Soil crust preparation and analysis system

Abstract

The present invention relates to the technical field of soil crust analysis, and more specifically to a soil crust preparation and analysis system. The system comprises soil preparation equipment, environmental data acquisition equipment, crust data acquisition equipment, and data processing equipment, wherein the environmental data acquisition equipment and the crust data acquisition equipment are both communicatively connected to the data processing equipment, wherein: the soil preparation equipment is used to prepare soil crust; the environmental data acquisition equipment is used to collect environmental time series data during the soil crust preparation process; the crust data acquisition equipment is used to collect crust data corresponding to the soil crust; and the data processing equipment is used to train a target causal analysis model based on the environmental time series data and the crust data. This provides a soil crust preparation and analysis system that can collect and analyze data.

Description

Technical Field

[0001] The present invention relates to the technical field of soil crust analysis, and in particular to a soil crust preparation and analysis system. Background Art

[0002] Soil crust is an important factor affecting soil water evaporation, erosion, and plant growth. Currently, research on soil crusts mainly focuses on the formation mechanism, physical and chemical properties of soil crusts, and their impact on soil erosion resistance.

[0003] However, existing soil crust preparation and detection equipment still faces many technical difficulties in practical applications. First, existing technologies lack in-situ real-time detection capabilities, and most equipment can only perform parameter detection through manual sampling, which not only increases the experimental cycle but also affects the real-time and accuracy of the data. In addition, the existing equipment lacks sufficient control accuracy over crust formation conditions (such as temperature, humidity, and salt gradients), and cannot accurately simulate the dynamic changes in the natural environment, resulting in experimental results that are not highly repeatable and stable. Finally, although some equipment can monitor certain characteristics of soil crusts, there is a lack of in-depth research on the dynamic relationship between crust characteristics and soil parameters, and it has failed to achieve comprehensive data analysis and intelligent regulation.

[0004] Therefore, providing a soil crust preparation and analysis system that can collect and analyze data has become an urgent problem to be solved. Summary of the Invention

[0005] In view of this, the present invention provides a soil crust preparation and analysis system to solve the problem of providing a soil crust preparation and analysis system that can collect and analyze data.

[0006] In a first aspect, the present invention provides a soil crust preparation and analysis system, which includes a soil preparation device, an environmental data acquisition device, a crust data acquisition device, and a data processing device. The environmental data acquisition device and the crust data acquisition device are both communicatively connected to the data processing device, wherein:

[0007] soil preparation equipment for preparing soil crusts;

[0008] Environmental data collection equipment, used to collect environmental time series data during the soil crust preparation process; the environmental time series data includes temperature time series data, humidity time series data, rainfall time series data, and rainfall salinity data;

[0009] Crust data acquisition equipment, used to collect crust data corresponding to soil crust; crust data includes crust image data, crust spectral data, crust electrochemical data, and crust mechanical properties;

[0010] Data processing equipment is used to train a target causal analysis model based on environmental time series data and crust data.

[0011] The soil crust preparation and analysis system provided in the embodiment of the present application includes a soil preparation device for preparing soil crust; an environmental data acquisition device for collecting environmental time series data during the soil crust preparation process, and capturing dynamic environmental changes during the crust preparation process, such as sudden changes in rainfall intensity and diurnal temperature fluctuations. The crust data acquisition device is used to collect crust data corresponding to the soil crust, thereby ensuring the accuracy of the crust data collected. Therefore, there is no need for manual sampling for parameter detection, which reduces the experimental cycle and also ensures the real-time and accuracy of the collected environmental time series data. The data processing device is used to train the target causal analysis model based on the environmental time series data and the crust data, thereby ensuring the accuracy of the target causal analysis model obtained. This enables an in-depth study of the dynamic relationship between crust characteristics and soil parameters. The soil crust preparation and analysis system realizes accurate data collection through equipment collaboration and deep causal modeling through algorithm innovation, thereby building a complete technical chain from experimental preparation to intelligent analysis. This provides a soil crust preparation and analysis system that can collect and analyze data. Its core value lies in pushing soil science research from "experience-driven" to "data-causal dual-driven", providing efficient technical tools for understanding the complex ecological processes of soil crust and addressing soil degradation problems. It has significant scientific significance and practical application potential.

[0012] In an optional embodiment, a data processing device is used to input the crust image data, crust spectral data, crust electrochemical data, and crust mechanical properties into a preset feature fusion model, perform feature fusion on the crust image data, crust spectral data, crust electrochemical data, and crust mechanical properties, and generate a target fusion feature;

[0013] The target fusion features and temperature time series data, humidity time series data, rainfall time series data and rainfall salinity data are input into the initial causal analysis network, and the initial causal analysis network is trained to obtain the target causal analysis model.

[0014] In an optional embodiment, a data processing device is used to input the crust image data, crust spectral data, crust electrochemical data, and crust mechanical properties into corresponding embedding layers to generate corresponding image feature vectors, spectral feature vectors, electrochemical feature vectors, and mechanical feature vectors;

[0015] Based on the multi-scale feature fusion network, the image feature vector, spectral feature vector, electrochemical feature vector and mechanical feature vector are extracted, and the corresponding sub-initial image feature vector, sub-initial spectral feature vector, sub-initial electrochemical feature vector and sub-initial mechanical feature vector at each scale are output;

[0016] Based on the attention mechanism, the sub-initial image feature vectors, sub-initial spectral feature vectors, sub-initial electrochemical feature vectors and sub-initial mechanical feature vectors corresponding to each scale are fused to generate target fusion features.

[0017] In an optional embodiment, the data processing device is used to calculate the real-time statistical features corresponding to the image feature vector, the spectral feature vector, the electrochemical feature vector, and the mechanical feature vector respectively;

[0018] The real-time statistical features corresponding to the image feature vector, spectral feature vector, electrochemical feature vector, and mechanical feature vector are input into the adaptive dilated convolution kernel generation mechanism based on the reinforcement learning algorithm;

[0019] Using each real-time statistical feature as state input, different initial empty convolution kernel parameters are dynamically generated; each initial empty convolution kernel parameter includes an initial convolution kernel size of different scales;

[0020] Based on the preset reward function, calculate the reward value corresponding to each initial empty convolution kernel parameter;

[0021] According to the reward value, select the target empty convolution kernel parameter with the largest reward value;

[0022] Based on the target convolution kernel sizes of different scales in the target empty convolution kernel parameters, feature extraction is performed on the image feature vector, spectral feature vector, electrochemical feature vector and mechanical feature vector, and the corresponding sub-initial image feature vector, sub-initial spectral feature vector, sub-initial electrochemical feature vector and sub-initial mechanical feature vector at each scale are output.

[0023] In an optional embodiment, the data processing device is used to perform pairwise combinations of the sub-initial image feature vectors, the sub-initial spectral feature vectors, and the sub-initial electrochemical feature vectors of the same scale to generate multiple cross-modal feature groups;

[0024] For each feature group, the feature group is input into the attention calculation network to calculate the correlation between the two sub-feature vectors in the feature group;

[0025] Based on the correlation, the attention weight matrix is output; the attention weight matrix is used to represent the degree of attention between two sub-feature vectors in the feature group;

[0026] Based on the attention weight matrix, the two sub-feature vectors in the feature group are updated;

[0027] This cycle is repeated to obtain the corresponding sub-updated image feature vector, sub-updated spectral feature vector, sub-updated electrochemical feature vector, and sub-updated mechanical feature vector at each scale.

[0028] Fusing each sub-updated image feature vector, each sub-updated spectral feature vector, each sub-updated electrochemical feature vector, and each sub-updated mechanical feature vector according to the data modality to obtain a fused sub-target image feature vector, sub-target spectral feature vector, sub-target electrochemical feature vector, and sub-target mechanical feature vector;

[0029] The sub-target image feature vector, sub-target spectral feature vector, sub-target electrochemical feature vector and sub-target mechanical feature vector are fused to generate target fusion features.

[0030] In an optional embodiment, the data processing device is used to generate a training set and a validation set based on the target fusion feature and the temperature time series data, the humidity time series data, the rainfall time series data, and the salinity data of the rainfall;

[0031] Input the training set into the input layer of the initial causal analysis network;

[0032] The hidden layer in the initial causal analysis network extracts features from the target fusion features and the temperature time series data, humidity time series data, rainfall time series data, and the salinity of rainfall, and outputs the target features.

[0033] The output layer in the initial causal analysis network outputs the virtual causal relationship between the crust data and the environmental time series data based on the target features;

[0034] Based on the virtual causal relationship and the real causal relationship, the loss value corresponding to the target loss function is calculated; the real causal relationship is determined by analyzing the environmental time series data and the soil crust data during the soil crust preparation process;

[0035] Update the parameters of the initial causal analysis network based on the loss value to obtain an initial causal analysis model;

[0036] The initial causal analysis model is verified based on the validation set, and when the accuracy of the initial causal analysis model is greater than or equal to a preset accuracy threshold, the target causal analysis model is obtained.

[0037] In an optional embodiment, the data processing device is used to construct a priori causal knowledge graph based on prior knowledge of the causal relationship between environmental factors and crusting data;

[0038] Mapping the virtual causal relationships output by the initial causal analysis network to the corresponding nodes and edges in the prior causal knowledge graph;

[0039] Detect conflicts between virtual causal relationships and prior causal knowledge graphs;

[0040] Calculate causal consistency loss based on the conflict between virtual causal relationships and the prior causal knowledge graph;

[0041] Classify the samples in the training set according to the causal relationship type between the crust data and the environmental time series data, and determine the number of samples corresponding to each category;

[0042] Based on the number of samples corresponding to each type, different weights are assigned to samples of different categories to determine the balanced loss function;

[0043] Based on causal consistency loss, balance loss function and mean square error loss function, the target loss function is constructed.

[0044] In an optional embodiment, the data processing device is further configured to obtain environmental change data corresponding to the target area within a preset time period before the current time, input the environmental change data into the target causal analysis model, and output future soil crust data corresponding to the target area;

[0045] Based on future soil crust data, output corresponding management measures.

[0046] In an optional embodiment, the soil preparation device includes:

[0047] A preparation rack comprises a support rack and a preparation container, wherein the preparation container is placed on the support rack;

[0048] A soil crust cultivation soil column is placed inside a preparation container and is used to prepare a soil crust;

[0049] A rainmaker is installed on the top surface of the preparation container and is used to rain rain on the soil crust cultivation soil column;

[0050] The pressure regulating nozzle is installed on the rain generator to simulate the amount of rainfall;

[0051] One end of the rainmaker is connected to the salt dissolution preparation device through a first water pump and a first valve, and the other end is connected to the pure water preparation device through a second water pump and a second valve; when the first valve is opened, the salt solution is sprayed onto the soil crust cultivation soil column; when the second valve is opened, the pure water is sprayed onto the soil crust cultivation soil column;

[0052] A solution collection bucket is placed inside the support frame and below the preparation container to collect the liquid flowing down from the soil crust cultivation soil column;

[0053] Heating equipment to change the ambient temperature during soil crust preparation.

[0054] In an optional embodiment, the crust data collection device includes:

[0055] A camera device for collecting soil crust image data corresponding to the soil crust;

[0056] Near-infrared spectroscopy equipment, used to collect crust spectral data;

[0057] An electrochemical sensor group is used to collect electrochemical data of the crust;

[0058] Soil mechanical sensors are used to collect the mechanical properties of soil crust.

[0059] The soil crust preparation and analysis system provided in the embodiment of the present application prepares soil crust based on soil preparation equipment, realizing a refined study of the soil crust formation process. Then, the environmental time series data in the soil crust preparation process and the crust data corresponding to the soil crust are input into a preset feature fusion model for feature fusion to generate target fusion features, thereby realizing feature extraction and feature fusion of the environmental time series data and the crust data, and ensuring the accuracy of the generated target fusion features. Then, based on the target fusion features and the environmental time series data, a training set and a validation set are generated; then, a target causal analysis model is obtained by training based on the training set and the validation set. The accuracy of the obtained target causal analysis model is guaranteed. Therefore, the future number of soil crusts corresponding to the target area can be determined based on the target causal analysis model, and then the corresponding management measures can be output. This ensures the correspondence between the output management measures and the future soil crust data. It achieves early response to soil crusts and reduces the negative impact of soil crusts on plant growth. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0061] Figure 1 2 is a schematic structural diagram of a soil crust preparation and analysis system according to an embodiment of the present invention;

[0062] Figure 2 is a schematic diagram of a process for obtaining a target causal analysis model through training according to an embodiment of the present invention;

[0063] Figure 3is a flowchart of predicting future soil crust data and outputting corresponding management measures according to an embodiment of the present invention;

[0064] Figure 4 2 is a schematic structural diagram of another soil crust preparation and analysis system according to an embodiment of the present invention.

[0065] Among them: soil preparation equipment 1; preparation bracket 11; soil crust cultivation soil column 12; rain generator 13; first water pump 14; first valve 15; salt dissolution preparation device 16; second water pump 17; second valve 18; pure water preparation device 19; solution collection barrel 110; heating equipment 111; pressure regulating nozzle 112; environmental data acquisition equipment 2; crust data acquisition equipment 3; data processing equipment 4. DETAILED DESCRIPTION

[0066] To make the purpose, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are 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 those skilled in the art without making creative efforts shall fall within the scope of protection of the present invention.

[0067] The present application embodiment provides a soil crust preparation and analysis system, such as Figure 1 As shown, the soil crust preparation and analysis system includes a soil preparation device 1, an environmental data acquisition device 2, a crust data acquisition device 3 and a data processing device 4. The environmental data acquisition device 2 and the crust data acquisition device 3 are both communicatively connected to the data processing device 4, wherein:

[0068] Soil preparation equipment 1 for preparing soil crust;

[0069] Environmental data collection equipment 2, used to collect environmental time series data during the soil crust preparation process; the environmental time series data includes temperature time series data, humidity time series data, rainfall time series data, and rainfall salinity data;

[0070] Crust data acquisition equipment 3, used to collect crust data corresponding to soil crust; the crust data includes crust image data, crust spectral data, crust electrochemical data, and crust mechanical properties;

[0071] The data processing device 4 is used to train a target causal analysis model based on the environmental time series data and the crust data.

[0072] Specifically, the soil preparation device 1 can produce a soil crust. During the soil crust preparation process, a temperature sensor (such as a thermocouple or thermistor sensor) in the environmental data acquisition device 2 monitors temperature changes in the preparation environment in real time. A humidity sensor (such as a capacitive or resistive humidity sensor) in the environmental data acquisition device 2 senses the water vapor content in the environment, converts the humidity information into an electrical signal, and transmits this signal to the data processing device 4 after signal processing, thereby recording dynamic changes in the ambient humidity. Optionally, a pressure sensor in the environmental data acquisition device 2 can measure the pressure in the pure water preparation device 19 in the soil preparation device 1 to determine rainfall time series data. The environmental data acquisition device 2 can also obtain the salt concentration in the salt dissolution preparation device 16 to determine the salinity data of the rainfall.

[0073] Crust data collection device 3 collects soil crust data corresponding to soil crusts. After receiving the environmental time-series data and crust data, data processing device 4 uses data processing and machine learning algorithms to train the target causal analysis model. First, the collected raw data undergoes preprocessing, including data cleaning (removing outliers and missing values), data standardization (unifying the data scale), and feature extraction (extracting key features from image, spectral, and other data) to convert the data into a format suitable for model training. Then, based on the preprocessed data, an appropriate causal analysis model (such as a structural causal model or causal forest) is selected for training. During training, the model learns the associations between environmental factors (such as temperature, humidity, rainfall, and salinity) and soil crust characteristics (such as image features, spectral features, electrochemical characteristics, and mechanical properties) to identify causal relationships. For example, the model may discover that changes in humidity have a significant causal effect on the formation rate and structure of crusts, or that rainfall salinity and the chemical composition of crusts are causally related. After a large amount of data training and optimization, the obtained target causal analysis model can predict the formation process and characteristic changes of soil crust under different environmental conditions, providing a scientific basis for soil science research, land management and ecological protection.

[0074] The soil crust preparation and analysis system provided in the embodiment of the present application includes a soil preparation device 1 for preparing soil crust; an environmental data acquisition device 2 for collecting environmental time series data during the soil crust preparation process, and capturing dynamic environmental changes during the crust preparation process, such as sudden changes in rainfall intensity and diurnal temperature fluctuations. The crust data acquisition device 3 is used to collect crust data corresponding to the soil crust, thereby ensuring the accuracy of the crust data collected. Therefore, there is no need for manual sampling for parameter detection, which reduces the experimental cycle and also ensures the real-time and accuracy of the collected environmental time series data. The data processing device 4 is used to train the target causal analysis model based on the environmental time series data and the crust data, thereby ensuring the accuracy of the target causal analysis model obtained. This enables in-depth research on the dynamic relationship between crust characteristics and soil parameters. The soil crust preparation and analysis system realizes accurate data collection through equipment collaboration and achieves causal deep modeling through algorithm innovation, thus building a complete technical chain from experimental preparation to intelligent analysis. This provides a soil crust preparation and analysis system that can collect and analyze data. Its core value lies in pushing soil science research from "experience-driven" to "data-causal dual-driven", providing efficient technical tools for understanding the complex ecological processes of soil crust and addressing soil degradation problems. It has significant scientific significance and practical application potential.

[0075] In an optional embodiment of the present application, Figure 2 As shown, “the data processing device 4 trains a target causal analysis model based on the environmental time series data and the crust data” may include the following steps:

[0076] In step S101, the crust image data, crust spectral data, crust electrochemical data and crust mechanical properties are input into a preset feature fusion model, and the crust image data, crust spectral data, crust electrochemical data and crust mechanical properties are subjected to feature fusion to generate target fusion features.

[0077] Specifically, the above step S101 may include the following steps:

[0078] In step S1011 , the crust image data, crust spectral data, crust electrochemical data, and crust mechanical properties are input into corresponding embedding layers to generate corresponding image feature vectors, spectral feature vectors, electrochemical feature vectors, and mechanical feature vectors.

[0079] Specifically, data processing device 4 converts the acquired crust image data into a uniformly dimensional image feature vector through an image embedding layer. The crust spectral data, crust electrochemical data, and crust mechanical property data are also converted into uniformly dimensional feature vectors that meet the model input requirements through the corresponding spectral embedding layer, electrochemical embedding layer, and mechanical embedding layer, respectively. Environmental time series data also needs to be processed through a dedicated time series embedding layer to ensure dimensional consistency with other data, preparing for subsequent fusion analysis.

[0080] Specifically, the image embedding layer captures image features at different scales through multiple convolutional layers (such as the residual block of ResNet). Typical convolution operations can be: .in, is the convolution kernel of the lth layer, is the feature map of the previous layer. The spatial features are compressed into a vector through global average pooling (GAP), and then the fully connected layer (FC) outputs a fixed-dimensional image feature vector (e.g., 512 dimensions): ImageFeature = FC(GAP(ConvOutput).

[0081] The spectral embedding layer captures the characteristics of the local wavelength range (such as the characteristic peak of a specific functional group) through a one-dimensional convolution kernel, and stacks multiple convolution layers to extract multi-scale spectral patterns: Where K is the convolution kernel size and t is the wavelength point index.

[0082] The self-attention mechanism captures the long-range dependencies between different wavelengths (such as the synergistic effects between characteristic peaks), which is suitable for complex spectral analysis: After position encoding, the spectral sequence is input into the Transformer, which outputs a context-aware spectral feature vector. The spectral features are compressed to a specified dimension (e.g., 256 dimensions) through a fully connected layer to highlight the contribution of key feature peaks.

[0083] The electrochemical embedding layer converts the crust electrochemical data curve into a time series or frequency domain vector (such as the real / imaginary part of the impedance after Fourier transformation), normalizing it to zero mean and unit variance. LSTM or GRU units are used to capture the evolution of potential / current over time: Where xt is the electrochemical measurement value (e.g., redox potential) at time t, and ht is the hidden state, integrating historical information. The hidden state or graph node embedding at the last moment is taken and passed through a fully connected layer to output an electrochemical feature vector (e.g., 128 dimensions).

[0084] The mechanical embedding layer normalizes the crust mechanical property data. For multi-stage test data (such as loading-unloading curves), it converts the data into a fixed-length sequence (e.g., interpolated to 100 time points). This is directly mapped through the fully connected layer. For example, the hardness value undergoes a linear transformation and is concatenated with other parameters: MechanicalFeature = W⋅[hardness; elastic modulus] + b. For force-displacement curves during loading, 1D convolution or Transformer is used to extract features such as plastic deformation and yield point: (i is the sample index, j is the feature index.) Finally, a fixed-dimensional mechanical feature vector (e.g., 64 dimensions) is output, retaining key mechanical performance indicators (e.g., peak force, fracture strain).

[0085] Step S1012: perform feature extraction on the image feature vector, spectral feature vector, electrochemical feature vector, and mechanical feature vector based on a multi-scale feature fusion network, and output the corresponding sub-initial image feature vector, sub-initial spectral feature vector, sub-initial electrochemical feature vector, and sub-initial mechanical feature vector at each scale.

[0086] Specifically, the above step S1012 may include the following steps:

[0087] In step a1, the real-time statistical features corresponding to the image feature vector, the spectral feature vector, the electrochemical feature vector, and the mechanical feature vector are input into an adaptive dilated convolution kernel generation mechanism based on a reinforcement learning algorithm.

[0088] Specifically, for the image feature vector, local variance is used to quantify texture roughness, gradient direction histogram is used to capture edge directionality, and tiny cracks on the crust surface are detected based on the Laplace operator response to obtain real-time statistical features corresponding to the image feature vector.

[0089] For the spectral feature vector, feature extraction is performed on the spectral feature vector to extract the absorption peak position, width, and intensity, and calculate the Pearson correlation coefficient between key bands. In addition, the spectral entropy of the spectral feature vector is calculated to measure the spectral complexity and information richness, and the real-time statistical characteristics corresponding to the spectral feature vector are obtained.

[0090] For the electrochemical characteristic vector, feature extraction is performed on the electrochemical characteristic vector, the ion concentration change rate and reaction rate are calculated, and parameters such as the peak current and peak potential of the voltammetric curve are extracted. In addition, based on autocorrelation analysis, the periodicity of the electrochemical process is quantified to obtain the real-time statistical characteristics corresponding to the electrochemical characteristic vector.

[0091] For the mechanical characteristic vector, feature extraction is performed on the mechanical characteristic vector to calculate the stress concentration coefficient, average stress, analyze the time dependence of the mechanical response, and evaluate the uniformity of the crust surface hardness to obtain the real-time statistical characteristics corresponding to the mechanical characteristic vector.

[0092] In step a2, different initial empty convolution kernel parameters are dynamically generated using the real-time statistical features as state input.

[0093] Among them, the parameters of each initial empty convolution kernel include initial convolution kernel sizes of different scales.

[0094] Specifically, the real-time statistical features corresponding to the image, spectral, electrochemical, and mechanical eigenvectors are concatenated to generate real-time statistical inputs. Based on the reinforcement learning agent's decision-making, the parameters of the initial empty convolution kernels of different scales are dynamically generated using these real-time statistical features as state inputs.

[0095] For the image feature vector, a series of convolution kernels with different void ratios may be generated to capture multi-scale information from subtle textures to overall structure. The generated convolution kernels with voids of different scales are used to perform convolution operations on the image feature vector. The convolution kernel with a small void ratio focuses on extracting the subtle texture information of the image, such as the tiny cracks on the surface of the crust, the fine texture of the particles, etc., and generates the corresponding sub-initial image feature vectors. These sub-initial vectors contain the detailed features of the image at a small scale. The convolution kernel with a large void ratio is used to extract the overall structure and macro features of the image, such as the overall shape of the crust, the texture distribution over a large area, etc., and also generates the corresponding sub-initial image feature vectors, covering the global information of the image at a large scale.

[0096] For the spectral feature vector, a dilated convolution kernel adapted to its frequency distribution characteristics is generated. The generated convolution kernel is not fixed. As the real-time statistical information of the input feature vector changes, the convolution kernel will also be updated in real time to adapt to the dynamic changes of the data. Dilated convolution kernels of different scales perform convolution operations according to the frequency distribution of the spectrum. The convolution kernel with a small dilation ratio focuses on the frequency bands in the spectrum that change dramatically and contain key information, extracts fine spectral features, and generates sub-initial spectral feature vectors. These vectors can reflect the detailed changes of the spectrum in the local frequency range. The convolution kernel with a large dilation ratio integrates a wider range of spectral information to obtain the overall trend and characteristics of the spectrum. The generated sub-initial spectral feature vector contains the large-scale frequency characteristics of the spectrum, which helps to analyze the overall characteristics of the spectrum.

[0097] The electrochemical eigenvectors are convolved using the generated dilated convolution kernel. Small-scale convolution kernels capture subtle changes in local electrochemical parameters, such as initial ion concentration, and the resulting sub-initial electrochemical eigenvectors accurately reflect the local electrochemical characteristics. Large-scale convolution kernels, on the other hand, integrate the changing trends of electrochemical parameters over a larger range from a more macroscopic perspective, generating sub-initial electrochemical eigenvectors that reflect the overall electrochemical changes.

[0098] The mechanical feature vectors are processed using a dilated convolution kernel. A small dilated convolution kernel extracts microscopic variations in mechanical parameters, such as local hardness differences and stress concentration points, generating corresponding sub-initial mechanical feature vectors that reflect the details of the mechanical characteristics. A large dilated convolution kernel captures the overall characteristics of the mechanical characteristics at larger scales, such as the overall hardness distribution and the mechanical stability of the structure. The generated sub-initial mechanical feature vectors contain macroscopic information about the mechanical characteristics.

[0099] Step a3: Calculate the reward value corresponding to each initial empty convolution kernel parameter based on the preset reward function.

[0100] Specifically, the reinforcement learning algorithm's adaptive dilated convolution kernel generation mechanism uses this statistical information as state input and, based on a pre-defined reward function, attempts to select different dilated convolution kernel parameters (such as dilation ratio and kernel size). The reward function is designed to encourage the generated convolution kernels to better capture the multi-scale information in the feature vector. If the generated convolution kernels are able to extract richer and more representative features in subsequent feature extraction, the agent will receive a higher reward; otherwise, it will receive a lower reward. Through continuous trial and error and learning, the agent gradually masters the optimal convolution kernel generation strategy.

[0101] Among them, the reward function can be: ;

[0102] Among them, R represents the reward value and α and β are weight coefficients used to adjust the relative importance of information entropy and feature importance in the reward function, and satisfy α+β=1, α,β∈[0,1]. For example, if the current task focuses more on obtaining rich feature details, the value of α can be appropriately increased; if more attention is paid to the representativeness of the features, the proportion of β can be increased. N represents the number of scales into which the feature vector is divided. Xi represents the feature vector extracted at the i-th scale. H(Xi) is the information entropy of the feature vector Xi at the i-th scale, which is used to measure the richness of the features. The calculation formula for information entropy is: Where p(xk) is the probability of the kth eigenvalue occurring in the eigenvector Xi. A larger information entropy value indicates richer and more disordered information contained in the eigenvector at that scale, i.e., a higher diversity of features. M is the number of predefined key feature indicators.

[0103] Yj represents the jth key characteristic index. For example, in soil crust research, key characteristic indicators may include mechanical parameters reflecting crust stability and electrochemical parameters related to salt content.

[0104] I(Yj) is used to measure the importance of the jth key feature indicator in the extracted feature vector. It can be determined by calculating the correlation coefficient between the key feature indicator and the extracted feature vector, such as the Pearson Correlation Coefficient: , where xk is the element in the extracted feature vector and x is its mean; yj,k is the element in the jth key feature indicator and yj is its mean. The value of I(Yj) ranges from -1 to 1. The closer the absolute value is to 1, the stronger the correlation between the key feature indicator and the extracted feature vector, that is, the more representative the extracted feature vector is of the key feature indicator. This reward function provides a quantitative reward standard for reinforcement learning agents by comprehensively considering feature richness (measured by information entropy) and feature representativeness (measured by the importance of key feature indicators).

[0105] Step a4: Select the target empty convolution kernel parameter with the largest reward value according to the reward value;

[0106] Specifically, the agent generates N = 20 candidate empty convolution kernel parameters (e.g., different combinations of k and d) in each iteration and calculates the reward value R for each candidate empty convolution kernel parameter. Non-maximum suppression (NMS) is used to filter similar parameters (e.g., only the one with higher reward is retained for k = 5, d = 2 and k = 5, d = 3) to reduce redundant evaluations.

[0107] Next, we select parameters whose reward value R is greater than the mean + σ (standard deviation) to avoid the noise caused by occasional high rewards. High-reward parameters (R > threshold) are stored in the experience pool, and the PPO algorithm is used to update the agent network, strengthening the probability of successful strategies.

[0108] Finally, the target empty convolution kernel parameters with the largest reward value are selected.

[0109] Step a5, based on the target convolution kernel sizes of different scales in the target empty convolution kernel parameters, feature extraction is performed on the image feature vector, spectral feature vector, electrochemical feature vector and mechanical feature vector, and the corresponding sub-initial image feature vector, sub-initial spectral feature vector, sub-initial electrochemical feature vector and sub-initial mechanical feature vector at each scale are output.

[0110] Specifically, for the image feature vector, a k×k two-dimensional dilated convolution is applied with a stride of 1 and a padding of (k−1) / 2 to preserve spatial resolution. Small scales (for example, k=3, d=1) are used to capture local details such as micropores and granular texture on the crust surface. Large scales (using k=7, d=3) extract the overall structure of the crust (such as the distribution of large-scale cracks).

[0111] For spectral and electrochemical feature vectors, we apply one-dimensional dilated convolution to capture long-range dependencies in the sequence. Small-scale convolution (e.g., k=3, d=1) focuses on subtle changes in specific wavelengths / potential ranges, while large-scale convolution (e.g., k=7, d=3) integrates trends across a wide frequency domain / long timeframe.

[0112] For the mechanical characteristic vector, if the mechanical characteristic vector is a scalar (such as hardness value), the scale effect is simulated through 1×1 convolution (fully connected layer); if the mechanical characteristic vector is a sequence (such as loading curve), one-dimensional convolution is used to extract the elastic and plastic segment features.

[0113] Finally, the sub-initial image feature vector, sub-initial spectral feature vector, sub-initial electrochemical feature vector and sub-initial mechanical feature vector corresponding to each scale are output.

[0114] For example, a three-scale feature vector is generated for each modality: The feature dimension is unified to 128 dimensions (controlled by the number of convolution kernels) to facilitate subsequent fusion.

[0115] In step S1013, based on the attention mechanism, feature fusion is performed on the sub-initial image feature vectors, sub-initial spectral feature vectors, sub-initial electrochemical feature vectors, and sub-initial mechanical feature vectors corresponding to each scale to generate target fusion features.

[0116] Specifically, the above step S1013 may include the following steps:

[0117] In step b1, the sub-initial image feature vectors, sub-initial spectral feature vectors, and sub-initial electrochemical feature vectors of the same scale are paired and combined to generate multiple cross-modal feature groups.

[0118] Specifically, the electronic device iterates over the sub-initial image feature vectors, sub-initial spectral feature vectors, and sub-initial electrochemical feature vectors at each scale. Then, the sub-initial image feature vectors, sub-initial spectral feature vectors, and sub-initial electrochemical feature vectors at the same scale are paired and combined to generate multiple cross-modal feature sets.

[0119] In step b2, for each feature group, the feature group is input into the attention calculation network to calculate the correlation between the two sub-feature vectors in the feature group.

[0120] Specifically, for each feature group, the feature group is input into the attention calculation network. The spatial-semantic branch of the attention calculation network calculates the correlation score corresponding to the two sub-feature vectors, and the causal branch calculates the causal correlation score corresponding to the two sub-feature vectors. Then, the correlation score corresponding to each sub-feature vector and the causal correlation score are combined to obtain the correlation between the two sub-feature vectors.

[0121] Step b3: Output the attention weight matrix based on the correlation.

[0122] Among them, the attention weight matrix is used to represent the degree of attention between two sub-feature vectors in the feature group.

[0123] Specifically, the electronic device outputs the correlation corresponding to the two sub-feature vectors in each feature group in the form of a matrix, and outputs an attention weight matrix.

[0124] Step b4: Update the two sub-feature vectors in the feature group based on the attention weight matrix.

[0125] Specifically, the electronic device "injects" the information of sub-feature vector A in the feature group into sub-feature vector B according to the attention weight matrix. The formula is: Where σ is the Sigmoid function, which controls the intensity of information injection;

[0126] Symmetrically, the information of sub-eigenvector B is “injected back” into sub-eigenvector A to ensure two-way interaction.

[0127] Then, the original feature sub-feature vector A+B is retained as the basis, and the updated features after interaction are superimposed to prevent the gradient from disappearing.

[0128] For example, when there is a strong causal correlation between the "salt absorption peak" in the spectral feature and the "crust crack" in the image feature, the forward update will enhance the crack-related dimension in the image feature, and the reverse update will enhance the response intensity of the salt peak in the spectral feature, forming a cross-modal causal feature enhancement.

[0129] Step b5, repeat this process to obtain the corresponding sub-updated image feature vector, sub-updated spectral feature vector, sub-updated electrochemical feature vector and sub-updated mechanical feature vector at each scale.

[0130] Specifically, three rounds of attention calculation and update iterations are performed on each scale feature group to gradually strengthen the causal dependencies between features. During the iteration process, the attention weight matrices of adjacent scales (such as small scale and medium scale) are forced to satisfy similarity constraints, as follows: This avoids conflicting attention allocations for the same modality pair at different scales (e.g., high image-spectrum correlation at small scales, but sudden unrelatedness at medium scales). This allows us to obtain the corresponding sub-updated image feature vectors, sub-updated spectral feature vectors, sub-updated electrochemical feature vectors, and sub-updated mechanical feature vectors at each scale.

[0131] Step b6, fuse each sub-updated image feature vector, each sub-updated spectral feature vector, each sub-updated electrochemical feature vector and each sub-updated mechanical feature vector according to the data modality to obtain the fused sub-target image feature vector, sub-target spectral feature vector, sub-target electrochemical feature vector and sub-target mechanical feature vector.

[0132] Specifically, for the sub-updated image feature vectors at each scale, non-local operations are used to integrate the spatial features at different scales (such as small-scale cracks + large-scale structures), highlight the multi-resolution texture of the crust surface, and obtain the fused sub-target image feature vector.

[0133] For the sub-updated spectral feature vectors at each scale, band grouping convolution is performed to aggregate the updated features of adjacent bands according to functional group categories (such as hydroxyl group and carbonate group) to enhance the chemical semantic expression and obtain the fused sub-target spectral feature vector.

[0134] Update the electrochemical feature vectors for each sub-target at each scale: Use the temporal convolutional network (TCN) to capture the time dependence of ion concentration changes and obtain the fused sub-target electrochemical feature vectors.

[0135] For the sub-updated mechanical feature vectors at each scale, the mechanical parameters at each scale (such as the mean of the elastic modulus and the variance of the hardness) are directly counted, the global mechanical properties are retained, and the fused sub-target mechanical feature vectors are obtained.

[0136] Step b7: fuse the sub-target image feature vector, the sub-target spectral feature vector, the sub-target electrochemical feature vector, and the sub-target mechanical feature vector to generate a target fusion feature.

[0137] Specifically, the electronic device can first fuse the modal pairs with strong causal association, fuse the sub-target spectral feature vector and the sub-target electrochemical feature vector to generate a "chemical-electrochemical feature";

[0138] Then, the sub-target image feature vector is fused with the chemical-electrochemical feature to generate the "structural-chemical feature";

[0139] Finally, the sub-target mechanical feature vector is fused with the structure-chemistry feature to generate the target fusion feature containing structure-chemistry-mechanics information.

[0140] The fusion logic is: TargetFusion = MLP(Concat(Struct-Chem, Mechanical)), where Struct-Chem is the fusion feature of image and spectral-electrochemical, and Mechanical is the mechanical feature. The multi-layer perceptron (MLP) is used to capture high-order causal relationships across modalities (such as "salt crystallization → crust hardness change").

[0141] In step S102 , the target fusion features and temperature time series data, humidity time series data, rainfall time series data, and rainfall salinity data are input into an initial causal analysis network, and the initial causal analysis network is trained to obtain a target causal analysis model.

[0142] Specifically, the above step S102 may include the following steps:

[0143] Step S1021: Generate a training set and a validation set based on the target fusion features and temperature time series data, humidity time series data, rainfall time series data, and rainfall salinity data.

[0144] Specifically, the electronic device can divide the target fusion features and temperature time series data, humidity time series data, rainfall time series data, and the salinity of rainfall in chronological order to ensure that the environmental data window [t−n,t] corresponds to the target fusion feature window [t,t+m] of the crust data, thereby reflecting the hysteresis effect.

[0145] Then, 70% of the dataset is divided into a training set and 30% is divided into a validation set.

[0146] Step S1022: input the training set into the input layer of the initial causal analysis network.

[0147] Specifically, the electronic device uses time position encoding (such as sinusoidal encoding) + causal label encoding (such as marking whether it is intervention data do(X)) to encode the temperature time series data, humidity time series data, and rainfall time series data in the training set. The formula is: ;

[0148] Then, the static target fusion features are concatenated with the dynamic environment time series features to form an input vector containing "crusting state - environment history".

[0149] The input layer of the initial causal analysis network contains two branches: the environmental time series branch, which processes sequence data such as temperature and humidity and outputs the dimension Denv, and the crust feature branch, which processes the target fusion feature and outputs the dimension Dfeat. The final input is Concat(Eenv,Ffusion), with the dimension Denv + Dfeat.

[0150] In step S1023, the hidden layer in the initial causal analysis network extracts features from the target fusion features and the temperature time series data, humidity time series data, rainfall time series data, and the salinity of rainfall, and outputs the target features.

[0151] Specifically, the hidden layer in the initial causal analysis network converts the environmental time series data Eenv=[T1,H1,R1,S1;T2,H2,R2,S2;...] (temperature, humidity, rainfall, salinity) into a fixed-dimensional feature vector and uses 1D convolution to extract time series patterns (such as sudden temperature drop, continuous high humidity, etc.).

[0152] Then, the control gate is calculated using the learnable weight matrix W1: σ(W1•Eenv), where σ is a Sigmoid function with output values between [0,1]. The control gate determines which time series features should be enhanced or suppressed (for example, only humidity changes directly related to crust cracks are allowed to pass). The control gate is element-wise multiplied by the convolution output (⊙) to obtain the direct causal feature F direct For example, a sudden increase in rainfall will directly lead to a change in the wetness of the crust surface, and this direct relationship of "rainfall → wetness" is selectively strengthened by the gating mechanism.

[0153] In addition, the hidden layer in the initial causal analysis network uses environmental variables (temperature, humidity, etc.), mediating variables (water content, salt concentration, etc.), and crust characteristics (image texture, spectral peak intensity, etc.) as nodes. Based on prior knowledge, possible causal relationships are predefined (such as temperature → water content, water content → salt migration), and the initial weight information of the edges corresponding to each node is determined.

[0154] Then, the node representation is iteratively updated through the graph convolutional network (GCN) to simulate the propagation process of causal effects: indirect =GCN(G(Eenv,Ffusion)). For example, information from the temperature node is gradually transmitted through the path of "temperature → moisture content → salt crystallization → spectral characteristics," ultimately affecting the spectral characteristics node of the crust. Compared to the first-order association of direct paths, indirect paths can model multi-step causal chains (e.g., temperature increase → moisture evaporation → increased salt concentration → change in crust hardness). GNN edge weights automatically learn the strength of each causal relationship (e.g., the weight of "temperature → moisture content" reflects the degree of influence of temperature changes on moisture).

[0155] Next, the direct path feature F direct and the indirect path feature F indirect, calculate the attention score: α=Softmax(MLP([Fdirect;Findirect])); α represents the relative importance of the direct path and the indirect path under the current input (e.g., under drought conditions, the indirect effect of temperature on crust hardness may be more important).

[0156] The two paths are dynamically fused based on the attention score to produce the target feature: Ftarget = α1 • Fdirect + α2 • Findirect. When rainfall suddenly increases, the model automatically increases the weight of the direct path (to capture the rapid response from rainfall to wetness); whereas, during prolonged drought conditions, the model increases the weight of the indirect path (to capture the slower process from temperature to evaporation to salt accumulation).

[0157] Step S1024: the output layer in the initial causal analysis network outputs the virtual causal relationship between the crust data and the environmental time series data based on the target features.

[0158] Specifically, the output layer of the initial causal analysis network uses a multi-layer perceptron (MLP) or linear regression model to map target features. For example, for each environmental variable (such as temperature T) and crust feature (such as the density of cracks in the image, Cimg), a causal effect value τ(T→Cimg) is calculated. This value quantifies the degree to which temperature changes affect the density of crust cracks. A positive value indicates a positive correlation (increased temperature leads to an increase in cracks), while a negative value indicates a negative correlation.

[0159] Traverse all combinations of environmental variables (temperature, humidity, rainfall, salinity) and crust characteristics (image, spectrum, electrochemical, mechanical data), calculate the causal effect value between each pair of variables, and form a causal effect matrix. For example, the causal effect matrix can be as follows: ;

[0160] Where T is temperature, H is humidity, Cimg is the image crack density, Cspec is the spectral salinity peak intensity, and τ(T→Cimg) is the causal effect of temperature on image crack density.

[0161] Finally, the output layer outputs the virtual causal relationship according to the causal effect matrix. For example, .

[0162] Step S1025 , based on the virtual causal relationship and the real causal relationship, calculate the loss value corresponding to the target loss function.

[0163] Among them, the true causal relationship was determined by analyzing the environmental time series data and crust data during the soil crust preparation process.

[0164] Specifically, based on the virtual causal relationship and the real causal relationship, the loss value corresponding to the target loss function is calculated.

[0165] Step S1026 , updating the parameters of the initial causal analysis network based on the loss value to obtain an initial causal analysis model.

[0166] Specifically, the electronic device updates parameters of the initial causal analysis network based on the loss value to obtain an initial causal analysis model.

[0167] In an optional embodiment of the present application, the data processing device 4 constructs a loss function based on the following steps, specifically including:

[0168] Step c1: construct a priori causal knowledge graph based on the prior knowledge of the causal relationship between environmental factors and crusting data.

[0169] Specifically, the electronic device can receive user-entered prior knowledge of the causal relationship between environmental factors and crusting data. For example, "increased humidity → decreased crusting air permeability" or "salt content → shift in spectral characteristic peaks" can be used. The electronic device then identifies the prior knowledge of the causal relationship between the environmental factors and the crusting data and extracts a causal relationship triplet: a cause node, a causal relationship, and a result node.

[0170] Then, a directed acyclic graph (DAG) model was used. Nodes included environmental variables (such as temperature and humidity) and crust characteristics (such as image texture and mechanical strength), and edges represented causal relationships (such as the direct effect of "temperature → moisture content" and the indirect effect of "moisture content → salt migration → spectral characteristics"). Each edge was assigned a priori causal strength (e.g., a weight range of -1 to 1 based on literature results, with negative values indicating negative causality) and a confidence level (e.g., a reliability score of 0.1-1.0 based on expert assessment).

[0171] Step c2: Map the virtual causal relationships output by the initial causal analysis network to the corresponding nodes and edges in the prior causal knowledge graph.

[0172] Specifically, the electronic device creates unique identifiers for environmental variables and crust characteristics (e.g., "T" for temperature, "C_spec_Salt" for spectral salt peak intensity), ensuring that the variable names in the model output exactly correspond to the graph nodes. For the virtual causal relationships output by the initial causal analysis network (e.g., "temperature → spectral moisture peak intensity"), the graph is searched for edges with the same starting and ending points. If such edges exist, the corresponding prior strength and confidence level are recorded. If not, the relationship is flagged as a newly discovered causal relationship or a potential conflict.

[0173] Step c3: Detect the conflict between the virtual causal relationship and the prior causal knowledge graph.

[0174] Specifically, check whether the initial causal analysis network predicts an inverse causal relationship (such as "crust hardness → rainfall"), which violates the physical logic of "environmental variables as causes and crust characteristics as effects".

[0175] Then, compare the virtual causal relationship τ predicted by the initial causal analysis network pred Correlation with prior knowledge τ prior , if satisfied or (δ is the tolerance threshold, such as 0.5), it is determined to be an intensity conflict.

[0176] For example, the prior knowledge suggests that "humidity → crust cracking" is positively correlated (τprior = 0.6). If the initial causal analysis network predicts a negative correlation (τpred = -0.3), a conflict is triggered.

[0177] In step c4, the causal consistency loss is calculated based on the conflict between the virtual causal relationship and the prior causal knowledge graph.

[0178] Specifically, high weights are assigned to directional conflicts (e.g., weight coefficient wdir = 10), intensity conflicts are weighted by confidence (e.g., for edges with a prior confidence of 0.9, the conflict loss is multiplied by 0.9), and causal consistency loss is calculated based on the conflict between the virtual causal relationship and the prior causal knowledge graph.

[0179] For example, the causal consistency loss function is as follows:

[0180]

[0181] in, is the conflict edge set, w uv is the edge weight, To avoid the minimum value with zero denominator, is the prior knowledge relevance, is a virtual causal relationship. When the model prediction is completely consistent with the prior, the loss is 0; the greater the conflict, the loss approaches wuv.

[0182] Step c5: classify the samples in the training set according to the causal relationship type between the crusting data and the environmental time series data, and determine the number of samples corresponding to each category.

[0183] Specifically, according to the direction of causal effect, it is divided into: positive correlation, negative correlation, and no significant correlation; according to the causal path, it is divided into: direct causality, indirect causality, and mixed causality (including both direct and indirect effects); according to the type of environmental variables, it is divided into: temperature correlation, humidity correlation, salt correlation, etc.

[0184] Each training sample is assigned to one of the above categories based on the combination of its corresponding environmental variables and crust characteristics. For example, "Temperature → Crust Hardness (Direct Positive Correlation)" belongs to the "Direct Causation - Positive Correlation - Temperature" category. When counting the number of samples in each category, for example, the "Indirect Causation" category may only account for 15% of the total data, indicating a class imbalance.

[0185] In step c6, different weights are assigned to samples of different categories based on the number of samples corresponding to each type, and a balanced loss function is determined.

[0186] Specifically, calculate the category weights , where N is the total number of samples, K is the number of categories, and nc is the number of samples in category c.

[0187] Example: If the number of samples in the "indirect causal" category nc = 150 and the total number of samples N = 1000, then wc = 3×1501000≈2.22 (assuming K = 3 categories).

[0188] Increasing the penalty for minority samples, the balanced loss function formula is: .

[0189] Where γ is the focusing parameter (e.g. 2.0), is the predicted probability of the model for sample i, and ci is the sample category.

[0190] Step c7: construct the target loss function based on the causal consistency loss, the balance loss function, and the mean square error loss function.

[0191] Specifically, based on the causal consistency loss, the balance loss function, and the mean square error loss function, the target loss function is constructed. The target loss function is as follows:

[0192] ,

[0193] in, is the mean square error loss, which measures the regression error between the predicted value and the true value of the crust feature to ensure the basic prediction ability of the model.

[0194] Step S1027 : verifying the initial causal analysis model based on the validation set. When the accuracy of the initial causal analysis model is greater than or equal to a preset accuracy threshold, obtaining a target causal analysis model.

[0195] Specifically, the electronic device inputs the verification set into the initial causal analysis model for verification, and verifies the verification set based on the target loss function. If the loss value corresponding to the target loss function is greater than or equal to the preset loss value threshold, then when it is determined that the accuracy of the initial causal analysis model is greater than or equal to the preset accuracy threshold, the target causal analysis model is obtained.

[0196] The soil crust preparation and analysis system provided in the embodiments of the present application inputs crust image data, crust spectral data, crust electrochemical data, and crust mechanical properties into corresponding embedding layers, generating corresponding image feature vectors, spectral feature vectors, electrochemical feature vectors, and mechanical feature vectors. This unifies the representation of different types of data (pixel matrix for images, wavelength intensity sequence for spectra) to facilitate subsequent fusion processing. Real-time statistical features corresponding to the image feature vectors, spectral feature vectors, electrochemical feature vectors, and mechanical feature vectors are calculated, ensuring the accuracy of the obtained real-time statistical features. The real-time statistical features corresponding to the image feature vectors, spectral feature vectors, electrochemical feature vectors, and mechanical feature vectors are input into an adaptive dilated convolution kernel generation mechanism based on a reinforcement learning algorithm. Using each real-time statistical feature as a state input, different initial dilated convolution kernel parameters are dynamically generated. This avoids the limitations of fixed-scale convolution (e.g., small kernels miss global trends, large kernels blur local details), improving the flexibility of feature extraction. Effective scales are selectively activated based on data characteristics, reducing redundant computations (e.g., skipping small-scale kernels for smoothed spectral data). Based on a preset reward function, the reward value corresponding to each initial empty convolution kernel parameter is calculated. Based on the reward value, the target empty convolution kernel parameters with the largest reward value are selected. The effectiveness of the convolution kernel is quantified through preset reward functions (such as information entropy, feature importance, and causal consistency), and the optimal parameters are selected to ensure that feature extraction takes into account information richness, causal relevance, and computational efficiency, avoiding overfitting of a single metric. Based on the target convolution kernel sizes at different scales in the target empty convolution kernel parameters, feature extraction is performed on image feature vectors, spectral feature vectors, electrochemical feature vectors, and mechanical feature vectors. The corresponding sub-initial image feature vectors, sub-initial spectral feature vectors, sub-initial electrochemical feature vectors, and sub-initial mechanical feature vectors at each scale are output, ensuring the accuracy of the sub-initial image feature vectors, sub-initial spectral feature vectors, sub-initial electrochemical feature vectors, and sub-initial mechanical feature vectors at each scale.

[0197] Then, the sub-initial image feature vectors, sub-initial spectral feature vectors, and sub-initial electrochemical feature vectors at the same scale are paired and combined to generate multiple cross-modal feature groups. This intermodal information exchange compensates for information missing from a single modality (for example, the lack of spatial location information in spectral data can be supplemented by image features). For each feature group, the feature group is input into the attention computation network, which calculates the correlation between the two sub-feature vectors within the feature group. Based on this correlation, an attention weight matrix is output. This avoids causal bias caused by unidirectional information flow and ensures the physical plausibility of the "environment-crust" causal chain. Based on the attention weight matrix, the two sub-feature vectors within the feature group are updated. This cycle continues to generate the corresponding sub-updated image feature vectors, sub-updated spectral feature vectors, sub-updated electrochemical feature vectors, and sub-updated mechanical feature vectors at each scale. By iteratively updating the feature vectors, high-order cross-modal causal relationships (such as the indirect path from "temperature → moisture → salt → structure") are gradually strengthened.

[0198] Next, each sub-updated image feature vector, each sub-updated spectral feature vector, each sub-updated electrochemical feature vector, and each sub-updated mechanical feature vector are fused according to data modality to obtain the fused sub-target image feature vector, sub-target spectral feature vector, sub-target electrochemical feature vector, and sub-target mechanical feature vector. The sub-target image feature vector, sub-target spectral feature vector, sub-target electrochemical feature vector, and sub-target mechanical feature vector are then fused to generate the target fusion feature. This gradual abstraction from the local (single modality, single scale) to the global (multimodal, multiscale) approach aligns with the multi-physical process characteristics of soil crust formation. The direct and indirect effects of environmental variables (such as temperature and rainfall) on crust characteristics are encoded into a unified feature space, supporting end-to-end learning of causal models.

[0199] Finally, training and validation sets were generated based on the target fusion features and temperature, humidity, rainfall, and salinity data. The training set was fed into the input layer of the initial causal analysis network. The hidden layer in the initial causal analysis network extracted features from the target fusion features, temperature, humidity, rainfall, and salinity data, and output the target features, ensuring the accuracy of the output target features. The output layer of the initial causal analysis network outputs a virtual causal relationship between the crusting data and the environmental time series data based on the target features. A priori causal knowledge graph was constructed based on prior knowledge of the causal relationship between environmental factors and crusting data. The virtual causal relationships output by the initial causal analysis network were mapped to corresponding nodes and edges in the priori causal knowledge graph. Conflicts between the virtual causal relationships and the priori causal knowledge graph were detected. Based on these conflicts, the causal consistency loss was calculated. The model-predicted virtual causal relationships were compared with domain prior knowledge (e.g., the "rainfall → crust moisture content" edge in the graph) to detect conflicts in directionality and intensity. This prevents the model from learning spurious associations that violate soil science principles (such as the inverse causal relationship between crust hardness and temperature). In scenarios where prior knowledge is reliable, graph information is leveraged to compensate for data deficiencies and improve model generalization. The training set samples are then classified according to the type of causal relationship between the crust data and the environmental time series data. The number of samples corresponding to each category is determined. Based on the number of samples corresponding to each category, different weights are assigned to samples of different categories, and a balanced loss function is determined. This prevents the model from favoring the majority class (such as direct causal samples) and ensures that features of the minority class (such as indirect causal samples) are fully learned. Simultaneously, prediction accuracy, causal consistency, and sample balance are optimized to enhance the model's robustness in complex scenarios. A target loss function is constructed based on the causal consistency loss, the balanced loss function, and the mean squared error loss function, ensuring its accuracy. The loss value corresponding to the target loss function is calculated based on the virtual and true causal relationships. The parameters of the initial causal analysis network are updated based on the loss value to obtain the initial causal analysis model, ensuring its accuracy. The initial causal analysis model is validated against the validation set. When the accuracy of the initial causal analysis model is greater than or equal to the preset accuracy threshold, the target causal analysis model is obtained. This avoids overfitting the training data and ensures the predictive reliability of the target causal analysis model in real-world environments (e.g., varying soil types and climate change). Errors discovered during the validation process are fed back into data collection or model structure design, forming a closed "train-validate-improve" loop. This ensures the accuracy of the resulting target causal analysis model.

[0200] In an optional embodiment of the present application, Figure 3 As shown, the data processing device 4 is further configured to perform the following steps:

[0201] Step S201: Obtain environmental change data corresponding to the target area within a preset time period before the current time, input the environmental change data into the target causal analysis model, and output future soil crust data corresponding to the target area.

[0202] Specifically, the electronic device may receive environmental change data of a target area within a preset time period before the current time, input by a user or sent by other devices. The environmental change data includes but is not limited to temperature data, humidity data, precipitation data, and salt content data in the precipitation.

[0203] The electronic device cleans the raw data (removes outliers, fills missing values), normalizes it, and organizes it into an input format recognizable by the model (such as a multi-dimensional feature matrix) in time series, and inputs it into the target causal analysis model.

[0204] The preprocessed environmental change data is then fed into a trained target causal analysis model (i.e., a causal analysis model optimized through pre-training, incorporating direct causal pathways, indirect causal pathways, and prior knowledge constraints). The target causal analysis model uses a direct causal mechanism decoupling network (CMDN) to separate the direct effects of environmental variables on soil crusts (e.g., rainfall directly affecting crust moisture content) from their indirect effects (e.g., temperature through water evaporation → salt crystallization → changes in crust mechanical strength). An attention mechanism is then used to dynamically weight the features of the direct and indirect pathways to generate target features that comprehensively reflect the causal relationship. Finally, the output layer of the target causal analysis model predicts soil crust data for future time periods. Soil crust data includes: physical properties (crust thickness, porosity, surface hardness, and crack density); chemical properties (salt content, organic matter content, and pH); spectral features (reflectance spectrum curve and characteristic band absorption intensity); and mechanical properties (compressive strength and shear strength).

[0205] Step S202: output corresponding management measures based on future soil crust data.

[0206] Specifically, the electronic device performs a multi-dimensional analysis of the future soil crust data output by the target causal analysis model, identifying key trends and outliers. For example, a significant increase in predicted crust thickness and a decrease in porosity could indicate a deterioration in soil permeability, impacting plant root growth; or a salt content exceeding a threshold could indicate an increased risk of soil salinization.

[0207] Then, combined with the ecological impact threshold of soil crust (such as agricultural production standards and ecological protection standards), the prediction results are divided into different risk levels (such as low, medium, and high risks).

[0208] Finally, based on the predicted crusting data and risk level, combined with the causal relationship of environmental factors, targeted management measures are generated. The following are typical scenarios and corresponding strategies:

[0209] Scenario 1: Increased soil crust thickness leads to soil compaction (direct cause and effect: rainfall → soil crust compaction). Short-term measures include mechanical soil loosening (such as shallow rotary tillage) to break up the surface layer of the crust and improve soil aeration; mulching with organic materials (such as straw and humus) reduces surface runoff and inhibits soil crust formation. Long-term measures include optimizing irrigation methods (such as drip irrigation instead of flooding) to reduce the impact of rainfall on the ground surface; and planting cover crops with well-developed root systems (such as legumes) to enhance soil structural stability through root activity.

[0210] Scenario 2: Salt accumulation triggers soil salinization (indirect causal relationship: temperature → water evaporation → salt migration and concentration). Short-term measures: Implement leaching improvement: Reduce surface soil salt concentration through freshwater irrigation; apply soil conditioners (such as gypsum) to adjust soil pH and promote salt leaching. Long-term measures: Develop drainage systems to prevent rising groundwater levels and the upward movement of salt; promote salt-tolerant crop varieties (such as Suaeda salsa) to reduce soil salt content through biosorption.

[0211] Scenario 3: Reduced mechanical strength of the soil crust leads to wind erosion risk (combined causal relationships: wind speed → surface wear + moisture content → decreased adhesion). Short-term measures: Spray soil stabilizers (such as polymers) to enhance the surface erosion resistance of the soil crust; temporarily cover with dust screens or lay gravel layers to reduce direct wind erosion. Long-term measures: Increase vegetation cover (such as planting windbreak strips) to stabilize the soil through plant roots; adjust land use patterns (such as returning farmland to forest and grassland) to reduce surface exposure.

[0212] Prioritize management measures based on risk level and implementation costs. For example, high-risk scenarios prioritize short-term emergency measures, while medium- and low-risk scenarios prioritize long-term ecological restoration. Furthermore, establish a closed-loop monitoring mechanism to regularly compare actual environmental data with model predictions to evaluate the effectiveness of measures and adjust strategies. For example, if the salt concentration after leaching and amelioration does not meet expectations, increase the amount of ameliorator or adjust the leaching frequency.

[0213] The soil crust preparation and analysis system provided in the embodiments of the present application, including the data processing device 4, is further configured to obtain environmental change data corresponding to a target area within a preset period of time prior to the current time, input the environmental change data into a target causal analysis model, and output future soil crust data corresponding to the target area, thereby ensuring the accuracy of the output future soil crust data corresponding to the target area. Based on the future soil crust data, corresponding management measures are output, thereby ensuring the correspondence between the output management measures and the future soil crust data.

[0214] In an optional embodiment of the present application, Figure 4 As shown, the soil preparation device 1 includes:

[0215] Preparation support 11, including a support frame and a preparation container, the preparation container is placed on the support frame;

[0216] A soil crust cultivation soil column 12 is placed inside the preparation container and is used to prepare the soil crust;

[0217] A rainmaker 13 is installed on the top surface of the preparation container to rain on the soil crust cultivation column 12;

[0218] The pressure regulating nozzle 112 is installed on the rain generator 13 to simulate the amount of rainfall;

[0219] One end of the rainmaker 13 is connected to the salt dissolution preparation device 16 through the first water pump 14 and the first valve 15, and the other end is connected to the pure water preparation device 19 through the second water pump 17 and the second valve 18; the first valve 15 is opened to spray the salt solution onto the soil crust cultivation column 12; the second valve 18 is opened to spray the pure water onto the soil crust cultivation column 12;

[0220] A solution collection bucket 110 is placed inside the support frame and below the preparation container to collect liquid flowing down from the soil crust cultivation soil column 12;

[0221] The heating device 111 is used to change the ambient temperature during the soil crust preparation process.

[0222] Specifically, rainfall control involves varying the water pressure through the pressure-regulating nozzle 112 to simulate rainfall intensity (e.g., light rain, moderate rain, or heavy rain). Salt gradient control involves the salt dissolution device 16 providing salt solutions of varying concentrations, with the first valve 15 controlling the proportion of the solution sprayed to simulate the salinity of actual rainfall (e.g., high-salt precipitation in coastal areas). The purified water device 19 provides a control group of low-salt / salt-free precipitation. Temperature control involves increasing or decreasing the ambient temperature through the heating device 111 (e.g., heating wires, temperature control system), thereby affecting the soil water evaporation rate, salt crystallization, and microbial activity, thereby regulating the physical and chemical properties of the soil crust (e.g., hardness and porosity). During the cultivation process, environmental data (e.g., temperature, humidity, rainfall, and salinity) and crust data (e.g., moisture content, mechanical strength, and spectral characteristics) are simultaneously monitored to provide training samples for the causal analysis model, analyzing the direct and indirect effects of various factors on crust formation (e.g., the causal path of "salt solution → soil salt migration → increased crust hardness").

[0223] The working process is detailed as follows:

[0224] Step 1: Device construction and soil column filling;

[0225] Place the soil crust cultivation column 12 into the preparation container. The soil column must be filled in layers according to the natural soil structure (such as a surface humus layer and a bottom sand layer) to ensure the authenticity of the simulation. A rain generator 13 is fixed to the center of the top surface of the preparation container, with the pressure regulating nozzle 112 facing the surface of the soil column to ensure uniform rainfall coverage. The solution collection bucket 110 is placed below the preparation container and connected to the drain port at the bottom of the container via a pipe to collect the permeate. The heating device 111 surrounds the outside or bottom of the preparation container, and the target temperature (such as 20°C, 30°C, etc.) is set by the temperature control system.

[0226] Step 2: Setting environment variable parameters;

[0227] Rainfall Simulation: Based on experimental requirements, the rainfall intensity can be preset using the pressure-regulated nozzle 112 (e.g., 50 mm / h for heavy rain). Salt Treatment Options: To simulate a high-salinity environment, open the first valve 15 and start the first water pump 14, pumping the salt solution from the salt dissolution preparation device 16 to the rainmaker 13 for spraying onto the soil column surface. To simulate low-salinity or natural precipitation, close the first valve 15, open the second valve 18, and pump purified water via the second water pump 17. Temperature Setting: Enter the target temperature (e.g., 35°C for simulating a high-temperature summer environment) via the control panel of the heating device 111. The device automatically maintains this constant temperature.

[0228] Step 3: Soil crust cultivation and data collection;

[0229] The rain generator 13 is activated to continuously spray a solution (salt solution or purified water) according to preset parameters, simulating a natural rainfall process. During rainfall, water infiltrates and evaporates from the soil surface. Salt migrates to the surface with the water and gradually crystallizes, causing soil particles to cement together and form a crust.

[0230] Collect data regularly: Environmental data: record real-time temperature, humidity, accumulated rainfall and solution salinity through temperature sensors, humidity sensors and flow meters.

[0231] Crust data: The mechanical strength of the crust was measured by a micro hardness tester, the chemical composition of the crust surface (such as salt content) was detected by a spectrometer, and the change in soil moisture content was measured by weighing.

[0232] The permeate flows into the solution collection barrel 110 through the bottom of the preparation container, and samples can be taken regularly to analyze the salt concentration, organic matter content, etc., to reflect the material migration process inside the soil.

[0233] Step 4: Multiple control experiments and model training;

[0234] By adjusting single variables (such as rainfall, salinity, and temperature), multiple controlled experiments were designed: Control group: purified water at 20°C with low rainfall, simulating a clean, low-humidity environment; Experimental group 1: a high-concentration saline solution at 30°C with high rainfall, simulating a hot, rainy coastal environment; and Experimental group 2: a low-concentration saline solution at 15°C with moderate rainfall, simulating a humid, temperate climate. Each experiment was repeated multiple times to accumulate crusting data under different environmental combinations. This data was then used to train causal analysis models (such as the CMDN network) to analyze direct and indirect causal relationships between environmental variables and crusting characteristics (e.g., the indirect path of "increased temperature → accelerated water evaporation → salt crystallization → increased crust hardness").

[0235] Step 5: Cleaning and maintenance of the device;

[0236] After the experiment, close all valves and pumps, drain any remaining solution from the pipes, and rinse the rain gauge 13 and nozzles with purified water to prevent blockage by salt crystallization. Clean the soil column residue from the preparation container and disinfect it for later use. Clean and dry the solution collection bucket 110 to prevent microbial growth. Check the temperature control system of the heating equipment 111 to ensure proper functioning and ensure the accuracy of the parameters for the next experiment.

[0237] The soil crust preparation and analysis system provided in the embodiments of the present application includes a preparation support 11, which includes a support frame and a preparation container, with the preparation container placed on the support frame. A soil crust cultivation column 12 is placed inside the preparation container and is used to prepare a soil crust. A rainmaker 13 is mounted on the top surface of the preparation container and is used to rain rain onto the soil crust cultivation column 12. A pressure-regulating nozzle 112 is mounted on the rainmaker 13 and is used to simulate the amount of rainfall. By adjusting the pressure parameters, the amount of rainfall can be precisely controlled (e.g., varying intensities such as light rain, moderate rain, and heavy rain), simulating the dynamic process of natural rainfall and studying the effects of rainfall intensity on soil crust formation (e.g., scouring force and changes in infiltration rate). One end of the rainmaker 13 is connected to a salt dissolution preparation device 16 via a first water pump 14 and a first valve 15, and the other end is connected to a purified water preparation device 19 via a second water pump 17 and a second valve 18. When the first valve 15 is opened, a salt solution is sprayed onto the soil crust cultivation column 12; when the second valve 18 is opened, purified water is sprayed onto the soil crust cultivation column 12. Spraying a salt solution can simulate the effects of high-salinity environments (such as saline-alkali land and coastal areas) on soil crust formation, allowing for investigation of the relationship between salt accumulation and crust hardness and air permeability. Spraying pure water can serve as a control experiment, isolating the effects of rainfall (which does not contain salt) on soil crust formation and enhancing the comparability of experimental data. A solution collection bucket 110, placed within the support frame and below the preparation container, collects liquid flowing from the soil crust cultivation column 12. A heating device 111 is used to alter the ambient temperature during the soil crust preparation process. This device, by integrating multivariable controllable simulation (rainfall, salinity, and temperature), a standardized experimental unit (soil column), and full-process data collection (liquid collection and environmental parameter monitoring), enables detailed research on the soil crust formation process. Combining scientific research, flexibility, and ease of operation, it provides an efficient experimental platform for both basic research and applied practice in the field of soil environment.

[0238] In an optional embodiment of the present application, Figure 4 As shown, the crust data acquisition device 3 includes:

[0239] A camera device 31 is used to collect soil crust image data corresponding to the soil crust;

[0240] Near infrared spectroscopy equipment 32, used to collect crust spectrum data;

[0241] an electrochemical sensor group 33 for collecting electrochemical data of the crust;

[0242] The soil mechanical sensor 34 is used to collect the mechanical properties of the soil crust.

[0243] Specifically, a high-resolution camera or microscope imaging system is used to capture images of the soil crust surface and cross-section. Image acquisition software controls shooting parameters (such as focal length, exposure time, and resolution) to obtain image data at different angles and magnifications. These images intuitively demonstrate the surface morphology, structural texture, pore distribution, and other characteristics of the crust, providing visual data for subsequent morphological analysis and change monitoring.

[0244] Crust spectral data collection: Spectrometers (such as near-infrared spectrometers and Raman spectrometers) are used to measure the soil crust's spectrum. The spectrometer emits light of specific wavelengths to illuminate the crust sample. The spectrometer captures the sample's absorption, reflection, and scattering characteristics of these wavelengths and converts them into spectral data. By analyzing this spectral data, information such as the chemical composition, mineral composition, and organic matter content of the crust can be obtained, revealing the crust's material composition and chemical properties.

[0245] Crust electrochemical data acquisition: Using an electrochemical workstation, electrode probes are placed in contact with the soil crust to measure electrochemical parameters such as potential and current. By applying different electrical signals (such as constant potential and constant current), the electrochemical response of the soil crust is recorded. This allows analysis of the soil crust's ion exchange capacity, redox characteristics, and charge transport properties, providing a deeper understanding of the soil crust's properties and reaction processes from an electrochemical perspective.

[0246] Crust mechanical properties collection: Mechanical testing instruments (such as micro-indenters and tensile testing machines) are used to test the mechanical properties of soil crusts. For example, a micro-indenter applies pressure to the crust surface using an indenter, measuring the relationship between indentation depth and pressure change, and calculating mechanical parameters such as the crust's hardness and elastic modulus. A tensile testing machine can measure the crust's tensile strength, ductility, and other properties, thereby obtaining data on the crust's behavior and performance under mechanical loads.

[0247] The soil crust preparation and analysis system provided in the embodiments of this application includes a camera 31 that collects crust image data corresponding to the soil crust; a near-infrared spectrometer 32 that collects crust spectral data; an electrochemical sensor group 33 that collects crust electrochemical data; and a soil mechanical sensor 34 that collects crust mechanical properties. Through the coordinated monitoring of multiple devices, including cameras, spectroscopy, electrochemistry, and mechanics, a comprehensive "appearance-chemistry-physics" characterization of soil crusts is achieved. This not only provides multi-source data input for causal analysis models but also enhances the accuracy, reliability, and timeliness of causal relationship inference through data cross-validation and dynamic time series analysis. This three-dimensional monitoring system is suitable for both controlled experiments under laboratory conditions (such as simulating different rainfall / temperature scenarios) and can be expanded to long-term field monitoring, providing strong technical support for the scientific research and practical management of soil crusts.

[0248] Although the embodiments of the present invention have been described with reference to the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention. Such modifications and variations are all within the scope defined by the appended claims.

Claims

1. A soil crust preparation and analysis system, characterized in that: The soil crust preparation and analysis system includes a soil preparation device, an environmental data acquisition device, a crust data acquisition device, and a data processing device. The environmental data acquisition device and the crust data acquisition device are both communicatively connected to the data processing device, wherein: The soil preparation equipment is used to prepare soil crust; The environmental data acquisition device is used to collect environmental time series data during the soil crust preparation process; the environmental time series data includes temperature time series data, humidity time series data, rainfall time series data, and rainfall salt content data; The crust data acquisition device is used to collect crust data corresponding to the soil crust; the crust data includes crust image data, crust spectral data, crust electrochemical data, and crust mechanical properties; The data processing device is used to input the crust image data, crust spectral data, crust electrochemical data and crust mechanical properties into corresponding embedding layers respectively, and generate corresponding image feature vectors, spectral feature vectors, electrochemical feature vectors and mechanical feature vectors; Based on the multi-scale feature fusion network, the image feature vector, spectral feature vector, electrochemical feature vector and mechanical feature vector are extracted, and the corresponding sub-initial image feature vector, sub-initial spectral feature vector, sub-initial electrochemical feature vector and sub-initial mechanical feature vector at each scale are output; The sub-initial image feature vectors, sub-initial spectral feature vectors, and sub-initial electrochemical feature vectors of the same scale are paired and combined to generate multiple cross-modal feature groups; For each feature group, the feature group is input into the attention calculation network to calculate the correlation between the two sub-feature vectors in the feature group; Based on the correlation, the attention weight matrix is output; the attention weight matrix is used to represent the degree of attention between two sub-feature vectors in the feature group; Based on the attention weight matrix, the two sub-feature vectors in the feature group are updated; This cycle is repeated to obtain the corresponding sub-updated image feature vector, sub-updated spectral feature vector, sub-updated electrochemical feature vector, and sub-updated mechanical feature vector at each scale. Fusing each sub-updated image feature vector, each sub-updated spectral feature vector, each sub-updated electrochemical feature vector, and each sub-updated mechanical feature vector according to the data modality to obtain a fused sub-target image feature vector, sub-target spectral feature vector, sub-target electrochemical feature vector, and sub-target mechanical feature vector; The sub-target image feature vector, sub-target spectral feature vector, sub-target electrochemical feature vector and sub-target mechanical feature vector are fused to generate target fusion features; The target fusion features and temperature time series data, humidity time series data, rainfall time series data and rainfall salinity data are input into the initial causal analysis network, and the initial causal analysis network is trained to obtain the target causal analysis model.

2. The soil crust preparation and analysis system according to claim 1, characterized in that: The data processing device is used to calculate the real-time statistical features corresponding to the image feature vector, the spectral feature vector, the electrochemical feature vector, and the mechanical feature vector respectively; Inputting the real-time statistical features corresponding to the image feature vector, the spectral feature vector, the electrochemical feature vector, and the mechanical feature vector into an adaptive dilated convolution kernel generation mechanism based on a reinforcement learning algorithm; Using each of the real-time statistical features as state input, dynamically generating different initial empty convolution kernel parameters; each of the initial empty convolution kernel parameters includes initial convolution kernel sizes of different scales; Based on the preset reward function, calculate the reward value corresponding to each initial empty convolution kernel parameter; According to the reward value, select the target empty convolution kernel parameter with the largest reward value; Based on the target convolution kernel sizes of different scales in the target empty convolution kernel parameters, feature extraction is performed on the image feature vector, the spectral feature vector, the electrochemical feature vector, and the mechanical feature vector, and the corresponding sub-initial image feature vector, the sub-initial spectral feature vector, the sub-initial electrochemical feature vector, and the sub-initial mechanical feature vector at each scale are output.

3. The soil crust preparation and analysis system according to claim 1, characterized in that: The data processing device is used to generate a training set and a validation set based on the target fusion feature and the temperature time series data, the humidity time series data, the rainfall time series data, and the salinity data of the rainfall; Inputting the training set into the input layer of the initial causal analysis network; The hidden layer in the initial causal analysis network performs feature extraction on the target fusion feature and the temperature time series data, the humidity time series data, the rainfall time series data, and the salinity of the rainfall, and outputs the target feature; The output layer in the initial causal analysis network outputs a virtual causal relationship between the crust data and the environmental time series data based on the target feature; Calculating a loss value corresponding to a target loss function based on the virtual causal relationship and the real causal relationship; the real causal relationship is determined by analyzing the environmental time series data and the soil crust data during the soil crust preparation process; updating the parameters of the initial causal analysis network based on the loss value to obtain an initial causal analysis model; The initial causal analysis model is verified based on the verification set, and when the accuracy of the initial causal analysis model is greater than or equal to a preset accuracy threshold, the target causal analysis model is obtained.

4. The soil crust preparation and analysis system according to claim 3, characterized in that: The data processing device is used to construct a priori causal knowledge graph based on the priori knowledge of the causal relationship between environmental factors and crusting data; Mapping the virtual causal relationship output by the initial causal analysis network to corresponding nodes and edges in the prior causal knowledge graph; detecting conflicts between the virtual causal relationship and the prior causal knowledge graph; Calculating causal consistency loss based on the conflict between the virtual causal relationship and the prior causal knowledge graph; Classifying the samples in the training set according to the causal relationship type between the crust data and the environmental time series data, and determining the number of samples corresponding to each category; Based on the number of samples corresponding to each type, different weights are assigned to samples of different categories to determine the balanced loss function; Based on the causal consistency loss, the balance loss function and the mean square error loss function, a target loss function is constructed.

5. The soil crust preparation and analysis system according to claim 1, characterized in that: The data processing device is further configured to obtain environmental change data corresponding to a target area within a preset time period before a current time, input the environmental change data into the target causal analysis model, and output future soil crust data corresponding to the target area; Based on the future soil crust data, corresponding management measures are output.

6. The soil crust preparation and analysis system according to claim 1, characterized in that: The soil preparation equipment includes: A preparation rack comprises a support rack and a preparation container, wherein the preparation container is placed on the support rack; A soil crust cultivation soil column is placed inside the preparation container and is used to prepare the soil crust; a rainmaker, mounted on the top surface of the preparation container, for raining rain on the soil crust cultivation column; A pressure regulating nozzle is installed on the rainmaker to simulate the amount of rainfall; One end of the rainmaker is connected to a salt dissolution preparation device via a first water pump and a first valve, and the other end is connected to a pure water preparation device via a second water pump and a second valve; the first valve is opened to spray a salt solution onto the soil crust cultivation column; the second valve is opened to spray pure water onto the soil crust cultivation column; A solution collection bucket is placed inside the support frame and below the preparation container to collect liquid flowing down from the soil crust cultivation soil column; A heating device is used to change the ambient temperature during the soil crust preparation process.

7. The soil crust preparation and analysis system according to claim 1, characterized in that: The crust data acquisition device comprises: A camera device, used for collecting the soil crust image data corresponding to the soil crust; A near-infrared spectroscopy device for collecting the crust spectral data; An electrochemical sensor group, used for collecting electrochemical data of the crust; The soil mechanical sensor is used to collect the mechanical properties of the soil crust.

Images