Crop planting intelligent decision reasoning system and method based on multi-source data fusion

Abstract

The application provides a crop planting intelligent decision reasoning system and method based on multi-source data fusion, and relates to the technical field of agricultural planting. The system comprises: a multi-source data fusion module for extracting key features of target crop growth from multi-source data; an intelligent sensing module for constructing an agricultural knowledge graph according to multi-source heterogeneous data of crop planting; and a reasoning decision module for making decision reasoning on multi-source data and key features according to the agricultural knowledge graph, a preset reasoning model and a decision model. Through the application, the defects of the traditional crop planting decision system, such as insufficient decision-making ability, low accuracy in the planting management process, and lack of cross-domain multi-source data collaborative processing mechanism, are solved, and the generalization ability of the system is insufficient.

Description

Technical Field

[0001] The present invention relates to the field of agricultural planting technology, and in particular to a crop planting intelligent decision-making reasoning system and method based on multi-source data fusion. Background Art

[0002] With the acceleration of global agricultural modernization, crop cultivation has gradually evolved towards intelligent and precise methods. Traditional agricultural production methods rely on farmers' experience and historical planting data, but are less adaptable to complex and changing climates and soil conditions. This often leads to delayed decision-making, wasted resources, and reduced crop yields. How to effectively integrate multi-source data and improve the accuracy of planting decisions has become a key challenge in agricultural production that urgently needs to be addressed.

[0003] Traditional crop planting decision-making systems often rely on single data sources or simple historical data, lacking effective integration and analysis of real-time, multi-dimensional data, making them incapable of adapting to the complex and ever-changing agricultural environment. This is particularly true when multiple variables interact (e.g., climate change, soil conditions, pest and disease development), where traditional systems' decision-making capabilities are significantly insufficient. This leads to low precision in planting management and difficulties in ensuring yield and quality. Furthermore, existing decision-making systems are mostly rule-based or expert systems, failing to fully leverage modern machine learning and intelligent reasoning technologies. This limits their ability to support decisions in complex agricultural scenarios. Furthermore, most decision-making systems are optimized for a single domain or specific crop, lacking cross-domain mechanisms for collaborative processing of multi-source data, resulting in insufficient generalization capabilities. Some decision-making systems can only effectively analyze climate or crop data for specific regions and are unable to adapt to the production needs of diverse regions and crops. Summary of the Invention

[0004] The present invention provides a crop planting intelligent decision-making and reasoning system and method based on multi-source data fusion, which is used to solve the problems of insufficient decision-making ability of traditional crop planting decision-making systems, resulting in low accuracy in the planting management process. In addition, most decision-making systems are optimized on a single field or specific crops, lacking a cross-domain multi-source data collaborative processing mechanism, resulting in insufficient generalization ability of the system.

[0005] The present invention provides a crop planting intelligent decision-making and reasoning system based on multi-source data fusion, comprising:

[0006] a multi-source data fusion module, configured to obtain multi-source data of a target crop and extract features from the multi-source data using a spatiotemporal convolutional network to obtain key growth characteristics of the target crop, wherein the multi-source data includes first meteorological data and first soil data;

[0007] An intelligent perception module, configured to construct an agricultural knowledge graph based on multi-source heterogeneous data of crop planting, wherein the multi-source heterogeneous data includes second meteorological data, second soil data, crop data, and expert experience knowledge text;

[0008] The inference and decision module is used to match the multi-source data with the agricultural knowledge graph according to the rules to obtain the recommended planting area of ​​the target crop, call the preset inference model to predict the yield of the key features to obtain the planting yield of the target crop in the recommended planting area, and call the preset decision model to make a decision on the multi-source data to obtain the planting planning action of the target crop.

[0009] In some embodiments, the calling of a spatiotemporal convolutional network to extract features from the multi-source data to obtain key features of target crop growth includes:

[0010] Calling the spatiotemporal convolutional network to perform temporal convolution processing and spatial convolution processing on the multi-source data respectively to obtain corresponding temporal convolution features and spatial convolution features;

[0011] Performing channel splicing processing on the temporal convolution features and the spatial convolution features to obtain the spatiotemporal correlation features of the multi-source data;

[0012] According to the feature weights assigned by the attention mechanism in the spatiotemporal convolutional network, a candidate feature with the largest feature weight is selected from the spatiotemporal correlation features;

[0013] An adversarial feature alignment network is called to perform feature alignment on the candidate features to obtain key features for target crop growth.

[0014] In some embodiments, constructing an agricultural knowledge graph based on multi-source heterogeneous data on crop planting includes:

[0015] Extracting entities and relationships from the expert experience knowledge text and the crop data to obtain corresponding entity nodes and relationships between entity nodes;

[0016] Extracting the temporal variation characteristics of the second meteorological data and the spatial characteristic graph of the second soil data, and embedding the temporal variation characteristics and the spatial characteristic graph into the entity node as node attributes;

[0017] Construct triples based on the entity nodes and the relationships to obtain a primary knowledge graph;

[0018] The expert experience knowledge text is converted into agricultural planting rules, and the agricultural planting rules are registered in the reasoning engine of the primary knowledge graph to obtain an agricultural knowledge graph.

[0019] In some embodiments, the rule matching of the multi-source data with the agricultural knowledge graph to obtain a recommended planting area for a target crop includes:

[0020] matching the agricultural planting rules in the agricultural knowledge graph with the first meteorological data and the first soil data to obtain an adaptability score of the candidate region under the agricultural planting rules, wherein the agricultural planting rules include temperature rules, soil moisture rules, and light rules for the target crop;

[0021] Weighting the adaptability scores under the temperature rule, soil moisture rule, and light rule according to the preset temperature weight, humidity weight, and light weight, respectively, and summing the weighted results to obtain the total adaptability score of the candidate area under the agricultural planting rule;

[0022] The candidate areas with the highest overall adaptation scores were identified as recommended planting areas for the target crops.

[0023] In some embodiments, calling a preset inference model to perform yield prediction on the key features to obtain the planting yield of the target crop in the recommended planting area includes:

[0024] Calling a preset LSTM model to perform preliminary prediction on the key features to obtain preliminary yield characteristic values ​​of the target crop, wherein the LSTM model is trained based on historical meteorological data and historical soil data of the target crop in the recommended planting area;

[0025] From the multi-source heterogeneous data, a temporal variation feature of the second meteorological data and a spatial feature map of the second soil data are obtained, and target growth cycle features of the crop data corresponding to the target crop are extracted;

[0026] Splicing the temporal variation feature, the spatial feature map, and the target growth cycle feature to obtain a spliced ​​feature;

[0027] The splicing features and the preliminary yield feature values ​​are input into a preset Transformer model to perform crop yield prediction, and the planting yield of the target crop in the recommended planting area is obtained.

[0028] In some embodiments, the preset decision model includes a DQN model, and the pre-training process of the DQN model includes:

[0029] Acquiring planting status data, including climate conditions, soil conditions, planting area, historical planting experience, and crop growth status;

[0030] At each time step, the planting state data is input into the DQN model for decision-making to obtain corresponding planting actions, wherein the planting actions include selecting a planting area, adjusting the planting time, and changing the fertilization and irrigation plan;

[0031] Under the planting state data, the crop yield, resource utilization efficiency, and environmental impact value are calculated according to the planting action, and the reward Q value of the DQN model in the time step is estimated based on the crop yield, the resource utilization efficiency, and the environmental impact value;

[0032] Calculating a loss function of the DQN model based on the reward Q value and the target Q value, wherein the target Q value is calculated based on the maximum Q value obtained in the next time step after the current time step and the reward Q value;

[0033] Backpropagation is performed in the DQN model through the loss function to update the parameters of the DQN model.

[0034] In some embodiments, the inference and decision-making module is also used to regularly adaptively update the DQN model and store the historical experience after each decision of the DQN model into a historical experience library. The historical experience library is used to provide the DQN model with decision-making actions based on historical experience when the DQN model is adaptively updated.

[0035] In some embodiments, the system further comprises:

[0036] a feature engineering module, configured to call a preset time series model to perform time series encoding on the second meteorological data, the second soil data, and the crop data included in the multi-source heterogeneous data, to obtain time series sequence features of multiple crop plantings;

[0037] Assigning corresponding feature weights to each time series feature through the attention mechanism in the time series model;

[0038] Weighting each time series feature according to the feature weight, and fusing the weighted results to obtain a feature subset of the multi-source heterogeneous data;

[0039] The node attributes of the entity nodes in the agricultural knowledge graph are updated using the feature subset.

[0040] In some embodiments, the system further comprises:

[0041] The visualization and human-computer interaction module is used to visualize the multi-source data of the target crop, the planting yield of the target crop in the recommended planting area, and the planting planning actions of the target crop, and perform decision reasoning based on the multi-source data of the crop to be decided that is input in real time to obtain the recommended planting area of ​​the crop to be decided, the planting yield in the recommended planting area, and the planting planning actions.

[0042] The present invention also provides a crop planting intelligent decision-making reasoning method based on multi-source data fusion, comprising:

[0043] Acquire multi-source data of a target crop, and call a spatiotemporal convolutional network to perform feature extraction on the multi-source data to obtain key growth characteristics of the target crop, wherein the multi-source data includes first meteorological data and first soil data;

[0044] Constructing an agricultural knowledge graph based on multi-source heterogeneous data of crop planting, wherein the multi-source heterogeneous data includes second meteorological data, second soil data, crop data, and expert experience knowledge text;

[0045] The multi-source data is matched with the agricultural knowledge graph by rules to obtain the recommended planting area of ​​the target crop, the preset inference model is called to predict the yield of the key features to obtain the planting yield of the target crop in the recommended planting area, and the preset decision model is called to make a decision on the multi-source data to obtain the planting planning action of the target crop.

[0046] The present invention also provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the computer program, it implements any of the above-described intelligent decision-making and reasoning methods for crop planting based on multi-source data fusion.

[0047] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-described intelligent decision-making and reasoning methods for crop planting based on multi-source data fusion.

[0048] The present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements any of the above-mentioned intelligent decision-making and reasoning methods for crop planting based on multi-source data fusion.

[0049] The intelligent crop planting decision-making and reasoning system and method based on multi-source data fusion, provided by this invention, utilizes multi-source data on target crops to make planting decisions. This overcomes the shortcomings of traditional planting decision-making systems, which are mostly optimized for a single field or specific crop, lack cross-domain multi-source data collaborative processing mechanisms, and have insufficient generalization capabilities. Furthermore, by combining agricultural knowledge graphs, pre-set reasoning models, and decision-making models to execute planting decisions for target crops, the system's planting decision-making capabilities are enhanced, improving the accuracy of the planting management process. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced one by one below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0051] Figure 1 It is a schematic diagram of the framework of the crop planting intelligent decision-making and reasoning system based on multi-source data fusion provided by the present invention.

[0052] Figure 2 It is a schematic diagram of the principle of the crop planting intelligent decision-making and reasoning system based on multi-source data fusion provided by the present invention.

[0053] Figure 3 It is a schematic diagram of the principle of the pre-training process of the DQN model provided by the present invention.

[0054] Figure 4 This is an application framework diagram of the crop planting intelligent decision-making reasoning method based on multi-source data fusion provided by the present invention.

[0055] Figure 5 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0056] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0057] The following describes the crop planting intelligent decision-making and reasoning system and method based on multi-source data fusion of the present invention with reference to the accompanying drawings. Figure 1 This is a schematic diagram of the framework of the crop planting intelligent decision-making and reasoning system based on multi-source data fusion provided by the present invention. Figure 1As shown, the crop planting intelligent decision reasoning system based on multi-source data fusion (hereinafter referred to as the system) comprises a multi-source data fusion module, an intelligent perception module, a reasoning decision module, a feature engineering module, a visualization and human-computer interaction module, which will be described one by one.

[0058] The multi-source data fusion module is used to obtain multi-source data of the target crop, and call a spatio-temporal convolution network to extract features from the multi-source data, so as to obtain key features of the growth of the target crop.

[0059] The multi-source data fusion module is mainly used for preprocessing the multi-source data of the target crop, extracting key features of the growth, and making reasoning and decision in the subsequent module. Here, the target crop is the crop specified by the user (such as a farmer or a planting personnel) for planting decision, which can be one crop or multiple crops, such as tomato, cucumber, rice, etc. The multi-source data includes first meteorological data and first soil data, which are generally obtained by meteorological sensors and soil measuring instruments, and the key features are generally soil moisture, precipitation, temperature, crop growth stage, light duration and photosynthetic efficiency, etc. These features can reflect the specific influence of meteorological, soil and other environmental factors on crop growth, and provide key basis for subsequent intelligent reasoning and decision.

[0060] The intelligent perception module is used to construct an agricultural knowledge graph according to the multi-source heterogeneous data of crop planting. The multi-source heterogeneous data includes second meteorological data, second soil data, crop data and expert experience knowledge text. The agricultural knowledge graph as a large agricultural decision knowledge base can make planting decisions for various crops.

[0061] As shown in Figure 2 The data source of the agricultural knowledge graph is the multi-source heterogeneous data of crop planting, wherein the second meteorological data includes high-frequency meteorological time series such as collected temperature, precipitation and light, the second soil data is the static and dynamic indexes of soil moisture, nutrients and pH of the land, and the crop data is the growth cycle characteristic data of different crop varieties, wherein the crop data should cover as many crop varieties as possible to enhance the generality of the agricultural knowledge graph. The expert experience knowledge text is the crop planting rules summarized by experts from the historical planting experience of various crops, which can be converted into agricultural planting rules, such as what planting action to perform under what meteorological, soil conditions, planting area and crop growth state for tomato, including planting area, adjusting planting time and changing fertilization and irrigation plan, etc., and what planting yield and resource utilization efficiency can be obtained. Such agricultural planting rules can enable the agricultural knowledge graph to have reasoning and decision-making ability for the planting management of the target crop.

[0062] When constructing an agricultural knowledge graph, entity nodes and their relationships are extracted from expert knowledge text and crop data to form triples of basic knowledge units, forming the framework of the agricultural knowledge graph. The features of secondary meteorological and soil data are then embedded into the node attributes of the entity nodes. Finally, a graph neural network is used to perform inference learning on the agricultural knowledge graph. Graph neural networks support joint training using entity attributes, graph structural relationships, and embedded spatiotemporal features, forming graph representation and prediction capabilities for agricultural tasks.

[0063] On this basis, agricultural planting rules formed from expert knowledge texts are incorporated into the joint optimization of a graph neural network in the form of triple constraints. This allows the data-driven reasoning process to be synergistically enhanced with the knowledge and experience of domain experts, improving the scientificity and reliability of the reasoning results. The resulting agricultural knowledge graph uniformly expresses the multidimensional data relationships related to crop cultivation and provides semantic support for subsequent reasoning tasks.

[0064] The inference and decision-making module is used to match multi-source data with agricultural knowledge graphs to obtain recommended planting areas for target crops, call preset inference models to predict yields of key features, obtain the planting yields of target crops in recommended planting areas, and call preset decision models to make decisions on multi-source data to obtain planting planning actions for target crops.

[0065] like Figure 2 As shown in the figure, the inference decision module uses the constructed agricultural knowledge graph, the preset inference model, and the decision model to simultaneously make planting decisions for the target crop. The planting decision process is divided into three stages. The first stage is rule-based reasoning, which uses the multi-source data of the target crop to match the agricultural planting rules in the agricultural knowledge graph to preliminarily determine the recommended planting area for the target crop. The second stage is the prediction of the inference model. Based on the recommended planting area, the yield of key features is predicted through the preset inference model to obtain the planting yield. The third stage is the planting decision planning of the decision model. According to the preset decision model, decisions are made directly on the multi-source data to obtain planting planning actions, which include selecting the planting area, adjusting the planting time, and changing the fertilization and irrigation plan.

[0066] Therefore, through the crop planting intelligent decision-making and reasoning system based on multi-source data fusion, the recommended planting areas, planting yields and planting planning actions can be output, which can provide users (such as farmers or planting personnel) with an optimal planting strategy for their target crops, allowing users to have a clear understanding of the planting areas, planting actions and final planting yields during the planting management process, and plan the planting process of the target crops accordingly.

[0067] The embodiment of the present application uses multi-source data of a target crop to make planting decisions, overcomes the defects of traditional planting decision systems that are mostly optimized in a single field or for a specific crop, lack a cross-field multi-source data collaborative processing mechanism, and have insufficient generalization ability. Moreover, the planting decisions of the target crop are executed in combination with an agricultural knowledge graph, a preset reasoning model, and a decision model, thereby enhancing the planting decision capability of the system and improving the precision in the planting management process.

[0068] In some embodiments, in the multi-source data fusion module, a spatio-temporal convolutional network is called to perform feature extraction on the multi-source data to obtain key features of the growth of the target crop. This can be achieved in the following manner, which is described in detail below.

[0069] First, a spatio-temporal convolutional network (STCN) is called to perform time convolution processing and space convolution processing on the multi-source data respectively to obtain corresponding time convolution features and space convolution features. The STCN model can simultaneously process the first meteorological data and the first soil data. Through time convolution processing, the time variation trend of the first meteorological data can be captured, and through space convolution processing, the differences between the soil and climate in different regions can be captured.

[0070] Then, the time convolution features and the space convolution features are subjected to channel concatenation processing to obtain spatio-temporal correlation features of the multi-source data. Here, the time convolution features and the space convolution features output by the time convolution processing and the space convolution processing are integrated through a fusion layer to ensure that the dependency relationship between the data sources is reflected. In specific implementation, a feature concatenation (Concatenation) method is adopted. The time convolution features and the space convolution features are concatenated in the channel dimension to form a new feature vector, and then a fully connected layer or a convolution layer is used to further learn the nonlinear relationship between the features, thereby extracting spatio-temporal correlation features in time and space, and capturing the influence of the linkage between soil moisture and precipitation on crop growth.

[0071] Next, according to the feature weight allocated by the attention mechanism in the spatio-temporal convolution network, the candidate features with the largest feature weight are screened out from the spatio-temporal correlation features. Here, the feature importance evaluation can be performed through the attention mechanism in the spatio-temporal convolution network when extracting the spatio-temporal correlation features. Because there can be multiple spatio-temporal correlation features, that is, there can be multiple cases of the influence of soil and weather on crop growth. Therefore, the attention mechanism in the spatio-temporal convolution network is directly used to assign a corresponding feature weight to each spatio-temporal correlation feature, so as to screen out the candidate features with the largest feature weight, that is, the features with the greatest influence on crop growth. And other features are removed step by step using the recursive feature elimination algorithm, and the features that most significantly improve the performance of the model are retained, and the selection of the features is based on the direct influence of the features on the predicted crop yield, health status and the like.

[0072] Finally, the candidate features are subjected to feature alignment by calling the adversarial feature alignment network, and the key features of the target crop growth are obtained. Because the key features come from multiple sources of data, there is a certain difference between different data sources, and the dimensions and modalities are different, so the candidate features extracted from the multiple sources of data need to be subjected to feature alignment. The embodiment of the application realizes feature alignment through an adversarial feature alignment network (Adversarial Feature Alignment Network).

[0073] First, the candidate features extracted are mapped to an embedding space, so as to have the same dimension, and preparation is made for subsequent feature alignment, which can be realized through a fully connected neural network. Then, the candidate features are input into the feature alignment network, and the feature alignment network includes a feature extractor and a domain discriminator. The purpose of the feature extractor is to extract discriminative features from different data sources and map them to a shared feature space, and the domain discriminator is used to distinguish from which data source the features come. After the candidate features are input into the feature alignment network, the aligned feature vector is directly output as the key feature of the target crop growth.

[0074] In the embodiment of the application, by performing feature extraction and feature fusion on the multiple sources of data, the feature factors that most influence the target crop growth can be screened out, and by feature alignment, the feature vectors can be ensured to be located in a unified feature space, the differences between different data sources are eliminated, and the real state of the crop growth environment is better reflected, thereby providing high-quality input data for subsequent intelligent reasoning and decision-making.

[0075] In some embodiments, in the intelligent perception module, an agricultural knowledge graph is constructed according to the multiple sources of heterogeneous data of crop planting, which can be realized in the following manner, which is specifically described below.

[0076] Firstly, the entities and relations are extracted from the expert experience knowledge text and crop data to obtain the corresponding entity nodes and the relations between the entity nodes. Here, the core entities in the expert experience knowledge text and the crop data are extracted by calling a named entity recognition model, including crop names, agronomic attributes, meteorological factors, etc., and are classified by type and concept level according to the knowledge field. And the semantic relations between entities are extracted by combining dependency syntax analysis method and pattern matching algorithm. These core entities and semantic relations are respectively used as entity nodes and relations between entity nodes in the agricultural knowledge graph.

[0077] In order to enable the agricultural knowledge graph to recognize feature expressions, the time series variation features of the second meteorological data and the spatial feature maps of the second soil data are extracted here. The time series variation features can be directly extracted by a spatio-temporal convolution network and a long short-term memory (LSTM) network, and the corresponding vector representation is output, while the spatial feature maps can also be extracted by spatial convolution through a spatio-temporal convolution network.

[0078] Next, the time series variation features and spatial feature maps are embedded as node attributes into the entity nodes. The time series variation features can be directly embedded as node attributes into the entity nodes to express the time series variation of the meteorological data, while the spatial feature maps are used to express the spatial distribution of the soil data and can be embedded into the entity nodes related to geography.

[0079] After embedding the node attributes into the entity nodes, the triplets are constructed according to the entity nodes and the relations to obtain the primary knowledge graph. That is, the structured knowledge triplets for agricultural applications are constructed to form the basic knowledge units in the form of “entity-relation-entity”, which integrates the structural relations, time evolution and spatial distribution features to form a multi-modal heterogeneous graph structure for agriculture.

[0080] After the construction of the primary knowledge graph, the graph neural network is generally used to reason and learn the primary knowledge graph, so that the primary knowledge graph can have the graph representation and prediction ability for agricultural tasks. However, on this basis, the expert experience knowledge text needs to be converted into agricultural planting rules, and the agricultural planting rules need to be registered into the reasoning engine of the primary knowledge graph to obtain the agricultural knowledge graph. The agricultural planting rules can participate in the reasoning learning of the graph neural network in the form of triplets, so that the data-driven reasoning process can be coordinated with the knowledge and experience of the field experts to enhance the reasoning ability of the agricultural knowledge graph.

[0081] When constructing a knowledge graph, this embodiment of the present invention utilizes expert knowledge text and crop data. Meteorological and soil data are embedded as node attributes within the knowledge graph, enabling the primary knowledge graph to perform agricultural decision-making reasoning based on multi-source data for target crops. Further incorporating expert knowledge into the knowledge graph enhances its reasoning capabilities and improves the scientificity and reliability of its inference results. The resulting agricultural knowledge graph uniformly represents multidimensional data relationships related to crop cultivation and provides semantic support for subsequent reasoning tasks.

[0082] The following describes the process of using the agricultural knowledge graph to make agricultural decisions in the reasoning and decision-making module. In the first stage of the reasoning and decision-making module, rule-based reasoning is performed to match multi-source data with the agricultural knowledge graph to obtain recommended planting areas for target crops. The following details the rule matching process.

[0083] First, the agricultural planting rules in the agricultural knowledge graph are matched with the first meteorological data and the first soil data to obtain the adaptability scores of the candidate regions under the agricultural planting rules. The agricultural knowledge graph can express agricultural planting rules for various crops. The multi-source data of the target crop is directly matched with these agricultural planting rules. Based on the matched agricultural planting rules, multiple candidate regions suitable for planting the target crop are determined.

[0084] Agricultural planting rules include temperature, soil moisture, and light for target crops. Based on the actual planting characteristics of the target crop, we select the corresponding planting area for each rule and calculate a suitability score. This example uses tomatoes as an example.

[0085] For the temperature rule, the temperature range suitable for tomato growth is 20°C-30°C. If the historical and predicted temperatures of a candidate area fall within this range, the candidate area meets the temperature condition. The predicted temperature T can then be used to calculate the adaptability score under the temperature rule. The formula is as follows:

[0086] (1)

[0087] For the soil moisture rule, the soil moisture data of the candidate area is directly evaluated to ensure that it meets the soil moisture requirements of tomatoes, that is, the moisture data must reach 60%-70% of the optimal humidity. If the soil moisture data of the candidate area is lower than 60% of the optimal humidity or greater than 70% of the optimal humidity, the adaptability score under the soil moisture rule will be The value is 0. If the humidity data is within 60%-70% of the optimal humidity, the score is based on how close the humidity data is to 65% of the optimal humidity. For example, if the humidity data is 65% of the optimal humidity, The highest value is 1, followed by 64% and 66% at 0.8, and so on, 69% and 60% at 0.2.

[0088] For the light rule, tomatoes need 6-8 hours of light per day. Here, we filter the candidate areas that meet this condition based on the light duration data of the candidate areas. If the light duration of the candidate area is 6-8 hours, the adaptability score under the light rule is Take 1, the rest Take 0 or use a selection method similar to the soil moisture rule.

[0089] Next is the evaluation process of the candidate area. According to the preset temperature weight, humidity weight and light weight, the adaptability scores under the temperature rule, soil moisture rule and light rule are weighted respectively, and the weighted results are summed to obtain the total adaptability score of the candidate area under the agricultural planting rule.

[0090] Temperature weight , humidity weight And lighting weight It is set according to the actual needs of the target crop and is used to measure the importance of the target crop planting decision. For example, tomato growth requires a higher temperature, so the larger the temperature weight is, the smaller it is. The adaptability score under the corresponding rule is weighted by the corresponding feature weight, and then the weighted results are summed to finally obtain the total adaptability score of the candidate area under the agricultural planting rule. , the formula is as follows:

[0091] (2)

[0092] Finally, the candidate regions with the highest overall adaptability scores are identified as recommended planting areas for the target crop. These recommended planting areas not only meet the target crop's climate adaptability but also provide an optimal growing environment during the growing season. Typically, multiple candidate regions are selected, and their overall adaptability scores are sorted in descending order. The candidate region or regions at the top of the list are selected as the recommended planting areas for the target crop.

[0093] In an embodiment of the present invention, an agricultural knowledge graph is used to match multi-source data of target crops, and a variety of agricultural planting rules therein are used to screen out the most suitable planting areas for the target crops, thereby making a reasonable preliminary plan for the planting decision of the target crops.

[0094] In the above embodiment, the recommended planting area for the target crop is preliminarily planned through the agricultural planting rules of the agricultural knowledge graph. Furthermore, in the second stage of the reasoning decision module, the prediction of the reasoning model calls the preset reasoning model to predict the yield of the key features, and the planting yield of the target crop in the recommended planting area is obtained, which is explained in detail below.

[0095] The preset inference models specifically include LSTM and Transformer models. First, the preset LSTM model is called to perform a preliminary prediction of key features, obtaining preliminary yield characteristic values ​​for the target crop. The LSTM model is pre-trained based on historical meteorological and soil data for the target crop within the recommended planting area. Therefore, the LSTM model can analyze the impact of dynamic changes in climate data, soil properties, and other factors on crop yield based on the key features of the target crop, thereby estimating the potential future yield of the target crop. The LSTM model's preliminary prediction process can be expressed as the following formula:

[0096] (3)

[0097] (4)

[0098] (5)

[0099] (6)

[0100] (7)

[0101] (8)

[0102] In the above formulas (3) to (8), represents the sigmoid function, t represents the time step, 、 Represents the hidden state at time step t and time step t-1, which is generated by the interaction of historical information and current input, and represents the "memory" of the LSTM unit about the tomato yield at time step t. The key features input at the current time step t are used to predict the tomato yield of a certain region in different seasons in the future, and to generate crop growth trends by combining weather data and historical records. 、 、 are the activation values ​​of the input gate, forget gate, and output gate at each time step t, which control the flow of information in the LSTM unit. They determine which information should be forgotten, retained, or output. Represents the state of the LSTM unit (cell state), which serves as the model memory storage. It is updated at each time step t and carries long-term dependency information through the control of the forget gate and input gate. 、 、 、 Represent the weight matrices of the input gate, forget gate, output gate and LSTM unit respectively, 、 、 、 They represent the input gate, forget gate, output gate and bias term of LSTM unit respectively. Represents the state update value of the LSTM unit, 、 Represents the output value of the LSTM unit at time step t and time step t-1.

[0103] Therefore, at the last time step of the LSTM model, the output value obtained by the LSTM unit is the preliminary yield characteristic value. However, this preliminary yield characteristic value cannot accurately represent the yield of the target crop because it does not consider the impact of temporal and spatial variations and crop growth characteristics on crop yield. Therefore, this embodiment of the present invention also combines the characteristics of multi-source heterogeneous data to further perform prediction.

[0104] For multi-source heterogeneous data, we obtain the temporal variation characteristics of the second meteorological data and the spatial feature map of the second soil data, and extract the target growth cycle characteristics of the crop data corresponding to the target crop. The process of extracting temporal variation characteristics and spatial feature maps can be referred to the process of building the agricultural knowledge graph using the intelligent perception module described above and will not be detailed here. The target growth cycle characteristics of the crop data corresponding to the target crop are the growth cycle characteristics of tomatoes, which can be extracted using a recurrent neural network model, such as an LSTM model.

[0105] The temporal variation features, spatial feature maps, and target growth cycle features are then concatenated to generate concatenated features. These concatenated features and preliminary yield feature values ​​are then input into a pre-set Transformer model to predict crop yields and determine the target crop's yield within the recommended planting area. In the Transformer model, the preliminary yield feature values ​​are normalized and used as a query vector. The concatenated features are then normalized to construct key and value vectors. An attention mechanism is then applied to the query, key, and value vectors, ultimately outputting the corresponding prediction result through a multi-layer perceptron. This prediction is the target crop's yield within the recommended planting area.

[0106] In an embodiment of the present invention, by executing the second stage of the reasoning decision module, the prediction of the reasoning model is used to further predict the planting yield of the target crop based on rule-based reasoning. In this way, the agricultural knowledge graph and the reasoning model are combined to further enhance the reasoning and prediction capabilities of the system.

[0107] In the third stage of the inference decision module, the planting decision planning of the decision model is executed, and the multi-source data is directly decided according to the preset decision model to obtain the planting planning action. Among them, the preset decision model includes the Deep Q-Network (DQN) model. The DQN model is pre-trained by reinforcement learning to dynamically optimize the crop planting planning strategy. The DQN model can perform corresponding planting actions according to the input planting status data, thereby obtaining corresponding rewards. The loss function is constructed through rewards to train the DQN model to ensure that the DQN model can make the best decision based on the planting status data every time. Figure 3 To introduce the pre-training process of the DQN model.

[0108] First, we acquire planting status data. This data, denoted as S, characterizes the current crop environment and includes climate conditions, soil conditions, planting area, historical planting experience, and crop growth status. Climate conditions include current temperature, humidity, and light intensity, which can be acquired in real time using sensors and remote sensing technology. Soil conditions include soil moisture, nutrient content, and pH. Planting area represents the size of the recommended planting area, while historical planting experience includes yield, soil changes, fertilization, and irrigation records for the target crop over previous planting cycles. Crop growth status describes the target crop's current growth stage, health, and the presence of pests and diseases.

[0109] At each time step t, the planting status data is input into the DQN model for decision-making, resulting in the corresponding planting action. The DQN model makes corresponding decisions based on the input planting status data, thereby generating a series of planting actions, represented as A, and the planting actions include selecting a planting area, adjusting the planting time, and changing the fertilization and irrigation plan. Selecting a planting area is to select the location for planting the target crop in different recommended planting areas, taking into account regional climate and soil differences. Adjusting the planting time is to determine when to sow or harvest the target crop based on climatic conditions to improve planting yield and resource utilization efficiency. Changing the fertilization and irrigation plan is to adjust the amount of irrigation and fertilizer according to the needs of the target crop and climate changes to ensure that the target crop has optimal growth conditions.

[0110] Next, based on the planting status data, the crop yield, resource utilization efficiency, and environmental impact value are calculated according to the planting actions, and the reward Q value of the DQN model is estimated based on the crop yield, resource utilization efficiency, and environmental impact value.

[0111] Here, the planting status data at time step t Next, the planting action generated by executing the DQN model decision Combined with the actual growth conditions of the target crop, three indicators can be estimated: crop yield, resource utilization efficiency, and environmental impact. Crop yield refers to the harvested yield of the target crop after planting, resource utilization efficiency refers to the efficiency of water and fertilizer utilization, and the environmental impact refers to the degree of pest and disease damage and environmental pollution. Furthermore, rewards can be assigned based on the values ​​of these three indicators. These rewards are determined by preset thresholds. If crop yield exceeds the preset yield threshold, a positive reward (e.g., 1) is awarded; otherwise, a negative reward (e.g., -1) is given. If resource utilization efficiency exceeds the preset efficiency threshold, a positive reward (e.g., 1) is awarded; otherwise, a negative reward (e.g., -1) is given. If the environmental impact value is below the preset impact threshold, a positive reward (e.g., 1) is awarded; otherwise, a negative reward (e.g., -1) is given.

[0112] Then, according to the different focuses of the indicators (because some target crop planting focuses on yield while some crop planting focuses on resource utilization efficiency or environmental impact), different reward weights are assigned to the reward values ​​of the three indicators respectively, and then the weighted indicators are summed to obtain the instantaneous reward value at the current time step t , the reward Q value of the DQN model is the expected cumulative reward obtained at each time step t. The instantaneous reward value of the current time step is The reward Q value of the next time step t+1 will be updated and calculated to obtain the reward Q value of the DQN model at the current time step t, which is recorded as , the update calculation is expressed as the following formula:

[0113] (9)

[0114] in, Represents an equal sign definition, indicating that the content before the equal sign is defined and updated to the content after the equal sign. Indicates the planting status data at time step t Execute the planting action The reward Q value obtained, Indicates that at the next time step t+1, in the next planting state data The reward Q value obtained by executing the planting action a is, Measuring planting action The long-term expected return under the current state. represents the immediate reward value obtained at the next time step t+1, and represents the immediate feedback on the tomato yield obtained by the system after the action is performed at the next time step t+1. is a discount factor that measures the importance of future rewards. The closer the discount factor is to 1, the more importance the system places on future rewards. Is the learning rate, which controls the step size of the reward Q value update.

[0115] Since the DQN model can perform multiple planting actions at time step t, there must be an optimal planting action that maximizes the reward Q value. The goal of DQN model training is to find the optimal planting action that maximizes the reward Q value.

[0116] When training the DQN model, the loss function of the DQN model is calculated based on the reward Q value and the target Q value, which is recorded as , the target Q value is calculated based on the maximum Q value and reward Q value obtained in the next time step of the current time step, The calculation formula is as follows:

[0117] (10)

[0118] In the above formula (10), It represents the immediate reward value obtained after executing the planting action a in the planting state data s at the current time step. It represents the reward Q value obtained after the planting action a is performed on the planting state data s at the current time step before the DQN model updates the parameters. Represents the parameters of the DQN model before updating, Indicates the next planting state data after the DQN model updates the parameters Execute optimal planting action The reward Q value obtained after Represents the updated parameters of the DQN model. represents the discount factor, and E represents the calculated mathematical expectation.

[0119] Finally, the loss function Back propagation is performed in the DQN model. During the back propagation process, the parameters of the DQN model are updated. After the parameters are updated, the DQN model performs the next time step to target the next planting state data. Make decisions.

[0120] In an embodiment of the present invention, a DQN model is trained through reinforcement learning for decision-making. When making decisions based on multi-source data of target crops, the optimal planting action is selected for output, thereby ensuring the optimal reward value. This enhances the decision-making ability of the system, and ensures that after the target crops execute the optimal planting action, the three indicators of crop yield, resource utilization efficiency, and environmental impact value are comprehensively optimized.

[0121] Considering that climate, soil, planting area, and other environmental factors are constantly changing during crop planting planning, DQN models often struggle to cope with these complex dynamics. Therefore, the inference and decision-making module in this embodiment of the present invention is also used to regularly adaptively update the DQN model and store the historical experience gained from each DQN model decision in a historical experience database. This historical experience database is used to provide the DQN model with planting action decisions during adaptive updates.

[0122] like Figure 3 As shown in the figure, during DQN model training, the state after each planting action is stored as historical experience in the historical experience library. This historical experience includes planting state data, the corresponding planting action, the immediate reward value, and the planting state data before the next planting action. When calculating the loss function, the relevant data is directly extracted from the historical experience library.

[0123] When the DQN model is regularly adaptively updated, the historical experience stored in these historical experience libraries can be used by the DQN model for retrospective analysis, thereby providing the DQN model with decision-making actions based on historical experience, thereby executing a training process to update the parameters of the DQN model and achieve adaptive updates.

[0124] In an embodiment of the present invention, a model self-learning mechanism is designed to adaptively update the DQN model periodically using historical experience from the training process. This enables the DQN model to adapt to the complex dynamic changes of multi-source data and still make optimal decisions, thereby improving the adaptability of system decision-making.

[0125] In some embodiments, as Figure 1 The illustrated crop planting intelligent decision-making and inference system based on multi-source data fusion also includes a feature engineering module. This module is used to invoke a preset time series model to perform time series encoding on the multi-source heterogeneous data, including the second meteorological data, the second soil data, and the crop data, to obtain time series features for multiple crop plantings. The time series model can specifically be an LSTM model. During feature extraction, the attention mechanism within the time series model assigns a corresponding feature weight to each time series feature. Each time series feature is then weighted according to the feature weight, and the weighted results are fused to obtain a feature subset of the multi-source heterogeneous data.

[0126] Here, the attention mechanism of the time series model is used to assign feature weights to evaluate the importance of each time series feature and achieve weighted enhancement of the features.

[0127] In the process of constructing the agricultural knowledge graph, multi-source heterogeneous data have been uniformly modeled and semantically structured, the extracted relevant features have been embedded into the entity nodes as node attributes, and agricultural experience rules have been integrated for graph neural network reasoning. Here, the node attributes of the entity nodes in the agricultural knowledge graph are updated through feature subsets. On the one hand, the old node attributes in the agricultural knowledge graph are updated, and on the other hand, the deep modeling and screening mechanism of the original multi-source heterogeneous data is supplemented.

[0128] In an embodiment of the present invention, the node attributes of entity nodes in the agricultural knowledge graph are updated through feature subsets, thereby further improving the expression accuracy and decision-making reasoning ability of the node attributes in the agricultural knowledge graph, and further realizing the synergistic enhancement of data-driven and knowledge-driven, thereby improving the reasoning ability of the agricultural knowledge graph.

[0129] In some embodiments, in order to enable the system to achieve intelligent decision-making, the system also includes a visualization and human-computer interaction module. The visualization and human-computer interaction module is implemented in two parts. One part is visualization, which is used to visualize the multi-source data of the target crop, the planting yield of the target crop in the recommended planting area, and the planting planning action of the target crop. That is to say, the decision results output by the reasoning decision module in the system are visualized, so that the user can clearly understand the planting decision of the target crop for reference and adjustment. Visualization can be achieved through a display interface or an external display device connected to the visualization and human-computer interaction module, and can be specifically displayed in the form of graphics or text. The other part is human-computer interaction, which is used to perform decision reasoning on the multi-source data of the crop to be decided that is input in real time, and obtain the recommended planting area of ​​the crop to be decided, the planting yield within the recommended planting area, and the planting planning action. The visualization and human-computer interaction module can obtain multi-source data of the crops to be decided that are input by the user into the system in real time. This means that the user needs the system to make a planting decision for the crops to be decided. At this time, the visualization and human-computer interaction module can trigger the system's multi-source data fusion module, intelligent perception module and reasoning decision module to start operating, thereby realizing decision reasoning. Finally, the reasoning decision module outputs the recommended planting area of ​​the crops to be decided, the planting yield within the recommended planting area and the planting planning action, and visualizes them.

[0130] The embodiment of the present invention realizes the visualization of decision reasoning results by designing a visualization and human-computer interaction module on the system, and on the one hand, realizes the visualization of decision reasoning results, so that users can clearly understand the planting decisions made by the system; on the other hand, it realizes human-computer interaction and can interact with users for data, meet users' decision needs for crops to be decided in real time, and improve user experience.

[0131] The following describes the crop planting intelligent decision-making and reasoning method based on multi-source data fusion provided by the present invention. The crop planting intelligent decision-making and reasoning method based on multi-source data fusion described below and the crop planting intelligent decision-making and reasoning system based on multi-source data fusion described above can be referenced to each other.

[0132] like Figure 4 As shown, the crop planting intelligent decision-making reasoning method based on multi-source data fusion includes steps 101, 102 and 103, which are described in detail below.

[0133] In step 101, multi-source data of a target crop is obtained, and a spatiotemporal convolutional network is called to extract features from the multi-source data to obtain key features of the growth of the target crop. The multi-source data includes first meteorological data and first soil data.

[0134] In step 102, an agricultural knowledge graph is constructed based on multi-source heterogeneous data of crop planting, where the multi-source heterogeneous data includes second meteorological data, second soil data, crop data, and expert experience knowledge text.

[0135] In step 103, the multi-source data is matched with the agricultural knowledge graph to obtain the recommended planting area of ​​the target crop, the preset inference model is called to predict the yield of the key features to obtain the planting yield of the target crop in the recommended planting area, and the preset decision model is called to make a decision on the multi-source data to obtain the planting planning action of the target crop.

[0136] It should be noted that the beneficial effects of the intelligent decision-making reasoning method for crop planting based on multi-source data fusion here correspond to the beneficial effects of the intelligent decision-making reasoning system for crop planting based on multi-source data fusion mentioned above. Therefore, the beneficial effects of the intelligent decision-making reasoning method for crop planting based on multi-source data fusion will not be repeated here.

[0137] Figure 5 An example of a physical structure diagram of an electronic device is shown below. Figure 5As shown, the electronic device can include a processor 510, a communications interface 520, a memory 530, and a communications bus 540, wherein the processor 510, the communications interface 520, and the memory 530 complete mutual communication through the communications bus 540. The processor 510 can invoke a logic instruction in the memory 530 to execute a crop planting intelligent decision reasoning method based on multi-source data fusion, the method comprising: acquiring multi-source data of a target crop, and invoking a spatio-temporal convolution network to perform feature extraction on the multi-source data to obtain key features of growth of the target crop, the multi-source data including first meteorological data and first soil data; constructing an agricultural knowledge graph according to multi-source heterogeneous data of crop planting, the multi-source heterogeneous data including second meteorological data, second soil data, crop data, and expert experience knowledge text; performing rule matching on the multi-source data and the agricultural knowledge graph to obtain a recommended planting area of the target crop, invoking a preset reasoning model to perform yield prediction on the key features to obtain a planting yield of the target crop in the recommended planting area, and invoking a preset decision model to perform decision-making on the multi-source data to obtain a planting planning action of the target crop.

[0138] In addition, the logic instruction in the memory 530 described above can be implemented in the form of a software function unit and sold or used as an independent product, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the methods described in various embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various program code storage media.

[0139] On the other hand, the present invention also provides a computer program product, which includes a computer program, which can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the crop planting intelligent decision-making and reasoning method based on multi-source data fusion provided by the above methods, the method including: obtaining multi-source data of the target crop, and calling a spatiotemporal convolutional network to extract features from the multi-source data to obtain key features of the target crop growth, the multi-source data including first meteorological data and first soil data; constructing an agricultural knowledge graph based on multi-source heterogeneous data of crop planting, the multi-source heterogeneous data including second meteorological data, second soil data, crop data and expert experience knowledge text; matching the multi-source data with the agricultural knowledge graph according to rules to obtain a recommended planting area for the target crop, calling a preset inference model to perform yield prediction on the key features to obtain the planting yield of the target crop in the recommended planting area, and calling a preset decision model to make decisions on the multi-source data to obtain a planting planning action for the target crop.

[0140] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the intelligent decision-making and reasoning method for crop planting based on multi-source data fusion provided by the above-mentioned methods, the method comprising: obtaining multi-source data of the target crop, and calling a spatiotemporal convolutional network to perform feature extraction on the multi-source data to obtain key features of the target crop growth, the multi-source data including first meteorological data and first soil data; constructing an agricultural knowledge graph based on multi-source heterogeneous data of crop planting, the multi-source heterogeneous data including second meteorological data, second soil data, crop data and expert experience knowledge text; performing rule matching on the multi-source data with the agricultural knowledge graph to obtain a recommended planting area for the target crop, calling a preset reasoning model to perform yield prediction on the key features to obtain the planting yield of the target crop in the recommended planting area, and calling a preset decision model to make a decision on the multi-source data to obtain a planting planning action for the target crop.

[0141] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0142] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0143] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A crop planting intelligent decision-making and reasoning system based on multi-source data fusion, characterized by: include: a multi-source data fusion module, configured to obtain multi-source data of a target crop and extract features from the multi-source data using a spatiotemporal convolutional network to obtain key growth characteristics of the target crop, wherein the multi-source data includes first meteorological data and first soil data; An intelligent perception module, configured to construct an agricultural knowledge graph based on multi-source heterogeneous data of crop planting, wherein the multi-source heterogeneous data includes second meteorological data, second soil data, crop data, and expert experience knowledge text; An inference and decision module is used to match the multi-source data with the agricultural knowledge graph to obtain a recommended planting area for the target crop, use a preset inference model to predict the yield of the key features to obtain the planting yield of the target crop in the recommended planting area, and use a preset decision model to make a decision on the multi-source data to obtain a planting planning action for the target crop; The calling of a preset inference model to perform yield prediction on the key features to obtain the planting yield of the target crop in the recommended planting area includes: Calling a preset LSTM model to perform preliminary prediction on the key features to obtain preliminary yield characteristic values ​​of the target crop, wherein the LSTM model is trained based on historical meteorological data and historical soil data of the target crop in the recommended planting area; From the multi-source heterogeneous data, a temporal variation feature of the second meteorological data and a spatial feature map of the second soil data are obtained, and target growth cycle features of the crop data corresponding to the target crop are extracted; Splicing the temporal variation feature, the spatial feature map, and the target growth cycle feature to obtain a spliced ​​feature; Inputting the splicing features and the preliminary yield feature values ​​into a preset Transformer model to perform crop yield prediction, thereby obtaining the planting yield of the target crop in the recommended planting area; The preset decision model includes a DQN model, and the pre-training process of the DQN model includes: Acquiring planting status data, including climate conditions, soil conditions, planting area, historical planting experience, and crop growth status; At each time step, the planting state data is input into the DQN model for decision-making to obtain corresponding planting actions, wherein the planting actions include selecting a planting area, adjusting the planting time, and changing the fertilization and irrigation plan; Under the planting state data, the crop yield, resource utilization efficiency, and environmental impact value are calculated according to the planting action, and the reward Q value of the DQN model in the time step is estimated based on the crop yield, the resource utilization efficiency, and the environmental impact value; Calculating a loss function of the DQN model based on the reward Q value and the target Q value, wherein the target Q value is calculated based on the maximum Q value obtained in the next time step after the current time step and the reward Q value; Backpropagation is performed in the DQN model through the loss function to update the parameters of the DQN model.

2. The crop planting intelligent decision-making and reasoning system based on multi-source data fusion according to claim 1 is characterized in that: The calling of the spatiotemporal convolutional network to extract features from the multi-source data to obtain key features of target crop growth, including: Calling the spatiotemporal convolutional network to perform temporal convolution processing and spatial convolution processing on the multi-source data respectively to obtain corresponding temporal convolution features and spatial convolution features; Performing channel splicing processing on the temporal convolution features and the spatial convolution features to obtain the spatiotemporal correlation features of the multi-source data; According to the feature weights assigned by the attention mechanism in the spatiotemporal convolutional network, a candidate feature with the largest feature weight is selected from the spatiotemporal correlation features; An adversarial feature alignment network is called to perform feature alignment on the candidate features to obtain key features for target crop growth.

3. The crop planting intelligent decision-making and reasoning system based on multi-source data fusion according to claim 1 is characterized in that: The agricultural knowledge graph is constructed based on multi-source heterogeneous data of crop planting, including: Extracting entities and relationships from the expert experience knowledge text and the crop data to obtain corresponding entity nodes and relationships between entity nodes; Extracting the temporal variation characteristics of the second meteorological data and the spatial characteristic graph of the second soil data, and embedding the temporal variation characteristics and the spatial characteristic graph into the entity node as node attributes; Construct triples based on the entity nodes and the relationships to obtain a primary knowledge graph; The expert experience knowledge text is converted into agricultural planting rules, and the agricultural planting rules are registered in the reasoning engine of the primary knowledge graph to obtain an agricultural knowledge graph.

4. The crop planting intelligent decision-making and reasoning system based on multi-source data fusion according to claim 1 is characterized in that: The method of matching the multi-source data with the agricultural knowledge graph to obtain the recommended planting area of ​​the target crop includes: matching the agricultural planting rules in the agricultural knowledge graph with the first meteorological data and the first soil data to obtain an adaptability score of the candidate region under the agricultural planting rules, wherein the agricultural planting rules include temperature rules, soil moisture rules, and light rules for the target crop; Weighting the adaptability scores under the temperature rule, soil moisture rule, and light rule according to the preset temperature weight, humidity weight, and light weight, respectively, and summing the weighted results to obtain the total adaptability score of the candidate area under the agricultural planting rule; The candidate areas with the highest overall adaptation scores were identified as recommended planting areas for the target crops.

5. The crop planting intelligent decision-making and reasoning system based on multi-source data fusion according to claim 1 is characterized in that: The reasoning and decision-making module is also used to regularly adaptively update the DQN model and store the historical experience after each decision of the DQN model in a historical experience library. The historical experience library is used to provide the DQN model with planting actions for decision-making based on historical experience when the DQN model is adaptively updated.

6. The crop planting intelligent decision-making and reasoning system based on multi-source data fusion according to claim 1 is characterized in that: The system further comprises: a feature engineering module, configured to call a preset time series model to perform time series encoding on the second meteorological data, the second soil data, and the crop data included in the multi-source heterogeneous data, to obtain time series sequence features of multiple crop plantings; Assigning corresponding feature weights to each time series feature through the attention mechanism in the time series model; Weighting each time series feature according to the feature weight, and fusing the weighted results to obtain a feature subset of the multi-source heterogeneous data; The node attributes of the entity nodes in the agricultural knowledge graph are updated using the feature subset.

7. The crop planting intelligent decision-making and reasoning system based on multi-source data fusion according to claim 1 is characterized in that: The system further comprises: The visualization and human-computer interaction module is used to visualize the multi-source data of the target crop, the planting yield of the target crop in the recommended planting area, and the planting planning actions of the target crop, and perform decision reasoning based on the multi-source data of the crop to be decided that is input in real time to obtain the recommended planting area of ​​the crop to be decided, the planting yield in the recommended planting area, and the planting planning actions.

8. A crop planting intelligent decision-making reasoning method based on multi-source data fusion, characterized by: include: Acquire multi-source data of a target crop, and call a spatiotemporal convolutional network to perform feature extraction on the multi-source data to obtain key growth characteristics of the target crop, wherein the multi-source data includes first meteorological data and first soil data; Constructing an agricultural knowledge graph based on multi-source heterogeneous data of crop planting, wherein the multi-source heterogeneous data includes second meteorological data, second soil data, crop data, and expert experience knowledge text; Matching the multi-source data with the agricultural knowledge graph to obtain a recommended planting area for the target crop, using a preset inference model to predict the yield of the key features to obtain the planting yield of the target crop in the recommended planting area, and using a preset decision model to make a decision on the multi-source data to obtain a planting planning action for the target crop; The calling of a preset inference model to perform yield prediction on the key features to obtain the planting yield of the target crop in the recommended planting area includes: Calling a preset LSTM model to perform preliminary prediction on the key features to obtain preliminary yield characteristic values ​​of the target crop, wherein the LSTM model is trained based on historical meteorological data and historical soil data of the target crop in the recommended planting area; From the multi-source heterogeneous data, a temporal variation feature of the second meteorological data and a spatial feature map of the second soil data are obtained, and target growth cycle features of the crop data corresponding to the target crop are extracted; Splicing the temporal variation feature, the spatial feature map, and the target growth cycle feature to obtain a spliced ​​feature; Inputting the splicing features and the preliminary yield feature values ​​into a preset Transformer model to perform crop yield prediction, thereby obtaining the planting yield of the target crop in the recommended planting area; The preset decision model includes a DQN model, and the pre-training process of the DQN model includes: Acquiring planting status data, including climate conditions, soil conditions, planting area, historical planting experience, and crop growth status; At each time step, the planting state data is input into the DQN model for decision-making to obtain corresponding planting actions, wherein the planting actions include selecting a planting area, adjusting the planting time, and changing the fertilization and irrigation plan; Under the planting state data, the crop yield, resource utilization efficiency, and environmental impact value are calculated according to the planting action, and the reward Q value of the DQN model in the time step is estimated based on the crop yield, the resource utilization efficiency, and the environmental impact value; Calculating a loss function of the DQN model based on the reward Q value and the target Q value, wherein the target Q value is calculated based on the maximum Q value obtained in the next time step after the current time step and the reward Q value; Backpropagation is performed in the DQN model through the loss function to update the parameters of the DQN model.

Images