A method and prediction system for constructing a farmland soil moisture content prediction model

Abstract

This invention discloses a method and system for constructing a farmland soil moisture content prediction model, belonging to the field of farmland water conservancy engineering technology. The method includes acquiring multi-source observation data of a target farmland area; constructing a domain knowledge graph based on farmland soil moisture domain knowledge, and determining key driving factors affecting soil moisture content changes and their correlations based on the domain knowledge graph; constructing knowledge constraints for constraining a physical information neural network model based on the key driving factors and their correlations; and training the physical information neural network model based on the multi-source observation data and the knowledge constraints to obtain a farmland soil moisture content prediction model. This invention solves the problems of difficulty in obtaining traditional model parameters and the lack of physical mechanism constraints in purely data-driven models, improving the accuracy, stability, and physical rationality of farmland soil moisture content prediction.

Description

Technical Field

[0001] This invention relates to the field of farmland water conservancy engineering technology, and in particular to a method for constructing a farmland soil moisture content prediction model and a prediction system. Background Technology

[0002] Soil moisture content is a crucial indicator affecting crop yield. Accurately acquiring and predicting changes in farmland soil moisture content is not only the scientific basis for developing precision irrigation and improving agricultural water resource utilization efficiency, but also an important basis for agricultural drought monitoring and disaster prevention and mitigation decision-making. However, soil moisture content exhibits high spatiotemporal heterogeneity, and its changes are influenced by the complex coupling of multiple factors, including meteorological conditions, soil physicochemical properties, vegetation cover, and field management practices. This presents a significant technical challenge to achieving high-precision, spatiotemporally continuous long-term prediction of soil moisture content. Current methods for predicting soil moisture content mainly rely on numerical simulations based on mechanistic models. Mechanistic models based on soil hydrodynamic equations (such as HYDRUS and TEC models) have well-defined physical mechanisms and can simulate the transport process of water in soil profiles. However, their application heavily depends on detailed soil hydraulic parameters (such as unsaturated hydraulic conductivity and soil moisture characteristic curves) and accurate initial boundary conditions. These parameters are difficult and costly to obtain in practical applications, and parameter uncertainties often lead to the accumulation of simulation errors, limiting the generalizability of the models. Currently, the automatic soil moisture monitoring stations deployed in farmland areas of my country provide multi-dimensional, time-series matched observation data on soil moisture, air temperature, air humidity, and air pressure. This data contains key information on vegetation growth status, water supply, and environmental stress, serving as an important data source for analyzing soil moisture changes and achieving accurate drought early warning. However, no technology yet fully integrates the aforementioned multi-source observation data to analyze the influencing factors of spatiotemporal soil moisture changes through knowledge reasoning and to construct an efficient farmland soil moisture content prediction model using knowledge-constrained PINN, thus achieving the goal of long-term prediction of farmland soil moisture content. Furthermore, existing technologies have shortcomings, including low prediction accuracy, poor spatiotemporal adaptability, reliance on difficult-to-obtain soil hydraulic parameters, and a lack of physical mechanism support. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a method and system for constructing a prediction model for farmland soil moisture content, aiming to solve the problems of existing technologies.

[0004] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a method for constructing a prediction model for farmland soil moisture content, including: Acquire multi-source observation data of the target farmland area; A domain knowledge graph is constructed based on knowledge of farmland soil moisture, and the key driving factors affecting changes in soil moisture content and the correlation between the key driving factors are determined based on the domain knowledge graph. Based on the key driving factors and their correlations, a knowledge constraint is constructed to constrain the physical information neural network model. Based on the multi-source observation data and the knowledge constraints, the physical information neural network model is trained to obtain a farmland soil moisture content prediction model; wherein, the farmland soil moisture content prediction model is used to output the soil moisture content prediction result of the target farmland area at the target soil layer depth when real-time observation data of the target farmland area is input.

[0005] In some optional embodiments, the method further includes: Preprocessing of the multi-source observation data includes at least one of the following: time-series alignment, missing value imputation, outlier removal, data cleaning, feature filtering, encoding, normalization, and dataset partitioning. The multi-source observation data includes soil factor data, meteorological environment data, and crop growth data.

[0006] In some optional embodiments, the construction of a domain knowledge graph based on farmland soil moisture domain knowledge includes: Obtain domain data on farmland soil moisture, and extract environmental factor knowledge points, soil hydrology knowledge points, and crop physiology knowledge points from the domain data; The environmental knowledge points, soil and hydrology knowledge points, and crop physiology knowledge points are used as knowledge graph nodes, and the association edges between the knowledge graph nodes are constructed to obtain the domain knowledge graph.

[0007] In some optional embodiments, determining the key driving factors affecting changes in soil moisture content and their correlations based on the domain knowledge graph includes: Candidate driving factors are determined based on the relationship between each node in the domain knowledge graph and the soil moisture content node; Based on the candidate driving factors, the degree of influence of each candidate driving factor on the change of soil moisture content is obtained; The key driving factors are selected based on the degree of influence, and the correlation between the key driving factors is determined.

[0008] In some optional embodiments, the physical information neural network model includes a meteorological-driven soil state subnetwork, a soil crop physiology subnetwork, and a crop response feature subnetwork; The knowledge constraints include structural constraints and loss constraints; The structural constraints include classifying the key driving factors into meteorological driving factors, soil state factors, and soil property factors according to their categories. Input the key driving factors of different categories into the corresponding sub-network branches; The connection methods between each sub-network branch are determined based on the correlation of the key driving factors, so that the network topology of the physical information neural network model matches the soil moisture transport mechanism.

[0009] In some optional embodiments, the loss constraint includes at least one of physical residual constraint, value residual constraint, and structural residual constraint; The physical residual constraints are constructed based on at least one of the following: soil water balance relationship, water infiltration relationship, evaporation relationship, root water absorption relationship and Richards equation; The residual constraint is constructed based on the reasonable range of soil moisture content corresponding to the soil type of the target farmland area. The structural residual constraints are constructed based on the correlation between the key driving factors.

[0010] In some optional embodiments, training the physical information neural network model based on the multi-source observation data and the knowledge constraints includes: The multi-source observation data are input into the corresponding sub-network branches of the physical information neural network model according to the categories of key driving factors to obtain the feature representations of each category of factors; The feature representations of each category of factors are fused through the fusion output layer of the physical information neural network model to obtain a comprehensive feature representation; Based on the comprehensive feature representation, output the predicted value of stratified soil moisture content; The data fitting loss is calculated based on the deviation between the predicted and measured soil moisture contents of the stratified soil, and the knowledge constraint loss is calculated based on the knowledge constraints. A total loss function is constructed based on the data fitting loss and the knowledge constraint loss, and the model parameters of the physical information neural network model are updated through backpropagation. During the training of the physical information neural network model, the weight of the knowledge constraint loss in the total loss function is dynamically adjusted.

[0011] Secondly, this application provides a method for predicting farmland soil moisture content, including: Acquire real-time observation data of the target farmland area; The real-time observation data is input into the farmland soil moisture content prediction model to obtain the soil moisture content prediction results of the target farmland area at the target soil depth. The farmland soil moisture content prediction model is obtained through the farmland soil moisture content prediction model construction method described above.

[0012] Thirdly, this application provides a system for predicting farmland soil moisture content, including: The data acquisition module is used to acquire multi-source observation data of the target farmland area, including soil factor data, meteorological environment data and crop growth data; The data preprocessing module is used to preprocess the multi-source observation data; The knowledge graph construction module is used to construct a domain knowledge graph based on knowledge of farmland soil moisture, and to determine the key driving factors affecting changes in soil moisture content and the correlation between the key driving factors based on the domain knowledge graph. A knowledge constraint construction module is used to construct knowledge constraints for constraining the physical information neural network model based on the key driving factors and the correlation between the key driving factors. The knowledge constraints include structural constraints and loss constraints. The model training module is used to train the physical information neural network model based on the standardized observation data and the knowledge constraints to obtain a farmland soil moisture content prediction model. The prediction output module is used to input real-time observation data of the target farmland area into the farmland soil moisture content prediction model and output the soil moisture content prediction results of the target farmland area at the target soil depth.

[0013] In some optional embodiments, the physical information neural network model includes a meteorological-driven soil state subnetwork, a soil crop physiology subnetwork, and a crop response feature subnetwork; The knowledge constraint construction module includes: A structural constraint unit is used to set up sub-networks of the physical information neural network model according to the category of the key driving factors, and to determine the connection mode between each sub-network according to the correlation between the key driving factors. The loss constraint unit is used to construct the knowledge constraint loss term of the physical information neural network model based on the soil moisture transport mechanism, the reasonable range of soil moisture content, and the correlation of key driving factors.

[0014] Through the above technical solutions, the beneficial effects of this invention are as follows: by acquiring multi-source time-series data on stratified soil moisture content and meteorological environment, and combining prior knowledge of soil moisture transport to determine the structure of key driving factors, the model input can more accurately reflect the main influencing factors of changes in farmland soil moisture content; the structure of key driving factors, the laws of soil moisture transport, and the reasonable range of moisture content corresponding to soil type are respectively transformed into structural constraints, physical constraints, and value constraints, and further a knowledge constraint loss is embedded in the PINN model training process, so that the model prediction results not only fit the observed data, but are also constrained by the physical laws of soil moisture, thereby reducing the risk of the prediction results deviating from the actual water transport laws. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention, and the embodiments in the accompanying drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating a method for constructing a farmland soil moisture content prediction model according to an embodiment of this application; Figure 2 A schematic diagram of the domain knowledge graph construction process provided in an embodiment of this application; Figure 3 This is a partial visualization diagram of a domain knowledge graph provided in an embodiment of this application; Figure 4 An architecture diagram of a knowledge-constrained PINN model provided in an embodiment of this application; Figure 5 A comparison diagram of 50-hour prediction results and actual measurements for various soil depths provided in an embodiment of this application; Figure 6 A comparison diagram of 120-hour prediction results and actual measurements for various soil depths provided in an embodiment of this application; Figure 7 A structural diagram of a farmland soil moisture content prediction system provided in an embodiment of this application; The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0017] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Equivalent substitutions or conventional modifications made by those skilled in the art based on the disclosure of the present invention without departing from the concept of the present invention should all fall within the scope of protection of the present invention.

[0018] This embodiment provides a method for constructing a farmland soil moisture content prediction model based on a knowledge-constrained PINN model. The method uses multi-source observation data of the target farmland area as a foundation, constructs a domain knowledge graph using knowledge of farmland soil moisture, identifies key driving factors affecting soil moisture content changes and their correlations based on the domain knowledge graph, and transforms these correlations into structural and loss constraints of the PINN model, thereby training a soil moisture content prediction model for stratified soil moisture content prediction.

[0019] like Figure 1 As shown, the method in this embodiment includes the following steps: S1, acquire multi-source observation data of the target farmland area; S2, construct a domain knowledge graph based on knowledge of farmland soil moisture, and determine the key driving factors affecting changes in soil moisture content and the correlation between the key driving factors based on the domain knowledge graph; S3, Based on the key driving factors and their correlation, construct knowledge constraints for constraining the physical information neural network model; S4. Based on the multi-source observation data and the knowledge constraints, the physical information neural network model is trained to obtain a farmland soil moisture content prediction model. The farmland soil moisture content prediction model is used to output the predicted soil moisture content of the target farmland area at the target soil layer depth when real-time observation data of the target farmland area is input.

[0020] In step S1, the target farmland area refers to the farmland monitoring area where soil moisture content prediction is required. This area can be a farmland demonstration area, agricultural experimental station, irrigation management area, or farmland water conservancy monitoring area equipped with automatic soil moisture monitoring equipment. The target farmland area can correspond to one or more monitoring stations and can have corresponding geographical location, soil type, crop type, soil depth, and sampling time range.

[0021] The multi-source observation data refers to a dataset obtained from the target farmland area that reflects changes in soil moisture content and its influencing factors. This multi-source observation data may include soil factor data, meteorological environmental data, and crop growth data. Specifically, soil factor data may include historical soil moisture content at different soil depths, soil type, soil texture, soil depth, and soil moisture change rate; meteorological environmental data may include air temperature, air humidity, air pressure, precipitation, light intensity, total radiation, wind speed, and wind direction; and crop growth data may include crop phenological stages, SIF fluorescence values, historical NDVI values, crop growth level, and crop water stress level. This data can be used as model training data or as input data for the trained model.

[0022] The stratified soil moisture content data can include the volumetric water content of soil at different soil depths such as 10cm, 20cm, 40cm, and 60cm, or other soil depths depending on the actual deployment depth of the observation equipment. Meteorological environmental data can include one or more of the following: air temperature, air humidity, air pressure, precipitation, light intensity, total radiation, wind speed, and wind direction. Multi-source observation data can also include auxiliary information such as sampling time, station number, soil type, soil depth, and crop growth stage.

[0023] Furthermore, the multi-source observation data is obtained through preprocessing, including: First, unifying the timestamps of data collected by different sensors or observation stations, and resampling or interpolating various types of data according to preset time intervals to ensure consistency between stratified soil moisture content data and meteorological environment data in the time dimension; Second, identifying and removing outliers from the multi-source observation data. For example, if the soil moisture content at a certain moment significantly exceeds the reasonable range for the corresponding soil type, or if the data mutation amplitude between adjacent sampling times exceeds a preset threshold, the data is identified as outlier data; Third, filling in missing data. For short-term missing data, linear interpolation, sliding window averaging, or similar data from adjacent time periods can be used to fill in the missing data; for long-term continuous missing data, the corresponding time period can be removed from the training samples to avoid affecting the stability of model training; Then, normalizing or standardizing data of different dimensions to ensure that variables such as air temperature, precipitation, light intensity, and soil moisture content are on a uniform numerical scale; Finally, dividing the standardized observation data into training set, validation set, and test set according to time order. The training set is used for learning the parameters of the Physical Information Neural Network (PINN) model, the validation set is used for adjusting the model parameters and determining convergence, and the test set is used for final prediction performance evaluation. Preprocessing of multi-source observation data ensures that data from different data sources, sampling frequencies, and units can form sample data that meets the model training requirements, thereby reducing the impact of outlier data, missing data, and unit differences on the stability of model training.

[0024] This application acquires multi-source observation data of the target farmland area and constructs a domain knowledge graph by combining knowledge of farmland soil moisture. This enables the model construction process to not only rely on historical observation data, but also to use knowledge of soil moisture transport, meteorological drivers, and crop water response to identify key driving factors affecting soil moisture content changes, thereby improving the correlation between model input factors and soil moisture content changes.

[0025] In step S2, after completing the acquisition and preprocessing of multi-source observation data, a domain knowledge graph is constructed based on knowledge of farmland soil moisture. The key driving factors influencing changes in soil moisture content and their correlations are then determined based on this domain knowledge graph. This domain knowledge graph transforms unstructured knowledge related to farmland irrigation, agricultural hydrology, crop water stress, and soil moisture transport into structured knowledge that can be utilized by the PINN model, thus providing a basis for subsequently constructing model structural constraints and loss function constraints.

[0026] The aforementioned knowledge of farmland soil moisture refers to knowledge content that reflects the changing patterns of soil moisture content, obtained from data on farmland irrigation, soil moisture transport, crop water stress, and meteorological environmental influences. Specifically, relevant data on farmland soil moisture can be obtained first. This data can then be segmented into text, and terms can be identified and synonyms merged to obtain knowledge points related to changes in soil moisture content. These knowledge points can then be used as graph nodes, and connections between nodes can be constructed based on co-occurrence, interaction, or hierarchical relationships, thereby obtaining a domain knowledge graph. Therefore, this application does not directly extract graph nodes from the data, but rather first identifies knowledge points from the data and then processes these identified knowledge points as graph nodes.

[0027] like Figure 2 As shown, the construction process of the domain knowledge graph includes graph design, text processing, knowledge extraction, relationship determination, graph generation, and graph verification steps.

[0028] First, a knowledge graph design model is constructed. This model defines the node types, relationship types, and knowledge extraction rules within the domain knowledge graph. In this embodiment, the knowledge graph design model includes at least three types of nodes: environmental factors, soil hydrological factors, and crop physiological and remote sensing factors; and three types of edges: hierarchical relationships, causal driving relationships, and related influence relationships. By pre-establishing the knowledge graph design model, the subsequent knowledge extraction process has clear target boundaries, avoiding the extraction of generalized information unrelated to soil moisture content prediction from literature.

[0029] Secondly, domain-specific data is acquired for constructing the domain knowledge graph. This data can include professional information related to agricultural hydrology, farmland irrigation, soil moisture transport, crop drought and flood stress, SIF remote sensing, vegetation index, and stress early warning. In one specific implementation, Web of Science and CNKI core databases can be selected as literature sources. Searches can be conducted using keywords or combinations of keywords such as drought stress, drought and flood stress, crop water stress, SIF fluorescence, soil moisture transport, soil moisture infiltration, evaporation, root water uptake, and stress early warning. Core literature is then selected based on citation frequency, journal level, and industry recognition. Simultaneously, industry standards for farmland irrigation, monographs on soil moisture mechanisms, and project field measurement reports can be used as supplementary data sources to improve the adaptability of the domain knowledge graph to actual farmland scenarios.

[0030] Then, the selected domain data undergoes text preprocessing. This preprocessing includes one or more of the following: text segmentation, irrelevant content removal, terminology standardization, synonym merging, and entity candidate word extraction. For example, "soil moisture content," "soil volumetric water content," and "layered soil moisture" are standardized into node expressions related to soil moisture content; "rainfall" and "precipitation amount" are standardized into precipitation amount nodes; and "sunlight," "solar radiation," and "total radiation" are standardized or have corresponding relationships established based on their actual meanings.

[0031] Furthermore, prompt words are generated based on a pre-defined graph design model, and a large language model is used to perform semantic understanding and knowledge extraction on the preprocessed text fragments. The knowledge extraction includes at least entity recognition, concept recognition, extraction of relationships between concepts, and knowledge tuple generation. Specifically, entity recognition is used to identify domain entities from domain data, including temperature, precipitation, air humidity, light intensity, soil type, stratified soil moisture content, soil moisture infiltration, evaporation, root water absorption, SIF fluorescence, NDVI, crop phenological stages, photosynthetic efficiency, and drought and flood stress levels; relationship extraction is used to determine the direction of interaction and relationship type between entities in each domain; and knowledge tuple generation is used to represent the extracted domain knowledge in a structured form of "head entity - relationship - tail entity".

[0032] This application first identifies environmental factor knowledge points, soil hydrological knowledge points, and crop physiological knowledge points from farmland soil moisture domain data. Then, it uses the identified knowledge points as knowledge graph nodes and constructs the association edges between the nodes. This can transform unstructured domain data into computable, associative, and structured knowledge that can be used to construct model constraints, thereby improving the operability of domain knowledge in model training.

[0033] In this embodiment, the extracted domain knowledge graph nodes mainly include the following three categories: The first category is environmental factor nodes. These environmental factor nodes are used to characterize the driving effect of the external environment on changes in soil moisture content, and may include one or more of the following: temperature, precipitation, air humidity, light intensity, air pressure, wind speed, wind direction, total radiation, and soil type. Among them, precipitation is usually related to soil moisture recharge, temperature and light intensity are usually related to evaporation and evapotranspiration, air humidity is related to evaporation intensity, and soil type is related to water retention capacity and water transport rate.

[0034] The second category is soil hydrological nodes. These nodes characterize soil moisture state and its transport processes, and may include one or more of the following: stratified soil moisture content, shallow soil moisture content, deep soil moisture content, soil moisture infiltration, soil moisture evaporation, root water uptake, soil moisture conduction, and soil water balance. Stratified soil moisture content serves as the target prediction object and also as historical state input for model training; soil moisture infiltration, evaporation, and root water uptake serve as physical constraints, used in the subsequent construction of the knowledge constraint loss term for the PINN model.

[0035] The third category consists of crop physiological and remote sensing nodes related to soil moisture changes. These nodes are used to characterize crop growth status and water stress status, and may include one or more of the following: SIF fluorescence, NDVI, crop phenological stage, photosynthetic efficiency, and drought / flood stress levels. These nodes are not independent of soil moisture content prediction but are used to reflect the impact of soil moisture changes on crop physiological status, or to assist in determining the relationship between soil moisture changes and drought / flood stress.

[0036] After identifying the nodes in the graph, the associated edges between the nodes are constructed. These associated edges include at least one of hierarchical relationships, causal relationships, and correlational influence relationships. Hierarchical relationships describe the hierarchical relationship between concepts. For example, drought and waterlogging stress can include drought stress and waterlogging stress; soil moisture transport can include water infiltration, evaporation, runoff, and root uptake; meteorological environmental factors can include temperature, precipitation, air humidity, and light intensity. Causal relationships describe the directional influence of one node on another. For example, precipitation increases shallow soil moisture content through infiltration; evaporation decreases surface soil moisture content; rising temperatures and increased light intensity increase evaporation or evapotranspiration, thus affecting soil moisture content; root uptake decreases root zone soil moisture content; drought stress inhibits crop photosynthetic efficiency.

[0037] The term "correlation relationship" describes a statistical or mechanistic correlation between nodes, but it is not necessarily limited to a direct causal relationship. For example, soil texture affects the rate of water transport; soil type affects the reasonable range of soil moisture content; there is a correlation between light intensity and SIF fluorescence intensity; and there is a time lag between changes in shallow soil moisture content and changes in deep soil moisture content.

[0038] For each associated edge, its relationship type, direction of influence, and association strength are recorded. The relationship type distinguishes between hierarchical, causal, or related influence relationships; the direction of influence indicates the direction of impact between nodes; and the association strength indicates the degree of influence of the relationship on soil moisture content prediction. In one specific implementation, the association strength can be determined by one or more of the following: the amount of documentary evidence, the frequency of documentary citations, the weight of expert rules, Bayesian inference results, or the statistical correlation of historical observation data.

[0039] After constructing the nodes and associated edges, a domain knowledge graph is formed. For example... Figure 3 As shown, the domain knowledge graph focuses on soil moisture content changes and structurally expresses the relationships between meteorological environment, soil hydrology, crop physiology, and remote sensing indicators. Through this domain knowledge graph, key driving factors directly or indirectly related to soil moisture content changes can be identified, and the interaction pathways between different driving factors can be clarified.

[0040] Furthermore, key driving factors are identified based on the domain knowledge graph. Specifically, this includes: taking the stratified soil moisture content node as the target node, calculating the association paths from each candidate node in the graph to the target node; calculating the degree of influence of the candidate node on the change of soil moisture content based on the association path length, association edge type, direction of action, and association strength; and identifying the candidate node as a key driving factor when the degree of influence of the candidate node is greater than a preset threshold.

[0041] The key driving factors refer to knowledge points or observed variables in the domain knowledge graph that are directly or indirectly related to the soil moisture content node and have an impact on changes in soil moisture content. When determining key driving factors, the soil moisture content node can be used as the target node. The association paths, edge types, directions of influence, and association strengths between other graph nodes and the target node can be analyzed. Based on the analysis results, graph nodes whose influence on changes in soil moisture content meets preset conditions are selected, and the knowledge points or observed variables corresponding to these graph nodes are determined as key driving factors. The key driving factors may include precipitation, air temperature, air humidity, light intensity, historical stratified soil moisture content, soil type, and target soil layer depth, etc.

[0042] The correlation between key driving factors refers to the direction of action, time lag of action, type of influence, or strength of correlation between different key driving factors. For example, precipitation affects the moisture content of shallow soil through infiltration, air temperature and light intensity affect the moisture content of surface soil through evaporation or evapotranspiration, changes in shallow soil moisture content precede changes in deep soil moisture content, and soil type affects the reasonable range of soil moisture content and the rate of water infiltration.

[0043] For example, when there is a causal driving path of "precipitation-infiltration-shallow soil moisture content" between the precipitation node and the shallow soil moisture content node, precipitation is identified as a meteorological driving factor; when there is a positive influence relationship between the temperature node and the light intensity node and the evaporation node, and a negative influence relationship between the evaporation node and the surface soil moisture content, temperature and light intensity are identified as key driving factors affecting the change of surface soil moisture content; when there is a correlation between the soil type node and the water transport rate and the reasonable range of moisture content, soil type is identified as a soil property factor; when there is a temporal continuity relationship between the historical stratified soil moisture content node and the future stratified soil moisture content node, historical stratified soil moisture content is identified as a soil state factor.

[0044] In one specific implementation, Bayesian inference can also be used to probabilistically evaluate each candidate driving factor. Specifically, using the change in soil moisture content as the target variable, and precipitation, temperature, air humidity, light intensity, historical soil moisture content, and soil type as conditional variables, the conditional probability or posterior probability of each conditional variable with respect to the change in soil moisture content is calculated, and key driving factors are screened based on the conditional probability or posterior probability. Bayesian inference reduces the subjectivity caused by relying solely on literature extraction, allowing the determination of key driving factors to consider both domain knowledge and observational data characteristics.

[0045] In this embodiment, the key driving factors ultimately determined may include precipitation, air temperature, air humidity, light intensity, air pressure, historical stratified soil moisture content, soil type, and soil depth. Based on the category of these key driving factors, they are divided into meteorological driving factors, soil state factors, and soil attribute factors. Meteorological driving factors include precipitation, air temperature, air humidity, light intensity, and air pressure; soil state factors include historical stratified soil moisture content and its rate of change; and soil attribute factors include soil type and soil depth.

[0046] Therefore, the domain knowledge graph is not only used to display the conceptual relationships in the field of farmland soil moisture, but also for the subsequent constraint construction of the PINN model. Specifically, the categories of key driving factors are used to determine the multi-branch input structure of the PINN model; the correlations between key driving factors are used to determine the feature fusion methods between branches; the mechanistic relationships of soil moisture infiltration, evaporation, root water absorption, and soil water balance are used to construct the physical residual loss; the relationship between soil type and reasonable moisture content range is used to construct the value constraint loss; and the time lag relationship between precipitation, evaporation, shallow moisture content, and deep moisture content is used to construct the structural residual loss.

[0047] By using the above method, this application determines candidate driving factors based on the relationship between each node in the domain knowledge graph and the soil moisture content node, and selects key driving factors based on the degree of influence of the candidate driving factors on soil moisture content changes. This can reduce the interference of factors that are weakly correlated with or irrelevant to soil moisture content changes on model training, thereby improving the model's ability to represent the main influencing factors.

[0048] In steps S3 and S4, based on the key driving factors and their correlations determined in step S2, a knowledge-constrained physical information neural network (PINN) model is constructed, and the PINN model is trained using standardized observation data and knowledge constraints to obtain a soil moisture content prediction model.

[0049] The knowledge constraints refer to the constraint information generated based on key driving factors and their relationships, used to limit the training process and prediction results of the physical information neural network model. The knowledge constraints can include structural constraints and loss constraints. Structural constraints are used to set the sub-network structure of the physical information neural network model according to the categories of key driving factors, and to determine the connection methods or feature fusion methods between sub-networks based on the relationships between key driving factors. Loss constraints are used to construct knowledge constraint loss terms during model training based on soil moisture transport mechanisms, reasonable ranges for soil moisture content, and the relationships between key driving factors.

[0050] The physical information neural network model refers to a prediction model that incorporates physical law constraints or domain knowledge constraints during the training process of a neural network model. In this embodiment, the physical information neural network model does not only fit data based on observed samples, but also uses multi-source observation data and knowledge constraints for training. This allows the farmland soil moisture content prediction model to balance data fitting accuracy and soil moisture transport patterns when outputting soil moisture content prediction results at the target soil depth.

[0051] like Figure 4 As shown, the knowledge-constrained PINN model includes an input layer, multiple knowledge-constrained sub-network modules, a fusion output layer, and a loss function constraint layer. The input layer receives multi-source observation data preprocessed in step S1. The multiple knowledge-constrained sub-network modules extract feature representations of different types of driving factors. The fusion output layer fuses the features output by each sub-network module and outputs the soil moisture content prediction result. The loss function constraint layer introduces physical mechanism constraints, value range constraints, and structural correlation constraints during model training.

[0052] The input layer receives input data including crop growth history data, soil factor data, and meteorological environment data. Crop growth history data may include one or more of the following: SIF fluorescence value, historical NDVI value, crop phenological stage, crop growth level, and crop water stress level, used to characterize the response of crop growth status to changes in soil moisture. Soil factor data may include one or more of the following: historical soil moisture content at different soil depths, soil type, soil texture, soil depth, soil moisture change rate, and neutron spectrometer-retrieved moisture data, used to characterize the soil moisture status and soil moisture retention capacity of the target farmland area. Meteorological environment data may include one or more of the following: air temperature, air humidity, precipitation, air pressure, light intensity, total radiation, wind speed, and wind direction, used to characterize the driving effect of external meteorological conditions on soil moisture replenishment, evaporation, and evapotranspiration processes.

[0053] The knowledge-constrained PINN model sets up multiple sub-network modules based on the categories of key driving factors and their relationships identified in the domain knowledge graph. Each sub-network module corresponds to a different physical or physiological process, used to embed causal relationships and action paths from the knowledge graph into the model structure. In one specific embodiment, the knowledge-constrained PINN model includes a meteorological-soil interaction module, a soil-crop physiology module, and a crop response module.

[0054] The meteorological-soil interaction module is used to simulate the influence of meteorological environmental factors on soil moisture status. This module receives meteorological environmental data and soil factor data, and extracts soil moisture change characteristics driven by meteorological conditions based on constraints such as water balance, infiltration, evaporation, and infiltration mechanisms. Precipitation positively replenishes the shallow soil moisture content through infiltration; air temperature, light intensity, and total radiation deplete the surface soil moisture content by influencing evaporation or evapotranspiration; soil type and soil depth affect the rate of water infiltration and soil water-holding capacity. Based on these relationships, the meteorological-soil interaction module performs a nonlinear mapping between meteorological environmental data and soil factor data to obtain soil state characteristics driven by meteorological conditions.

[0055] In terms of specific network structure, the meteorological-soil interaction module may include several fully connected layers, normalization layers, and nonlinear activation layers. Meteorological environmental data and soil factor data are respectively fed into their corresponding feature extraction branches, and then fused through a feature fusion layer to obtain the first soil state feature.

[0056] The soil-crop physiology module is used to simulate the influence of soil moisture status on crop physiological status. This module receives soil factor data and crop growth history data, and extracts the characteristics of the impact of soil moisture changes on crop physiological status based on mechanisms such as root water uptake, crop water stress, and photosynthetic response. When the root zone soil moisture content decreases below a preset water stress threshold, the crop may experience water stress, which in turn affects photosynthetic efficiency, SIF fluorescence intensity, or NDVI trends. When the soil moisture content is within a suitable range, the crop physiological status is relatively stable. Based on these relationships, the soil-crop physiology module jointly models historical stratified soil moisture content, soil type, soil depth, and crop growth history data to obtain crop physiological response characteristics.

[0057] In terms of specific network structure, the soil-crop physiology module may include a soil state subnetwork and a crop response subnetwork. The soil state subnetwork is used to extract stratified soil moisture content and its rate of change characteristics, while the crop response subnetwork is used to extract response characteristics such as SIF fluorescence, NDVI, or crop phenological stage. The two are linked and fused into a second soil state feature.

[0058] The crop response module simulates the crop response process under the combined influence of meteorological environment and soil moisture status. This module receives historical crop growth data, meteorological environment data, and soil state characteristics output by the aforementioned modules, and extracts crop response features. Light intensity, temperature, and air humidity in the meteorological environment data affect crop photosynthesis and evapotranspiration; soil moisture status affects crop water supply and the degree of drought and flood stress; historical crop growth data reflects the crop's response to changes in water conditions. Based on these relationships, the crop response module obtains crop response features through nonlinear mapping. These crop response features can be used as auxiliary features in soil moisture content prediction, or as auxiliary outputs to verify the model's ability to represent crop water stress.

[0059] The fusion output layer is used to fuse the features output by the meteorological-soil interaction module, the soil-crop physiology module, and the crop response module to obtain a comprehensive feature representation, and output the stratified soil moisture content prediction results of the target farmland area based on the comprehensive feature representation.

[0060] In one specific implementation, the fusion output layer employs splicing fusion, weighted fusion, or attention fusion to fuse the first soil state features, the second soil state features, and the crop response features. The fusion weights can be set based on the association strength determined by the domain knowledge graph in step S2, or they can be adaptively learned during model training.

[0061] The output of the fusion output layer includes predicted soil moisture content values ​​at one or more soil depths. For example, the fusion output layer can output predicted soil moisture content sequences at different depths such as 10cm, 20cm, 40cm, and 60cm over a preset time period. For scenarios requiring auxiliary judgment by combining remote sensing or crop response information, the fusion output layer can also simultaneously output SIF or NDVI predicted values ​​as auxiliary prediction results, but the main prediction object in this embodiment is the stratified soil moisture content.

[0062] To ensure that the model's predictions conform to the patterns of soil moisture transport in farmland, a total loss function is constructed during the training process. This total loss function includes data fitting loss and knowledge constraint loss.

[0063] Data fitting loss This is used to measure the deviation between the predicted soil moisture content output by the model and the measured soil moisture content. Specifically, the mean squared error (MSE) can be used to calculate the data fitting loss.

[0064] Where N is the number of training samples. Let i be the measured soil moisture content corresponding to the i-th sample. This is the predicted soil moisture content value for the i-th sample output by the model.

[0065] Knowledge constraint loss includes at least one of physical residual loss, value residual loss, and structural residual loss. Physical residual loss is constructed based on soil water balance relationships, water infiltration relationships, evaporation relationships, root water absorption relationships, or Richards equations. The physical residual loss increases when the soil moisture content change trend output by the model does not conform to physical laws such as precipitation recharge, evaporation consumption, root water absorption, and water infiltration. Value residual loss is constructed based on the reasonable range of soil moisture content values ​​corresponding to the soil type of the target farmland area. When the predicted soil moisture content value output by the model is lower than the reasonable lower limit or higher than the reasonable upper limit, the value residual loss is calculated based on the degree of boundary violation. Structural residual loss is constructed based on causal relationships, hierarchical relationships, and related influence relationships in the domain knowledge graph. For example, if a causal relationship exists in the knowledge graph that "increased precipitation increases shallow soil moisture content through infiltration," then the structural residual is calculated when the model's prediction results show a change that is significantly opposite to this relationship within a preset lag time. If an interlayer transmission relationship exists in the knowledge graph that "changes in shallow soil moisture content precede changes in deep soil moisture content," then the structural residual is calculated when the model outputs changes in deep soil moisture content earlier than changes in shallow soil moisture content.

[0066] In one specific implementation, the physical residual loss, the value residual loss, and the structural residual loss can be uniformly represented as the knowledge constraint loss. The total loss function is expressed as:

[0067] During training, standardized crop growth history data, soil factor data, and meteorological environment data are input into the knowledge-constrained PINN model. The model first extracts soil state features driven by meteorology through the meteorological-soil interaction module, extracts the influence features of soil moisture on crop physiological state through the soil-crop physiology module, and extracts crop response features under the combined effects of meteorological environment and soil moisture through the crop response module. Then, the features of each module are fused through the fusion output layer to output stratified soil moisture content prediction values.

[0068] After obtaining the predicted values ​​of soil moisture content in each layer, the data fitting loss and knowledge constraint loss are calculated. Backpropagation is then performed based on the total loss function to update the network weights and bias parameters in each sub-network module and the fused output layer. The model optimization algorithm can employ Adam optimization, stochastic gradient descent, or other gradient optimization algorithms.

[0069] In the early stages of model training, the weight of knowledge constraint loss in the total loss function can be increased, allowing the model to prioritize learning the structural relationships between soil moisture transport mechanisms, reasonable ranges of soil moisture content, and key driving factors. In the later stages of model training, the weight of knowledge constraint loss can be appropriately reduced to improve the model's ability to fit measured observation data.

[0070] For example, the weights of the knowledge constraint loss can be dynamically decayed according to the training epochs: Where β(epoch) is the knowledge constraint loss weight at the epoch-th training round. is the initial weight, k is the decay coefficient, and epoch is the current training round.

[0071] When the total loss function stabilizes, the validation set error no longer decreases significantly in several consecutive training rounds, or the preset maximum number of training rounds is reached, training is stopped, and the current model parameters are saved to obtain the trained knowledge-constrained PINN soil moisture content prediction model.

[0072] Through the above steps, the knowledge-constrained PINN model can improve the accuracy, stability, and physical rationality of the stratified soil moisture prediction results by being constrained by the mechanism of farmland soil moisture transport, the reasonable range of soil moisture content, and the correlation of key driving factors while using multi-source observation data for nonlinear fitting.

[0073] Furthermore, the performance of the trained soil moisture content prediction model is validated using a validation set or a test set. Validation metrics may include the coefficient of determination (R²) and the root mean square error (RMSE).

[0074] R² is used to characterize the fit between the model's predicted values ​​and the measured values. The closer R² is to 1, the better the model's prediction performance. RMSE is used to characterize the average deviation between the predicted values ​​and the measured values. The smaller the RMSE, the smaller the prediction error.

[0075] In one specific implementation, when the model validation results satisfy R²≥0.80 and RMSE≤5%, the model is determined to meet the requirements for predicting farmland soil moisture content; when the model validation results do not meet the above conditions, the number of network layers, number of nodes, learning rate, loss weights or key driving factor screening thresholds of the PINN model are adjusted, and the model is retrained until the validation results meet the requirements.

[0076] After successful model validation, the stratified soil moisture content data and meteorological environment data collected in real time in the target farmland area are processed into standardized real-time observation data according to the preprocessing method in step S1. Then, the standardized real-time observation data are input into the trained PINN model according to the categories of meteorological driving factors, soil state factors, and soil attribute factors. Based on the learned soil moisture transport mechanism, the correlation of key driving factors, and historical time series variation patterns, the PINN model outputs the predicted soil moisture content at different soil depths in the target farmland area.

[0077] The soil moisture content prediction results can include stratified soil moisture content time series data for one or more prediction periods in the future. For example, the model can output soil moisture content prediction curves at different soil depths such as 10cm, 20cm, 40cm, and 60cm in the next 50 hours, or it can output long-term soil moisture content prediction curves at different soil depths in the next 120 hours.

[0078] like Figure 5 As shown, in the 50-hour prediction scenario, the predicted soil moisture content curves for each soil layer depth output by the model have a high degree of consistency with the measured curves, indicating that the knowledge-constrained PINN model can fit the soil moisture content change trend in a short period of time well.

[0079] like Figure 6 As shown, in the 120-hour prediction scenario, the soil moisture content prediction curves at each soil layer depth output by the model can still maintain the ability to track the changing trend of the measured curves, indicating that the structural constraints and loss constraints constructed by the domain knowledge graph can improve the stability and physical rationality of the PINN model in long-term prediction.

[0080] Furthermore, the soil moisture content prediction results can be used for farmland drought risk assessment, precision irrigation decision-making, dynamic monitoring of farmland moisture, regional drought and flood early warning, and farmland water resource allocation. For example, when it is predicted that the soil moisture content at a certain soil depth is lower than the threshold required for crop growth within a preset time period, an irrigation prompt can be output; when it is predicted that the soil moisture content in multiple layers continues to decrease and reaches the drought risk threshold, a drought risk warning can be output.

[0081] In a specific application scenario, the target farmland area is a demonstration farmland area in different regions of Henan Province. Multiple ZY1700 type automatic regional soil moisture monitoring instruments are deployed in the fields of these demonstration farmland areas. This instrument is based on the principle that the intensity of fast cosmic ray neutrons near the Earth's surface is inversely proportional to the surface soil moisture content. It measures the intensity of fast cosmic ray neutrons using a neutron probe erected above the ground surface and then inverts the soil moisture status over a relatively large area.

[0082] The ZY1700 regional soil moisture automatic monitoring instrument can reflect the soil moisture status of the entire layer from the surface to a depth of tens of centimeters in a surrounding area of ​​about 600 acres. It can also automatically monitor various factors such as regional soil moisture, wind speed, wind direction, air temperature, relative humidity, air pressure, rainfall, and total radiation around the clock.

[0083] In this embodiment, data collected by automatic soil moisture monitoring instruments in Wangzhuang Village, Junxian County, and Anyang Academy of Agricultural Sciences from December 2024 to August 2025 were selected for model construction and validation. First, the data was processed, time-series aligned, outliers removed, missing values ​​filled, normalized, and features filtered to obtain standardized observation data. Then, a domain knowledge graph was constructed based on the mechanism of farmland soil moisture change, and key driving factors such as precipitation, air temperature, air humidity, light intensity, historical stratified soil moisture content, soil type, and soil depth were determined through associative reasoning. Next, a multi-branch knowledge-constrained PINN model was constructed based on the key driving factors, and the model was jointly trained using data fitting loss, physical residual loss, value residual loss, and structural residual loss. Finally, the trained model was used to predict soil moisture content at different depths for 50 hours and 120 hours.

[0084] Therefore, by combining the domain knowledge graph with the PINN model, this embodiment enables the model to not only learn the nonlinear variation patterns in multi-source observation data, but also to be constrained by the soil moisture transport mechanism and the reasonable range of soil moisture content, thereby improving the accuracy, stability and physical interpretability of farmland soil moisture content prediction.

[0085] This application also provides a method for predicting farmland soil moisture content, including acquiring real-time observation data of a target farmland area; The real-time observation data is input into the farmland soil moisture content prediction model to obtain the soil moisture content prediction results of the target farmland area at the target soil depth. The farmland soil moisture content prediction model is obtained through the farmland soil moisture content prediction model construction method described above.

[0086] like Figure 7 As shown, this application also provides a farmland soil moisture content prediction system. The farmland soil moisture content prediction system is used to execute the farmland soil moisture content prediction method described in the foregoing embodiments. The data acquisition, data preprocessing, domain knowledge graph construction, knowledge constraint construction, PINN model training, and prediction output processes, which have been detailed in the foregoing method embodiments, will not be repeated in this system embodiment. The following mainly describes the system's module composition and the cooperation relationships between the modules.

[0087] The farmland soil moisture content prediction system includes a data acquisition module, a data preprocessing module, a knowledge graph construction module, a knowledge constraint construction module, a model training module, and a prediction output module.

[0088] The data acquisition module is used to acquire multi-source observation data of the target farmland area, which includes at least stratified soil moisture content data and meteorological environmental data. The data acquisition module may include a multi-source observation data acquisition unit and a real-time observation data access unit. The multi-source observation data acquisition unit is used to collect or receive historical multi-source observation data of the target farmland area for model training and validation; the real-time observation data access unit is used to access real-time observation data collected from the target farmland area for the trained soil moisture content prediction model to perform real-time or future time-period predictions.

[0089] The data preprocessing module is connected to the data acquisition module and is used to preprocess the multi-source observation data to obtain standardized observation data. The data preprocessing module may include a time-series alignment unit and a cleaning and normalization unit. The time-series alignment unit is used to align data collected from different sensors, observation stations, or data sources according to a unified timestamp. The cleaning and normalization unit is used to perform at least one of the following processes on the multi-source observation data: outlier removal, missing value imputation, data cleaning, feature filtering, encoding, and normalization, so that the processed data meets the input format requirements for subsequent model training and prediction.

[0090] The knowledge graph construction module is used to construct a domain knowledge graph based on knowledge of farmland soil moisture, and to determine the key driving factors and their relationships affecting changes in soil moisture content based on the domain knowledge graph. The knowledge graph construction module may include a domain knowledge parsing unit and a key driving factor and relationship determination unit. Specifically, the domain knowledge parsing unit analyzes domain data related to farmland irrigation, soil moisture transport, crop water stress, and meteorological environmental impacts, extracting environmental factor nodes, soil hydrological nodes, and crop physiological and remote sensing nodes related to changes in soil moisture, and constructing hierarchical relationships, causal driving relationships, or related influence relationships between nodes; the key driving factor and relationship determination unit determines the key driving factors and their relationships based on the association paths, edge types, directions of action, and association strengths between each node and the soil moisture content node in the domain knowledge graph.

[0091] The knowledge constraint construction module is connected to the knowledge graph construction module and is used to construct knowledge constraints for constraining the PINN model based on the key driving factors and their relationships. The knowledge constraints include structural constraints and loss constraints. The knowledge constraint construction module may include a structural constraint unit and a loss constraint unit. Specifically, the structural constraint unit is used to set the sub-network branches of the PINN model according to the categories of the key driving factors, and to determine the connection method or feature fusion method between each sub-network branch based on the relationships between the key driving factors; the loss constraint unit is used to construct the knowledge constraint loss term of the PINN model based on the soil moisture transport mechanism, the reasonable range of soil moisture content, and the relationships between the key driving factors.

[0092] In one specific implementation, the structural constraint unit classifies key driving factors into at least two categories: meteorological driving factors, soil state factors, and soil attribute factors, and inputs the key driving factors of different categories into the corresponding sub-network branches in the PINN model. The knowledge constraint loss term constructed by the loss constraint unit may include at least one of physical residual loss, value residual loss, and structural residual loss; wherein, physical residual loss is used to constrain the prediction results to conform to soil water transport mechanisms such as soil water balance, water infiltration, evaporation, root water uptake, or Richards equation; value residual loss is used to constrain the prediction results to be within the reasonable moisture content range corresponding to the soil type of the target farmland area; and structural residual loss is used to constrain the prediction results to conform to the direction of interaction, time lag of interaction, or correlation strength between key driving factors in the knowledge graph.

[0093] The model training module is connected to both the data preprocessing module and the knowledge constraint construction module. It is used to train the PINN model based on the standardized observation data and the knowledge constraints to obtain a soil moisture content prediction model. The model training module may include a forward inference unit and a loss calculation and parameter update unit. The forward inference unit inputs the standardized observation data into the PINN model and outputs stratified soil moisture content prediction values ​​through sub-network branches and a fusion output layer in the PINN model. The loss calculation and parameter update unit calculates the data fitting loss based on the deviation between the stratified soil moisture content prediction values ​​and the measured soil moisture content, calculates the knowledge constraint loss based on the knowledge constraints, constructs a total loss function based on the data fitting loss and the knowledge constraint loss, and updates the model parameters of the PINN model through backpropagation.

[0094] In one specific implementation, the model training module is further used to dynamically adjust the weight of the knowledge constraint loss during the PINN model training process. In the early stages of training, the weight of the knowledge constraint loss in the total loss function is increased, enabling the PINN model to prioritize learning soil moisture transport mechanisms, reasonable ranges for soil moisture content, and the structural relationships between key driving factors. In the later stages of training, the weight of the knowledge constraint loss in the total loss function is decreased, allowing the PINN model to further improve its ability to fit measured observation data.

[0095] The prediction output module is connected to the model training module and is used to input real-time observation data of the target farmland area into the trained soil moisture content prediction model, and output the predicted soil moisture content of the target farmland area at one or more soil depths. The prediction output module may include a stratified soil moisture content prediction unit and a result verification and application unit. The stratified soil moisture content prediction unit is used to output predicted soil moisture content values ​​or prediction curves at different soil depths within a preset future time period; the result verification and application unit is used to verify the prediction results based on evaluation indicators such as the coefficient of determination R² and the root mean square error RMSE, and to apply the prediction results to farmland drought risk assessment, precision irrigation decision-making, farmland water dynamic monitoring, regional drought and flood early warning, or farmland water resource allocation.

[0096] In actual operation, the data acquisition module first acquires historical multi-source observation data and / or real-time observation data of the target farmland area; the data preprocessing module performs time-series alignment, cleaning, and normalization on the multi-source observation data to obtain standardized observation data; the knowledge graph construction module determines key driving factors and their correlations based on knowledge of farmland soil moisture; the knowledge constraint construction module transforms the key driving factors and their correlations into structural constraints and loss constraints of the PINN model; the model training module trains the PINN model based on standardized observation data and knowledge constraints to obtain a soil moisture content prediction model; and the prediction output module uses the trained soil moisture content prediction model to output soil moisture content prediction results at different soil depths.

[0097] Therefore, the farmland soil moisture content prediction system provided in this embodiment can correspond to the aforementioned farmland soil moisture content prediction method. It integrates multi-source observation data processing, domain knowledge graph construction, knowledge constraint generation, PINN model training, and prediction result output into a complete system, so that the soil moisture content prediction results are simultaneously constrained by observation data and soil moisture transport mechanisms, thereby improving the accuracy, stability, and physical rationality of the prediction results.

[0098] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0099] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method of constructing a farmland soil water content prediction model, characterized by, include: Acquire multi-source observation data of the target farmland area; A domain knowledge graph is constructed based on knowledge of farmland soil moisture, and the key driving factors affecting changes in soil moisture content and the correlation between the key driving factors are determined based on the domain knowledge graph. Based on the key driving factors and their correlations, a knowledge constraint is constructed to constrain the physical information neural network model. Based on the multi-source observation data and the knowledge constraints, the physical information neural network model is trained to obtain a farmland soil moisture content prediction model; wherein, the farmland soil moisture content prediction model is used to output the soil moisture content prediction result of the target farmland area at the target soil layer depth when real-time observation data of the target farmland area is input.

2. The method according to claim 1, wherein The method further includes: Preprocessing of the multi-source observation data includes at least one of the following: time-series alignment, missing value imputation, outlier removal, data cleaning, feature filtering, encoding, normalization, and dataset partitioning. The multi-source observation data includes soil factor data, meteorological environment data, and crop growth data.

3. The method for constructing a farmland soil moisture content prediction model according to claim 1, characterized in that, The construction of a domain knowledge graph based on farmland soil moisture domain knowledge includes: Obtain domain data on farmland soil moisture, and extract environmental factor knowledge points, soil hydrology knowledge points, and crop physiology knowledge points from the domain data; The environmental knowledge points, soil and hydrology knowledge points, and crop physiology knowledge points are used as knowledge graph nodes, and the association edges between the knowledge graph nodes are constructed to obtain the domain knowledge graph.

4. The method for constructing a farmland soil moisture content prediction model according to claim 1, characterized in that, The process of determining the key driving factors affecting changes in soil moisture content and their correlations based on the domain knowledge graph includes: Candidate driving factors are determined based on the relationship between each node in the domain knowledge graph and the soil moisture content node; Based on the candidate driving factors, the degree of influence of each candidate driving factor on the change of soil moisture content is obtained; The key driving factors are selected based on the degree of influence, and the correlation between the key driving factors is determined.

5. The method for constructing a farmland soil moisture content prediction model according to claim 1, characterized in that, The physical information neural network model includes a meteorological-driven soil state subnetwork, a soil crop physiology subnetwork, and a crop response feature subnetwork. The knowledge constraints include structural constraints and loss constraints; The structural constraints include classifying the key driving factors into meteorological driving factors, soil state factors, and soil property factors according to their categories. Input the key driving factors of different categories into the corresponding sub-network branches; The connection methods between each sub-network branch are determined based on the correlation of the key driving factors, so that the network topology of the physical information neural network model matches the soil moisture transport mechanism.

6. The method for constructing a farmland soil moisture content prediction model according to claim 5, characterized in that, The loss constraints include at least one of physical residual constraints, value residual constraints, and structural residual constraints; The physical residual constraints are constructed based on at least one of the following: soil water balance relationship, water infiltration relationship, evaporation relationship, root water absorption relationship and Richards equation; The residual constraint is constructed based on the reasonable range of soil moisture content corresponding to the soil type of the target farmland area. The structural residual constraints are constructed based on the correlation between the key driving factors.

7. The method for constructing a farmland soil moisture content prediction model according to claim 1, characterized in that, The training of the physical information neural network model based on the multi-source observation data and the knowledge constraints includes: The multi-source observation data are input into the corresponding sub-network branches of the physical information neural network model according to the categories of key driving factors to obtain the feature representations of each category of factors; The feature representations of each category of factors are fused through the fusion output layer of the physical information neural network model to obtain a comprehensive feature representation; Based on the comprehensive feature representation, output the predicted value of stratified soil moisture content; The data fitting loss is calculated based on the deviation between the predicted and measured soil moisture contents of the stratified soil, and the knowledge constraint loss is calculated based on the knowledge constraints. A total loss function is constructed based on the data fitting loss and the knowledge constraint loss, and the model parameters of the physical information neural network model are updated through backpropagation. During the training of the physical information neural network model, the weight of the knowledge constraint loss in the total loss function is dynamically adjusted.

8. A method for predicting soil moisture content in farmland, characterized in that, include: Acquire real-time observation data of the target farmland area; The real-time observation data is input into the farmland soil moisture content prediction model to obtain the soil moisture content prediction results of the target farmland area at the target soil depth. The farmland soil moisture content prediction model is obtained by the farmland soil moisture content prediction model construction method according to any one of claims 1 to 7.

9. A system for predicting farmland soil moisture content, characterized in that, include: The data acquisition module is used to acquire multi-source observation data of the target farmland area, including soil factor data, meteorological environment data and crop growth data; The data preprocessing module is used to preprocess the multi-source observation data; The knowledge graph construction module is used to construct a domain knowledge graph based on knowledge of farmland soil moisture, and to determine the key driving factors affecting changes in soil moisture content and the correlation between the key driving factors based on the domain knowledge graph. A knowledge constraint construction module is used to construct knowledge constraints for constraining the physical information neural network model based on the key driving factors and the correlation between the key driving factors. The knowledge constraints include structural constraints and loss constraints. The model training module is used to train the physical information neural network model based on the standardized observation data and the knowledge constraints to obtain a farmland soil moisture content prediction model. The prediction output module is used to input real-time observation data of the target farmland area into the farmland soil moisture content prediction model and output the soil moisture content prediction results of the target farmland area at the target soil depth.

10. The farmland soil moisture content prediction system according to claim 9, characterized in that, The physical information neural network model includes a meteorological-driven soil state subnetwork, a soil crop physiology subnetwork, and a crop response feature subnetwork. The knowledge constraint construction module includes: A structural constraint unit is used to set up sub-networks of the physical information neural network model according to the category of the key driving factors, and to determine the connection mode between each sub-network according to the correlation between the key driving factors. The loss constraint unit is used to construct the knowledge constraint loss term of the physical information neural network model based on the soil moisture transport mechanism, the reasonable range of soil moisture content, and the correlation of key driving factors.

Images