A progressive method for generating three-dimensional pore structure of soil aggregates

Abstract

A progressive generation method for the three-dimensional pore structure of soil aggregates. The invention belongs to the field of information-based image processing technology, and specifically to the field of soil aggregate structural feature analysis technology. The invention solves the problem that the existing generative adversarial network cannot retain the distribution law of the original soil aggregate pore characteristics. In view of the problem that the existing generative model cannot retain the distribution law of the original soil aggregate pore characteristics, the generative adversarial network framework is adaptively improved, which can automatically and quickly generate a large number of relatively clear, visually realistic and diverse three-dimensional soil aggregate pore structures; the one-time generation process of the original generative adversarial network is changed to a progressive generation process, which can capture the overall distribution characteristics of the pore structure at a specific resolution, rather than just focusing on the specific detailed characteristics of a single sub-sample, thereby achieving a better generation effect.

Description

Technical Field

[0001] The invention belongs to the technical field of information-based image processing, and in particular to the technical field of soil aggregate structure characteristic analysis. Background Art

[0002] Soil supports about 95% of global agricultural production. It is an important resource for human survival and a key component of the earth's ecosystem. Soil quality is of great significance to maintaining biodiversity, stabilizing agricultural production and protecting the environment. Soil quality assessment needs to be based on baselines or reference values ​​in order to accurately determine the effectiveness of management measures. Generally, soil responds slowly to changes in land use and management methods, and it is often more complicated to monitor changes in soil quality before irreversible damage occurs. Soil quality assessment requires the determination of some soil sensitive properties that can reflect the ecological service capacity of soil. Soil structure is a key factor in assessing soil quality and can be used as an indicator to evaluate soil quality. Soil structure plays a vital role in regulating soil ecological functions. It directly affects water infiltration and retention, gas exchange, soil organic matter and nutrient cycling, and soil microbial diversity and activity. In addition, soil structure is affected by many factors, including environmental factors (such as parent material, topography, climate), biological factors (such as soil animals, plant roots, microbial activities) and human factors (such as tillage and mechanical compaction).

[0003] As the basic unit of soil structure, soil aggregates play an important role in regulating the physical, chemical and biological processes of soil, and their stability is an important indicator of the quality of soil structure. There is a very complex pore system inside soil aggregates, which directly affects the porosity and connectivity of the soil, determines the movement and exchange capacity of soil water, gas and nutrients, as well as the various physical, chemical and biological processes occurring inside the aggregates, and regulates various functions of the soil. A good soil aggregate structure can promote root development, enhance the effective use of water and nutrients, and thus improve soil quality and maintain soil health. Therefore, conducting soil quality-related analysis based on soil aggregate structure is of great significance to maintaining ecosystem health and improving agricultural productivity.

[0004] At present, the traditional method of obtaining the three-dimensional pore structure characteristic images of soil aggregates cannot directly observe soil aggregates. Since soil aggregates are usually located below the soil surface, have complex structural characteristics and are affected by many factors, direct observation and image acquisition become difficult. Existing studies extract color, geometry, texture and other features from CT images, and then establish classification or regression models with dependent variables, and have achieved certain results, but this mode of artificial design and feature extraction has certain limitations and limited generalization ability.

[0005] Advanced artificial intelligence data mining technology based on machine learning can identify potential distribution patterns in data sets, which is conducive to in-depth exploration of potential laws and connections in soil environmental data, and brings new perspectives to soil environmental science research. At present, machine learning has been widely used in soil classification, soil property prediction, digital soil mapping, soil pollution monitoring and soil quality (health) evaluation, and has achieved good results. In recent years, with the significant improvement of computer computing power and the continuous updating of algorithms and theories, the application cost and difficulty of machine learning have continued to decrease, and the accuracy of model methods has continued to improve, which is conducive to the application of machine learning methods in soil disciplines to carry out more in-depth and extensive research. Deep learning is an important branch of machine learning and a hot topic in the field of artificial intelligence. As a complex machine learning algorithm, deep learning aims to automatically extract multi-layer feature representations from data through a series of nonlinear transformations, relying on nonlinear information processing mechanisms, and using a combination of supervised and unsupervised training methods to extract and transform feature structures and data, and further realize the fitting of complex data relationships of samples. At present, it has been widely used in the monitoring and prediction of soil temperature and humidity, soil organic carbon, soil nutrients, etc. The deep cross-integration of machine learning methods and soil science can significantly improve the ability to analyze and interpret soil data, and improve the scientificity and accuracy of soil property prediction and soil management decisions. This interdisciplinary research method provides strong data support and methodological guarantees for revealing the complex relationship between soil properties and environmental factors, optimizing soil management practices, and supporting sustainable agricultural development.

[0006] In order to solve the problem that a large amount of data is required for model training, generative adversarial networks came into being. Generative adversarial networks (GANs) have achieved great success in image generation. Through adversarial training of generative models and discriminative models, the distribution of training data is learned, and potential information in the data is mined to generate new images with similar features. However, the original GAN ​​has disadvantages such as unstable training and low resolution of generated images. In recent years, many improved models have been proposed. Among them, 3D-porous-GAN is used to reconstruct digital cores from a single 3D image. This method successfully simulates and reconstructs complex 3D rock microstructures. E-GAN introduces evolutionary learning in the GAN generator, converts the adversarial training process into an evolutionary problem, and improves the stability of the training process and the performance of the generator.

[0007] Generating the three-dimensional pore structure of soil aggregates can learn the distribution of real soil data and generate a more realistic soil structure. It helps to explore the diversity and characteristic distribution of soil structure, including different pore shapes, size distribution, connectivity, etc., and deepen the understanding of the complexity of soil structure. Many studies have shown that generative adversarial networks have made significant progress in the field of rock and porous media microstructure generation, but the existing generative adversarial networks cannot retain the distribution law of the original soil aggregate pore characteristics. Summary of the invention

[0008] In order to solve the problem that the existing generative adversarial network cannot retain the distribution law of the original soil aggregate pore characteristics, the present invention proposes a progressive generation method of the three-dimensional pore structure of soil aggregates, which includes the following steps:

[0009] S1, preprocessing the original soil aggregate CT scan image data, separating the pore part in the image, and constructing a three-dimensional pore structure voxel matrix data set of soil aggregates;

[0010] S2, performing random sampling of the soil aggregate three-dimensional pore structure dataset to obtain the soil aggregate three-dimensional pore substructure dataset;

[0011] S3, downsampling the three-dimensional pore substructure in the soil aggregate three-dimensional pore substructure dataset to obtain a soil aggregate three-dimensional pore substructure downsampled dataset;

[0012] S4. Construct a progressive reconstruction generative adversarial network model, which is divided into a generative adversarial network branch and a reconstruction network branch;

[0013] S5. Training a progressive reconstruction generative adversarial network model based on a soil aggregate three-dimensional pore substructure dataset to obtain an improved progressive reconstruction generative adversarial network model, and using the improved progressive reconstruction generative adversarial network model to generate the three-dimensional pore structure of soil aggregates.

[0014] Furthermore, the step S1 is specifically as follows: intercepting a crack-free area in the original soil aggregate CT scan image as a feature domain image, performing normalization processing, noise elimination processing, and smoothing processing on the intercepted feature domain image in sequence, and then performing threshold segmentation on the processed feature domain image to separate the pore part in the feature domain image, generate a binary image, and convert the binary image into a voxel matrix, which is a voxel matrix data set of the three-dimensional pore structure of soil aggregates.

[0015] Further, in step S2, when random sampling is performed on the soil aggregate three-dimensional pore structure data set, according to The voxel matrix specification is used to extract the soil aggregate 3D pore structure dataset, where The length, width and height of the voxel matrix are all n.

[0016] Further, adopt h The three-dimensional pore substructure is downsampled by different sampling multiples to obtain the downsampled dataset of the three-dimensional pore substructure of soil aggregates. ,in, They represent data sets corresponding to different sampling multiples.

[0017] Further, in step S4, the generative adversarial network branch includes a generative model and discriminant model ;

[0018] The target loss function of the generative adversarial network branch is:

[0019]

[0020] In the formula, To generate the model, is the discriminant model, A dataset of soil aggregate 3D pore substructure The real pore structure, A dataset of soil aggregate 3D pore substructure The mathematical expectation of the real pore structure in Generate model for input A normally distributed random variable, To generate the mathematical expectation of the pore structure;

[0021] The generative model Depend on The deconvolution layer consists of is a hyperparameter, generating a model To the front The deconvolution output of the layer is batch normalized and ReLU nonlinearly activated to generate the model Perform Tanh nonlinear activation on the deconvolution output of the last layer;

[0022] The discriminant model Depend on The discriminant model is composed of layers of convolutional layers. To the front The convolution output of the layer is spectrally normalized and LeakyReLU nonlinearly activated; discriminant model The convolution output of the last layer is activated by Sigmoid nonlinearity. The expression of Sigmoid function is as follows:

[0023] ; In the formula, For the discriminant model The convolution output of the last layer.

[0024] Further, in step S4, the reconstructed network branch includes generating a model and discriminant model ;

[0025] The loss function of the reconstruction network branch consists of two parts: the pore structure quality loss function and the adversarial loss function, and each generation model and discriminant model Combination of Each has its corresponding loss function; The pore structure quality loss function is:

[0026] ;

[0027] In the formula, express The pore structure quality loss function is Represents a downsampled dataset of the three-dimensional pore substructure of soil aggregates The number of mesoporous structures, Represents a downsampled dataset of the three-dimensional pore substructure of soil aggregates The The pore structure Represents a downsampled dataset of the three-dimensional pore substructure of soil aggregates The The pore structure

[0028] The adversarial loss function is:

[0029] ;

[0030] In the formula, To generate the model, is the discriminant model, A dataset of soil aggregate 3D pore substructure The real pore structure, A dataset of soil aggregate 3D pore substructure The mathematical expectation of the real pore structure in A dataset of soil aggregate 3D pore substructure The real pore structure, A dataset of soil aggregate 3D pore substructure The mathematical expectation of the real pore structure in ;

[0031] The generative model All by Layer convolutional layer composition, generating model To the front The convolution output of the layer is batch normalized and LeakyReLU nonlinearly activated to generate the model Perform Tanh nonlinear activation on the convolution output of the last layer;

[0032] The discriminant model All by Layer convolution layer composition, discriminant model To the front The convolution output of the layer is batch normalized and LeakyReLU nonlinearly activated to discriminate the model Sigmoid nonlinear activation is performed on the convolution output of the last layer.

[0033] Furthermore, the training of the progressive reconstruction generative adversarial network model is specifically as follows: improving the effect of the three-dimensional pore substructure generated by the progressive reconstruction generative adversarial network model, optimizing the network framework structure and hyperparameters, until the progressive reconstruction generative adversarial network model generates a three-dimensional pore structure of soil aggregates that matches the characteristic distribution of the real three-dimensional pore substructure.

[0034] The beneficial effects of the method of the present invention are:

[0035] (1) Aiming at the problem that the existing generative model cannot preserve the distribution law of the original soil aggregate pore characteristics, the present invention makes adaptive improvements to the generative adversarial network framework, which can automatically and quickly generate a large number of relatively clear, visually realistic, and diverse soil aggregate pore three-dimensional structures; the one-time generation process of the original generative adversarial network is changed to a progressive generation process. The distribution characteristics of the soil pore structure data show stronger overall regularity than ordinary RGB image data. The amount of information of a single pore structure sub-sample is relatively small, while the overall distribution of multiple sub-samples of the same soil type and the same resolution shows a high consistency. Based on this, the progressive generation process can capture the overall distribution characteristics of the pore structure at a specific resolution, rather than just focusing on the specific detailed characteristics of a single sub-sample, thereby achieving a better generation effect.

[0036] (2) The present invention can effectively extract the distribution law of pore characteristics of soil aggregates and provide a reference method for the generation of three-dimensional pore structure of soil aggregates.

[0037] (3) The progressive reconstruction generative adversarial network provided by the present invention can provide new ideas and technical assistance for other three-dimensional image data enhancement problems in the field of agricultural informatization technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Flow chart of the method for generating three-dimensional pore structure of soil aggregates based on progressive reconstruction generative adversarial network;

[0039] Figure 2 Generate a basic framework diagram of the adversarial network for progressive reconstruction;

[0040] Figure 3 Schematic diagram of the generative adversarial network branches in the progressive reconstruction generative adversarial network;

[0041] Figure 4 Schematic diagram of the reconstruction network branches in the progressive reconstruction generative adversarial network. DETAILED DESCRIPTION

[0042] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the protection scope of the present invention.

[0043] Embodiment 1,

[0044] This embodiment provides a method for gradually generating a three-dimensional pore structure of soil aggregates. Figure 1 As shown, the method comprises the following steps:

[0045] S1, preprocessing the original soil aggregate CT scan image data, separating the pore part in the image, and constructing a three-dimensional pore structure voxel matrix data set of soil aggregates;

[0046] S2, performing random sampling of the soil aggregate three-dimensional pore structure dataset to obtain the soil aggregate three-dimensional pore substructure dataset;

[0047] S3, downsampling the three-dimensional pore substructure in the soil aggregate three-dimensional pore substructure dataset to obtain a soil aggregate three-dimensional pore substructure downsampled dataset;

[0048] S4. Construct a progressive reconstruction generative adversarial network model, which is divided into a generative adversarial network branch and a reconstruction network branch;

[0049] S5. Training a progressive reconstruction generative adversarial network model based on a soil aggregate three-dimensional pore substructure dataset to obtain an improved progressive reconstruction generative adversarial network model, and using the improved progressive reconstruction generative adversarial network model to generate the three-dimensional pore structure of soil aggregates.

[0050] Step S1 is specifically as follows:

[0051] ImageJ software was used to capture the middle area of ​​the soil aggregate CT scan image ( Voxels) are used as the feature domain. When selecting the feature domain, in order to reduce the influence of external factors such as edge effects and beam hardening, obvious fissures and the sample edge area need to be avoided. Then, the two-dimensional image stack is imported into ImageJ software, and the following operations are performed for preprocessing the image segmentation. First, the Enhance Contrast command is used to normalize the image to eliminate the brightness difference between different CT slice images. Then, the Despeckle command is used to remove the noise in the image. Finally, the Median operation is used to achieve a smoothing operation, making the pore structure in the image clearer and more accurate. When performing the image segmentation operation, the Threshold operation is used for image threshold segmentation to separate the pore part in the image. A binary image is generated, where the white area represents pores and the black area represents the soil matrix, and they are converted into a voxel matrix. The soil aggregate three-dimensional pore structure voxel matrix dataset is composed of the voxel matrix.

[0052] Step S2 is specifically as follows:

[0053] Based on the representative elementary volume cutting theory, select a size of The voxel matrix is used as the substructural unit for analysis. Here, n < N, where N is the overall size and n is the size of the small block randomly selected from the whole. The voxel matrix of this specification is sufficient to effectively capture and retain the substructural geometric characteristics and the interactions between substructures. According to The specification of the voxel matrix, intercept the soil aggregate three-dimensional pore structure voxel matrix dataset, and use the interval random sampling method to extract the submatrix as the soil aggregate three-dimensional pore substructure dataset .

[0054] Step S3 is specifically as follows:

[0055] Adopt h different sampling multiples to downsample the three-dimensional pore substructure to obtain the soil aggregate three-dimensional pore substructure downsampled dataset , where respectively represent the datasets corresponding to different sampling multiples.

[0056] The progressive reconstruction generative adversarial network model in Step S4 is as Figure 2 shown, including a generative adversarial network branch and a reconstruction network branch;

[0057] The generative model in the generative adversarial network branch is responsible for generating a rough version of the pore structure, which is then input into the reconstruction network branch. The reconstruction network branch contains multiple generative models with the same structure, and each generative model is preceded by an upsampling layer. The data in the generative adversarial network branch is input into the reconstruction network branch and then passes through a combination of h upsampling layers and generative models in sequence. By improving the resolution of the generated pore structure layer by layer, the detailed features of the pore structure at different resolutions are gradually refined and restored.

[0058] The generative adversarial network branch uses a deconvolutional neural network and a convolutional neural network as the main frameworks of the generative model and the discriminative model, respectively. Figure 3 As shown:

[0059] The generative model is used to learn the latent distribution of real pore structures and generate pseudo-pore structures that are indistinguishable from real pores;

[0060] The discriminant model is used to learn to distinguish true and false pore structures and calculate the probability that the pore structure comes from the real training set;

[0061] The target loss function of the generative adversarial network branch is:

[0062]

[0063] In the formula, To generate the model, is the discriminant model, A dataset of soil aggregate 3D pore substructure The real pore structure, A dataset of soil aggregate 3D pore substructure The mathematical expectation of the real pore structure in Generate model for input A normally distributed random variable, To generate the mathematical expectation of the pore structure;

[0064] The generative model Depend on The layer is composed of deconvolution layers, k is a hyperparameter obtained by experimental tests;

[0065] Generative Models The input is: a normally distributed random variable ;

[0066] The generative model To the front The deconvolution output of the layer is batch normalized and ReLU nonlinearly activated to generate the model The deconvolution output of the last layer is subjected to a Tanh nonlinear activation to generate a low-resolution representation of the 3D pore substructure of soil aggregates.

[0067] The discriminant model Depend on Layer convolutional layer composition;

[0068] Discriminant model The input is: Generate model The resulting pore structure or soil aggregate 3D pore substructure dataset Real pore structure ;

[0069] The discriminant model To the front The convolution output of the layer is spectrally normalized and LeakyReLU nonlinearly activated; discriminant model The convolution output of the last layer is activated by Sigmoid nonlinearity to calculate the probability that the input pore structure is classified as true; the expression of the Sigmoid function is as follows:

[0070] ;

[0071] In the above formula, For the discriminant model The convolution output of the last layer.

[0072] The reconstructed network branch is composed of Generative model and discriminative model with the same group structure Constructed and use convolutional neural network as the main framework of the generation model and the discriminant model;

[0073] The loss function of the reconstruction network branch consists of two parts: the pore structure quality loss function and the adversarial loss function, as shown in Figure 4 As shown;

[0074] in, The pore structure quality loss function is:

[0075] ;

[0076] In the formula, Represents a downsampled dataset of the three-dimensional pore substructure of soil aggregates The number of mesoporous structures, Represents a downsampled dataset of the three-dimensional pore substructure of soil aggregates The The pore structure Represents a downsampled dataset of the three-dimensional pore substructure of soil aggregates The A pore structure.

[0077] The adversarial loss function is:

[0078] ;

[0079] In the formula, To generate the model, is the discriminant model, A dataset of soil aggregate 3D pore substructure The real pore structure, A dataset of soil aggregate 3D pore substructure The mathematical expectation of the real pore structure in A dataset of soil aggregate 3D pore substructure The real pore structure, A dataset of soil aggregate 3D pore substructure The mathematical expectation of the real pore structure in ;

[0080] The generative model All by Layer convolutional layer composition;

[0081] Generative Models The input is: soil aggregate 3D pore substructure dataset Real pore structure ;

[0082] Real pore structure Input to the generative model After that, we first perform linear upsampling on it and modify The size of the soil aggregates makes it similar to the 3D pore substructure dataset Real pore structure The size is consistent, and then the pore structure after upsampling Input to the first convolutional layer of the generative model;

[0083] The generative model To the front The convolution output of the layer is batch normalized and LeakyReLU nonlinearly activated to generate the model Tanh nonlinear activation is performed on the convolution output of the last layer to generate the three-dimensional pore substructure of soil aggregates at the corresponding resolution.

[0084] The discriminant model All by Layer convolutional layer composition;

[0085] Discriminant model The input is: Generate model The resulting pore structure or soil aggregate 3D pore substructure dataset Real pore structure ;

[0086] The discriminant model To the front The convolution output of the layer is batch normalized and LeakyReLU nonlinearly activated to discriminate the model The convolution output of the last layer is activated by Sigmoid nonlinearity to calculate the probability that the input pore structure is classified as true.

[0087] In step S5, training the progressive reconstruction generative adversarial network model specifically includes: improving the effect of the three-dimensional pore substructure generated by the progressive reconstruction generative adversarial network model, optimizing the network framework structure and hyperparameters, until the progressive reconstruction generative adversarial network model generates a three-dimensional pore structure of soil aggregates that matches the characteristic distribution of the real three-dimensional pore substructure;

[0088] The hyperparameters of the optimized progressive reconstruction generative adversarial network model are: batch size is 16, loss function optimizer is Adam, learning rate is 0.0002, Beta1 is 0.5, and number of iterations is 100; the batch size is the number of samples selected for one training; the Beta1 is the exponential decay rate of the first-order moment estimate; the training of the progressive reconstruction generative adversarial network model can be implemented using the python-based pytorch deep learning module.

[0089] Embodiment 2,

[0090] This embodiment is a further explanation of the embodiment 1, and the embodiment 1 is further explained by a specific experimental method.

[0091] Step S1:

[0092] ImageJ software was used to intercept the four middle parts (300×300×300 voxels) of the soil aggregate CT scan image as feature domains. When selecting the feature domain, in order to reduce the influence of external factors including edge effect and radiation hardening, obvious cracks and sample edge areas were avoided. The two-dimensional image stack was then imported into ImageJ software, and the following operations were performed for image segmentation preprocessing. First, the Enhance Contrast command was used to normalize the image to eliminate the brightness difference between different CT slice images; then the Despeckle command was used to eliminate the image noise; finally, the Median operation was used to perform a smoothing operation to make the pore structure in the image clearer and more accurate. When performing image segmentation operations, the Threshold operation was used to perform image threshold segmentation to separate the pore part in the image. Four groups of 300 binary images were generated, in which the white area represented the pores and the black area represented the soil matrix, and they were converted into four Voxel matrix, which is the voxel matrix dataset of soil aggregate three-dimensional pore structure.

[0093] Step S2:

[0094] The size is selected based on the representative basic volume cutting theory. The voxel matrix is ​​used as the substructure unit for analysis. The voxel matrix of this specification is sufficient to effectively capture and preserve the geometric characteristics of the substructures and the interactions between the substructures. The voxel matrix specifications were used to intercept the soil aggregate three-dimensional pore structure voxel matrix dataset, and 2000 sub-matrices were extracted as the soil aggregate three-dimensional pore substructure dataset using the interval random sampling method.

[0095] Step S3:

[0096] In order to train the progressive reconstruction generative adversarial network model, it is necessary to downsample the three-dimensional pore substructure in the soil aggregate three-dimensional pore substructure dataset to obtain the soil aggregate three-dimensional pore substructure downsampled dataset. , and . A dataset of downsampled soil aggregate 3D pore substructure The data in the above table is the collection of soil aggregate three-dimensional pore substructure data. The 3D pore substructure of the voxel is downsampled to The three-dimensional pore substructure of a voxel;

[0097] A dataset of down-sampled soil aggregate 3D pore substructure The data in the above table is the collection of soil aggregate three-dimensional pore substructure data. The 3D pore substructure of the voxel is downsampled to The three-dimensional pore substructure of a voxel.

[0098] A dataset of down-sampled soil aggregate 3D pore substructure The data in the above table is the collection of soil aggregate three-dimensional pore substructure data. The 3D pore substructure of the voxel is downsampled to The three-dimensional pore substructure of a voxel.

[0099] Step S4:

[0100] (1) Generative adversarial network branch:

[0101] The generative model in the generative adversarial network branch consists of 4 deconvolution layers;

[0102] The number of channels in the input layer of the generative model is 100; from deconvolution layers 1 to 3, the number of channels output by each layer is 128, 64, and 32 respectively; the number of output channels of deconvolution layer 4 is 1; the convolution kernel size of the first deconvolution layer is , filled with 0, with a step size of 1; the convolution kernel size of the second deconvolution layer is , the padding is 1, the stride is 2; the convolution kernel size of the third deconvolution layer is , the padding is 1, the stride is 2; the convolution kernel size of the fourth deconvolution layer is , fill is 1, stride is 2;

[0103] The generative model performs batch normalization and ReLU nonlinear activation on the deconvolution outputs of the first three layers, and performs Tanh nonlinear activation on the deconvolution outputs of the last layer to generate Low-resolution representation of the three-dimensional pore substructure of soil aggregates.

[0104] The discriminative model consists of 4 convolutional layers in the opposite order of the generative model;

[0105] The input of the first convolutional layer is: Low-resolution representation of the three-dimensional pore substructure of soil aggregates;

[0106] The number of channels in the input layer of the discriminant model is 1; from convolutional layers 1 to 3, the number of channels output by each layer is 32, 64, and 128 respectively; the number of output channels of convolutional layer 4 is 1; the convolution kernel size of the first convolutional layer is , the padding is 1, the stride is 2; the convolution kernel size of the second convolution layer is , the padding is 1, the stride is 2; the convolution kernel size of the third convolution layer is , the padding is 1, the stride is 2; the convolution kernel size of the fourth convolution layer is , filled with 0 and stride of 1.

[0107] (2) Reconstructing network branches:

[0108] The reconstruction network branch consists of 3 groups of generative models and discriminative models with the same structure. Constructed and use convolutional neural network as the main framework of the generation model and the discriminant model; The input is a sampling dataset of soil aggregate 3D pore substructure The pore structure in The true pore structure in the soil is derived from the soil aggregate 3D pore substructure downsampling dataset ; The input is a sampling dataset of soil aggregate 3D pore substructure The pore structure in The true pore structure in the soil is derived from the soil aggregate 3D pore substructure downsampling dataset ; The input is a sampling dataset of soil aggregate 3D pore substructure The pore structure in The actual pore structure in comes from the soil aggregate 3D pore substructure dataset.

[0109] The loss function of the reconstruction network branch consists of two parts: image quality loss and adversarial loss.

[0110] in, The pore structure quality loss function is:

[0111]

[0112] In the formula, Represents a downsampled dataset of the three-dimensional pore substructure of soil aggregates The number of mesoporous structures, Represents a downsampled dataset of the three-dimensional pore substructure of soil aggregates The The pore structure Represents a downsampled dataset of the three-dimensional pore substructure of soil aggregates The A pore structure.

[0113] in, The pore structure quality loss function is:

[0114]

[0115] In the formula, Represents a downsampled dataset of the three-dimensional pore substructure of soil aggregates The number of mesoporous structures, Represents a downsampled dataset of the three-dimensional pore substructure of soil aggregates The The pore structure Represents a downsampled dataset of the three-dimensional pore substructure of soil aggregates The A pore structure.

[0116] in, The pore structure quality loss function is:

[0117]

[0118] In the formula, Represents a downsampled dataset of the three-dimensional pore substructure of soil aggregates The number of mesoporous structures, Represents a downsampled dataset of the three-dimensional pore substructure of soil aggregates The The pore structure Represents the first A pore structure.

[0119] , and The adversarial loss function is:

[0120] ;

[0121] In the formula, To generate the model, is the discriminant model, is the real pore structure. is the mathematical expectation of the real pore structure, Generate model for input A normally distributed random variable, To generate the mathematical expectation of the pore structure;

[0122] The generative model , and They are all composed of 5 convolutional layers;

[0123] The first convolutional layer has 1 input channel and 16 output channels; the second to fourth convolutional layers have 16 input and output channels; the fifth convolutional layer has 16 input channels and 1 output channel; all convolutional kernel sizes are set to , filled with 0, step size is 1;

[0124] The generative model performs batch normalization and LeakyReLU nonlinear activation on the convolution outputs of the first four layers, and performs Tanh nonlinear activation on the convolution output of the last layer to generate a three-dimensional pore substructure of soil aggregates with corresponding resolution.

[0125] The discriminant model , and They are all composed of 5 convolutional layers;

[0126] The first convolutional layer has 1 input channel and 16 output channels; the second to fourth convolutional layers have 16 input and output channels; the fifth convolutional layer has 16 input channels and 1 output channel; all convolutional kernel sizes are set to , filled with 0, step size is 1;

[0127] The discriminant model performs batch normalization and LeakyReLU nonlinear activation on the convolution outputs of the first four layers, and performs Sigmoid nonlinear activation on the convolution output of the last layer to calculate the probability that the input pore structure is classified as true.

[0128] Step S5:

[0129] After completing the construction of the progressive reconstruction generative adversarial network model, it is necessary to further optimize the hyperparameters of the model to achieve the best effect in the generation of the three-dimensional pore structure of soil aggregates; S5 The hyperparameters of the optimized progressive reconstruction generative adversarial network model are: batch size is 16, loss function optimizer is Adam, learning rate is 0.0002, Beta1 is 0.5, and number of iterations is 100; the batch size is the number of samples selected for one training; Beta1 is the exponential decay rate of the first-order moment estimate; The training of the progressive reconstruction generative adversarial network model can be implemented using the python-based pytorch deep learning module.

Claims

1. A progressive method for generating three-dimensional pore structures of soil aggregates, It is characterized in that The method comprises the following steps: S1, preprocessing the original soil aggregate CT scan image data, separating the pore part in the image, and constructing a three-dimensional pore structure voxel matrix data set of soil aggregates; S2, performing random sampling of the soil aggregate three-dimensional pore structure dataset to obtain the soil aggregate three-dimensional pore substructure dataset; S3, downsampling the three-dimensional pore substructure in the soil aggregate three-dimensional pore substructure dataset to obtain a soil aggregate three-dimensional pore substructure downsampled dataset; S4. Construct a progressive reconstruction generative adversarial network model, which is divided into a generative adversarial network branch and a reconstruction network branch; In step S4, the generative adversarial network branch includes a generative model and discriminant model ; The target loss function of the generative adversarial network branch is: In the formula, To generate the model, is the discriminant model, A dataset of soil aggregate 3D pore substructure The real pore structure, A dataset of soil aggregate 3D pore substructure The mathematical expectation of the real pore structure in Generate model for input A normally distributed random variable, To generate the mathematical expectation of the pore structure; The generative model Depend on The deconvolution layer is composed of layers. is a hyperparameter, generating a model To the front The deconvolution output of the layer is batch normalized and ReLU nonlinearly activated to generate the model Perform Tanh nonlinear activation on the deconvolution output of the last layer; The discriminant model Depend on The discriminant model is composed of layers of convolutional layers. To the front The convolution output of the layer is batch normalized and LeakyReLU nonlinearly activated; discriminant model The convolution output of the last layer is activated by Sigmoid nonlinearity. The expression of Sigmoid function is as follows: ; In the formula, For the discriminant model The convolution output of the last layer; The reconstruction network branch includes a generation model and discriminant model ; The loss function of the reconstruction network branch consists of two parts: the pore structure quality loss function and the adversarial loss function, and each generation model and discriminant model Combination of Each has its corresponding loss function; The pore structure quality loss function is: ； In the formula, express The pore structure quality loss function is Represents a downsampled dataset of the three-dimensional pore substructure of soil aggregates The number of mesoporous structures, Represents a downsampled dataset of the three-dimensional pore substructure of soil aggregates The The pore structure Represents a downsampled dataset of the three-dimensional pore substructure of soil aggregates The The pore structure The adversarial loss function is: ； In the formula, To generate the model, is the discriminant model, A dataset of soil aggregate 3D pore substructure The real pore structure, A dataset of soil aggregate 3D pore substructure The mathematical expectation of the real pore structure in A dataset of soil aggregate 3D pore substructure The real pore structure, A dataset of soil aggregate 3D pore substructure The mathematical expectation of the real pore structure in ; The generative model All by Layer convolutional layer composition, generating model To the front The convolution output of the layer is batch normalized and LeakyReLU nonlinearly activated to generate the model Perform Tanh nonlinear activation on the convolution output of the last layer; The discriminant model All by Layer convolution layer composition, discriminant model To the front The convolution output of the layer is batch normalized and LeakyReLU nonlinearly activated to discriminate the model Perform Sigmoid nonlinear activation on the convolution output of the last layer; S5. Training a progressive reconstruction generative adversarial network model based on a soil aggregate three-dimensional pore substructure dataset to obtain an improved progressive reconstruction generative adversarial network model, and using the improved progressive reconstruction generative adversarial network model to generate the three-dimensional pore structure of soil aggregates.

2. The method for progressively generating a three-dimensional pore structure of soil aggregates according to claim 1, It is characterized in that The step S1 specifically comprises: intercepting a crack-free area in the original soil aggregate CT scan image as a feature domain image, performing normalization processing, noise elimination processing, and smoothing processing on the intercepted feature domain image in sequence, and then performing threshold segmentation on the processed feature domain image to separate the pore part in the feature domain image to generate a binary image, and converting the binary image into a voxel matrix, and the voxel matrix constitutes a three-dimensional pore structure voxel matrix data set of soil aggregates.

3. The method for progressively generating a three-dimensional pore structure of soil aggregates according to claim 2, It is characterized in that In step S2, when random sampling is performed on the soil aggregate three-dimensional pore structure data set, according to The voxel matrix specification is used to extract the soil aggregate 3D pore structure dataset, where The length, width and height of the voxel matrix are .

4. The method for progressively generating a three-dimensional pore structure of soil aggregates according to claim 3, It is characterized in that In step S3, use h The three-dimensional pore substructure is downsampled by different sampling multiples to obtain the downsampled dataset of the three-dimensional pore substructure of soil aggregates. ,in, They represent data sets corresponding to different sampling multiples.

5. The method for progressively generating a three-dimensional pore structure of soil aggregates according to claim 4, It is characterized in that In step S5, training the progressive reconstruction generative adversarial network model specifically includes: improving the effect of the three-dimensional pore substructure generated by the progressive reconstruction generative adversarial network model, optimizing the network framework structure and hyperparameters, until the progressive reconstruction generative adversarial network model generates a three-dimensional pore structure of soil aggregates that matches the characteristic distribution of the real three-dimensional pore substructure.

Images