss function for the Turkey V2 adaptive weight map are used to achieve refined reconstruction of high-weight candidate regions.

[0067] Read the strongly constrained reconstruction result final.mat, the original normalized data origin.mat, and the two-stage energy map energy.mat. Select the anomalous residuals and denote them as R_sel, corresponding to R2 in the triple residual system; in this example, final represents anomalous enhancement reconstruction, so R_sel = final - origin. If the connected model defines final as background reconstruction, then R_sel = origin - final is used instead. Call the configuration to fix the residual_direction and check whether the positive high values ​​correspond to the bright kernel. Here, R2 is the residual number. Then, Gaussian difference enhancement is performed on R_sel, with the standard deviation of the Gaussian kernel at small scale σ_small = 4 pixels and the standard deviation of the Gaussian kernel at large scale σ_large = 18 pixels; the enhancement result is normalized to [0,1] and denotes it as the selected anomalous map fo_anom.

[0068] Step S20: Perform high-resolution bright kernel patch detection on the selected abnormal map, extract candidate patches through morphological processing and connected component filtering, and output an integer labeled map.

[0069] Morphological processing is performed using scipy.ndimage. `survey_mask` represents the valid survey area mask; if no separate mask file exists, it is set to a fully True array of the same shape as `fo_anom`. First, `valid_mask = np.isfinite(fo_anom) & survey_mask` is created; if `valid_mask` contains no pixels, the integer label map `labels_raw` is directly taken as a matrix of all zeros. In other cases, `thr = np.percentile(fo_anom[valid_mask],99.2)` is calculated, and `binary = valid_mask & (fo_anom>=thr)` is set. The structuring element is `structure = generate_binary_structure(2,1)`, i.e., a two-dimensional four-connected cross structure; `binary_opening` and `binary_closing` are each executed once, and after each execution, they are ANDed with `valid_mask`. `scipy.ndimage.label` uses the same structure. Connected components with an area less than 120 pixels are set to zero, and the remaining connected components are renumbered, outputting the integer label map `labels_raw`. If the pixel size is Δx×Δy, then the real area corresponding to 120 pixels is 120×Δx×Δy. Based on the data range of this embodiment, each image typically retains 5 to 30 candidate anomalous patches.

[0070] Step S30: Perform a three-branch fast discriminant filtering on the integer label map based on the crustal abundance Clark value to remove background noise patches that are lower than the crustal background abundance level, and output the Clark-sensing label map.

[0071] This step uses the original concentration field by default. The program denormalizes origin.mat to conc_real according to the scaling parameters saved during training, so that it is in the same concentration unit as the Clark value; final.mat should not be used directly to replace the original concentration. If the project explicitly requires the reconstructed concentration field for discrimination, field_source = reconstruction and the scaling parameters must be written to the results file. The abundance configuration records the data system, crustal level and version, and can be selected from Taylor-McLennan (1985), Wedbohr (1995) or Rudnick-Gao (2003). In this embodiment, Cu uses the Rudnick-Gao upper crustal value of 28 ppm, factor = 1.1, so threshold = 30.8 ppm (approximately 31 ppm). Each patch only enters the following three judgments based on a limited value in conc_real:

[0072] First option: If all finite concentration values ​​within a patch are not lower than the threshold, retain the patch and do not calculate the mean.

[0073] Second: When all finite concentration values ​​within a patch are below the threshold, the patch number is set to zero.

[0074] The third approach: If only some finite concentration values ​​are below the threshold, calculate the average of all finite concentration values ​​of the patch; if the average is below the threshold, set it to zero; otherwise, retain the original label.

[0075] When a patch has no finite concentration value, its label is set to zero, and clarke_no_data is written to the filter record. After processing, the Clarke-sensing label map labels_clarke is output; the remaining patches use the integer labels in labels_raw.

[0076] Step S40: Perform image border effect detection on each patch in the Clark perception label map, and exclude patches whose bounding box edge is less than the image boundary distance from the mineralization potential evaluation to obtain a set of valid patches.

[0077] In this embodiment, the maximum scale set value Lmax = 20 pixels, and the base margin base_margin = 20 pixels. Therefore, the preset border margin border_margin = max(base_margin, ⌈Lmax⌉) = 20 pixels. For each patch in the Clarke-aware label map labels_clarke, calculate the bounding box coordinates ymin, ymax, xmin, and xmax, and calculate dtop = ymin, dbottom = H-1-ymax, dleft = xmin, and dright = W-1-xmax. If any distance is less than border_margin, the patch is marked as border_excluded and not included in the fitting process; the remaining patches proceed to step S50.

[0078] Step S50: Perform soliton radial fitting on each patch in the set of effective patches, extract the centroid of the patch and calculate the radial distance and energy value, use a multi-scale soliton model to perform grid search fitting, and obtain the maximum goodness of fit of each patch as the soliton morphology matching score Q.

[0079] For each patch that passes the border detection, perform the following sub-steps:

[0080] a. Extract the centroid coordinates (c_y, c_x) of the current patch. This embodiment uses square pixels, and the radial distance is calculated as r = √[(y-c_y)]. 2 +(x-c_x) 2 Calculate; if the horizontal and vertical pixel dimensions of the grid are Δx and Δy respectively, then use r = √{[(y-c_y)Δy] 2 +[(x-c_x)Δx] 2}, r and scale L use the same unit of length;

[0081] b. Extract the energy value v of the patch pixels. When the two-stage energy map is available, v comes from energy.mat; when it is unavailable, v is taken from the corresponding pixel in the selected anomaly map fo_anom. First, delete NaN and ±Inf. In this embodiment, a preset tolerance ε = 10 is used. -6 When there are fewer than 3 finite values ​​or when vmax-vmin ≤ ε, set Q = 0 and record insufficient_energy, then do not perform scale search. In other cases, normalize to [0,1] using vals = (v-vmin) / (vmax-vmin);

[0082] c. In this embodiment, the side length of the square cell is denoted as δ, and the scale parameter set is L∈{2,4,6,8,10,15,20}δ; when the code is in pixels, δ=1. The energy has been normalized, so A0=1, and the soliton radial template is represented as: pred(r;L) =sech 2 (r / L) = 1 / cosh 2 (r / L);

[0083] d. Calculate the pred corresponding to each L, and divide the pred by its maximum value. Calculate the goodness-of-fit R only on the pixels corresponding to finite energy values. 2 = 1 - SSR / max(SST, ε), where SSR=Σ(vals-pred) 2 SST = Σ(vals - mean(vals)) 2 ε=10 -6 ;

[0084] e. The soliton morphological matching score of the patch is taken as Q = max(0, maxL R). 2 (L)), the code variable is denoted as best_r2. The larger Q is, the better the radial energy sequence of the patch is compared with the sech of this embodiment. 2 The closer the template is, the less calibrated the Q point sample frequency is. The patches recorded as insufficient_energy retain the screening status and Q=0, but do not proceed to the Qmin and Qmax calculations in step S60.

[0085] Step S60: Perform cross-patch global normalization on the soliton morphology matching scores of the effective patches with a fitting state of ok, and convert the original Q value into an intra-batch relative mineralization potential score. The pixels of the patches excluded by Clarke discrimination, border effect detection or energy validity check are assigned a value of zero, and pixels outside the survey area are retained as no data.

[0086] Let Q be the number of patches with a fitted state of "ok". i Qmax and Qmin are the maximum and minimum values ​​of this batch, respectively. If there are no scoreable patches, the effective survey area is written as 0; if there is only one patch or Qmax-Qmin≤10, the effective survey area is written as 0. -6 When the score is 0.5, the score for that map patch is set to 0.5; for the remaining map patches, the score is set to 0.5. i = (Q i -Qmin) / (Qmax-Qmin) is calculated. Patches excluded by insufficient_energy, Clarke discrimination screening, and border effect detection are kept at 0. The score ∈ [0,1] is only used for sorting in this batch and is not interpreted as the occurrence probability after ore point frequency calibration.

[0087] Step S70: Save the relative mineralization potential score as a mineralization potential score map in a preset data format, and output a visualization file with an array of patch records for sorting of target areas in the same batch.

[0088] The relative mineralization potential score is stored as a single-precision floating-point matrix (float32). The patch record array contains the patch number, Q value, and in-batch score annotation. `scipy.io.savemat` is called to write `score_map.astype(np.float32)` to `{element}_sech2_potential.mat`. The same MAT file stores `blob_id`, `shape_score_Q`, `batch_score`, `Qmin`, `Qmax`, `validity_code`, `field_source`, `residual_direction`, and `L_set`. PNGs with patch numbers and scores are written to the same final results directory; the reporting program does not read intermediate debugging directories.

[0089] Furthermore, the method for evaluating the mineralization potential of geochemical anomaly patches based on soliton fitting also includes applying the mineralization potential scoring map to the multi-element comprehensive mineral exploration target area ranking:

[0090] For each of the multiple target elements, perform steps S10 to S70 respectively to obtain the corresponding single-element mineralization potential rating map;

[0091] For spatially overlapping patches with consistent grids, in the effective element set E of cell x x S_multi(x)=Σe∈E x w e S e (x) / Σe∈E x w e Perform weighted fusion, where w e ≥0 and Σe∈E x w e >0; When no weight is preset, take w for E target elements. e =1 / E, for example, Cu, Au, As, and Sb each take 0.25 when there is no preset weight; or it can be in the set of valid elements E x Inner pressing S_multi(x)=maxe∈E x S e (x) takes the maximum value per pixel; E x When the value is empty, the cell is recorded as no data, without using zero to represent it; thus, a multi-element comprehensive mineralization potential rating map is obtained.

[0092] Based on the numerical values ​​of the multi-element comprehensive mineralization potential scoring map S_multi, a descending target area list is generated. Each record in the list includes the element combination, patch number, xmin, xmax, ymin, ymax, fusion method, weight, single element score, and comprehensive score.

[0093] To verify the effectiveness of the method of this invention, calculations were performed using Cu elemental data. In a test area of ​​a metallogenic belt in Southwest China, the high anomaly zone delineated by the traditional coefficient of variation method contained mechanically transported enrichment along the river system. According to the original verification records of the project, the anomaly verification matching rate of the traditional method was approximately 45%. The test records of this method listed 5 representative patches, of which 3 were included in the fitting, 1 was deleted during Clarke screening, and 1 was excluded during border inspection:

[0094] Patch #1 (known hydrothermal alteration core area): best_r2=0.82, score within the batch=1.00, listed as the highest priority in this batch;

[0095] Patch #2 (known skarn contact zone): best_r2=0.68, score within batch=0.66, ranked second;

[0096] Patch #3 (Scattered anomaly in a structurally fractured zone): best_r2=0.41, intra-batch score=0.00, lowest sorted.

[0097] Patch #4 (background residual area): The patch's mean value is below the Clarke threshold, so it is deleted in S30, score=0;

[0098] Patch #5 (Tile splicing edge anomaly): Insufficient margin, excluded in S40, score=0.

[0099] Subsequent engineering records show that mineralization clues were observed in both patch #1 and patch #2. The anomaly verification matching rate is calculated as "number of verified patches showing mineralization clues / total number of verified patches × 100%"; the original verification table record for the same project shows 82.5%, and the numerator, denominator, verification date, and project type in the table constitute the basis for recalculating this value. This value is only used for this test area and is not extrapolated to a fixed accuracy for other areas.

[0100] Figure 2 The original geochemical grid, R2 enhancement results, bright core patch numbers, Clarke screening results, 20-pixel border effective area, and sech are presented in sequence. 2 Radial fitting curve, R of each patch 2The chart includes in-batch normalized scores, score plots, final potential plots, and comparisons of masks and horizontal profiles before and after screening. In the illustrated batches, the R² values ​​for #1, #2, and #3 are 0.922, 0.490, and 0.000, respectively, corresponding to relative scores of 1.00, 0.53, and 0.00. This illustrated batch, along with the aforementioned Cu engineering records, is used to illustrate the process and engineering verification, respectively.

[0101] The pseudocode for the core algorithm is as follows:

[0102] Algorithm 1: Fast Three-Branch Discrimination Based on Clarke Value

[0103] INPUT: labels_raw(H,W), conc_real(H,W) [inverse normalized from the original input and in the same units as clarke_val], valid_mask(H,W), clarke_val, factor=1.1

[0104] threshold ← factor × clarke_val

[0105] valid_conc ← valid_mask ∧ isfinite(conc_real)

[0106] out ← labels_raw.copy()

[0107] FOR lb IN unique(labels_raw, lb>0):

[0108] pm ← (labels_raw == lb)

[0109] pv ← conc_real[pm ∧ valid_conc]

[0110] IF len(pv)==0: out[pm] ← 0; RECORD("clarke_no_data"); CONTINUE

[0111] IF all(pv >= threshold): CONTINUE # Reserved

[0112] IF all(pv < threshold): out[pm] ← 0; CONTINUE # Delete

[0113] IF mean(pv) < threshold: out[pm] ← 0 # Delete

[0114] OUTPUT: out (labels_clarke)

[0115] Algorithm 2: Border Effect Removal

[0116] INPUT: labels_clarke(H,W), base_margin=20, L_max=20

[0117] border_margin ← max(base_margin, ceil(L_max)); H,W ← labels_clarke.shape

[0118] valid_blobs ← []

[0119] FOR lb IN unique(labels_clarke, lb>0):

[0120] yy, xx ← where(labels_clarke == lb)

[0121] IF yy.min() < border_margin OR

[0122] yy.max() >= H-border_margin OR

[0123] xx.min() < border_margin OR

[0124] xx.max() >= W-border_margin:

[0125] PRINT("border-excluded")

[0126] CONTINUE

[0127] valid_blobs.append(lb)

[0128] OUTPUT: valid_blobs

[0129] Algorithm 3: sech 2 Multiscale radial fitting

[0130] FUNCTION sech2(x): RETURN 1 / cosh(x)^2

[0131] INPUT: labels_clarke(H,W), valid_blobs, energy_map(H,W), Δx=1, Δy=1, K={2,4,6,8,10,15,20}, L_set={k√(ΔxΔy)|k∈K}, ε=1e-6

[0132] blob_records ← []

[0133] FOR lb IN valid_blobs:

[0134] yy_all,xx_all ← where(labels_clarke == lb); cy,cx ← mean(yy_all),mean(xx_all)

[0135] keep ← isfinite(energy_map[yy_all,xx_all]); yy,xx,v ← yy_all[keep],xx_all[keep],energy_map[yy_all,xx_all][keep]

[0136] IF len(v)<3 OR max(v)-min(v)≤ε: blob_records.append((lb,0,"insufficient_energy")); CONTINUE

[0137] r ← sqrt(((yy-cy)×Δy) 2 +((xx-cx)×Δx) 2 ); vals ← (v-min(v)) / (max(v)-min(v))

[0138] best_r2 ← 0

[0139] FOR L IN L_set:

[0140] pred ← sech2(r / L) / max(sech2(r / L))

[0141] R2 ← 1 − Σ(vals-pred) 2 / max(Σ(vals-mean(vals)) 2 , ε)

[0142] best_r2 ← max(best_r2,R2)

[0143] blob_records.append((lb,max(0,best_r2),"ok"))

[0144] OUTPUT: blob_records

[0145]

Algorithm 4: Intra-batch Normalized Score Output

[0146] INPUT: labels_clarke(H,W), valid_mask(H,W), blob_records=[(lb,Q,status), ...], ε=1e-6

[0147] score_records ← [record for record in blob_records if record.status=="ok"]; score_map ← full(H,W,NaN); score_map[valid_mask] ← 0

[0148] IF len(score_records)==0: OUTPUT score_map,blob_records; STOP

[0149] q_all ← [Q for (lb,Q,status) in score_records]; q_min ← min(q_all);q_max ← max(q_all); denom ← q_max-q_min

[0150] FOR (lb,Q,status) IN score_records:

[0151] yy, xx ← where(labels_clarke == lb)

[0152] IF denom > 1e-6:

[0153] score_map[yy,xx] ← (Q-q_min) / denom

[0154] ELSE:

[0155] score_map[yy,xx] ← 0.5

[0156] OUTPUT: score_map[valid_mask]∈[0,1], score_map[¬valid_mask]=NaN,blob_records

[0157] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any changes made based on the design principles of the present invention, or any non-creative modifications made thereon, shall fall within the scope of protection of the present invention.

Claims

1. A method for evaluating the mineralization potential of geochemical anomaly patches based on soliton fitting, characterized in that, Includes the following steps: S10. Obtain a selected anomaly residual map of geochemical gridded data, perform Gaussian difference enhancement on the selected anomaly residual map and normalize it to a preset interval to obtain a selected anomaly map; S20. Perform high-resolution bright kernel patch detection on the selected anomaly map, extract candidate patches through morphological processing and connected component filtering, and output an integer labeled map. S30. Perform a three-branch fast discriminant filtering on the integer label map based on the crustal abundance Clarke value to remove background noise patches that are lower than the crustal background abundance level and output the Clarke-sensing label map. S40. Perform image border effect detection on each patch in the Clark perception label map, and exclude patches whose bounding box edge is less than the image boundary distance from the preset border margin from the mineralization potential evaluation to obtain a set of valid patches. S50. Perform soliton radial fitting on each of the effective soliton sets, extract the centroid of the soliton and calculate the radial distance and energy value, use a multi-scale soliton model to perform grid search fitting, and obtain the maximum goodness of fit of each soliton as the soliton morphology matching score Q. S60. Perform cross-patch global normalization on the soliton morphology matching scores of effective patches with a fitting status of ok, and convert the original Q value into an intra-batch relative mineralization potential score. Among them, the patch pixels excluded by Clarke discrimination, border effect detection or energy validity check are assigned a value of zero, and the pixels outside the survey area are retained as no data. S70. Save the relative mineralization potential score as a mineralization potential score map in a preset data format, and output a visualization file with an array of patch records for sorting of target areas in the same batch.

2. The method for evaluating the mineralization potential of geochemical anomaly patches based on soliton fitting according to claim 1, characterized in that, The process of obtaining the selected anomaly residual map of geochemical gridded data in step S10 includes: Geochemical gridded data is input into a two-stage progressive autoencoder, which includes a first-stage weakly pre-trained network and a second-stage refined network. The first-stage weak pre-trained network uses the Charbonnier loss function to learn the geochemical background laws; the second-stage refined network uses a multinomial loss function that includes soliton constraints and adaptive weight graphs to refine the reconstruction of mineral-induced anomaly regions. The residual direction is determined based on the output semantics of the second-stage refined network: when the output is the background reconstruction result, the original input data is subtracted from the output; when the output is the anomaly enhancement result, the original input data is subtracted from the output; the resulting positive anomaly residual is used as the selected anomaly residual map.

3. The method for evaluating the mineralization potential of geochemical anomaly patches based on soliton fitting according to claim 2, characterized in that, The process of performing high-resolution bright kernel patch detection on the selected anomaly map in step S20 includes: Only a limited number of pixels located within the effective measurement area are counted. A preset high proportion quantile is used as the threshold, and pixels not lower than the threshold are recorded as bright kernel pixels. The same four-connected structuring element is used to perform one morphological opening operation and one morphological closing operation on the brightness kernel binary map; after each operation, the invalid survey area pixels are reset to zero; Connected components in the processed image are extracted using the four-connectivity rule. Connected components with an area smaller than a preset pixel threshold are filtered out. The remaining connected components are then reassigned with consecutive integer labels to generate an integer label map. When there are no finite pixels in the effective measurement area, the integer label map is an all-zero map.

4. The method for evaluating the mineralization potential of geochemical anomaly patches based on soliton fitting according to claim 3, characterized in that, The process of performing a three-branch fast discriminant filtering on the integer label map based on the crustal abundance Clarke value in step S30 includes: Obtain standard crustal abundance data corresponding to the target element; denormalize the original geochemical concentration field to the same concentration unit as the Clark value, use only the limited concentration values ​​within the patch for discrimination, and use the product of the Clark value and the preset factor as the discrimination threshold; For each patch in the integer label map, perform the first discrimination: if there are no pixels in the patch below the discrimination threshold, then retain the patch and skip the mean calculation; If all pixels within a patch are below the discrimination threshold, then perform the second discrimination and delete the patch; If only some finite concentration values ​​within a patch are lower than the discrimination threshold, then the third discrimination is performed to calculate the mean of all finite concentration values ​​for that patch; if the mean is lower than the discrimination threshold, the patch is deleted; if the mean is not lower than the discrimination threshold, the patch is retained; if there are no finite concentration values ​​within a patch, the patch is deleted and a state of no data is recorded. The labels of the deleted patches are reset to zero, generating a Clark perception label map.

5. The method for evaluating the mineralization potential of geochemical anomaly patches based on soliton fitting according to claim 4, characterized in that, In step S40, the process of performing image border effect detection on each patch in the Clarke-sensory labeling map includes: Traverse each patch in the Clark perception labeling map and calculate the edge coordinates of the patch's bounding box in the vertical and horizontal directions; Determine whether the distance between the above edge coordinates and the corresponding boundary of the image is less than the border margin; if the distance of any edge is less than the border margin, then the patch is determined to be a border-truncated patch. The border-truncated patches are excluded from the Clark-aware label map to obtain a set of valid patches.

6. The method for evaluating the mineralization potential of geochemical anomaly patches based on soliton fitting according to claim 5, characterized in that, The process of performing soliton radial fitting on each of the valid patch sets in step S50 includes: Extract the centroid coordinates of the current patch and calculate the radial distance from each pixel in the patch to the centroid. After removing non-finite energy values, normalize the remaining energy values ​​to a preset range. If there are fewer than 3 finite energy values ​​or the range is not greater than the preset tolerance ε, set the soliton morphology matching score Q of the patch to zero. A set of scale parameters based on the soliton solution of the KdV nonlinear partial differential equation is constructed; the amplitude A0 = 1 is set for the normalized energy value, and the soliton radial template is expressed as: pred(r;L) = sech 2 (r / L) = 1 / cosh 2 (r / L), where L is the scale parameter and r is the radial distance; Calculate a predicted value for each L in the set of scale parameters, and normalize the sequence of predicted values; Calculate the goodness of fit R between the normalized predicted value and the normalized energy value. 2 = 1 - SSR / max(SST, ε), where SSR is the residual sum of squares, SST is the total sum of squares, and ε is a positive number to prevent the denominator from being zero; Corresponding each scale parameter to R 2 The maximum value is compared with zero, and the larger of the two values ​​is taken as the soliton morphological matching score Q of the patch.

7. The method for evaluating the mineralization potential of geochemical anomaly patches based on soliton fitting according to claim 6, characterized in that, The process of performing cross-patch global normalization in step S60 includes: Collect the soliton morphological matching scores Q of the effective patches with the fitted state of ok, and calculate the maximum value Qmax and the minimum value Qmin. When Qmax-Qmin is greater than the tolerance ε, the intra-batch relative mineralization potential score of the patch is calculated according to score = (Q-Qmin) / (Qmax-Qmin); When the number of scalable patches is zero, the effective survey area is assigned a value of zero and no data is retained outside the survey area; when the number of scalable patches is one, or Qmax-Qmin is not greater than ε, the relative mineralization potential score of the scalable patches is assigned a preset median value; patches whose fitting state is not OK are kept as zero.

8. The method for evaluating the mineralization potential of geochemical anomaly patches based on soliton fitting according to claim 7, characterized in that, The process of saving the relative mineralization potential score as a mineralization potential score map in a preset data format and outputting a visualization file with an array of patch records in step S70 includes: The relative mineralization potential score is stored as a single-precision floating-point matrix; the patch record array contains the patch number, Q value, and in-batch score annotation; the matrix and record array are written to a configurable final results directory, and the heat map serves as a visualization copy; The final results directory and the intermediate results directory generated during the evaluation process use different paths and are independent of each other; the reporting program only reads the final results directory.

9. The method for evaluating the mineralization potential of geochemical anomaly patches based on soliton fitting according to claim 8, characterized in that, When the two-stage energy map produced by the two-stage progressive autoencoder is unavailable, the selected anomaly map is directly input as an energy proxy field into step S50 for calculating the radial energy value.

10. The method for evaluating the mineralization potential of geochemical anomaly patches based on soliton fitting according to claim 9, characterized in that, This also includes applying the mineralization potential scoring map to the multi-element comprehensive mineral exploration target area ranking: For each of the multiple target elements, perform steps S10 to S70 respectively to obtain the corresponding single-element mineralization potential rating map; For spatially overlapping patches with consistent grids, the single-element mineralization potential score maps of each pixel are weighted and merged or the maximum value is taken to generate a multi-element comprehensive mineralization potential score map. A descending list of target areas is generated based on the numerical values ​​of the multi-element comprehensive mineralization potential scoring map. Each record in the list must specify at least the element combination, patch number, grid range, fusion method, and score.

Images