A sediment source tracing method based on multi-dimensional morphological characteristics of heavy minerals

Abstract

The present application belongs to the technical field of the cross of sedimentology and artificial intelligence, and particularly relates to a sediment source tracking method based on multi-dimensional morphological characteristics of heavy minerals. The present application establishes a "high roundness-fresh fracture" combined discrimination mechanism, accurately captures the key geological evidence of "highly round particle surface superimposed fresh fracture surface" through a double-head neural network, establishes a standard geological fingerprint logic library, and combines neural network training standardization "multi-cycle" heavy mineral characteristics, so as to improve the extraction depth of mineral morphological characteristics without relying on expensive geochemical tests, effectively distinguish the old sedimentary rock re-circulation components and the far-source direct transport components, realize the quantitative source tracking of sediment dynamics based on the population statistical law, and provide more objective, macroscopic and repeatable quantitative data support for basin analysis, paleogeographic reconstruction and oil and gas reservoir prediction.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of sedimentology and artificial intelligence, and particularly relates to a sediment source tracing method based on heavy mineral multi-dimensional morphological characteristics. BACKGROUND

[0002] Heavy minerals, as terrigenous detrital minerals with specific gravity greater than 2.85 in sedimentary rocks, are regarded as important tracers for recording geological history evolution due to their stable physical and chemical properties. In the fields of sedimentary basin analysis, oil and gas reservoir prediction, and paleogeographic environment reconstruction, heavy mineral source tracing is the core and basic work. By analyzing the combination characteristics, surface morphology and chemical composition of heavy minerals, geologists can trace the parent rock properties, transport path and environmental conditions during deposition of sediments.

[0003] For a long time, heavy mineral source tracing mainly relies on manual identification under a microscope and micro-area geochemical analysis. The traditional optical microscope identification method is the most classic basic method, which uses a polarizing microscope or a solid microscope to rely on naked eye observation and statistics of the color, crystal form, roundness and surface characteristics of particles. However, this method has significant limitations, and the results are extremely dependent on the personal experience and subjective judgment of the identifier. The data between different identifiers often lacks repeatability and comparability. In addition, the manual identification work is intensive and inefficient, which is difficult to meet the demand of modern geological exploration for high-throughput analysis of massive data, and cannot accurately quantify the micro-morphological characteristics.

[0004] With the development of analysis and testing technology, the micro-area geochemistry and isotopic chronology method based on laser ablation inductively coupled plasma mass spectrometry (LA-ICP-MS) (such as zircon U-Pb dating) has gradually become the mainstream high-precision method for source analysis. This method can accurately trace the age and type of parent rock by measuring the crystallization age or trace element fingerprint of the mineral. However, the geochemical method has a natural technical blind spot in solving complex sediment dynamics problems, especially in identifying “sedimentary multi-cycle” sedimentary processes. In the geological history, old sedimentary strata are often subjected to erosion after tectonic uplift, and the heavy mineral particles are transported and deposited again in new basins. For such multi-cycle origin heavy minerals, their geochemical properties and isotopic ages are completely consistent with the first cycle particles directly derived from the ancient crystalline bedrock, and chemical means cannot distinguish whether the mineral is “fresh component just stripped from the parent rock” or “residual component repeatedly transported in the geological history”. This lack of identification ability for multi-cycle processes can easily lead to misjudgment of near-source re-deposited source as far-source primary minerals, resulting in significant deviation in the reconstruction of paleocurrent system and paleogeographic pattern.

[0005] In recent years, some studies have begun to introduce digital image processing techniques to calculate the roundness, length-to-short-axis ratio and other geometric parameters of particles from conventional two-dimensional micrographs, in an attempt to quantify the transport distance. However, existing image analysis methods are mostly limited to geometric measurements of the external contour shape of particles, and seriously ignore the in-depth exploration of the micro-texture features of the particle surface. In actual geological processes, the key evidence for distinguishing "primary" and "reworked" sources, such as the "fresh fracture" or "crack" superimposed on the round particle surface, and the "micro-impact crater" indicating a specific sedimentary medium, are often manifested as surface texture structures rather than changes in edge contours. Conventional single imaging modes cannot effectively highlight and quantify these fine surface fingerprints, resulting in analysis results that simply rely on roundness data being unable to effectively distinguish the reworked components of ancient sedimentary rocks from the primary minerals of distant sources, and being difficult to meet the needs of fine geological research. In summary, the existing technology in the field of heavy mineral source tracing urgently needs an intelligent analysis method that can correctly recognize the morphology and texture of heavy minerals, effectively identify multiple cycles of sources, and objectively quantify the micro-fingerprint of the sedimentary environment. SUMMARY

[0006] The main purpose of the present application is to provide a sediment source tracing method based on the multi-dimensional morphological characteristics of heavy minerals, aiming to solve the technical problems in the prior art that the micro-zone geochemical method is difficult to identify multiple cycles of sedimentary processes, the traditional manual identification is low in efficiency and strong in subjectivity, and the conventional two-dimensional image analysis has projection errors and lacks micro-texture quantification. The present application establishes a quantitative mapping relationship between the micro-morphological fingerprint of heavy minerals and the macro-geological process, thereby realizing the automatic, high-throughput and objective intelligent tracing of the properties of the sediment source, the transport history and the sedimentary environment.

[0007] To achieve the above-mentioned purpose, the present application comprises the following steps:

[0008] S1, multi-modal two-dimensional digital imaging of heavy minerals: using a microscopic imaging device equipped with double light modes to obtain images of heavy mineral particles. The clear contour silhouette of the particles is collected by transmission light mode, the surface texture and color information of the particles are collected by ring lateral light mode, and the images under the two modes are spatially registered and associated to construct a multi-modal image dataset containing accurate geometric boundaries and surface micro-structures.

[0009] S2, quantitative extraction of multi-dimensional morphological fingerprint: feature extraction is performed on the multi-modal images. Based on the transmission light silhouette, the ratio of the average value of the curvature radius of each convex angle of the particle to the maximum inscribed circle radius is calculated using the Wadell roundness calculation logic, the roundness feature is extracted, and the roundness index is calculated; the ring lateral light texture image is used to identify the fracture, crack and micro-impact crater structure on the surface of the particle, and the color space mean value is extracted, thereby constructing a multi-dimensional feature vector containing roundness, fracture, crack, surface impact mark and apparent color.

[0010] S3, intelligent tracking based on double-head neural network: first, a "standardized geological fingerprint logic library" is constructed as an intelligent discrimination benchmark, and then the multi-dimensional feature vector is input into a pre-trained double-head neural network provenance analysis model. The first branch of the model is configured for provenance property discrimination, which distinguishes far-source primary transported provenance from paleosedimentary rock re-cycling provenance by joint analysis of roundness and fracture characteristics; the second branch of the model is configured for transport maturity inversion, which outputs the transport distance estimation value based on surface impact marks and apparent color characteristics of minerals.

[0011] S4, intelligent application and visualization of samples: based on the trained double-head neural network model, the samples to be tested are standardized and pre-processed, and the morphological parameters are automatically extracted; through the classification and regression branches of the network, the mineral, cycle, environmental attribute discrimination results of the particles and the transport maturity after de-normalization are output synchronously, and a four-dimensional geological attribute prediction vector of single particle level is constructed accordingly; finally, based on kernel density estimation, the transport maturity probability distribution curve is drawn, and a comprehensive geological interpretation report integrating micro-particle attribute tracing and macro-transport path evolution is automatically generated.

[0012] Further, in step S1, the collection logic of the double light mode is preferably: first, turn on the transmission bottom light source mode to obtain high-contrast particle silhouette; then switch to the ring-shaped side light source mode to collect surface images containing high-frequency texture information.

[0013] Further, in step S2, the fracture and crack feature extraction can use edge detection algorithm or texture analysis algorithm, which quantitatively represents the fragmentation degree of the particle surface by counting the proportion of high-frequency edge pixels or specific texture regions in the image.

[0014] Further, in step S3, the double-head neural network model uses a multi-task joint loss function for optimization in the training stage. Specifically, the provenance property discrimination branch can use a classification loss function, and the regression branch can use a regression loss function, and the model converges through weighted summation.

[0015] Further, in step S4, the generation of the intelligent provenance analysis report is based on the statistical aggregation of batch heavy mineral particle analysis results. The model not only outputs the independent discrimination results of single particles, but is also configured to statistically analyze all particle data captured in the same sample and visualize the results.

[0016] Compared with the prior art, the beneficial effects of the present application are:

[0017] 1. Solve the problem of identifying "multi-cycle" provenance in sedimentary geology. The invention innovatively establishes a "high roundness-fresh fracture" combined discrimination mechanism, uses a double-head neural network to accurately capture the key geological evidence of "highly rounded particle surface superimposed fresh fracture surface", establishes a standard geological fingerprint library, and combines neural network training to standardize "multi-cycle" heavy mineral characteristics, thereby effectively distinguishing old sedimentary rock re-circulation components from far-source direct transport components without relying on expensive geochemical tests, filling the gap in the identification of multi-cycle sedimentary processes in the prior art.

[0018] 2. Solve the contradiction between contour accuracy and surface texture in two-dimensional image analysis. The invention uses transmitted light silhouette to ensure the geometric accuracy of Wadell roundness calculation, and uses ring-shaped lateral light to highlight surface microscopic fractures and environmental impact marks, improving the extraction depth of mineral morphological characteristics.

[0019] 3. Realize quantitative provenance tracking based on group statistical law of sediment dynamics. The invention uses a deep learning model to realize high-throughput automated analysis of particles. This data processing method based on statistical large sample effectively eliminates the influence of single-particle accidental error on geological conclusions, providing more objective, macroscopic and repeatable quantitative data support for basin analysis, paleogeographic reconstruction and oil and gas reservoir prediction. BRIEF DESCRIPTION OF DRAWINGS

[0020] The invention will be further described in detail below in conjunction with the drawings and specific embodiments.

[0021] Figure 1 is a flowchart of the invention;

[0022] Figure 2 is a zircon morphological image in the invention;

[0023] Figure 3 is a schematic diagram of Wadell roundness calculation geometry in the invention;

[0024] Figure 4 is a schematic diagram of heavy mineral surface fracture and microscopic texture characteristics in the invention;

[0025] Figure 5 is a network architecture diagram of the double-head neural network model in the invention;

[0026] Figure 6 is a transport maturity probability distribution curve in the invention. DETAILED DESCRIPTION

[0027] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0028] like Figure 1 As shown, this invention provides a sediment provenance tracing method based on the multidimensional morphological characteristics of heavy minerals. The specific implementation steps of this method will be described in detail below in logical order:

[0029] S1, Multimodal Two-Dimensional Digital Imaging of Heavy Minerals

[0030] The core objective is to construct a high-quality image dataset that includes both precise geometric contours and rich surface textures, laying the foundation for subsequent multidimensional feature extraction. In the specific implementation process, firstly, the pre-processed heavy mineral samples are uniformly adhered to a transparent glass slide and placed on the automated stage of the microscopic imaging device. To avoid morphological misjudgment caused by particle stacking, it is necessary to ensure that the particles are distributed as discretely as possible in a single layer within the field of view. Subsequently, the image acquisition system initiates a preset automated temporal logic to perform time-division multimodal acquisition of the target particles within the same field of view.

[0031] The first stage of the acquisition process is the transmitted light contour acquisition mode. First, the transmitted bottom light source located below the stage is turned on, and the light intensity is adjusted to a high saturation state, so that the background of the microscopic field of view appears as a uniform, bright white light, while the solid, opaque heavy mineral particles appear as clear black silhouettes. At this point, a high-resolution industrial camera is triggered to perform the first exposure, acquiring a high-contrast transmitted light grayscale image. The significant advantage of this image is that it can completely eliminate the interference of particle surface color variations, transparency differences, or high-reflection points on the edge extraction algorithm, thus obtaining extremely sharp, continuous, and closed particle geometric boundaries. This high-quality contour data is a prerequisite for subsequent high-precision Wadell roundness calculations.

[0032] The second stage of the acquisition process is the ring-shaped side-light texture acquisition mode. After acquiring the first image, the stage position is kept strictly locked, the transmitted bottom light source is automatically turned off, and the low-angle ring-shaped side light source around the objective lens is simultaneously turned on. Under side-lighting conditions, the light passes over the particle surface at an oblique angle. Utilizing the principle of light and shadow imaging, the microscopic pits, impact scratches, and fresh fractures with sharp edges on the particle surface form obvious shadow contrasts, thereby significantly enhancing the visibility of high-frequency texture information. Figure 2 At this point, the camera is triggered to perform a second exposure, magnifying the texture details of the fracture crack. Figure 3The process involves obtaining surface morphology images containing rich texture details and true color information. Finally, image processing algorithms are used to perform strict spatial registration and index association based on pixel coordinates on the images of the same particle acquired under the two lighting modes mentioned above, generating a particle-specific "contour-texture" multimodal image pair, which is then stored in the original image database for use in subsequent steps.

[0033] S2, Quantitative Extraction of Multidimensional Morphological Fingerprints

[0034] Extraction of S21 roundness

[0035] After image acquisition, this embodiment first preprocesses and extracts geometric features from the transmitted light silhouette image obtained in step S1. First, the optimal global threshold is automatically calculated using Otsu's maximum inter-class variance method. To ensure the adaptability of the segmentation, the algorithm iterates through all possible gray levels. Calculate the inter-class variance between foreground and background Select the one with the largest variance The optimal threshold is calculated using the following formula:

[0036]

[0037] Where ω0 and ω1 represent the occurrence probabilities of background and foreground pixels, respectively, and µ0 and µ1 represent the corresponding average gray levels. After converting the grayscale image into a binary image using this threshold, morphological closing operations are applied to the binary image—first dilation, then erosion—to eliminate the hole noise caused by transparent mineral inclusions and obtain a solid connected component mask. Subsequently, the Suzuki contour tracking algorithm is used to extract the external boundary coordinate sequence of the particles. Its mathematical expression is:

[0038]

[0039] Where N is the total number of contour pixels. x i , y i For the first i The pixel coordinates of the edge points.

[0040] After obtaining the precise contour coordinates, this embodiment executes the core Wadell roundness calculation program. First, it calculates the maximum inscribed circle radius of the particle. R The system performs a Euclidean distance transformation on the solid binary mask image, calculates the straight-line distance from each foreground pixel to the nearest background pixel, and generates a distance map. The global maximum value in this map is the distance to the nearest background pixel. R Next, the radius of curvature of each convex corner of the particle is calculated. A convex hull algorithm is used to identify all convex corner vertices on the contour. For each identified convex corner...j Extract local contour segments near its vertices (including...) K (Number of coordinate points), and use the least squares method to fit the best approximation circle. This step aims to minimize the fitting error. The objective function is defined as:

[0041]

[0042] in,( x k , y k ) represents the coordinates of a local contour point. a j , b j () is the center of the fitted circle. r j For the first j The system iterates through the entire contour, filtering out all radii of curvature smaller than the radius of the largest inscribed circle. R Let the total number of these effective convex angles be denoted as . M ,like Figure 4 As shown.

[0043] Finally, the geometric roundness is calculated according to the mathematical formula defined by Wadell. p w This formula is defined as the ratio of the arithmetic mean of the curvature radii of all effective convex angles to the radius of the largest inscribed circle, and its calculation formula is expressed as follows:

[0044]

[0045] in, r j For the first j The radius of curvature of each convex angle M The number of effective convex angles, R The radius of the largest inscribed circle. The calculated value is... p w The value ranges from 0 to 1. The closer the value is to 1, the rounder the particles are, and the closer the value is to 0, the more angular the particles are. This calculation process is entirely based on geometric mathematical derivation, eliminating the subjective error of manual estimation.

[0046] S22 Gradient-based fracture density and texture contrast calculation

[0047] While performing geometric contour calculations under transmitted light, the reflected light image obtained in step S1 is first processed to extract physical structure features. The color image is first converted to a single-channel grayscale image, and for any pixel ( u , v ), its grayscale value I (u , v The calculation uses a weighted average method:

[0048]

[0049] in, R ( u , v ) is a pixel ( u , v The red channel component value; G ( u , v ) is a pixel ( u , v The green channel component value; B ( u , v ) is a pixel ( u , v The blue channel component value;

[0050] Subsequently, the Sobel operator is used to perform convolution operations on the image to extract the horizontal gradient. G x and vertical gradient G y The Sobel convolution kernel is defined as follows:

[0051]

[0052] Based on the above two components G x and G y Calculate the total gradient magnitude for each pixel. :

[0053]

[0054] The system sets an adaptive threshold. T g The total number of edge pixels whose gradient magnitude is greater than the threshold is counted. And calculate the fracture density. D frac Its formula is:

[0055]

[0056] in, Q total This refers to the total number of pixels within the image area containing heavy mineral particles.

[0057] Furthermore, for the identification of surface micro-impact craters, a gray-level co-occurrence matrix is ​​constructed, and texture contrast is calculated. Ccon to characterize the roughness degree of the texture:

[0058]

[0059] wherein, L is the total number of gray levels after quantization of the image; h is the row number of the gray co-occurrence matrix; k is the column number of the gray co-occurrence matrix; P h k is the normalized co-occurrence matrix probability.

[0060] Based on the probability distribution, the Haralick texture feature extraction algorithm is used to calculate the contrast representing the depth change of the texture gully, the energy measuring the uniformity of the image gray distribution, the entropy representing the confusion degree of the texture, and the inverse difference moment describing the local texture uniformity. Finally, in order to retain the original physical properties of the multi-dimensional features for the neural network to mine the nonlinear relationship, the above texture statistical features and the geometric features of the fracture density calculated in the previous step are spliced into a high-dimensional surface micro-feature vector V surface , which will be used as the input layer data of the deep neural network model in the subsequent steps.

[0061]

[0062] wherein D frac is the fracture density, C con is the texture contrast, Energy is the energy measuring the uniformity of the image gray distribution, Entropy is the entropy representing the confusion degree of the texture, Homogeneity is the inverse difference moment describing the local texture uniformity.

[0063] S23 apparent color feature extraction based on HSV space

[0064] After extracting the physical texture features, further color analysis is performed on the reflected light image to quantify the chemical oxidation degree of the particles. The original RGB image is converted to the HSV color space, and the saturation S sat is defined , and the saturation calculation formula is:

[0065]

[0066] The system calculates the arithmetic mean of the saturation of all pixels in the particle region ​​, which is directly related to the thickness and continuity of the particle surface oxide film, as a key chemical indicator to determine whether it has experienced long-term surface exposure or sedimentation interruption.

[0067] Finally, in order to realize the deep fusion of multi-modal features, the system cascades the sphericity p w , the surface micro-feature vector generated by S22 V surface , and the apparent color feature extracted in this step. In this way, a multi-dimensional feature vector containing geometric sphericity, fracture density, texture contrast and apparent color information is constructed V com , which will serve as the standard input data for the double-head neural network model in the subsequent S3 step.

[0068] S3, Intelligent source tracking based on double-head neural network

[0069] S31 Construction of standardized geological fingerprint logic library

[0070] Step S3 aims to build an intelligent sediment source analysis model using deep learning technology. Before executing the analysis process for unknown samples, a "standardized geological fingerprint logic library" is first constructed as an intelligent discrimination benchmark. The establishment of this logic library aims to convert abstract geological deposition rules into computer-readable digital mapping relationships, and its construction process follows the two principles of "double end-member control" and "typical environment constraint".

[0071] First, construct a physical sample sub-library. This embodiment selects zircon, monazite, apatite, rutile, and sphene as five typical heavy minerals with significant differences in weathering resistance as standard end-members: zircon and rutile represent "high stability end-members" for recording long-term evolution information of multiple cycles and ancient source areas; apatite, monazite and sphene represent "low stability end-members" for recording single transport and recent hydrodynamic modification information. For the above-mentioned minerals, standard samples are collected in four typical environments with known geological backgrounds: (1) near-source primary samples (collected from river mouths of granite body disintegration, representing zero-distance transport); (2) far-source transport samples (collected from the lower reaches of long rivers, representing long-distance water rounding); (3) multiple cycle samples (collected from ancient sandstone strata, representing re-deposited components); (4) aeolian samples (collected from desert dunes, representing air medium impact).

[0072] Second, construct a digital feature vector sub-library. Use the multi-modal imaging technology described in step S1 and the feature extraction algorithm described in step S2 to perform full-scan and calculation on the above standard physical samples. Each standard particle is converted into a multi-dimensional feature vector containing geometric sphericity p w , fracture density Dfrac , texture contrast C con and multi-dimensional feature vector of apparent color information V com In order to establish the logical mapping relationship, senior sedimentologists are invited to manually check the images, and senior sedimentologists are invited to build a four-level multi-dimensional geological label system:

[0073] (1) First-level label - mineral phase ( L min ), calibrate the types of granular minerals ( L min = {zircon, apatite, rutile, …}), used to correct the anti-erosion difference coefficient of different minerals;

[0074] (2) Second-level label - genetic phase ( L cyc ), determine the particle cycle attribute ( L cyc = {primary / no cycle, multi-cycle / redeposition}), used to eliminate the interference of inherited rounding characteristics;

[0075] (3) Third-level label - sedimentary facies ( L env ), identify the dominant transport medium and rough environment ( L env = {water-near source, water-remote source, aeolian dune}), assist in analyzing the causes of surface microtexture; among them, "aeolian" corresponds to particles with specific surface disc-shaped pits, and "water" is further divided into near-source and remote-source according to the degree of rounding;

[0076] (4) Four-level label - transport facies ( L dist ), used to semi-quantitatively represent the transport maturity of particles, and defines a dimensionless transport index with a value range of [0, 1]. Among them, the value of the near-source / primary standard sample I trans is distributed in the interval [0, 0.3] to represent low maturity, the medium transport distance sample is distributed in the transition interval [0.3, 0.7], and the remote / long distance transport standard sample is distributed in the interval [0.7, 1.0] to represent high maturity; This index label will be used as the supervised true value of the neural network regression branch, aiming to train the model to learn the continuous evolution rule from "angular" to "highly rounded", rather than predicting the absolute physical distance.

[0077] Finally, the above data is stored by using vector database technology, and a logical index based on cosine similarity is established. For any two standard vectors A and B in the logical library, the logical similarity of A and B is defined as S sim As follows:

[0078]

[0079] The significance of constructing the logical index lies in that, on the one hand, the high-dimensional topological structure of the vector space is utilized to realize fast nearest neighbor retrieval of a large number of standard particles; on the other hand, when the subsequent neural network outputs a low confidence prediction, the “abnormal outlier particle” deviating from the standard geological cluster can be identified by comparing and retrieving the similarity index, so as to ensure that the provenance discrimination process has both statistical regularity and interpretability of petrology physical significance.

[0080] In order to further improve the robustness of the model to particle rotation, illumination change and scale difference, the embodiment performs an online data enhancement strategy before inputting the image into the network. Since the orientation of the heavy mineral particle on the object table is random, and the sedimentology attribute has rotation invariance, a random geometric transformation is performed on the image in the original training sample set Ω train , the original image is defined as , the enhanced image is , and the transformation operation set Φ includes random rotation , random horizontal flip and random scaling factor . The mathematical expression of the enhancement operation is:

[0081]

[0082] Through this step, the effective size of the training data is artificially expanded by 10-20 times, the network is forced to learn the morphological features of the particle essence rather than the accidental pixel position features, thereby effectively preventing the model from overfitting on limited geological samples.

[0083] S32 construction and training based on deep convolutional neural network

[0084] After obtaining the standard data set, the embodiment constructs a deep convolutional neural network (CNN) based on a transfer learning architecture. In order to adapt to the network input requirements, the multi-modal images obtained in step S1 are first uniformly adjusted to a fixed size of 224x224 pixels, and normalized. The network backbone selects a pre-trained ResNet50 on ImageNet to extract deep abstract features of the image. At the same time, in order to fuse the explicit physical features calculated in step S2, the network introduces a feature concatenation layer to concatenate the image feature vector extracted by ResNet with the geometric / texture numerical feature vector Cascade fusion is performed. Specifically, a 2048-dimensional deep feature vector output by the ResNet50 global average pooling layer is extracted, and the numerical feature vector A concatenation operation is performed in the feature channel dimension. The fused multi-modal feature vector is then connected to two parallel task branches to form a "double-headed" topology, and the overall structure is as shown in Figure 5 .

[0085] To optimize the model parameters, the embodiment designs a multi-task joint loss function. The first branch is a source attribute classification branch, which is used to output the geological attribute probability of the particle ("primary / regenerated" cycle attribute or mineral species). This branch uses a multi-class cross-entropy loss function for supervised training.

[0086] Suppose there are N b samples in a Batch during training, and the total number of categories is M . For the i th sample, its true category label is y i,c , the probability distribution predicted by the model classification branch is , and the total number of categories is C , then the classification loss L cls is calculated as follows:

[0087]

[0088] The second branch is a transport maturity regression branch, which aims to quantitatively invert the equivalent transport maturity index of particles from the source area to the deposition area based on the extracted multi-modal features. This branch uses a mean square error loss function for supervised training to minimize the deviation between the predicted maturity and the true label. Suppose the true transport index label of the i th sample is I i (0-1), and the predicted value output by the model regression branch is , then the regression loss L reg is calculated as follows:

[0089]

[0090] To optimize both the source attribute classification and the transport maturity prediction tasks, the embodiment constructs a multi-task joint total loss function L total . By introducing balance hyperparameters λ1 and λ2, the weights of the two task branches in gradient backpropagation are dynamically adjusted to prevent one task from dominating the training process. The total loss function is defined as follows:

[0091]

[0092] After establishing the multi-task joint loss function, the embodiment adopts the gradient-based error back propagation algorithm combined with the Adam adaptive moment estimation algorithm to iteratively update the network parameters. During the initialization of the training process, the data set is divided into several batches. For the first iteration, the system first performs forward propagation to calculate the loss value, and then calculates the total loss function L total The gradient vector of all weight parameters θ in the network is g t The calculation formula of the gradient is:

[0093]

[0094] Wherein θ represents the partial derivative operator of the parameter. In order to overcome the defects that the traditional stochastic gradient descent method is easy to fall into local optimal solution or oscillation in the later training period, the embodiment introduces the first-order moment estimation m t and the second-order moment estimation v t to smooth the gradient. The iterative update formulas of the two moment variables are as follows:

[0095]

[0096]

[0097] Wherein, β1 and β2 are the exponential decay rates of the first and second moments, respectively, represents the element-wise square of the gradient. In order to eliminate the influence of the bias in the early training, the bias-corrected moment estimation values and are further calculated. Finally, the network parameters θ are updated according to the calculated moment estimation values, and the update rule is as follows:

[0098]

[0099] Wherein, η is the global learning rate, is the smoothing term to prevent the denominator from being zero. By repeatedly executing the above "forward calculation-backward derivation-parameter update" loop until the loss function converges, the model parameters are continuously optimized.

[0100] After each round of training iteration, the reserved validation set is used to monitor the model performance to determine the best model parameter freezing time. For the classification branch, the F 1 score is used as a comprehensive evaluation index to balance the precision and recall. Let TP ​This represents the number of true positive samples (i.e., correctly identified feature particles). FP This represents the number of false positive samples. FN If the number of false negative samples is [not specified], then [then the number of false negative samples is specified]. F The formula for calculating a fraction is:

[0101]

[0102] For the regression branch, the coefficient of determination is used. R 2 The degree of fit between predicted and actual values. Let the validation set contain... N val One sample, If the actual transport maturity is the average value, then R 2 The calculation formula is:

[0103]

[0104] The system is configured with an early stop mechanism, which is triggered when the validation set... The score exceeds the preset threshold and The model is considered converged when the preset threshold is exceeded and the metric no longer shows significant improvement within five consecutive periods. At this point, the current weight parameters θ are automatically saved. best The network was then frozen, marking the end of the model training phase and signifying that the model now possessed the practical capability to process unknown geological samples.

[0105] S4. Intelligent Application and Visualization of Samples

[0106] In this embodiment, after completing model training and parameter freezing in step S3, step S4 is executed. This step aims to utilize the trained dual-head neural network model to perform fully automated source property analysis and quantitative tracing of unknown geological samples collected in the field, following the analytical logic of "from single-particle micro-quantification to population macro-statistics".

[0107] Data standardization preprocessing of S41 test samples

[0108] First, multimodal imaging and feature extraction are performed on the unknown particles to be tested, identical to steps S1-S2, to generate the feature vector to be tested. To eliminate dimensional differences and adapt to the input distribution of the neural network, the mean and standard deviation obtained from the training set in step S3 are used to perform Z-score standardization on the feature vector to be tested. The standardization calculation formula is as follows:

[0109]

[0110] In the formula, Indicates the number of samples to be tested The standardized values ​​of the 3D features xj The original sample is the first j dimensional feature extraction value; μ j and σ j The global statistical mean and standard deviation of the training data set in step S3 must be strictly used on the first j dimensional feature. This constraint ensures that the feature space distribution benchmark of the test data and the model training data is consistent, which is the premise of ensuring the reasoning accuracy of the model.

[0111] S42 inference and calculation of double-headed neural network

[0112] The preprocessed test sample batch is input into the double-headed neural network model with frozen weights in step S3, and the output results of the classification branch and the regression branch are obtained respectively by using the forward propagation algorithm.

[0113] (1) The output of the classification branch is mapped to a probability distribution by the Softmax function, and the index with the maximum probability is selected as the geological property category of the particle.

[0114] (2) The regression branch directly outputs a scalar value between 0 and 1, which is the model's predicted transport index , representing the relative transport maturity of the particle from the source area to the deposition area, without the need for physical distance reverse normalization conversion.

[0115] S43 constructing a single particle's source prediction vector

[0116] Instead of directly performing simple statistical averaging on the data, the system first constructs a dedicated source prediction vector for each independent particle in the field of view ;

[0117] This vector is a key data structure that connects micro-prediction and macro-interpretation, and its content strictly corresponds to the four-level geological label system defined in step S3. The mathematical structure is expressed as:

[0118]

[0119] In the formula, the four components of the vector respectively represent the comprehensive discrimination results of the model for the first particle:

[0120] is the mineral phase category (such as "zircon", "apatite", etc.) identified by the model;

[0121] is the cycle attribute (such as "primary non-cyclic" or "multi-cycle redeposition") identified by the model;

[0122] The sedimentary transport environment inferred by the model (such as "aquatic-proximal" or "aeolian").

[0123] This is the dimensionless transport index for model regression inversion.

[0124] The system will extract the sample N s The vector convergence of individual particles generates a holographic geological attribute matrix for the sample, serving as the data foundation for subsequent geological tracing.

[0125] S44 Material Transport Maturity Probability Distribution Curve Plotting

[0126] For the transport maturity index sequence output by the model ;

[0127] Kernel density estimation (KDE) was used to plot continuous probability distribution curves of sediment transport maturity to identify major river convergence patterns and source characteristics within the watershed. This was done within the transport index space. probability density function at The estimation formula is:

[0128]

[0129] In the formula, N s This represents the total number of effective particles in the sample to be tested. h The bandwidth parameter is used to control the smoothness of the curve. In this embodiment, the Scott rule is used to adaptively determine the optimal bandwidth. K ( ) represents the kernel function. In this embodiment, the standard Gaussian kernel function is selected, and its specific form is as follows:

[0130]

[0131] Generation of S45 Source Prediction Visualization Report

[0132] Finally, based on the "point-to-surface" analysis results, a comprehensive geological interpretation report is automatically generated: on the one hand, it lists the data of each particle in the form of a data table. Vectors enable traceability and outlier detection at the micro-level; on the other hand, they are used to visualize data through graphs. The evolution curve, with the horizontal axis position of the peak indicating the transport maturity characteristics of the sediment: if the main peak is located in the interval [0, 0.3], it indicates rapid near-source deposition; if the main peak is located in the interval [0.7, 1.0], it indicates long-distance transport from a distant source; if the main peak is located in the interval [0.3, 0.7], it indicates medium-distance transport. Figure 6 Bimodal or multimodal morphology reveals a complex sedimentary process within the basin involving the mixed supply of sediments at different maturity levels.

[0133] The above examples are only used to illustrate the technical solutions of the present application, but not to limit the present application; although the present application has been described in detail with reference to the foregoing examples, it should be understood by those skilled in the art that the technical solutions recorded in the foregoing examples can be modified, or some or all of the technical features can be replaced equivalently; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application, and they should be covered within the protection scope of the claims and the specification of the present application.

Claims

1. A sedimentary provenance tracing method based on the multidimensional morphological characteristics of heavy minerals, characterized in that, Includes the following steps: S1. Multimodal two-dimensional digital imaging of heavy minerals: Images of heavy mineral particles are acquired using a microscopic imaging device equipped with dual illumination modes. The images were captured in a transmitted light mode to obtain clear outline silhouettes of particles, and in a ring side light mode to obtain surface texture and color information of particles. The images in the two modes were spatially registered and correlated to construct a multimodal image dataset containing precise geometric boundaries and surface microstructures. S2. Quantitative extraction of multidimensional morphological fingerprints: Feature extraction is performed on the multimodal image; based on transmitted light silhouettes, Wadell roundness calculation logic is used to calculate the ratio of the average radius of curvature of each convex corner of the particle to the radius of the largest inscribed circle, extract roundness features, and calculate the roundness index. P w Based on the annular lateral light texture image, the structure of fracture, crack and micro impact pit on the particle surface is identified, and the color space mean is extracted to construct a multi-dimensional feature vector containing geometric roundness, fracture density, texture contrast and apparent color information. S3. Intelligent tracking based on dual-head neural network: First, a standardized geological fingerprint logic library is constructed as the intelligent discrimination benchmark. Then, the multi-dimensional feature vector is input into the pre-trained dual-head neural network source analysis model. The first branch of the model is configured for source property discrimination. By jointly analyzing roundness and fracture characteristics, the model distinguishes between distant primary transported sources and ancient sedimentary recycle sources. The second branch of the model is configured for transport maturity inversion, which outputs an estimate of transport distance based on surface impact marks and mineral apparent color characteristics. S4. Intelligent application and visualization of samples: Based on the trained dual-head neural network model, the sample to be tested is subjected to standardized preprocessing and automated extraction of morphological parameters; through the classification and regression branches of the network, the mineral, cycle, and environmental attribute discrimination results of the particles and the transport maturity after inverse normalization are output simultaneously, and a four-dimensional geological attribute prediction vector at the single-particle level is constructed accordingly; finally, the transport maturity probability distribution curve is plotted based on kernel density estimation, and a comprehensive geological interpretation report integrating microscopic particle attribute tracing and macroscopic transport path evolution is automatically generated.

2. The sediment provenance tracing method based on the multidimensional morphological characteristics of heavy minerals as described in claim 1, characterized in that, The roundness index mentioned in step S2 P w The calculation formula is: in, r j For the first j The radius of curvature of each convex angle; M The number of effective convex angles; R The radius of the largest inscribed circle; the calculated value is... p w The value is between 0 and 1. The closer the value is to 1, the rounder the particles are, and the closer the value is to 0, the more distinct the edges of the particles are.

3. The sedimentary provenance tracing method based on the multidimensional morphological characteristics of heavy minerals according to claim 1, characterized in that, In step S2, the specific process for calculating fracture density and texture contrast is as follows: While performing geometric contour calculations under transmitted light, the reflected light image obtained in step S1 is first processed to extract physical structure features. The color image is first converted to a single-channel grayscale image. For any pixel ( u , v ), its grayscale value I ( u , v The calculation uses a weighted average method: in, R ( u , v ) is a pixel ( u , v The red channel component value; G ( u , v ) is a pixel ( u , v The green channel component value; B ( u , v ) is a pixel ( u , v The blue channel component value; Subsequently, the Sobel operator is used to perform convolution operations on the image to extract the horizontal gradient. G x and vertical gradient G y The Sobel convolution kernel is defined as follows: Based on the above two components G x and G y Calculate the total gradient magnitude for each pixel. : The system sets an adaptive threshold. T g The total number of edge pixels whose gradient magnitude is greater than the threshold is counted. And calculate the fracture density. D frac Its formula is: in, Q total This refers to the total number of pixels within the image area containing heavy mineral particles; Furthermore, for the identification of surface micro-impact craters, a gray-level co-occurrence matrix is ​​constructed, and texture contrast is calculated. C con To characterize the roughness of the texture: in, L It is the total number of gray levels after image quantization; h These are the row numbers of the gray-level co-occurrence matrix; k These are the column numbers of the gray-level co-occurrence matrix; P ( h , k ) represents the normalized co-occurrence matrix probability.

4. The sedimentary provenance tracing method based on the multidimensional morphological characteristics of heavy minerals according to claim 1, characterized in that, In step S2, the method for calculating the apparent color information is as follows: performing color analysis on the reflected light image to quantify the degree of chemical oxidation of the particles; converting the original RGB image to the HSV color space, and focusing on extracting the saturation that can indicate the degree of hematite staining. S sat ,definition ,but S sat The formula for calculating saturation is: 。 5. The sedimentary provenance tracing method based on the multidimensional morphological characteristics of heavy minerals according to claim 1, characterized in that, In step S3, the specific steps for constructing the standardized geological fingerprint logical library are as follows: First, a physical sample sub-library is constructed, selecting typical heavy minerals with significant differences in weathering tolerance as standard endmembers, and collecting standard samples in typical environments with known geological backgrounds; second, a digital feature vector sub-library is constructed, based on S1 and S2, performing a full scan and calculation on the collected standard samples, converting each standard particle into a vector containing geometric roundness. p w fracture density D frac Texture contrast C con and the multidimensional feature vector of apparent color information V com The images were manually verified to construct a four-level multidimensional geological labeling system: the first-level label is the mineral phase, which identifies the types of grain minerals and is used to correct the difference coefficient of abrasion resistance of different minerals; the second-level label is the genetic phase, which determines the cyclic properties of grains and is used to eliminate the interference of inherited rounding features. The third-level labeled sedimentary phases identify the dominant transport medium and coarse environment, aiding in the analysis of the origin of surface microtexture; the fourth-level labeled transport phases are used to semi-quantitatively characterize the transport maturity of particles; finally, vector database technology is used to store the above data.

Images