Soil regulation and control decision analysis method and system based on dynamic detection

Abstract

The invention discloses a soil regulation and control decision analysis method and system based on dynamic detection, and the method comprises the steps: constructing a graph structure between soil regions through feature multi-modal data by employing a GNN graph neural network, enabling nodes to represent a field block, enabling edges to represent a spatial relation between adjacent field blocks, and outputting a spatial embedded vector of each node; coding historical time sequence data of each node in the characteristic multi-modal data based on a Bi-GRU bidirectional gating circulation unit, and outputting time sequence characteristics; tensor outer product is carried out on the space embedded vector and the time sequence features through a feature cross encoder, importance weights of different areas are dynamically distributed through a self-attention mechanism, and a soil state prediction value is output; and displaying a thermodynamic diagram of spatial distribution and temporal evolution of soil moisture, salinity and pollution risks through a GIS platform based on the soil state prediction value. And the accuracy and efficiency of soil regulation and control are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of agricultural intelligence technology, in particular to a soil regulation decision analysis method and system based on dynamic detection. BACKGROUND

[0002] With the development of agricultural modernization, the relationship between soil health and crop growth is increasingly valued. Traditional soil monitoring relies on fixed sensors or manual sampling, which has problems such as data update lag and static regulation strategy. Traditional methods are difficult to perceive the rapid changes of soil environment in real time, resulting in lagging regulation strategy and low efficiency. Existing systems mostly monitor single parameters separately, lacking comprehensive analysis of multi-dimensional data such as soil pH, salt, pollutants, and microbial activity, making the analysis result inaccurate and the basis for regulation decision insufficient. SUMMARY

[0003] The purpose of the present application is to solve the above problems, and a soil regulation decision analysis method and system based on dynamic detection are designed.

[0004] To achieve the above purpose, the technical scheme of the present application is as follows: further, in the soil regulation decision analysis method based on dynamic detection, the soil regulation decision analysis method comprises the following steps: Collecting multi-modal data, performing data cleaning and outlier deletion on the multi-modal data, and using a sliding window algorithm to extract time series features from the data to obtain feature multi-modal data; According to the feature multi-modal data, a GNN graph neural network is used to construct a graph structure between soil regions, a node represents a field, and an edge represents the spatial relationship between adjacent fields, and a spatial embedding vector of each node is output; Based on the Bi-GRU bidirectional gated recurrent unit, the historical time series data of each node in the feature multi-modal data is encoded, and a time series feature is output; The spatial embedding vector and the time series feature are tensor outer producted through a feature cross encoder, the importance weight of different regions is dynamically allocated through a self-attention mechanism, and a soil state prediction value is output; Based on the soil state prediction value, a heat map of the spatial distribution and time evolution of soil moisture, salt, and pollution risk is displayed on a GIS platform.

[0005] Further, in the soil regulation decision analysis method based on dynamic detection, the collecting multi-modal data, performing data cleaning and outlier deletion on the multi-modal data, and using a sliding window algorithm to extract time series features from the data to obtain feature multi-modal data comprises: Collecting physical, chemical and biological data of soil, simultaneously acquiring pollution index data in soil, and combining meteorological data and crop growth state data to obtain multi-modal data; The missing data in the multi-modal data is filled by using the mean value, the data exceeding 1.5 times of the quartile range is determined as an abnormal value by the box plot method, and the abnormal value is deleted to obtain the cleaned multi-modal data; After the processed multi-modal data is standardized, the sliding window algorithm is used, the window size is set to 24 hours and the step size is set to 1 hour according to the time resolution of the data, the time series features of the data are extracted, and the characteristic multi-modal data is obtained.

[0006] Further, in the soil regulation decision analysis method based on dynamic detection, the GNN graph neural network is used to construct the graph structure between the soil regions according to the characteristic multi-modal data, the node represents the field block, the edge represents the spatial relationship between the adjacent field blocks, and the spatial embedding vector of each node is output, including: According to the soil type, terrain and planted crop factors, the soil region is divided into field blocks with a size of 10m*10m as nodes; The spatial relationship between adjacent field blocks is determined by the GIS geographic information system, and if the field blocks have a common boundary, they are determined to be adjacent, and the edge is constructed; The hidden layer dimension of the GCN graph convolution network is set to 64, the learning rate is set to 0.005, and the iteration number is set to 500 times, the GNN graph neural network is used to process the characteristic multi-modal data, the graph structure between the soil regions is constructed, and the spatial embedding vector of each node is output.

[0007] Further, in the soil regulation decision analysis method based on dynamic detection, the Bi-GRU bidirectional gated recurrent unit is used to encode the historical time series data of each node in the characteristic multi-modal data, and the time series features are output, including: Two layers of Bi-GRU bidirectional gated recurrent units are set, the hidden layer dimension is set to 128, and the characteristic multi-modal data within 30 days is used as the historical time series data; The length of the historical time series data sequence of each node is combined with the periodicity of the soil state change and the effectiveness of the data; The bidirectional structure of Bi-GRU respectively encodes information from the forward and reverse directions of the time series, the forward encoding captures the evolution trend of the soil state over time, and the reverse encoding focuses on the potential influence of future state on the current state; An attention mechanism is added to the Bi-GRU, and the time series features are output, which contain the dynamic change characteristics of each field node in the historical time series.

[0008] Further, in the soil regulation decision analysis method based on dynamic detection, the importance weight of different regions is dynamically allocated by the self-attention mechanism, and the soil state prediction value is output, including: The spatial embedding vector and temporal features are dimensionally adjusted, and their outer product is performed to obtain the cross-feature tensor. The cross-feature tensor is taken as input and mapped into three vectors through a fully connected layer: query vector, key vector, and value vector. The query vector represents the features of the field that needs to be focused on, while the key vector describes the feature attributes of all fields. By calculating the similarity between the query vector and the key vector, the degree of association between each field and the field of focus is obtained. Then, after normalization by the softmax function, the weight coefficients are obtained.

[0009] Furthermore, in the aforementioned soil regulation decision analysis method based on dynamic detection, the step of performing a tensor outer product of spatial embedding vectors and temporal features through a feature cross-encoder, dynamically allocating importance weights for different regions through a self-attention mechanism, and outputting predicted soil state values ​​further includes: The feature cross encoder introduces a multi-scale fusion strategy, which extracts cross features at different spatial scales by performing convolution operations on the tensor outer product feature matrix by setting convolution kernels of different sizes.

[0010] Furthermore, in the aforementioned soil regulation decision analysis method based on dynamic detection, the step of displaying a heat map of the spatial distribution and temporal evolution of soil moisture, salinity, and pollution risk through a GIS platform based on the predicted soil state values ​​includes: Based on the predicted soil condition values, and combined with crop growth requirements, soil type characteristics, and agricultural production standards, an initial control plan should be determined, which should include at least irrigation measures, irrigation time, irrigation volume, fertilization measures, fertilization volume, and fertilization method.

[0011] Furthermore, in the soil regulation decision analysis system based on dynamic detection, the soil regulation decision analysis system includes the following modules: The data acquisition and processing module is used to acquire multimodal data, perform data cleaning and outlier removal on the multimodal data, and extract time-series features from the data using a sliding window algorithm to obtain feature multimodal data. The spatial vector acquisition module is used to construct a graph structure between soil regions using a GNN graph neural network based on the feature multimodal data. Nodes represent fields, edges represent the spatial relationship between adjacent fields, and the module outputs the spatial embedding vector of each node. The time-series feature acquisition module is used to encode the historical time-series data of each node in the feature multimodal data based on the Bi-GRU bidirectional gated cyclic unit and output the time-series features; The soil state prediction module is used to perform tensor outer product of spatial embedding vector and temporal features through feature cross encoder, dynamically allocate importance weights of different regions through self-attention mechanism, and output soil state prediction value. The soil heat map display module is used to display the spatial distribution and temporal evolution of soil moisture, salinity, and pollution risk through a GIS platform based on the predicted soil condition values.

[0012] Furthermore, in the soil regulation decision analysis system based on dynamic detection, the soil state prediction module includes the following sub-modules: The extraction submodule is used to introduce a multi-scale fusion strategy in the feature cross encoder. By setting convolution kernels of different sizes to perform convolution operations on the tensor outer product feature matrix, cross features at different spatial scales are extracted.

[0013] Furthermore, in the soil regulation decision analysis system based on dynamic detection, the soil heat map display module includes the following sub-modules: The output submodule is used to perform spatial interpolation processing on the predicted values ​​of discrete fields using ordinary kriging interpolation to generate continuous surface raster data; the predicted data is then sliced ​​according to time granularity, and a corresponding heat map raster is generated for each time slice to establish a GIS display platform.

[0014] Its beneficial effects are as follows: 1. It fully considers the spatial heterogeneity and temporal dynamic changes of soil, avoiding the blindness of traditional experience-based regulation, making regulation measures more aligned with actual soil needs, and improving the precision of soil regulation. 2. Furthermore, while ensuring the regulation effect, it controls resource consumption within a reasonable range, effectively reducing the waste of agricultural resources such as water and fertilizer. 3. It improves operational efficiency, indirectly reducing production costs. In summary, this technical solution can significantly improve the economic benefits of agricultural production and bring more income to farmers. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 This is a schematic diagram of the first embodiment of the soil regulation decision analysis method based on dynamic detection in this invention. Figure 2 This is a schematic diagram of the second embodiment of the soil regulation decision analysis method based on dynamic detection in this invention. Figure 3 This is a schematic diagram of the first embodiment of the soil regulation decision analysis system based on dynamic detection in this invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0019] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 As shown, a soil regulation decision analysis method based on dynamic detection includes the following steps: Step 101: Collect multimodal data, perform data cleaning and outlier removal on the multimodal data, and use the sliding window algorithm to extract time-series features from the data to obtain feature multimodal data; Specifically, in this embodiment, physical, chemical and biological data of the soil are collected, and pollution index data in the soil are obtained. Combined with meteorological data and crop growth status data, multimodal data are obtained. Missing data in multimodal data are filled using the mean. Data exceeding 1.5 times the interquartile range are identified as outliers using box plots. Outliers are then deleted to obtain cleaned multimodal data. After standardizing the processed multimodal data, the sliding window algorithm is used to extract temporal features from the data by setting the window size to 24 hours and the step size to 1 hour, based on the temporal resolution of the data.

[0020] Specifically, (a) Data collection Soil physical data: Real-time monitoring was conducted using soil sensors, including TDR soil moisture meters and soil bulk density meters. Sensors were deployed at a density of one sensor per 50-100 square meters, covering a soil layer of 0-60 cm, depending on the field size and soil heterogeneity. Data collected included soil texture (sand, silt, and clay content, in %) and bulk density (in g / cm³). 3 The data, including porosity (%) and water content (%), were collected once per hour.

[0021] Soil chemical data: Soil sampling analysis was conducted monthly. Sampling points were distributed in a quincunx or serpentine pattern, with 5-10 mixed samples collected from each field at a depth of 0-20 cm. Measurements included pH (dimensionless), organic matter content (g / kg), total nitrogen (g / kg), available phosphorus (mg / kg), and available potassium (mg / kg), analyzed using instruments such as atomic absorption spectrometer and spectrophotometer.

[0022] Soil biological data: Soil samples are collected quarterly for laboratory analysis, using the same collection methods as for chemical data. Data on microbial community structure (bacterial and fungal species and quantities, in CFU / g) and enzyme activities (urease, phosphatase, etc., in mg / g·h) are obtained and measured using high-throughput sequencing, colorimetric methods, etc.

[0023] Pollution index data: The content of heavy metals (lead, cadmium, mercury, etc., in mg / kg), pesticide residues (in mg / kg), and petroleum pollutants (in mg / kg) in the soil are tested using professional testing instruments. Samples are collected every six months, and the testing methods refer to relevant national standards.

[0024] Meteorological data: Daily temperature (°C), precipitation (mm), light intensity (lux), wind speed (m / s), humidity (%), etc., are obtained from the meteorological station database. The data is obtained once a day.

[0025] Crop growth status data: A combination of drone remote sensing (once a week) and field observation (every 3 days). Drones equipped with multispectral cameras acquire data such as crop height (in cm), stem diameter (in cm), leaf area index (dimensionless), and leaf color (reflected by vegetation index); field observations record crop growth stages (sowing period, seedling stage, flowering period, etc.).

[0026] (ii) Data Processing Data cleaning and outlier removal: For missing data, if the missing percentage is less than 5%, imputation is performed using the mean or median; if the missing percentage is large, interpolation is used. For outliers, the Laida criterion (3σ rule) is applied, that is, data values ​​deviating from the mean by more than 3 times the standard deviation are considered outliers; or the box plot method is used to identify data values ​​exceeding 1.5 times the interquartile range as outliers and remove them.

[0027] Data standardization / normalization: Standardize (Z-score standardization) or normalize (min-max normalization) the processed multimodal data to make the data be of the same order of magnitude, which is convenient for subsequent model training.

[0028] Temporal feature extraction: Using the sliding window algorithm, the window size is set to 24 hours and the step size is 1 hour, based on the temporal resolution of the data, to extract temporal features and obtain multimodal feature data.

[0029] Step 102: Construct a graph structure between soil regions using a GNN graph neural network based on the feature multimodal data. Nodes represent fields, edges represent the spatial relationship between adjacent fields, and the spatial embedding vector of each node is output. Specifically, in this embodiment, the soil area is divided into 10m×10m plots as nodes based on soil type, topography, and crop planting factors; The spatial relationship between adjacent fields is determined by a GIS geographic information system. If fields share a common boundary, they are considered adjacent and edges are constructed. The hidden layer dimension of the GCN graph convolutional network is set to 64, the learning rate is set to 0.005, and the number of iterations is set to 500. The GNN graph neural network is used to process the feature multimodal data, construct the graph structure between soil regions, and output the spatial embedding vector of each node.

[0030] Specifically, Field division: Based on factors such as soil type, topography, and crops planted, the soil area is divided into fields ranging from 10m×10m to 50m×50m as nodes. Edge construction: The spatial relationship between adjacent fields is determined using a Geographic Information System (GIS). Fields are considered adjacent and an edge is constructed if they share a common boundary or are less than 5 meters apart. Simultaneously, soil property similarity is considered; an edge is also constructed when the similarity of soil properties (pH value, organic matter content, etc.) between two fields is greater than 0.8. GNN Model Selection and Parameter Settings: We selected GCN (Graph Convolutional Network), set the hidden layer dimensions to 64 and 128, the learning rate to 0.005, and the number of iterations to 500. The GNN graph neural network processes the multimodal feature data, constructs a graph structure between soil regions, and outputs the spatial embedding vector for each node.

[0031] Step 103: Encode the historical time series data of each node in the feature multimodal data based on the Bi-GRU bidirectional gated recurrent unit, and output the time series features; Specifically, in this embodiment, two layers of Bi-GRU bidirectional gated recurrent units are set, the hidden layer dimension is set to 128, and the feature multimodal data within 30 days is used as historical time series data. Combining the periodicity of soil condition changes and the validity of the data, the length of the historical time series data sequence for each node is determined; Bi-GRU's bidirectional structure encodes information from the forward and backward directions of the time series, respectively. Forward encoding captures the evolution trend of soil state over time, while backward encoding focuses on the potential impact of future state on the present. An attention mechanism is added to Bi-GRU to output time-series features, which include the dynamic changes of each field node in the historical time series.

[0032] Specifically, Data input preprocessing: For the historical time-series data of each field node, time dimension alignment is performed before inputting it into the Bi-GRU. Since the collection frequencies of different data types vary (soil physical data hourly, meteorological data daily), all data needs to be converted to the same time granularity. Here, hourly units are chosen, and low-frequency data are augmented using linear interpolation to ensure data continuity over time. Simultaneously, dimensionality checks are performed to ensure consistent feature dimensions at each time step, avoiding model training issues caused by data dimensionality mismatch. Bi-GRU Network Hierarchical Design: A two-layer Bi-GRU structure is adopted, with the first hidden layer having a dimension of 128 and the second hidden layer having a dimension of 64. The first layer is responsible for capturing local temporal features in the data, such as short-term fluctuations in soil moisture content and the impact of daily temperature changes on soil condition. The second layer, based on the first layer, further extracts global temporal patterns, such as the cumulative effect of continuous rainfall on soil nutrient loss and the phased change trend of soil fertility during the crop growth cycle. The reset gate and update gate parameters in the network are trained using the backpropagation algorithm. The reset gate determines the degree to which historical information is ignored, and the update gate controls the proportion of historical information carried over to the current state, thereby achieving effective filtering and retention of key temporal information. Sequence length setting: Considering the periodicity of soil condition changes and the validity of the data, the historical time series data sequence length for each node is set to 720 hours (i.e., 30 days). This is because the main physicochemical properties of the soil (organic matter content, pH value) change relatively slowly in the short term, while a 30-day time span can cover the key stages of crop growth and a complete precipitation-evaporation cycle, which is sufficient to reflect the dynamic changes in soil conditions. For initial data of less than 30 days, a forward imputation method is used, using the earliest available data to supplement the beginning of the sequence to ensure consistency in sequence length. Temporal feature output processing: The bidirectional structure of Bi-GRU encodes information from both the forward (from past to present) and backward (from present to past) directions of the time series. Forward encoding captures the evolutionary trend of soil state over time, while backward encoding focuses on the potential impact of future state on the present (when predicting future precipitation, backward encoding can strengthen the correlation between current soil moisture content and future precipitation). The outputs of the two Bi-GRU layers are concatenated to obtain a temporal feature vector with a dimension of 128 (64+64). This vector contains the dynamic change characteristics of each field node in the historical time series, reflecting both short-term fluctuations and long-term trends.

[0033] Step 104: Perform tensor outer product between spatial embedding vector and temporal features using a feature cross encoder, dynamically allocate importance weights for different regions using a self-attention mechanism, and output the predicted soil state value. Specifically, in this embodiment, the spatial embedding vector and temporal features are dimensionally adjusted, and an outer product operation is performed on the spatial embedding vector and temporal features to obtain a cross-feature tensor. The cross-feature tensor is taken as input and mapped into three vectors through a fully connected layer: query vector, key vector, and value vector. The query vector represents the features of the field that needs to be focused on, while the key vector describes the feature attributes of all fields. By calculating the similarity between the query vector and the key vector, the degree of association between each field and the field of focus is obtained. Then, after normalization by the softmax function, the weight coefficients are obtained.

[0034] The feature cross encoder introduces a multi-scale fusion strategy, which extracts cross features at different spatial scales by performing convolution operations on the tensor outer product feature matrix by setting convolution kernels of different sizes.

[0035] Specifically, Tensor outer product operation implementation: When performing the tensor outer product of spatial embedding vectors (from GNN output, dimension 64) and temporal features (from Bi-GRU output, dimension 128), the dimensions of the two vectors are first adjusted to ensure they match correctly during the outer product operation. The result of the outer product operation is a 64×128 two-dimensional tensor. Each element of this tensor represents the degree of cross-correlation between spatial and temporal features. For example, the correlation between the spatial location of a field (closeness to irrigation water source) and the changes in soil nutrients in that field over a specific time period, or the matching degree between the spatial relationship of adjacent fields and their soil moisture content change trends during the same period. In this way, the originally independent spatial and temporal information is fused into a feature matrix containing interactive relationships. Self-attention mechanism weight allocation: The core of the self-attention mechanism is to calculate the importance weights of different field areas in the current soil state prediction. First, the feature matrix obtained by tensor outer product is used as input and mapped to three vectors through a fully connected layer: query vector, key vector, and value vector, all with a dimension of 256. The query vector represents the features of the field currently of interest, and the key vector describes the feature attributes of all fields. By calculating the similarity between the query vector and the key vector (using cosine similarity), the correlation between each field and the currently interested field is obtained. After normalization by the softmax function, the weight coefficients are obtained. The magnitude of the weight coefficients reflects the degree of influence of different regions on the current prediction result. For example, when predicting soil nitrogen content, fields closer to fertilization areas will have higher weights, while fields far from pollution sources may have lower weights when predicting soil pollutant concentrations. Multi-scale feature fusion: To avoid information loss caused by single-scale feature intersection, the feature cross encoder also introduces a multi-scale fusion strategy. By setting convolution kernels of different sizes (3×3 and 5×5) to perform convolution operations on the tensor outer product feature matrix, cross features at different spatial scales are extracted. These features are then concatenated with the weighted features output by the self-attention mechanism to form a more comprehensive fused feature. The 3×3 convolution kernel is used to capture feature interactions within local regions, while the 5×5 convolution kernel is used to extract spatial-temporal correlation information over a larger range, such as the synergistic relationship between soil nutrient migration and temporal changes among multiple adjacent fields. Soil condition prediction output: The fused features are passed through a fully connected layer (output dimension 10) to obtain the predicted soil condition value for each field node. The predicted indicators include 10 key soil parameters such as soil moisture content, pH value, available phosphorus content, and lead content. The activation function of the fully connected layer is a linear activation function to ensure that the predicted values ​​can cover the actual value range of the parameters, providing an accurate quantitative basis for the generation of subsequent control schemes.

[0036] Step 105: Based on the predicted soil condition values, display a heat map showing the spatial distribution and temporal evolution of soil moisture, salinity, and pollution risk through a GIS platform.

[0037] Specifically, in this embodiment, the predicted values ​​of discrete fields are spatially interpolated using ordinary Kriging interpolation to generate continuous surface raster data; the predicted data are then sliced ​​according to time granularity, and a corresponding heat map raster is generated for each time slice to establish a GIS display platform.

[0038] Specifically, Based on the predicted soil condition values, a heat map showing the spatial distribution and temporal evolution of soil moisture, salinity, and pollution risk is displayed through a GIS platform. This process includes the following sub-steps: Predictive data spatial correlation and format conversion Extract the field ID, prediction timestamp, and corresponding parameter values ​​(moisture content, salinity concentration, pollution risk index) from the soil condition prediction values, and associate them with the preset field spatial attribute table (including latitude and longitude coordinates and field boundary vector data) to generate a structured dataset with spatial index. The correlated data is converted into a GIS-compatible format, with spatial data in Shapefile format (including .prj projection files, coordinate system 1 is WGS84), attribute data stored in DBF tables, and time series data exported as CSV format with time dimension fields marked.

[0039] GIS platform basic layer construction Create a base map layer in the GIS platform, load the remote sensing image base map (resolution not less than 30m) of the study area and the vector data of administrative division boundaries as a spatial reference benchmark. Three thematic layers were created for soil moisture, salinity, and pollution risk. Each layer was associated with a spatial dataset of corresponding parameters, and a layer visibility toggle control was set. Spatial interpolation and heatmap rendering parameter configuration The predicted values ​​of discrete fields are spatially interpolated using ordinary kriging interpolation to generate continuous surface raster data (raster size set to 10m×10m). During the interpolation process, a semi-variogram model (using a spherical model by default) and cross-validation parameters (error threshold ≤5%) are set. Configure a color mapping scheme for the heat map: use a blue-cyan-green gradient for moisture content (0-20% threshold), a yellow-orange-red gradient for salt concentration (0-5g / kg threshold), and a green-yellow-red-purple gradient for pollution risk (risk level 1-5), and generate corresponding color scale legends. Time-series animation frame generation The predicted data is sliced ​​according to time granularity (such as hour, day, week), and each time slice generates a corresponding heatmap raster. Bind the time attribute fields of each parameter layer to the time axis control in the GIS platform, set the time step (default 1 day) and animation playback rate (10 frames / second) to generate a dynamic evolution time axis. Interactive features and visualization optimization Add a spatial query control to allow users to click on any location to obtain the predicted parameter value, time series curve, and field attribute information for that point. Configure layer transparency (default 70%), border highlight (field boundary line width 0.5pt, color dark gray), and legend floating display function to optimize the overlay effect of heat map and base map.

[0040] Results Output and Dynamic Display Export static heatmaps as TIFF format (300dpi), including coordinate information and legend; export dynamic time evolution results as GIF animations (duration ≤ 60 seconds) and interactive HTML maps (supports online viewing). Generate a visualization report that includes a description of the spatial distribution characteristics of each parameter, an analysis of its temporal evolution trend, and annotations of key areas (such as high-risk areas).

[0041] Its beneficial effects are as follows: 1. It fully considers the spatial heterogeneity and temporal dynamic changes of soil, avoiding the blindness of traditional experience-based regulation, making regulation measures more aligned with actual soil needs, and improving the accuracy and efficiency of soil regulation. 2. Furthermore, while ensuring the regulation effect, it controls resource consumption within a reasonable range, effectively reducing the waste of agricultural resources such as water and fertilizer. 3. It improves operational efficiency, indirectly reducing production costs. In summary, this technical solution can significantly improve the economic benefits of agricultural production and bring more income to farmers.

[0042] Please see Figure 2 In the soil regulation decision analysis method based on dynamic detection, the historical time series data of each node in the feature multimodal data is encoded based on the Bi-GRU bidirectional gated recurrent unit, and the output time series features include the following steps: Step 201: Set up a 2-layer Bi-GRU bidirectional gated recurrent unit, set the hidden layer dimension to 128, and use the feature multimodal data within 30 days as historical time series data; Step 202: Combining the periodicity of soil state changes and the validity of the data, determine the length of the historical time series data sequence for each node; Step 203: The bidirectional structure of Bi-GRU encodes information from the forward and backward directions of the time series, respectively. Forward encoding captures the evolution trend of soil state over time, while backward encoding focuses on the potential impact of future state on the present. Step 204: Add an attention mechanism to Bi-GRU to output time-series features, which include the dynamic changes of each field node in the historical time series.

[0043] The above describes embodiments of the soil regulation decision analysis method based on dynamic detection of the present invention. Please refer to [link / reference]. Figure 3 In the soil regulation decision analysis system based on dynamic detection, the soil regulation decision analysis system includes the following modules: The data acquisition and processing module is used to acquire multimodal data, perform data cleaning and outlier removal on the multimodal data, and use the sliding window algorithm to extract time-series features from the data to obtain feature multimodal data. The spatial vector acquisition module is used to construct a graph structure between soil regions using a GNN graph neural network based on feature multimodal data. Nodes represent fields, edges represent the spatial relationship between adjacent fields, and the module outputs the spatial embedding vector of each node. The temporal feature acquisition module is used to encode the historical temporal data of each node in the feature multimodal data based on the Bi-GRU bidirectional gated recurrent unit and output the temporal features. The soil state prediction module is used to perform tensor outer product of spatial embedding vector and temporal features through feature cross encoder, dynamically allocate importance weights of different regions through self-attention mechanism, and output soil state prediction value. The soil heat map display module is used to display the spatial distribution and temporal evolution of soil moisture, salinity, and pollution risk through a GIS platform based on predicted soil condition values.

[0044] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A soil regulation decision analysis method based on dynamic detection, characterized in that, The soil regulation decision analysis method includes the following steps: Multimodal data is collected, and the multimodal data is cleaned and outlier removed. The sliding window algorithm is used to extract time-series features from the data to obtain feature multimodal data. Based on the aforementioned feature multimodal data, a graph structure between soil regions is constructed using a GNN graph neural network. Nodes represent fields, edges represent the spatial relationships between adjacent fields, and the spatial embedding vector of each node is output. The historical time-series data of each node in the feature multimodal data is encoded based on the Bi-GRU bidirectional gated cyclic unit to output the time-series features; The spatial embedding vector and temporal features are multiplied by a feature cross encoder, and the importance weights of different regions are dynamically assigned through a self-attention mechanism to output the predicted soil state value. Based on the predicted soil condition values, a heat map showing the spatial distribution and temporal evolution of soil moisture, salinity, and pollution risk is displayed through a GIS platform.

2. The soil regulation decision analysis method based on dynamic detection as described in claim 1, characterized in that, The process involves collecting multimodal data, cleaning and removing outliers from the multimodal data, and extracting temporal features using a sliding window algorithm to obtain feature-rich multimodal data, including: Collect physical, chemical and biological data of the soil, and obtain data on pollution indicators in the soil. Combine this with meteorological data and crop growth status data to obtain multimodal data. Missing data in multimodal data are filled using the mean. Data exceeding 1.5 times the interquartile range are identified as outliers using box plots. Outliers are then deleted to obtain cleaned multimodal data. After standardizing the processed multimodal data, the sliding window algorithm is used to extract temporal features from the data by setting the window size to 24 hours and the step size to 1 hour, based on the temporal resolution of the data.

3. The soil regulation decision analysis method based on dynamic detection as described in claim 1, characterized in that, The process involves constructing a graph structure between soil regions using a GNN (Graph Neural Network) based on the multimodal data, where nodes represent fields and edges represent spatial relationships between adjacent fields. The output is a spatial embedding vector for each node, including: Based on soil type, topography, and crop planting factors, the soil area is divided into 10m×10m plots as nodes; The spatial relationship between adjacent fields is determined by a GIS geographic information system. If fields share a common boundary, they are considered adjacent and edges are constructed. The hidden layer dimension of the GCN graph convolutional network is set to 64, the learning rate is set to 0.005, and the number of iterations is set to 500. The GNN graph neural network is used to process the feature multimodal data, construct the graph structure between soil regions, and output the spatial embedding vector of each node.

4. The soil regulation decision analysis method based on dynamic detection as described in claim 1, characterized in that, The Bi-GRU-based bidirectional gated recurrent unit encodes the historical time-series data of each node in the feature multimodal data and outputs time-series features, including: Two layers of Bi-GRU bidirectional gated recurrent units are set up, and the hidden layer dimension is set to 128. The feature multimodal data within 30 days is used as historical time series data. Combining the periodicity of soil condition changes and the validity of the data, the length of the historical time series data sequence for each node is determined; Bi-GRU's bidirectional structure encodes information from the forward and backward directions of the time series, respectively. Forward encoding captures the evolution trend of soil state over time, while backward encoding focuses on the potential impact of future state on the present. An attention mechanism is added to Bi-GRU to output time-series features, which include the dynamic changes of each field node in the historical time series.

5. The soil regulation decision analysis method based on dynamic detection as described in claim 1, characterized in that, The process involves performing a tensor outer product between the spatial embedding vector and temporal features using a feature cross-encoder, dynamically assigning importance weights to different regions through a self-attention mechanism, and outputting a predicted soil state value, including: The spatial embedding vector and temporal features are dimensionally adjusted, and their outer product is performed to obtain the cross-feature tensor. The cross-feature tensor is taken as input and mapped into three vectors through a fully connected layer: query vector, key vector, and value vector. The query vector represents the features of the field that needs to be focused on, while the key vector describes the feature attributes of all fields. By calculating the similarity between the query vector and the key vector, the degree of association between each field and the field of focus is obtained. Then, after normalization by the softmax function, the weight coefficients are obtained.

6. The soil regulation decision analysis method based on dynamic detection as described in claim 1, characterized in that, The method of performing tensor outer product between spatial embedding vectors and temporal features through a feature cross-encoder, dynamically assigning importance weights to different regions through a self-attention mechanism, and outputting soil state prediction values ​​also includes: The feature cross encoder introduces a multi-scale fusion strategy, which extracts cross features at different spatial scales by performing convolution operations on the tensor outer product feature matrix by setting convolution kernels of different sizes.

7. The soil regulation decision analysis method based on dynamic detection as described in claim 1, characterized in that, The process of displaying a heat map showing the spatial distribution and temporal evolution of soil moisture, salinity, and pollution risk based on the predicted soil condition values ​​via a GIS platform includes: The predicted values ​​of discrete fields are spatially interpolated using ordinary kriging to generate continuous surface raster data. The predicted data are then sliced ​​according to time granularity, and a corresponding heat map raster is generated for each time slice to establish a GIS display platform.

8. A soil regulation decision analysis system based on dynamic detection, characterized in that, The soil regulation decision analysis system includes the following modules: The data acquisition and processing module is used to acquire multimodal data, perform data cleaning and outlier removal on the multimodal data, and extract time-series features from the data using a sliding window algorithm to obtain feature multimodal data. The spatial vector acquisition module is used to construct a graph structure between soil regions using a GNN graph neural network based on the feature multimodal data. Nodes represent fields, edges represent the spatial relationship between adjacent fields, and the module outputs the spatial embedding vector of each node. The time-series feature acquisition module is used to encode the historical time-series data of each node in the feature multimodal data based on the Bi-GRU bidirectional gated cyclic unit and output the time-series features; The soil state prediction module is used to perform tensor outer product of spatial embedding vector and temporal features through feature cross encoder, dynamically allocate importance weights of different regions through self-attention mechanism, and output soil state prediction value. The soil heat map display module is used to display the spatial distribution and temporal evolution of soil moisture, salinity, and pollution risk through a GIS platform based on the predicted soil condition values.

9. The soil regulation decision analysis system based on dynamic detection as described in claim 8, characterized in that, The soil condition prediction module includes the following sub-modules: The extraction submodule is used to introduce a multi-scale fusion strategy in the feature cross encoder. By setting convolution kernels of different sizes to perform convolution operations on the tensor outer product feature matrix, cross features at different spatial scales are extracted.

10. The soil regulation decision analysis system based on dynamic detection as described in claim 8, characterized in that, The soil heat map display module includes the following sub-modules: The output submodule is used to perform spatial interpolation processing on the predicted values ​​of discrete fields using ordinary kriging interpolation to generate continuous surface raster data; it also performs time-series slicing on the predicted data according to time granularity, generating a corresponding heat map raster frame for each time slice, and establishing a GIS display platform.

Images