Soil crust preparation and analysis system

Abstract

The invention relates to the technical field of soil crust analysis, in particular to a soil crust preparation and analysis system. Comprising a soil preparation device, an environment data acquisition device, a crust data acquisition device and a data processing device, the environment data acquisition device and the crust data acquisition device are both in communication connection with the data processing device, and the soil preparation device is used for preparing soil crust; the environment data acquisition equipment is used for acquiring environment time sequence data in a soil crust preparation process; the crust data acquisition equipment is used for acquiring crust data corresponding to soil crust; and the data processing equipment is used for training to obtain a target causal analysis model based on the environment time sequence data and the crust data. The soil crust preparation and analysis system provided by the invention can be used for collecting and analyzing data.

Description

Technical Field

[0001] The present invention relates to the technical field of soil crust analysis, and particularly 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, the research on soil crust mainly focuses on the formation mechanism, physical and chemical properties of the crust, and its impact on soil erosion resistance.

[0003] However, the existing soil crust preparation and detection equipment still faces many technical problems in practical applications. First, the existing technology lacks the function of in-situ real-time detection. Most equipment can only detect parameters through manual sampling, which not only increases the experimental period but also affects the timeliness and accuracy of data. In addition, the control accuracy of the existing equipment for the crust formation conditions (such as temperature, humidity, salt gradient, etc.) is insufficient, and it is impossible to accurately simulate the dynamic changes in the natural environment, resulting in the experimental results not having high repeatability and stability. Finally, although some equipment can monitor certain characteristics of soil crust, there is a lack of in-depth research on the dynamic relationship between crust characteristics and soil parameters, and the comprehensive analysis and intelligent control of data cannot be achieved.

[0004] Therefore, it has become an urgent problem to provide a soil crust preparation and analysis system that can collect and analyze data. 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 collection device, a crust data collection device, and a data processing device. The environmental data collection device and the crust data collection device are both communicatively connected to the data processing device, wherein: The soil preparation device is used to prepare soil crust; The environmental data collection 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 the salt content data of rainfall; The crust data collection device is used to collect crust data corresponding to the soil crust; the crust data includes crust image data, crust spectral data, crust electrochemistry data, and crust mechanical properties; The data processing device is used to train a target causal analysis model based on the environmental time-series data and the crust data.

[0007] The soil crust preparation and analysis system provided by the embodiments of the present application includes a soil preparation device for preparing soil crust; an environmental data acquisition device for acquiring environmental time-series data during the soil crust preparation process to capture dynamic environmental changes during the crust preparation process, such as sudden changes in rainfall intensity and diurnal temperature fluctuations. A crust data acquisition device for acquiring crust data corresponding to the soil crust ensures the accuracy of the acquired crust data. Therefore, there is no need for manual sampling for parameter detection, which reduces the experimental period and also ensures the real-time and accuracy of the acquired environmental time-series data. A data processing device for training a target causal analysis model based on the environmental time-series data and crust data ensures the accuracy of the obtained target causal analysis model. Thus, an in-depth study of the dynamic relationship between crust characteristics and soil parameters is achieved. The soil crust preparation and analysis system realizes precise data acquisition through device collaboration and in-depth causal modeling through algorithm innovation, constructing a complete technical chain from experimental preparation to intelligent analysis. It 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-causality dual-driven", providing an efficient technical tool for understanding the complex ecological processes of soil crust and addressing soil degradation problems, with significant scientific significance and practical application potential.

[0008] In an alternative embodiment, the data processing device is configured to input the crust image data, crust spectral data, crust electrochemical data, and crust mechanical properties into a preset feature fusion model to perform feature fusion on the crust image data, crust spectral data, crust electrochemical data, and crust mechanical properties, and generate target fusion features; Input the target fusion features and temperature time-series data, humidity time-series data, rainfall time-series data, and salt content data of rainfall into an initial causal analysis network, and train the initial causal analysis network to obtain a target causal analysis model.

[0009] In an alternative embodiment, the data processing device is configured to input the crust image data, crust spectral data, crust electrochemical data, and crust mechanical properties into corresponding embedding layers respectively to generate corresponding image feature vectors, spectral feature vectors, electrochemical feature vectors, and mechanical feature vectors; Based on a multi-scale feature fusion network, perform feature extraction on the image feature vector, spectral feature vector, electrochemical feature vector, and mechanical feature vector, and output 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; Based on the attention mechanism, perform feature fusion on 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 to generate target fusion features.

[0010] In an alternative embodiment, a data processing device is configured to calculate real-time statistical features corresponding to an image feature vector, a spectral feature vector, an electrochemical feature vector, and a mechanical feature vector respectively; Input 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 into an adaptive dilated convolution kernel generation mechanism based on a reinforcement learning algorithm; Taking each real-time statistical feature as a state input, dynamically generate different initial empty convolution kernel parameters; different scales of initial convolution kernel sizes are included in each initial empty convolution kernel parameter; Based on a preset reward function, calculate the reward values corresponding to each initial empty convolution kernel parameter; According to the reward values, select the target empty convolution kernel parameter with the largest reward value; Based on the different scales of target convolution kernel sizes in the target empty convolution kernel parameter, perform feature extraction on the image feature vector, the spectral feature vector, the electrochemical feature vector, and the mechanical feature vector, 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.

[0011] In an alternative embodiment, a data processing device is configured to pair and combine the sub-initial image feature vectors, sub-initial spectral feature vectors, and sub-initial electrochemical feature vectors of the same scale in pairs to generate multiple groups of cross-modal feature groups; For each feature group, input the feature group into an attention calculation network to calculate the correlation between the two sub-feature vectors in the feature group; Based on the correlation, output an attention weight matrix; the attention weight matrix is used to characterize the degree of attention between the two sub-feature vectors in the feature group; Based on the attention weight matrix, update the two sub-feature vectors in the feature group; 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; Fuse the sub-updated image feature vectors, sub-updated spectral feature vectors, sub-updated electrochemical feature vectors, and sub-updated mechanical feature vectors 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; Fuse the sub-target image feature vector, sub-target spectral feature vector, sub-target electrochemical feature vector, and sub-target mechanical feature vector to generate a target fusion feature.

[0012] In an alternative embodiment, a data processing device is configured to generate a training set and a validation set based on the target fusion features, temperature time-series data, humidity time-series data, rainfall time-series data, and the salt content data of rainfall. Input the training set into the input layer of the initial causal analysis network. The hidden layer in the initial causal analysis network extracts features from the target fusion features, temperature time-series data, humidity time-series data, rainfall time-series data, and the salt content data of rainfall, and outputs target features. 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. Calculate the loss value corresponding to the target loss function based on the virtual causal relationship and the true causal relationship; the true causal relationship is determined by analyzing the environmental time-series data and the crust data during the soil crust preparation process. Update the parameters of the initial causal analysis network based on the loss value to obtain the initial causal analysis model. Validate 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 the preset accuracy threshold, obtain the target causal analysis model.

[0013] In an alternative embodiment, a data processing device is configured to construct a prior causal knowledge graph based on the prior knowledge of the causal relationship between environmental factors and crust data. Map the virtual causal relationship output by the initial causal analysis network to the corresponding nodes and edges in the prior causal knowledge graph. Detect the conflict between the virtual causal relationship and the prior causal knowledge graph. Calculate the causal consistency loss based on the conflict between the virtual causal relationship and the prior causal knowledge graph. 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. Assign different weights to the samples of different categories based on the number of samples corresponding to each type, and determine the balanced loss function. Construct the target loss function based on the causal consistency loss, the balanced loss function, and the mean squared error loss function.

[0014] In an alternative embodiment, the data processing device is further configured to obtain the environmental change data within a preset duration before the current time corresponding to the target area, input the environmental change data into the target causal analysis model, and output the future soil crust data corresponding to the target area. Output the corresponding management measures based on the future soil crust data.

[0015] In an alternative embodiment, the soil preparation device includes: A preparation bracket, including a support frame and a preparation container, with the preparation container placed on the support frame; A soil crust cultivation soil column, placed inside the preparation container for preparing the soil crust; A rainfall device, installed on the top surface inside the preparation container for raining on the soil crust cultivation soil column; A pressure regulating nozzle, installed on the rainfall device for simulating the amount of rainfall; One end of the rainfall device 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, a salt solution is sprayed onto the soil crust cultivation soil column; when the second valve is opened, pure water is sprayed onto the soil crust cultivation soil column; A solution collection bucket, placed inside the support frame and below the preparation container for collecting the liquid flowing from the soil crust cultivation soil column; A heating device for changing the environmental temperature during the soil crust preparation process.

[0016] In an alternative embodiment, the crust data acquisition device includes: A camera device for acquiring the crust image data corresponding to the soil crust; A near-infrared spectroscopy device for acquiring the crust spectral data; An electrochemistry sensor group for acquiring the crust electrochemistry data; A soil mechanics sensor for acquiring the crust mechanical properties.

[0017] The soil crust preparation and analysis system provided by the embodiments of the present application prepares the soil crust based on the soil preparation device, realizing a refined study of the soil crust formation process. Then, the environmental time-series data during 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, 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 trained based on the training set and the validation set. This ensures the accuracy of the obtained target causal analysis model. Thus, the future soil crust number corresponding to the target area can be determined based on the target causal analysis model, and then corresponding management measures can be output. This ensures the correspondence between the output management measures and the future soil crust data. It realizes early response to the soil crust and reduces the negative impact of the soil crust on plant growth. Description of the Drawings

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for use in the description of the specific embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0019] Figure 1 is a schematic structural diagram of a soil crust preparation and analysis system according to an embodiment of the present invention; Figure 2 is a schematic flowchart of training to obtain a target causal analysis model according to an embodiment of the present invention; Figure 3 is a schematic flowchart of predicting future soil crust data and outputting corresponding management measures according to an embodiment of the present invention; Figure 4 is a schematic structural diagram of another soil crust preparation and analysis system according to an embodiment of the present invention.

[0020] Wherein: soil preparation device 1; preparation support 11; soil crust cultivation soil column 12; rainfall device 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 bucket 110; heating device 111; pressure regulating nozzle 112; environmental data acquisition device 2; crust data acquisition device 3; data processing device 4. Specific Embodiments

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0022] The embodiments of the present application provide a soil crust preparation and analysis system, as Figure 1 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: The soil preparation device 1 is used to prepare the soil crust; An environmental data acquisition device 2 is used to acquire environmental time-series data during the preparation process of soil crust. The environmental time-series data includes temperature time-series data, humidity time-series data, rainfall time-series data, and the salt content data of rainfall. A crust data acquisition device 3 is used to acquire crust data corresponding to the soil crust. The crust data includes crust image data, crust spectral data, crust electrochemistry data, and crust mechanical properties. A data processing device 4 is used to train a target causal analysis model based on the environmental time-series data and the crust data.

[0023] Specifically, the soil preparation device 1 can prepare soil crust. During the preparation process of the soil crust, the temperature change of the preparation environment is monitored in real time through a temperature sensor (such as a thermocouple or a thermistor sensor) in the environmental data acquisition device 2. The 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 after signal processing, it is also transmitted to the data processing device 4 to record the dynamic change of the environmental humidity. Optionally, the pressure sensor in the environmental data acquisition device 2 can collect the pressure in the pure water preparation device 19 in the soil preparation device 1 to determine the 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 salt content data of rainfall.

[0024] The crust data acquisition device 3 can acquire the crust data corresponding to the soil crust. After receiving the environmental time-series data and the crust data, the data processing device 4 trains the target causal analysis model using data processing and machine learning algorithms. First, preprocess the collected raw data, including data cleaning (removing outliers, handling missing values), data standardization (unifying the data scale), feature extraction (extracting key features from image, spectral, etc. data), etc., to convert the data into a format suitable for model training. Then, based on the preprocessed data, select a suitable causal analysis model (such as a structural causal model, causal forest, etc.) for training. During the training process, the model learns the association between environmental factors (temperature, humidity, rainfall, and salt content, etc.) and soil crust characteristics (image features, spectral features, electrochemical features, mechanical properties, etc.) to identify the causal relationships. For example, the model may find that humidity changes have a significant causal impact on the formation speed and structure of the crust, or there is a causal link between the rainfall salt content and the chemical composition of the crust. 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, etc.

[0025] The soil crust preparation and analysis system provided by the embodiments of the present application. The soil preparation device 1 prepares the soil crust; the environmental data acquisition device 2 is used to acquire the environmental time-series data during the soil crust preparation process, capture the 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 acquire the crust data corresponding to the soil crust, ensuring the accuracy of the acquired crust data. Therefore, there is no need for manual sampling for parameter detection, reducing the experimental period and also ensuring the real-time and accuracy of the acquired environmental time-series data. 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, ensuring the accuracy of the obtained target causal analysis model. Thus, an in-depth study of the dynamic relationship between the crust characteristics and soil parameters is realized. The soil crust preparation and analysis system realizes precise data acquisition through device collaboration and in-depth causal modeling through algorithm innovation, constructing a complete technical chain from experimental preparation to intelligent analysis. It realizes 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-causality dual-driven", providing an efficient technical tool for understanding the complex ecological processes of soil crusts and addressing soil degradation problems, with significant scientific significance and practical application potential.

[0026] In an alternative embodiment of the present application, as Figure 2 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: Step S101, 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 target fusion features.

[0027] Specifically, the above step S101 may include the following steps: Step S1011, input the crust image data, crust spectral data, crust electrochemical data, and crust mechanical properties into the corresponding embedding layers respectively to generate corresponding image feature vectors, spectral feature vectors, electrochemical feature vectors, and mechanical feature vectors.

[0028] Specifically, the data processing device 4 can convert the acquired crust image data into an image feature vector of a unified dimension through the image embedding layer; the crust spectral data, crust electrochemical data, and crust mechanical property data are also respectively converted into unified dimension feature vectors that meet the model input requirements through the corresponding spectral embedding layer, electrochemical embedding layer, and mechanical embedding layer. The environmental time series data also needs to be processed through a dedicated time series embedding layer to make its dimension consistent with other data, preparing for subsequent fusion analysis.

[0029] Specifically, the image embedding layer captures image features at different scales through multiple convolutional layers (such as the residual blocks of ResNet). Among them, the typical convolution operation can be: . Among them, is the l-th layer convolution kernel, is the previous layer feature map. The spatial features are compressed into a vector through global average pooling (GAP), and then output a fixed-dimension image feature vector (such as 512 dimensions) through a fully connected layer (FC): ImageFeature = FC(GAP(ConvOutput)).

[0030] The spectral embedding layer captures features in a local wavelength range (such as the characteristic peaks of specific functional groups) through a one-dimensional convolution kernel, and stacks multiple convolutional layers to extract multi-scale spectral patterns: . Where K is the convolution kernel size and t is the wavelength point index.

[0031] Capture the long-range dependencies between different wavelength points (such as the synergistic effect between characteristic peaks) through the self-attention mechanism, which is suitable for complex spectral analysis: , the spectral sequence is input into the Transformer after position encoding, and a context-aware spectral feature vector is output. The spectral features are compressed to a specified dimension (such as 256 dimensions) through a fully connected layer to highlight the contribution of key characteristic peaks.

[0032] The electrochemical embedding layer converts the crust electrochemical data curve into a time series or frequency domain vector (such as the real part / imaginary part of the impedance after Fourier transform), and normalizes it to zero mean and unit variance. Use LSTM or GRU units to capture the evolution law of potential / current over time: Where xt is the electrochemical measurement value (such as the redox potential) at time t, and ht is the hidden state, integrating historical information. Take the hidden state or graph node embedding at the last moment, and output an electrochemical feature vector (such as 128 dimensions) through a fully connected layer.

[0033] The mechanical embedding layer normalizes the mechanical property data of the crust. If it is multi-stage test data (such as loading-unloading curves), it is converted into a sequence of fixed length (such as interpolated to 100 time points). It is directly mapped through a fully connected layer. For example, after the hardness value is linearly transformed, it is concatenated with other parameters: MechanicalFeature = W⋅[hardness; elastic modulus] + b. If it is the force-displacement curve during the loading process, 1D convolution or Transformer is used to extract features such as plastic deformation and yield point: (where i is the sample index and j is the feature index). Finally, a mechanical feature vector of fixed dimension (such as 64 dimensions) is output, retaining the key indicators of mechanical properties (such as peak force and fracture strain).

[0034] Step S1012: Based on the multi-scale feature fusion network, 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.

[0035] Specifically, the above step S1012 may include the following steps: Step a1: Input the real-time statistical features corresponding to the image feature vector, spectral feature vector, electrochemical feature vector, and mechanical feature vector into the adaptive dilated convolution kernel generation mechanism based on the reinforcement learning algorithm.

[0036] Specifically, for the image feature vector, local variance is used to quantify texture roughness, histogram of oriented gradients is used to capture edge directionality, and the Laplacian operator response is used to detect tiny cracks on the crust surface, obtaining the real-time statistical features corresponding to the image feature vector.

[0037] For the spectral feature vector, feature extraction is performed on the spectral feature vector, the positions, widths, and intensities of absorption peaks are extracted, and the Pearson correlation coefficient between key bands is calculated. In addition, the spectral entropy of the spectral feature vector is calculated to measure spectral complexity and information richness, obtaining the real-time statistical features corresponding to the spectral feature vector.

[0038] For the electrochemical feature vector, feature extraction is performed on the electrochemical feature vector, the ion concentration change rate and reaction rate are calculated, and parameters such as 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, obtaining the real-time statistical features corresponding to the electrochemical feature vector.

[0039] For the mechanical feature vector, feature extraction is performed on the mechanical feature vector, the stress concentration coefficient and average stress are calculated, the time dependence of the mechanical response is analyzed, and the uniformity of the hardness of the crust surface is evaluated, obtaining the real-time statistical features corresponding to the mechanical feature vector.

[0040] Step a2: Using each real-time statistical feature as the state input, dynamically generate different initial empty convolutional kernel parameters.

[0041] Among them, each initial empty convolutional kernel parameter includes initial convolutional kernel sizes of different scales.

[0042] Specifically, splice the real-time statistical features corresponding to the image feature vector, spectral feature vector, electrochemical feature vector, and mechanical feature vector respectively to generate a real-time statistical input. According to the decision of the reinforcement learning agent, using each real-time statistical feature as the state input, dynamically generate different-scale initial empty convolutional kernel parameters.

[0043] For the image feature vector, a series of convolutional kernels with different dilation rates may be generated to capture multi-scale information from fine textures to overall structures. Use the generated different-scale dilated convolutional kernels to perform convolutional operations on the image feature vector. The convolutional kernel with a small dilation rate focuses on extracting the fine texture information of the image, such as the tiny cracks on the crust surface and the fine textures of particles, and generates corresponding sub-initial image feature vectors, which contain the detailed features of the image at a small scale. The convolutional kernel with a large dilation rate is used to extract the overall structure and macroscopic features of the image, such as the overall shape of the crust and the large-area texture distribution, and also generates corresponding sub-initial image feature vectors, covering the global information of the image at a large scale.

[0044] For the spectral feature vector, dilated convolutional kernels adapted to its frequency distribution characteristics are correspondingly generated. The generated convolutional kernels are not fixed. As the real-time statistical information of the input feature vector changes, the convolutional kernels will also be updated in real time to adapt to the dynamic changes of the data. Different-scale dilated convolutional kernels perform convolutional operations according to the frequency distribution of the spectrum. The convolutional kernel with a small dilation rate focuses on the frequency bands with drastic changes and containing key information in the spectrum, extracts fine spectral features, and generates sub-initial spectral feature vectors, which can reflect the detailed changes of the spectrum within a local frequency range. The convolutional kernel with a large dilation rate integrates a wider range of spectral information, obtains the overall trend and features of the spectrum, and the generated sub-initial spectral feature vectors contain the frequency features of the spectrum at a large scale, which helps to analyze the overall characteristics of the spectrum.

[0045] For the electrochemical feature vector, perform convolutional operations according to the generated dilated convolutional kernels. The small-scale convolutional kernel captures the subtle changes of electrochemical parameters such as the initial ion concentration in a local area, and the generated sub-initial electrochemical feature vectors can accurately reflect the local electrochemical characteristics. The large-scale convolutional kernel, from a more macroscopic perspective, integrates the change trends of electrochemical parameters in a larger range, and the generated sub-initial electrochemical feature vectors reflect the overall electrochemical change situation.

[0046] For the mechanical feature vector, a dilated convolutional kernel is used to process the mechanical feature vector. A convolutional kernel with a small dilation rate extracts the changes in mechanical parameters at a micro scale, such as local hardness differences and stress concentration points, and generates corresponding sub-initial mechanical feature vectors, reflecting the details of mechanical features. A convolutional kernel with a large dilation rate obtains the overall features of mechanical features at a larger scale, such as the overall hardness distribution and the mechanical stability of the structure, and the generated sub-initial mechanical feature vectors contain the macroscopic information of mechanical features.

[0047] Step a3, based on a preset reward function, calculate the reward values corresponding to each initial dilated convolutional kernel parameter.

[0048] Specifically, the adaptive dilated convolutional kernel generation mechanism of the reinforcement learning algorithm takes these statistical information as state inputs, and according to a preset reward function, attempts to select different dilated convolutional kernel parameters (such as dilation rate, kernel size, etc.). The design of the reward function aims to encourage the generated convolutional kernel to better capture the multi-scale information in the feature vector. If the generated convolutional kernel can extract richer and more representative features in subsequent feature extraction, the agent will obtain a higher reward; otherwise, it will get a lower reward. Through continuous trial and error and learning, the agent gradually masters the optimal convolutional kernel generation strategy.

[0049] Among them, the reward function can be: ; Among them, R represents the reward value, 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 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, used to measure the richness of features. The calculation formula of information entropy is: . Where p(xk) is the probability that the k-th eigenvalue appears in the feature vector Xi. The larger the information entropy value, the richer and more disordered the information contained in the feature vector at this scale, that is, the higher the diversity of features. M is the number of predefined key feature indicators.

[0050] Yj represents the j-th key feature indicator. For example, in the study of soil crust, the key feature indicators may include mechanical parameters reflecting the stability of the crust, electrochemistry parameters related to salt, etc.

[0051] I(Yj) is used to measure the importance degree of the j-th key feature index in the extracted feature vector. It can be determined by calculating the correlation coefficient between the key feature index 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 j-th key feature index, and yj is its mean. The value range of I(Yj) is between -1 and 1. The closer the absolute value is to 1, the stronger the correlation between the key feature index and the extracted feature vector, that is, the better the representativeness of the extracted feature vector for the key feature index. By comprehensively considering the feature richness (measured by information entropy) and feature representativeness (measured by the importance of key feature indices), this reward function provides a quantitative reward criterion for the reinforcement learning agent.

[0052] Step a4, according to the reward value, select the target empty convolution kernel parameter with the maximum reward value; Specifically, the agent generates N = 20 candidate empty convolution kernel parameters (such as different combinations of k and d) in each iteration, and calculates the reward value R of each candidate empty convolution kernel parameter. Non-maximum suppression (NMS) is used to filter similar parameters (such as only retaining the one with higher reward between k = 5, d = 2 and k = 5, d = 3), reducing redundant evaluation.

[0053] Then, select the parameters with a reward value R higher than the mean + σ (standard deviation) to avoid the influence of noise with accidentally high rewards. Store the high-reward parameters (R > threshold) in the experience pool, and update the agent network through the PPO algorithm to enhance the probability of successful strategies.

[0054] Finally, select the target empty convolution kernel parameter with the maximum reward value.

[0055] Step a5, based on the target convolution kernel sizes of different scales in the target empty convolution kernel parameter, perform feature extraction on the image feature vector, spectral feature vector, electrochemical feature vector, and mechanical feature vector, 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.

[0056] Specifically, for the image feature vector: Apply a k×k two-dimensional dilated convolution with a stride of 1 and a padding of (k−1) / 2 to retain the spatial resolution. Use small scales (for example, k = 3, d = 1) to capture local details such as micropores and particle textures on the surface of the crust; large scales (using k = 7, d = 3): Extract the overall structure of the crust (such as the distribution of large-area cracks).

[0057] For the spectral feature vector and the electrochemical feature vector: Apply one-dimensional dilated convolution to capture the long-range dependencies in the sequence. Among them, small scales (e.g., k = 3, d = 1) focus on the fine changes in specific wavelength / potential intervals; large scales (e.g., k = 7, d = 3) integrate the trend features in a wide frequency domain / long time range.

[0058] For the mechanical feature vector, if the mechanical feature vector is a scalar (such as a hardness value), simulate the scale effect through 1×1 convolution (fully connected layer); if the mechanical feature vector is a sequence (such as a loading curve), use one-dimensional convolution to extract the elastic segment and plastic segment features.

[0059] Finally, output 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.

[0060] Exemplarily, generate three-scale feature vectors for each modality: The feature dimension is unified to 128 dimensions (controlled by the number of convolutional kernels) for subsequent fusion.

[0061] Step S1013, based on the attention mechanism, perform feature fusion on 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 to generate the target fusion features.

[0062] Specifically, the above step S1013 may include the following steps: Step b1, pair the sub-initial image feature vectors, sub-initial spectral feature vectors, and sub-initial electrochemical feature vectors of the same scale pairwise to generate multiple groups of cross-modal feature groups.

[0063] Specifically, the electronic device traverses the sub-initial image feature vectors, sub-initial spectral feature vectors, and sub-initial electrochemical feature vectors at each scale. Then, pair the sub-initial image feature vectors, sub-initial spectral feature vectors, and sub-initial electrochemical feature vectors of the same scale pairwise to generate multiple groups of cross-modal feature groups.

[0064] Step b2, for each feature group, input the feature group into the attention calculation network to calculate the correlation between the two sub-feature vectors in the feature group.

[0065] Specifically, for each feature group, input the feature group into the attention calculation network, calculate the correlation score corresponding to the two sub-feature vectors through the spatial-semantic branch in the attention calculation network, and calculate the causal correlation score corresponding to the two sub-feature vectors through the causal branch. Then, fuse the correlation score and causal correlation score corresponding to each sub-feature vector to obtain the correlation between the two sub-feature vectors.

[0066] Step b3: Output an attention weight matrix based on the correlation.

[0067] Among them, the attention weight matrix is used to represent the degree of attention between two sub - feature vectors in the feature group.

[0068] Specifically, the electronic device outputs the correlation corresponding to two sub - feature vectors in each feature group in the form of a matrix, and outputs the attention weight matrix.

[0069] Step b4: Update the two sub - feature vectors in the feature group based on the attention weight matrix.

[0070] 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 information injection intensity; Symmetrically, the information of sub - feature vector B is "back - injected" into sub - feature vector A to ensure two - way interaction.

[0071] Then, keep the original feature sub - feature vectors A + B as the basis, and superimpose the updated features after interaction to prevent gradient disappearance.

[0072] Exemplarily, when there is a strong causal association between the "salt absorption peak" in the spectral feature and the "crust crack" in the image feature, the forward update will enhance the dimensions related to the crack in the image feature, and the reverse update will enhance the response intensity of the salt peak in the spectral feature, forming cross - modal causal feature enhancement.

[0073] Step b5: Loop like this 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.

[0074] Specifically, perform 3 rounds of attention calculation - update iterations on the feature groups at each scale to gradually enhance the causal dependence relationship between features. During the iteration process, force the attention weight matrices of adjacent scales (such as small scale and medium scale) to satisfy the similarity constraint. The formula is: . Avoid contradictory attention allocation for the same modality pair at different scales (such as highly correlated image - spectrum at small scale and suddenly uncorrelated at medium scale). Thus, 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.

[0075] Step b6: Fuse the sub-updated image feature vectors, sub-updated spectral feature vectors, sub-updated electrochemical feature vectors, and sub-updated mechanical feature vectors according to the data modality to obtain the fused sub-target image feature vectors, sub-target spectral feature vectors, sub-target electrochemical feature vectors, and sub-target mechanical feature vectors.

[0076] Specifically, for the sub-updated image feature vectors at each scale, non-local operations are adopted to integrate the spatial features at different scales (such as small-scale cracks + large-scale structures), highlight the multi-resolution textures on the crust surface, and obtain the fused sub-target image feature vectors.

[0077] For the sub-updated spectral feature vectors at each scale, through band-grouped convolution, the updated features of adjacent bands are aggregated according to the functional group categories (such as hydroxyl group, carbonate group) to enhance the chemical semantic expression and obtain the fused sub-target spectral feature vectors.

[0078] For the sub-updated electrochemical feature vectors at each scale: Use a temporal convolutional network (TCN) to capture the time dependence of ion concentration changes and obtain the fused sub-target electrochemical feature vectors.

[0079] For the sub-updated mechanical feature vectors at each scale, directly calculate the mechanical parameters at each scale (such as the mean value of elastic modulus, variance of hardness), retain the global mechanical properties, and obtain the fused sub-target mechanical feature vectors.

[0080] Step b7: Fuse the sub-target image feature vectors, sub-target spectral feature vectors, sub-target electrochemical feature vectors, and sub-target mechanical feature vectors to generate the target fusion features.

[0081] Specifically, the electronic device can first fuse the strongly causally related modality pairs, fuse the sub-target spectral feature vectors and sub-target electrochemical feature vectors to generate "chemical-electrochemical features"; Then, fuse the result of the sub-target image feature vectors and the chemical-electrochemical features to generate "structure-chemical features"; Finally, fuse the sub-target mechanical feature vectors and the structure-chemical features to generate the target fusion features containing structure-chemical-mechanical information.

[0082] Among them, the fusion logic is: TargetFusion = MLP(Concat(Struct-Chem,Mechanical)) where Struct-Chem is the fusion feature of the image and spectral-electrochemical, and Mechanical is the mechanical feature. The multi-layer perceptron (MLP) is used to capture the cross-modal high-order causal relationships (such as "salt crystallization → change in crust hardness").

[0083] Step S102: Input the target fusion feature, temperature time-series data, humidity time-series data, rainfall time-series data, and salt content data of rainfall into the initial causal analysis network, and train the initial causal analysis network to obtain the target causal analysis model.

[0084] Specifically, the above step S102 may include the following steps: Step S1021: Generate a training set and a validation set based on the target fusion feature, temperature time-series data, humidity time-series data, rainfall time-series data, and salt content data of rainfall.

[0085] Specifically, the electronic device can divide the target fusion feature, temperature time-series data, humidity time-series data, rainfall time-series data, and salt content data 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 lag effect.

[0086] Then, 70% of the data in the dataset is divided into the training set, and 30% is divided into the validation set.

[0087] Step S1022: Input the training set into the input layer of the initial causal analysis network.

[0088] Specifically, the electronic device uses time position encoding (such as sine curve encoding) + causal marker encoding (such as marking whether it is intervention data do(X)) for the temperature time-series data, humidity time-series data, and rainfall time-series data in the training set. The formula is: ; Then, splice the static target fusion feature and the dynamic environmental time-series feature to form an input vector containing "crust state - environmental history".

[0089] Among them, 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 the output dimension is Denv, and the crust feature branch, which processes the target fusion feature, and the output dimension is Dfeat. The final input is Concat(Eenv,Ffusion), and the dimension is Denv+Dfeat.

[0090] Step S1023: The hidden layer in the initial causal analysis network extracts features from the target fusion feature, temperature time-series data, humidity time-series data, rainfall time-series data, and salt content data of rainfall, and outputs the target feature.

[0091] 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, salt content) into a feature vector of a fixed dimension, and uses 1D convolution to extract time-series patterns (such as sudden drops in temperature, persistent high humidity, etc.).

[0092] Then, the control gate is calculated through the learnable weight matrix W1: σ(W1 • Eenv), where σ is the Sigmoid function and the output value is between [0, 1]. The control gate determines which time-series features should be enhanced or suppressed (e.g., only allowing humidity changes directly related to crust cracks to pass through). The control gate is multiplied element-wise (⊙) with the convolution output to obtain the direct causal feature F direct . For example, a sudden increase in rainfall will directly cause a change in the wetness of the crust surface, and this direct relationship of "rainfall → wetness" is selectively strengthened by the gating mechanism.

[0093] In addition, the hidden layer in the initial causal analysis network takes environmental variables (temperature, humidity, etc.), intermediate variables (water content, salt concentration, etc.), and crust features (image texture, spectral peak intensity, etc.) as nodes, and predefines possible causal relationships (such as temperature → water content, water content → salt migration) based on prior knowledge to determine the initial weight information of the edges corresponding to each node.

[0094] Then, the node representations are iteratively updated through a graph convolutional network (GCN) to simulate the propagation process of causal effects: F indirect = GCN(G(Eenv, Ffusion)) For example: The information of the temperature node is gradually transmitted through the path of "temperature → water content → salt crystallization → spectral features", and finally affects the spectral feature node of the crust. Compared with the first-order association of the direct path, the indirect path can model multi-step causal chains (such as temperature increase → water evaporation → salt concentration increase → crust hardness change). The edge weights of the GNN automatically learn the strength of each causal relationship (e.g., the weight of "temperature → water content" reflects the degree of influence of temperature changes on water).

[0095] Next, for the direct path feature F direct and the indirect path feature F indirect , the attention score is calculated: 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).

[0096] Dynamically fuse two paths according to the attention scores to obtain the target feature: Ftarget = α1•Fdirect + α2•Findirect. When the rainfall suddenly increases, the model will automatically increase the weight of the direct path (capturing the rapid response of "rainfall → humidity"); under long-term drought conditions, the model will increase the weight of the indirect path (capturing the slow process of "temperature → water evaporation → salt accumulation").

[0097] 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 feature.

[0098] Specifically, the output layer in the initial causal analysis network uses a multi-layer perceptron (MLP) or a linear regression model to map the target feature. For example, for each environmental variable (such as temperature T) and crust feature (such as image crack density Cimg), calculate the causal effect value τ(T→Cimg). This value quantifies the "degree of influence of temperature change on the crust crack density", a positive value indicates a positive correlation (an increase in temperature leads to an increase in cracks), and a negative value indicates a negative correlation.

[0099] Traverse all combinations of environmental variables (temperature, humidity, rainfall, salt content) and crust features (image, spectrum, electrochemistry, mechanical data), calculate the causal effect value between each pair of variables, and form a causal effect matrix. Exemplarily, the causal effect matrix can be as follows: ; where, T is the temperature, H is the humidity, Cimg is the image crack density, Cspec is the spectral salt peak intensity, and τ(T→Cimg) is the causal effect value of temperature on the image crack density.

[0100] Finally, the output layer outputs the virtual causal relationship according to the causal effect matrix. Exemplarily, .

[0101] Step S1025, calculate the loss value corresponding to the target loss function based on the virtual causal relationship and the real causal relationship.

[0102] Among them, the real causal relationship is determined by analyzing the environmental time series data and the crust data during the soil crust preparation process.

[0103] Specifically, calculate the loss value corresponding to the target loss function based on the virtual causal relationship and the real causal relationship.

[0104] Step S1026, update the parameters of the initial causal analysis network based on the loss value to obtain the initial causal analysis model.

[0105] Specifically, the electronic device updates the parameters of the initial causal analysis network based on the loss value to obtain the initial causal analysis model.

[0106] In an alternative embodiment of the present application, the data processing device 4 constructs a loss function based on the following steps, which specifically include: Step c1, construct a prior causal knowledge graph based on the prior knowledge of the causal relationship between environmental factors and scaling data.

[0107] Specifically, the electronic device can receive the prior knowledge of the causal relationship between environmental factors and scaling data input by the user. For example, such as "humidity increase → decrease in scaling air permeability" and "salt content → spectral characteristic peak shift". Then, identify the prior knowledge of the causal relationship between environmental factors and scaling data, and extract causal relationship triples, namely cause nodes, causal relationships, and result nodes.

[0108] Then, use a directed acyclic graph (DAG) for modeling. The nodes include environmental variables (temperature, humidity, etc.) and scaling characteristics (image texture, mechanical strength, etc.), and the edges represent causal relationships (such as the direct effect of "temperature → water content" and the indirect effect of "water content → salt migration → spectral characteristics"). Assign prior causal strength (such as setting the weight range according to the literature results: -1 to 1, where a negative value represents a negative causal relationship) and confidence (such as the reliability score 0.1 - 1.0 evaluated by experts) to each edge.

[0109] Step c2, map the virtual causal relationship output by the initial causal analysis network to the corresponding nodes and edges in the prior causal knowledge graph.

[0110] Specifically, the electronic device establishes unique identifiers for environmental variables and scaling characteristics (such as "T" represents temperature and "C_spec_Salt" represents the spectral salt peak intensity) to ensure that the variable names output by the model exactly correspond to the graph nodes. For the virtual causal relationship output by the initial causal analysis network (such as "temperature → spectral water peak intensity"), search for the edge with the same starting and ending points in the graph. If it exists, record the corresponding prior strength and confidence; if not, mark it as a newly discovered causal relationship or a potential conflict.

[0111] Step c3, detect the conflict between the virtual causal relationship and the prior causal knowledge graph.

[0112] Specifically, check whether the initial causal analysis network predicts an inverse causal relationship (such as "scaling hardness → rainfall"), which violates the physical logic of "environmental variables as the cause and scaling characteristics as the result".

[0113] Then, compare the virtual causal relationship τ pred predicted by the initial causal analysis network with the prior knowledge correlation τ prior , if it satisfies or If (δ is the tolerance threshold, e.g., 0.5), it is determined as an intensity conflict.

[0114] Exemplarily, prior knowledge believes that "humidity → crust crack" is positively correlated (τprior = 0.6). If the initial causal analysis network predicts a negative correlation (τpred = -0.3), a conflict is triggered.

[0115] Step c4: Calculate the causal consistency loss based on the conflict between the virtual causal relationship and the prior causal knowledge graph.

[0116] Specifically, assign a high weight to the directional conflict (e.g., the weight coefficient wdir = 10), and weight the intensity conflict by confidence (e.g., for an edge with a prior confidence of 0.9, the conflict loss is multiplied by 0.9). Calculate the causal consistency loss based on the conflict between the virtual causal relationship and the prior causal knowledge graph.

[0117] Exemplarily, the causal consistency loss function is as follows:

[0118] Among them, is the set of conflict edges, w uv is the edge weight, is a minimum value to avoid a zero denominator, is the prior knowledge correlation, is the 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.

[0119] Step c5: 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.

[0120] Specifically, classify by the direction of the causal effect: positive correlation, negative correlation, no significant association; classify by the causal path: direct causation, indirect causation, mixed causation (including both direct and indirect effects); classify by the type of environmental variable: temperature-related, humidity-related, salinity-related, etc.

[0121] For each training sample, assign it to the above categories according to the combination of its corresponding environmental variable and crust feature. For example, "temperature → crust hardness (direct positive correlation)" belongs to "direct causation - positive correlation - temperature category". Count the number of samples of each type. For example, the samples in the "indirect causation" category may only account for 15% of the total data, resulting in an issue of class imbalance.

[0122] Step c6: Based on the number of samples corresponding to each type, assign different weights to samples of different categories and determine the balanced loss function.

[0123] Specifically, calculate the class weights , where N is the total number of samples, K is the number of classes, and nc is the number of samples in class c.

[0124] Example: If the number of samples nc in the "indirect causality" class is 150 and the total number of samples N is 1000, then wc = 3×150 / 1000 ≈ 2.22 (assuming K = 3 classes).

[0125] Increase the penalty for minority class samples. The balanced loss function formula is: .

[0126] 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 class.

[0127] Step c7, based on the causal consistency loss, balanced loss function, and mean squared error loss function, construct the objective loss function.

[0128] Specifically, based on the causal consistency loss, balanced loss function, and mean squared error loss function, construct the objective loss function. The objective loss function is as follows:

[0129] , where, is the mean squared error loss, which measures the regression error between the predicted value and the true value of the crust feature and ensures the basic prediction ability of the model.

[0130] Step S1027, verify 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 the preset accuracy threshold, obtain the target causal analysis model.

[0131] Specifically, the electronic device inputs the validation set into the initial causal analysis model for verification, verifies the validation set based on the objective loss function. If the loss value corresponding to the objective loss function is greater than or equal to the preset loss value threshold, it is determined that when the accuracy of the initial causal analysis model is greater than or equal to the preset accuracy threshold, obtain the target causal analysis model.

[0132] The soil crust preparation and analysis system provided by the embodiments of the present application inputs the crust image data, crust spectral data, crust electrochemical data, and crust mechanical properties into the corresponding embedding layers respectively to generate corresponding image feature vectors, spectral feature vectors, electrochemical feature vectors, and mechanical feature vectors. The representation forms of different types of data (pixel matrices of images, wavelength intensity sequences of spectra) are unified to facilitate subsequent fusion processing. The real-time statistical features corresponding to the image feature vector, spectral feature vector, electrochemical feature vector, and mechanical feature vector are calculated, ensuring the accuracy of the obtained real-time statistical features. 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 hollow convolution kernel generation mechanism based on the reinforcement learning algorithm, and each real-time statistical feature is used as the state input to dynamically generate different initial hollow convolution kernel parameters. The limitations of fixed-scale convolution are avoided (such as small kernels missing global trends and large kernels blurring local details), and the flexibility of feature extraction is improved; the effective scales are selectively activated according to the data characteristics to reduce redundant calculations (such as skipping small-scale kernels for smooth spectral data). Based on the preset reward function, the reward values corresponding to each initial hollow convolution kernel parameter are calculated, and according to the reward values, the target hollow convolution kernel parameter with the largest reward value is selected. The effectiveness of the convolution kernel is quantified through the preset reward function (such as information entropy, feature importance, causal consistency), and the optimal parameters are screened to ensure that feature extraction takes into account information richness, causal relevance, and computational efficiency, avoiding overfitting of a single index. Based on the target convolution kernel sizes of different scales in the target hollow 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 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 corresponding sub-initial image feature vectors, sub-initial spectral feature vectors, sub-initial electrochemical feature vectors, and sub-initial mechanical feature vectors output at each scale.

[0133] Then, 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 pairwise to generate multiple groups of cross-modal feature groups. Through information interaction between modalities, the information loss of a single modality is compensated (for example, the spectral data lacks spatial position information, which can be supplemented by image features). For each feature group, the feature group is input into the attention calculation network to calculate the correlation corresponding to the two sub-feature vectors in the feature group; based on the correlation, an attention weight matrix is output. To avoid the causal direction deviation caused by unidirectional information flow and ensure the physical rationality of the "environment - crust" causal chain. Based on the attention weight matrix, the two sub-feature vectors in the feature group are updated; repeating this process, 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 are obtained. By iteratively updating the feature vectors, the high-order causal relationship across modalities (such as the indirect path of "temperature → moisture → salinity → structure") is gradually strengthened.

[0134] Next, the sub-updated image feature vectors, sub-updated spectral feature vectors, sub-updated electrochemical feature vectors, and sub-updated mechanical feature vectors are fused according to the data modality to obtain the fused sub-target image feature vectors, sub-target spectral feature vectors, sub-target electrochemical feature vectors, and sub-target mechanical feature vectors; the sub-target image feature vectors, sub-target spectral feature vectors, sub-target electrochemical feature vectors, and sub-target mechanical feature vectors are fused to generate the target fusion feature. Abstracting from local (single modality and single scale) to global (multi-modal and multi-scale) gradually conforms to the multi-physical process characteristics of soil crust formation. Encoding the direct / indirect effects of environmental variables (such as temperature, rainfall) on crust features into a unified feature space supports the end-to-end learning of the causal model.

[0135] Finally, based on the target fusion features, temperature time-series data, humidity time-series data, rainfall time-series data, and salt content data of rainfall, training sets and validation sets are generated. The training sets are input into the input layer of the initial causal analysis network; the hidden layer in the initial causal analysis network extracts features from the target fusion features, temperature time-series data, humidity time-series data, rainfall time-series data, and salt content data of rainfall, and outputs target features, ensuring the accuracy of the output target features. 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; a prior causal knowledge graph is constructed based on the prior knowledge of the causal relationship between environmental factors and crust data; the virtual causal relationship output by the initial causal analysis network is mapped to the corresponding nodes and edges in the prior causal knowledge graph; the conflict between the virtual causal relationship and the prior causal knowledge graph is detected; based on the conflict between the virtual causal relationship and the prior causal knowledge graph, the causal consistency loss is calculated. The virtual causal relationship predicted by the model is compared with the domain prior knowledge (such as the edge "rainfall → crust water content" in the graph), and conflicts such as directionality and intensity are detected. Avoid the model learning spurious associations that violate soil science principles (such as the inverse causal relationship of "crust hardness → temperature"). In scenarios where the prior knowledge is reliable, the graph information is used to make up for the lack of data and improve the generalization ability of the model. Then, according to the causal relationship type between the crust data and the environmental time-series data, the samples in the training set are classified, and the number of samples corresponding to each category is determined. Based on the number of samples corresponding to each type, different weights are assigned to the samples of different categories, and a balanced loss function is determined. Prevent the model from being biased towards the majority class (such as direct causal samples) and ensure that the features of the minority class (such as indirect causality) are fully learned; at the same time, optimize the prediction accuracy, causal consistency, and sample balance, and improve the robustness of the model in complex scenarios. Based on the causal consistency loss, the balanced loss function, and the mean square error loss function, an objective loss function is constructed, ensuring the accuracy of the constructed objective loss function. Based on the virtual causal relationship and the true causal relationship, the loss value corresponding to the objective loss function is calculated; based on the loss value, the parameters of the initial causal analysis network are updated to obtain an initial causal analysis model, ensuring the accuracy of the obtained initial causal analysis model. The initial causal analysis model is verified based on the validation set. When the accuracy of the initial causal analysis model is greater than or equal to the preset accuracy threshold, a target causal analysis model is obtained. Avoid overfitting the training data and ensure the prediction reliability of the target causal analysis model in the real environment (such as different soil types and climate changes). The errors found during the verification process can be fed back to data collection or model structure design to form a closed loop of "training - verification - improvement". Ensure the accuracy of the obtained target causal analysis model.

[0136] In an alternative embodiment of the present application, as Figure 3 shown, the data processing device 4 is further configured to perform the following steps: Step S201: Obtain the environmental change data within a preset time period before the current time corresponding to the target area, input the environmental change data into the target causal analysis model, and output the future soil crust data corresponding to the target area.

[0137] Specifically, the electronic device can receive the environmental change data of the target area within a preset time period before the current time input by the user or sent by other devices. Among them, the environmental change data includes, but is not limited to, temperature data, humidity data, precipitation data, and the salt content data in the precipitation.

[0138] The electronic device cleans the original data (removes outliers and fills in missing values), performs normalization processing, and organizes it into an input format recognizable by the model (such as a multi-dimensional feature matrix) according to the time series, and inputs it into the target causal analysis model.

[0139] Then, input the preprocessed environmental change data into the trained target causal analysis model (that is, the causal analysis model optimized through previous training, which integrates direct causal paths, indirect causal paths, and prior knowledge constraints). The target causal analysis model separates the direct impact of environmental variables on soil crust (such as rainfall directly affecting the water content of the crust) and the indirect impact (such as temperature through water evaporation → salt crystallization → change in crust mechanical strength) through the causal mechanism decoupling network (CMDN). Then, use the attention mechanism to dynamically weight the features of the direct path and the indirect path to generate target features that comprehensively reflect the causal relationship. Finally, predict the soil crust data in the future time period through the output layer of the target causal analysis model. The soil crust data includes: Physical properties: crust thickness, porosity, surface hardness, crack density; Chemical properties: salt content, organic matter content, pH value; Spectral characteristics: reflectance spectral curve, absorption intensity of characteristic bands; Mechanical properties: compressive strength, shear strength, etc.

[0140] Step S202: Output the corresponding management measures based on the future soil crust data.

[0141] Specifically, the electronic device performs multi-dimensional analysis on the future soil crust data output by the target causal analysis model to identify key change trends and outliers. For example: If it is predicted that the crust thickness increases significantly and the porosity decreases, it may indicate that the soil air permeability deteriorates, affecting the growth of plant roots; If the salt content exceeds the threshold, it may indicate an increase in the risk of soil salinization.

[0142] Then, combined with the ecological impact thresholds of soil crust (such as agricultural production standards, ecological protection standards), divide the prediction results into different risk levels (such as low, medium, and high risks).

[0143] Finally, based on the predicted crust data and risk levels, combined with the causal relationships of environmental factors, targeted management measures are generated. The following are typical scenarios and corresponding strategies: Scenario 1: The increase in crust thickness leads to soil compaction (direct causality: rainfall → crust compaction). Short-term measures: Use mechanical soil loosening (such as shallow rotary tillage) to break the surface layer of the crust and improve soil aeration; cover with organic materials (such as straw, humus) to reduce surface runoff erosion and inhibit crust formation. Long-term measures: Optimize the irrigation method (such as replacing flood irrigation with drip irrigation) to reduce the impact of rainwater on the ground surface; plant cover crops with well-developed roots (such as legumes) to enhance soil structure stability through root activities.

[0144] Scenario 2: Salt accumulation causes soil salinization (indirect causality: temperature → water evaporation → salt migration and enrichment). Short-term measures: Implement leaching improvement: Lower the salt concentration in the surface soil by irrigating with fresh water; apply soil amendments (such as gypsum) to adjust the soil pH value and promote salt leaching. Long-term measures: Establish a drainage system to prevent the rise of the groundwater level from causing salt to move upward; promote salt-tolerant crop varieties (such as Suaeda salsa) to reduce the soil salt content through biological absorption.

[0145] Scenario 3: The reduction in the mechanical strength of the crust leads to wind erosion risk (mixed causality: wind speed → surface wear of the crust + water content → decrease in bonding force). Short-term measures: Spray soil solidifying agents (such as polymer) to enhance the erosion resistance of the crust surface; temporarily cover with dust-proof nets or lay gravel layers to reduce direct wind erosion. Long-term measures: Increase the vegetation coverage (such as planting windbreak grass belts) to fix the soil through the roots of vegetation; adjust the land use pattern (such as returning farmland to forests and grasslands) to reduce the degree of surface exposure.

[0146] According to the risk levels and the implementation costs of the measures, the management measures are prioritized. For example, short-term emergency measures are preferably adopted for high-risk scenarios, and long-term ecological restoration is emphasized for medium- and low-risk scenarios. In addition, a closed-loop monitoring mechanism is established to regularly compare the actual environmental data with the model prediction results, evaluate the effectiveness of the measures, and adjust the strategies. For example, if the salt concentration does not reach the expected level after leaching improvement, the dosage of the amendment can be increased or the leaching frequency can be adjusted.

[0147] The data processing device 4 of the soil crust preparation and analysis system provided by the embodiment of the present application is further configured to obtain environmental change data within a preset time period before the current time corresponding to the target area, input the environmental change data into the target causal analysis model, and output future soil crust data corresponding to the target area, 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.

[0148] In an alternative embodiment of the present application, such asFigure 4 As shown in Figure 4 , the soil preparation device 1 includes: A preparation support 11, including a support frame and a preparation container, with the preparation container placed on the support frame; A soil crust cultivation soil column 12, placed inside the preparation container, for preparing the soil crust; A rain shower 13, installed on the top surface inside the preparation container, for raining on the soil crust cultivation soil column 12; A pressure regulating nozzle 112, installed on the rain shower 13, for simulating the amount of rainfall; One end of the rain shower 13 is connected to a salt dissolution preparation device 16 through a first water pump 14 and a first valve 15, and the other end is connected to a pure water preparation device 19 through 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 soil column 12; when the second valve 18 is opened, pure water is sprayed onto the soil crust cultivation soil column 12; A solution collection bucket 110, placed inside the support frame and below the preparation container, for collecting the liquid flowing down from the soil crust cultivation soil column 12; A heating device 111, for changing the environmental temperature during the soil crust preparation process.

[0149] Specifically, rainfall control: By changing the water outlet pressure through the pressure regulating nozzle 112, the simulation of rainfall intensity (such as light rain, moderate rain, heavy rain) is achieved. Salt gradient control: The salt dissolution preparation device 16 provides salt solutions with different concentrations, and the proportion of the sprayed salt solution is controlled through the first valve 15 to simulate the salt content in actual rainfall (such as high-salt precipitation in coastal areas); the pure water preparation device 19 provides low-salt / salt-free precipitation for the control group. Temperature regulation principle: The heating device 111 (such as heating wire, temperature control system) affects the soil water evaporation rate, salt crystallization process and microbial activities by increasing or decreasing the environmental temperature, and then regulates the physical and chemical properties of the crust (such as hardness, porosity). During the cultivation process, environmental data (temperature, humidity, rainfall, salt content) and crust data (water content, mechanical strength, spectral characteristics, etc.) are synchronously monitored to provide training samples for the causal analysis model and analyze the direct / indirect effects of various factors on crust formation (such as the causal path of "salt solution → soil salt migration → increased crust hardness").

[0150] The detailed working process is as follows: Step 1: Device setup and soil column filling; Place the soil crust cultivation soil column 12 into the preparation container. The soil column needs to be filled in layers according to the natural soil structure (such as the surface humus layer and the bottom sandy layer) to ensure simulation authenticity. The rainfall simulator 13 is fixed at the center of the top surface of the preparation container, and the pressure regulating nozzle 112 faces the soil column surface to ensure uniform rainfall coverage. The solution collection bucket 110 is placed below the preparation container and is connected to the bottom drainage port of the container through a pipeline for collecting the infiltration liquid. 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 through the temperature control system.

[0151] Step 2: Set environmental variable parameters; Rainfall simulation: According to the experimental requirements, preset the rainfall intensity through the pressure regulating nozzle 112 (such as simulating heavy rain at 50 mm / h). Selection of salt treatment: If a high-salt environment needs to be simulated, open the first valve 15 and start the first water pump 14 to transport the salt solution in the salt dissolution preparation device 16 to the rainfall simulator 13 and spray it onto the soil column surface; if a low-salt or natural precipitation needs to be simulated, close the first valve 15 and open the second valve 18 to transport pure water through the second water pump 17. Temperature setting: Input the target temperature through the control panel of the heating device 111 (such as setting it to 35 °C when simulating a high-temperature environment in summer), and the device automatically maintains a constant temperature.

[0152] Step 3: Cultivate soil crust and collect data; Start the rainfall simulator 13 and continuously spray the solution (salt solution or pure water) according to the preset parameters to simulate the natural rainfall process. During the rainfall process, the water on the soil surface infiltrates and evaporates, and the salt migrates to the surface with the water and gradually crystallizes, promoting the cementation of soil particles to form a crust.

[0153] Collect data regularly: Environmental data: Record the real-time temperature, humidity, cumulative rainfall, and solution salt content through temperature sensors, humidity sensors, and flow meters.

[0154] Crust data: Measure the mechanical strength of the crust through a micro hardness tester, detect the chemical composition of the crust surface (such as salt content) through a spectrometer, and measure the change in soil column water content through the weighing method.

[0155] The infiltration liquid flows into the solution collection bucket 110 through the bottom of the preparation container, and samples can be taken regularly to analyze the salt concentration, organic matter content, etc. in it to reflect the material migration process inside the soil.

[0156] Step 4: Conduct multiple groups of control experiments and model training; By adjusting a single variable (such as rainfall, salt concentration, temperature), multiple groups of control experiments are designed: Control group: pure water + 20°C + low rainfall, simulating a clean and low-humidity environment; Experimental group 1: high-concentration salt solution + 30°C + high rainfall, simulating a coastal high-temperature rainstorm environment; Experimental group 2: low-salt solution + 15°C + medium rainfall, simulating a temperate humid environment. Each experiment is repeated multiple times to accumulate crust data under different environmental combinations for training a causal analysis model (such as the CMDN network) to analyze the direct / indirect causal relationships between environmental variables and crust characteristics (such as the indirect path of "temperature increase → accelerated water evaporation → salt crystallization → increased crust hardness").

[0157] Step 5: Equipment cleaning and maintenance; After the experiment, close all valves and pumps, drain the residual solution in the pipeline, and rinse the rain shower 13 and nozzles with pure water to prevent salt crystallization from clogging. Clean the soil column residue in the preparation container, disinfect it, and reserve it for future use; clean the solution collection bucket 110 and let it dry to prevent microbial growth. Check whether the temperature control system of the heating equipment 111 is normal to ensure the accuracy of experimental parameters for the next experiment.

[0158] The soil crust preparation and analysis system provided by the embodiment of the present application includes a preparation support 11, which comprises a support frame and a preparation container, and the preparation container is placed on the support frame. A soil crust cultivation soil column 12 is placed inside the preparation container and is used for preparing the soil crust. A rain shower 13 is installed on the top surface inside the preparation container and is used for raining on the soil crust cultivation soil column 12. A pressure regulating nozzle 112 is installed on the rain shower 13 and is used for simulating the magnitude of rainfall. By adjusting the pressure parameters, the magnitude of rainfall can be precisely controlled (such as different intensities of light rain, moderate rain, heavy rain, etc.), the dynamic process of natural rainfall can be simulated, and the influence of rainfall intensity on the formation of soil crust can be studied (such as changes in scouring force and infiltration rate). One end of the rain shower 13 is connected to a salt dissolution preparation device 16 through a first water pump 14 and a first valve 15, and the other end is connected to a pure water preparation device 19 through a second water pump 17 and a second valve 18; when the first valve 15 is opened, a salt solution is sprayed on the soil crust cultivation soil column 12; when the second valve 18 is opened, pure water is sprayed on the soil crust cultivation soil column 12. When spraying the salt solution, the influence of a high-salt environment (such as saline-alkali land, coastal areas) on the soil crust can be simulated, and the relationship between salt accumulation and crust hardness and air permeability can be studied; when spraying pure water, it can be used as a control experiment to isolate the effect of rainfall itself (without salt) on the crust and improve the comparability of experimental data. A solution collection bucket 110 is placed inside the support frame and below the preparation container and is used for collecting the liquid flowing down from the soil crust cultivation soil column 12; a heating device 111 is used to change the environmental temperature during the soil crust preparation process. This device realizes the refined research on the soil crust formation process through the organic combination of multi-variable controllable simulation (rainfall, salt, temperature), standardized experimental units (soil columns), and full-process data collection (liquid collection, environmental parameter monitoring), and has the characteristics of scientificity, flexibility, and convenient operation, providing an efficient experimental platform for basic research and application practice in the field of soil environment.

[0159] In an alternative embodiment of the present application, as Figure 4 shown, the crust data acquisition device 3 includes: A camera device 31, which is used for acquiring the crust image data corresponding to the soil crust; A near-infrared spectroscopy device 32, which is used for acquiring the crust spectral data; An electrochemistry sensor group 33, which is used for acquiring the crust electrochemistry data; A soil mechanics sensor 34, which is used for acquiring the mechanical properties of the crust.

[0160] Specifically, a high-resolution camera or microscope imaging system is used to photograph the surface and profile of the soil crust. The shooting parameters (such as focal length, exposure time, resolution) are controlled by image acquisition software to obtain image data at different angles and magnifications. These images visually present the surface morphology, structural texture, pore distribution and other characteristics of the crust, providing visual materials for subsequent morphological analysis and change monitoring.

[0161] Collection of crust spectral data: A spectrometer (such as a near-infrared spectrometer, Raman spectrometer) is used to measure the spectrum of the soil crust. The spectrometer emits light of a specific wavelength to irradiate the crust sample, and the absorption, reflection and scattering characteristics of the sample to light of different wavelengths are captured by the spectrometer and converted into spectral data. By analyzing the spectral data, information such as the chemical composition, mineral composition, and organic matter content of the crust can be obtained, revealing the material composition and chemical properties of the crust.

[0162] Collection of crust electrochemical data: With the help of an electrochemical workstation, an electrode probe is used to contact the soil crust to measure electrochemical parameters such as potential and current. By applying different electrical signals (such as constant potential, constant current), the electrochemical response of the crust is recorded, so as to analyze the ion exchange ability, redox characteristics, charge transport properties, etc. of the crust, and deeply understand the properties and reaction processes of the crust from the electrochemical perspective.

[0163] Collection of crust mechanical properties: A mechanical testing instrument (such as a micro-indentation instrument, tensile testing machine) is used to test the mechanical properties of the soil crust. For example, the micro-indentation instrument applies pressure to the surface of the crust through an indenter, measures the relationship between the indentation depth and the pressure change, and calculates mechanical parameters such as the hardness and elastic modulus of the crust; the tensile testing machine can measure the tensile strength, ductility and other characteristics of the crust, and obtain the behavior and performance data of the crust under mechanical action.

[0164] In the soil crust preparation and analysis system provided by the embodiment of the present application, the imaging device 31 collects the crust image data corresponding to the soil crust; the near-infrared spectral device 32 collects the crust spectral data; the electrochemical sensor group 33 collects the crust electrochemical data; the soil mechanical sensor 34 collects the crust mechanical properties. Through the collaborative monitoring of multiple devices such as imaging, spectroscopy, electrochemistry, and mechanics, a full-dimensional characterization of the soil crust in terms of "appearance - chemistry - physics" is achieved. This not only provides multi-source data input for the causal analysis model, but also enhances the accuracy, reliability and timeliness of causal relationship inference through data cross-validation and dynamic time-series analysis. Such a three-dimensional monitoring system is applicable to both controllable experiments under laboratory conditions (such as simulating different rainfall / temperature scenarios) and can be extended to long-term field monitoring, providing strong technical support for the scientific research and practical management of soil crusts.

[0165] Although embodiments of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the present invention, and such modifications and variations fall 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 soil preparation equipment, environmental data acquisition equipment, crust data acquisition equipment, and data processing equipment. The environmental data acquisition equipment and the crust data acquisition equipment are both communicatively connected to the data processing equipment, where: The soil preparation equipment is used to prepare soil crust; The environmental data acquisition equipment is used to acquire 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 salt content data of rainfall; The crust data acquisition equipment is used to acquire crust data corresponding to the soil crust. The crust data includes crust image data, crust spectral data, crust electrochemistry data, and crust mechanical properties; The data processing equipment is used to train a target causal analysis model based on the environmental time-series data and the crust data.

2. The soil crust preparation and analysis system according to claim 1, wherein The data processing equipment is used to input the crust image data, the crust spectral data, the crust electrochemistry data, and the crust mechanical properties into a preset feature fusion model, perform feature fusion on the crust image data, the crust spectral data, the crust electrochemistry data, and the crust mechanical properties, and generate target fusion features; Input the target fusion features, the temperature time-series data, the humidity time-series data, the rainfall time-series data, and the salt content data of rainfall into an initial causal analysis network, and train the initial causal analysis network to obtain the target causal analysis model.

3. The soil crust preparation and analysis system according to claim 2, characterized in that, The data processing equipment is used to respectively input the crust image data, the crust spectral data, the crust electrochemistry data, and the crust mechanical properties into corresponding embedding layers to generate corresponding image feature vectors, spectral feature vectors, electrochemistry feature vectors, and mechanical feature vectors; Based on a multi-scale feature fusion network, perform feature extraction on the image feature vector, the spectral feature vector, the electrochemistry feature vector, and the mechanical feature vector, and output corresponding sub-initial image feature vectors, sub-initial spectral feature vectors, sub-initial electrochemistry feature vectors, and sub-initial mechanical feature vectors at each scale; Based on an attention mechanism, perform feature fusion on the sub-initial image feature vectors, sub-initial spectral feature vectors, sub-initial electrochemistry feature vectors, and sub-initial mechanical feature vectors corresponding to each scale to generate the target fusion features.

4. The soil crust preparation and analysis system according to claim 3, characterized in that, The data processing equipment is used to calculate real-time statistical features corresponding to the image feature vector, the spectral feature vector, the electrochemistry feature vector, and the mechanical feature vector respectively; Input the real-time statistical features corresponding to the image feature vector, the spectral feature vector, the electrochemistry feature vector, and the mechanical feature vector respectively into an adaptive dilated convolution kernel generation mechanism based on a reinforcement learning algorithm; Use each real-time statistical feature as a state input to dynamically generate different initial dilated convolution kernel parameters. Each of the initial dilated convolution kernel parameters includes initial convolution kernel sizes of different scales; Based on a preset reward function, calculate reward values corresponding to each initial dilated convolution kernel parameter; Select the target empty convolution kernel parameter with the largest reward value according to the reward value; Based on the target convolution kernel sizes of different scales in the target empty convolution kernel parameter, perform feature extraction on the image feature vector, the spectral feature vector, the electrochemical feature vector, and the mechanical feature vector, 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.

5. The soil crust preparation and analysis system according to claim 3, wherein The data processing device is used to pair and combine the sub-initial image feature vector, the sub-initial spectral feature vector, and the sub-initial electrochemical feature vector of the same scale in pairs to generate multiple groups of cross-modal feature groups; For each of the feature groups, input the feature group into the attention calculation network to calculate the correlation between the two sub-feature vectors in the feature group; Based on the correlation, output the attention weight matrix; The attention weight matrix is used to characterize the degree of attention between the two sub-feature vectors in the feature group; Based on the attention weight matrix, update the two sub-feature vectors in the feature group; Repeat this process to obtain the sub-updated image feature vector, sub-updated spectral feature vector, sub-updated electrochemical feature vector, and sub-updated mechanical feature vector corresponding to each scale; Fuse the sub-updated image feature vectors, sub-updated spectral feature vectors, sub-updated electrochemical feature vectors, and sub-updated mechanical feature vectors 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; 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 the target fusion feature.

6. The soil crust preparation and analysis system according to claim 2, wherein The data processing device is used to generate a training set and a validation set based on the target fusion feature, the temperature time series data, the humidity time series data, the rainfall time series data, and the salt content data of the rainfall; Input the training set into the input layer of the initial causal analysis network; The hidden layer in the initial causal analysis network extracts features from the target fusion feature, the temperature time series data, the humidity time series data, the rainfall time series data, and the salt content data of the rainfall, and outputs the target feature; 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 feature; Based on the virtual causal relationship and the real causal relationship, calculate the loss value corresponding to the target loss function; the real causal relationship is determined by analyzing the environmental time series data and the crust data during the preparation of the soil crust; Update the parameters of the initial causal analysis network based on the loss value to obtain the initial causal analysis model; Validate the initial causal analysis model based on the validation set, and obtain the target causal analysis model when the accuracy rate of the initial causal analysis model is greater than or equal to a preset accuracy threshold.

7. The soil crust preparation and analysis system according to claim 6, characterized in that, The data processing device is configured to construct a prior causal knowledge graph based on prior knowledge of the causal relationship between environmental factors and crust data; Map the virtual causal relationship output by the initial causal analysis network to the corresponding nodes and edges in the prior causal knowledge graph; Detect the conflict between the virtual causal relationship and the prior causal knowledge graph; Calculate the causal consistency loss based on the conflict between the virtual causal relationship and the prior causal knowledge graph; 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; Based on the number of samples corresponding to each type, assign different weights to samples of different categories to determine the balanced loss function; Construct a target loss function based on the causal consistency loss, the balanced loss function, and the mean square error loss function.

8. 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 within a preset time period before the current time corresponding to the target area, input the environmental change data into the target causal analysis model, and output future soil crust data corresponding to the target area; Output corresponding management measures based on the future soil crust data.

9. The soil crust preparation and analysis system according to claim 1, wherein The soil preparation device includes: A preparation support, including a support frame and a preparation container, and the preparation container is placed on the support frame; A soil crust cultivation soil column, placed inside the preparation container, for preparing soil crust; A rainfall device, installed on the top surface inside the preparation container, for raining on the soil crust cultivation soil column; A pressure regulating nozzle, installed on the rainfall device, for simulating the magnitude of rainfall; One end of the rainfall device is connected to a salt dissolution preparation device through a first water pump and a first valve, and the other end is connected to a pure water preparation device through a second water pump and a second valve; when the first valve is opened, a salt solution is sprayed on the soil crust cultivation soil column; when the second valve is opened, pure water is sprayed on the soil crust cultivation soil column; A solution collection bucket, placed inside the support frame and below the preparation container, for collecting the liquid flowing down from the soil crust cultivation soil column; A heating device for changing the environmental temperature during the soil crust preparation process.

10. The soil crust preparation and analysis system according to claim 1, characterized in that, The crust data acquisition device includes: A camera device for acquiring the crust image data corresponding to the soil crust; A near-infrared spectroscopy device for acquiring the crust spectral data; An electrochemical sensor group for acquiring the crust electrochemical data; A soil mechanics sensor for acquiring the mechanical properties of the crust.

Images