Deep fertilization efficiency prediction method and device, electronic equipment and storage medium

Abstract

The invention provides a deep fertilization efficiency prediction method and device, electronic equipment and a storage medium, and the method comprises the steps: dynamically extracting a domain knowledge sub-graph related to the fertilization efficiency through the semantic association characteristics of an agricultural knowledge graph, and screening out nodes related to the fertilization efficiency through triple screening conditions; then, performing feature screening based on the domain knowledge sub-graph features, constructing feature selection samples formed by feature values of all initial features in an initial feature set corresponding to multiple deep fertilization modes, performing feature screening in combination with the feature compression capability of an automatic encoder, and screening redundant features; according to the method, the efficiency prediction features highly related to the fertilization efficiency are selected from the initial feature set, so that an intelligent prediction model for fertilization efficiency prediction is constructed, and the problem of feature latitude disaster caused by difficulty in determining effective features for fertilization prediction and the problem of prediction precision reduction caused by more redundant features are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of agricultural data processing, and particularly to a method, device, electronic device and storage medium for predicting deep fertilization efficiency. Background Art

[0002] Deep fertilization is a fertilization method that directly applies fertilizers deep into the soil near the root distribution area of crops. Its core goal is to improve the nutrient absorption efficiency of crops by accurately positioning the fertilizer location, while reducing nutrient loss caused by surface volatilization or rain erosion. Compared with traditional surface broadcasting, deep fertilization can significantly reduce the ammonia volatilization loss of nitrogen fertilizer by 20 - 40%, and the phosphorus fixation rate by 15 - 30%, which has important application value in water-saving agriculture and precision agriculture. However, due to the difference in the vertical permeability coefficient of different soil textures can reach 3 orders of magnitude (such as sandy soil 10 -3 cm / s vs clay 10 -6 cm / s), which directly affects the fertilizer migration path, and there are significant interspecific differences in crop root architecture (such as the dense layer of wheat fibrous roots is 0 - 20 cm vs the deep roots of fruit trees can reach 2 m). Moreover, for every 10 mm / h increase in rainfall intensity in environmental factors, the risk of surface fertilizer loss can be increased by 35%. Therefore, an accurate efficiency prediction model can help farmers optimize the configuration in dimensions such as fertilizer type selection, fertilization amount, and fertilization depth.

[0003] Current prediction schemes are mainly divided into three categories: (1) predicting fertilization efficiency based on nutrient migration models, which has the problem of high modeling complexity; (2) real-time monitoring schemes based on sensor networks, which have the problem of high deployment costs; and (3) prediction models based on machine learning. Among them, although the prediction models based on machine learning can integrate multi-source data such as meteorology, soil, and agronomy, they face the curse of dimensionality - typical input features include nearly a hundred-dimensional features such as environmental parameters, crop growth indicators, and fertilizer attributes, resulting in the training time of machine learning models increasing rapidly with the number of features and being prone to overfitting problems. More seriously, the probability of pseudo-correlation between irrelevant features (such as soil conductivity far from the root zone) and the target variable in high-dimensional space is relatively high, resulting in a significant decrease in the prediction accuracy of the model. Summary of the Invention

[0004] The present invention provides a method, device, electronic device and storage medium for predicting deep fertilization efficiency to solve the defect of poor prediction accuracy in the prior art.

[0005] The present invention provides a method for predicting deep fertilization efficiency, including:

[0006] Based on the agricultural knowledge graph, obtain the domain knowledge subgraph corresponding to the fertilization efficiency; each node in the agricultural knowledge graph corresponds to an agricultural concept in the agricultural text, and the edge attributes connecting each node include the agricultural text identifiers to which the nodes connected by the corresponding edges belong;

[0007] Based on the domain knowledge subgraph, determine the initial feature set for predicting the fertilization efficiency, and obtain the feature values of various deep fertilization methods for different cultivated lands and different crops corresponding to each initial feature in the initial feature set;

[0008] Based on the feature values of various deep fertilization methods corresponding to each initial feature in the initial feature set, select the efficiency prediction features from the initial feature set, and based on the efficiency prediction features, predict the efficiency of the current deep fertilization method for a specific cultivated land and a specific crop;

[0009] Among them, the obtaining of the domain knowledge subgraph corresponding to the fertilization efficiency based on the agricultural knowledge graph includes:

[0010] Starting from the target node corresponding to the fertilization efficiency, traverse along the connecting edges. If the influence degree of the currently traversed node on the target node is higher than the first preset threshold, the number of hops between the currently traversed node and the target node is less than the second preset threshold, and the number of agricultural texts spanned by the currently traversed node and the target node is less than the third preset threshold, then add the currently traversed node to the domain knowledge subgraph and continue to traverse along the connecting edges of the currently traversed node.

[0011] According to a deep fertilization efficiency prediction method provided by the present invention, the influence degree of the currently traversed node on the target node is calculated based on the following method:

[0012] Obtain the sequence of passing nodes included in the path between the currently traversed node and the target node;

[0013] Based on the co-occurrence times of any adjacent nodes in the sequence of passing nodes in each agricultural text and the total number of agricultural texts, determine the influence degree between any adjacent nodes;

[0014] Based on the influence degree between any adjacent nodes in the sequence of passing nodes, determine the influence degree of the currently traversed node on the target node.

[0015] According to a deep fertilization efficiency prediction method provided by the present invention, the selection of the efficiency prediction features from the initial feature set based on the feature values of various deep fertilization methods corresponding to each initial feature in the initial feature set includes:

[0016] Based on the feature values of various deep fertilization methods corresponding to each initial feature in the initial feature set, construct multiple feature selection samples;

[0017] Select samples based on the features to train the encoder and decoder in the autoencoder, obtaining a trained first autoencoder;

[0018] Based on the encoder in the trained first autoencoder and the weight configuration layer, construct a feature selection model; wherein, the parameters of the weight configuration layer are used to adjust the weights corresponding to each initial feature;

[0019] Train the feature selection model based on the feature selection samples, obtaining the weights corresponding to each initial feature configured by the weight configuration layer;

[0020] Based on the weights corresponding to each initial feature configured by the weight configuration layer, select the initial features with non-zero weights from the initial feature set as the efficiency prediction features.

[0021] According to a method for predicting deep fertilization efficiency provided by the present invention, the training of the encoder and decoder in the autoencoder based on the feature selection samples includes:

[0022] Input any feature selection sample into the autoencoder, obtaining the feature encoding vector corresponding to the any feature selection sample output by the encoder in the autoencoder and the feature reconstruction result reconstructed by the decoder in the autoencoder based on the feature encoding vector;

[0023] Obtain the Laplacian matrix of the any feature selection sample and the Laplacian matrix of the feature reconstruction result;

[0024] Based on the similarity between the any feature selection sample and the feature reconstruction result and the similarity between the Laplacian matrix of the any feature selection sample and the Laplacian matrix of the feature reconstruction result, adjust the model parameters of the encoder and decoder in the autoencoder.

[0025] According to a method for predicting deep fertilization efficiency provided by the present invention, the training of the feature selection model based on the feature selection samples includes:

[0026] Based on the current parameters of the weight configuration layer of the feature selection model and any feature selection sample, determine the input feature sample;

[0027] Based on the encoder of the feature selection model, encode the input feature sample, obtaining the feature encoding vector corresponding to the input feature sample;

[0028] Based on the encoder of the feature selection model, encode the any feature selection sample, obtaining the feature encoding vector corresponding to the any feature selection sample;

[0029] Determine the L1 norm of the current parameters of the encoder of the feature selection model and the current parameters of the weight configuration layer of the feature selection model;

[0030] Based on the difference between the feature encoding vector corresponding to the input feature sample and the feature encoding vector corresponding to any one of the feature selection samples, as well as the L1 norm of the current parameters of the encoder of the feature selection model and the current parameters of the weight configuration layer of the feature selection model, adjust the parameters of the weight configuration layer and the encoder in the feature selection model.

[0031] According to a deep fertilization efficiency prediction method provided by the present invention, the efficiency prediction of the current deep fertilization method for a specific cultivated land and a specific crop based on the efficiency prediction features includes:

[0032] Construct a fertilization efficiency prediction model based on the encoder and the regression layer of the feature selection model;

[0033] Wherein, the input of the encoder of the feature selection model includes the feature values of the efficiency prediction features corresponding to the current deep fertilization method for a specific cultivated land and a specific crop, and the regression layer is used to perform regression calculation based on the feature encoding vector output by the encoder of the feature selection model to obtain the fertilization efficiency of the current deep fertilization method.

[0034] The present invention also provides a deep fertilization efficiency prediction device, including:

[0035] A knowledge extraction unit, configured to obtain a domain knowledge subgraph corresponding to fertilization efficiency based on an agricultural knowledge graph; each node in the agricultural knowledge graph corresponds to an agricultural concept in agricultural texts, and the edge attributes connecting each node include the agricultural text identifiers to which the nodes connected by the corresponding edges belong;

[0036] An initial feature extraction unit, configured to determine an initial feature set for performing fertilization efficiency prediction based on the domain knowledge subgraph, and obtain the feature values of each initial feature in the initial feature set corresponding to multiple deep fertilization methods for different cultivated lands and different crops;

[0037] A feature selection and prediction unit, configured to select efficiency prediction features from the initial feature set based on the feature values of each initial feature in the initial feature set corresponding to the multiple deep fertilization methods, and perform efficiency prediction on the current deep fertilization method for a specific cultivated land and a specific crop based on the efficiency prediction features;

[0038] Wherein, the obtaining of the domain knowledge subgraph corresponding to fertilization efficiency based on the agricultural knowledge graph includes:

[0039] Starting from the target node corresponding to the fertilization efficiency, traverse along the connecting edges. If the influence degree of the currently traversed node on the target node is higher than the first preset threshold, the number of hops between the currently traversed node and the target node is less than the second preset threshold, and the number of agricultural texts spanned by the currently traversed node and the target node is less than the third preset threshold, then add the currently traversed node to the domain knowledge subgraph and continue traversing along the connecting edges of the currently traversed node.

[0040] According to a deep fertilization efficiency prediction device provided by the present invention, the influence degree of the currently traversed node on the target node is calculated based on the following method:

[0041] Obtain the sequence of passing nodes included in the path between the currently traversed node and the target node;

[0042] Based on the co-occurrence times of any adjacent nodes in the sequence of passing nodes in each agricultural text and the total number of agricultural texts, determine the influence degree between the any adjacent nodes;

[0043] Based on the influence degree between any adjacent nodes in the sequence of passing nodes, determine the influence degree of the currently traversed node on the target node.

[0044] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the deep fertilization efficiency prediction method as described in any one of the above.

[0045] The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the deep fertilization efficiency prediction method as described in any one of the above.

[0046] The present invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the deep fertilization efficiency prediction method as described in any one of the above.

[0047] A method, device, electronic device, and storage medium for predicting deep fertilization efficiency provided by the present invention dynamically extract domain knowledge subgraphs related to fertilization efficiency through the semantic association characteristics of an agricultural knowledge graph, where nodes related to fertilization efficiency can be screened through triple screening conditions; subsequently, feature screening is performed based on the characteristics of the domain knowledge subgraph. By constructing a feature selection sample formed by the feature values of each initial feature in the initial feature set corresponding to multiple deep fertilization methods and combining the feature compression ability of an autoencoder for feature screening, redundant features are screened out, and efficiency prediction features highly relevant to fertilization efficiency are selected from the initial feature set, thereby constructing an intelligent prediction model for fertilization efficiency prediction, overcoming the problem of feature dimension disaster caused by the difficulty in determining effective features for fertilization prediction and the problem of decreased prediction accuracy caused by a large number of redundant features. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0049] Figure 1 is a schematic flowchart of the deep fertilization efficiency prediction method provided by the present invention;

[0050] Figure 2 is a schematic flowchart of the efficiency prediction feature screening method provided by the present invention;

[0051] Figure 3 is a schematic structural diagram of the deep fertilization efficiency prediction device provided by the present invention;

[0052] Figure 4 is a schematic structural diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0053] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions in the present invention in conjunction with the drawings in the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments in the present invention belong to the scope of protection of the present invention.

[0054] Figure 1 is a schematic flowchart of the deep fertilization efficiency prediction method provided by the present invention, as Figure 1 shown, the method includes:

[0055] Step 110: Based on the agricultural knowledge graph, obtain the domain knowledge sub-graph corresponding to the fertilization efficiency; each node in the agricultural knowledge graph corresponds to an agricultural concept in the agricultural text, and the edge attributes connecting each node include the agricultural text identifier to which the nodes connected by the corresponding edge belong;

[0056] Step 120: Based on the domain knowledge sub-graph, determine the initial feature set for fertilization efficiency prediction, and obtain the feature values of each initial feature in the initial feature set corresponding to various deep fertilization methods for different cultivated lands and different crops;

[0057] Step 130: Select the efficiency prediction features from the initial feature set based on the feature values of each initial feature in the initial feature set corresponding to the various deep fertilization methods, and perform efficiency prediction on the current deep fertilization method for a specific cultivated land and a specific crop based on the efficiency prediction features;

[0058] Among them, the obtaining of the domain knowledge sub-graph corresponding to the fertilization efficiency based on the agricultural knowledge graph includes:

[0059] Starting from the target node corresponding to the fertilization efficiency, traverse along the connecting edges. If the influence degree of the currently traversed node on the target node is higher than the first preset threshold, the number of hops between the currently traversed node and the target node is less than the second preset threshold, and the number of agricultural texts spanned by the currently traversed node and the target node is less than the third preset threshold, then add the currently traversed node to the domain knowledge sub-graph, and continue to traverse along the connecting edges of the currently traversed node.

[0060] Here, the core of the embodiment of the present invention is to dynamically extract the domain knowledge sub-graph related to the fertilization efficiency through the semantic association characteristics of the knowledge graph, and perform feature screening based on the sub-graph features to construct an intelligent prediction model for fertilization efficiency prediction, so as to overcome the problems of feature dimension disaster caused by the difficulty in determining effective features for fertilization prediction and the decrease in prediction accuracy caused by many redundant features.

[0061] Among them, structured knowledge can be extracted from agricultural texts in the agricultural science literature database to construct an agricultural knowledge graph covering fields such as soil science, crop physiology, and fertilizer chemistry. Through natural language processing technology, entity extraction is performed on agricultural concepts in agricultural texts to form a knowledge graph containing nodes corresponding to multiple agricultural concepts such as soil components (such as organic matter content, pH value), crop characteristics (such as root morphology, nutrient absorption cycle), fertilization parameters (such as furrow depth, fertilizer slow release property), etc. The connecting edges connecting adjacent nodes record the metadata of the two nodes' source documents through the edge attribute field, including the agricultural text identifier and the paragraph position, etc. In some embodiments, the agricultural knowledge graph can be stored using a Neo4j graph database.

[0062] In some embodiments, after entity extraction of agricultural concepts in agricultural texts, nodes in the agricultural knowledge graph are formed. For any two agricultural concepts from the same agricultural text, in order to determine whether a connection edge needs to be constructed between the nodes corresponding to these two agricultural concepts, coreference resolution can be performed on the agricultural text to obtain multiple coreference sets corresponding to the agricultural text, where each coreference set contains entity mentions (entity mention) that point to the same entity. For any two of the above-mentioned agricultural concepts, if any one of the agricultural concepts is in a certain coreference set, the position of the entity mention with the earliest position in the agricultural text in the corresponding coreference set is determined as the starting position of the corresponding agricultural concept. If any one of the agricultural concepts is not in any coreference set, its own position in the agricultural text is used as the starting position of the agricultural concept. The target text segment between the starting positions of the two agricultural concepts is obtained, entity mentions in the target text segment are sequentially extracted to obtain an entity mention sequence, and the entity mention sequence is compressed based on the coreference set (other entity mentions between two entity mentions belonging to the same coreference set can be deleted) to obtain a compressed mention sequence. If the length of the compressed mention sequence is less than a preset value, it can be determined that a connection edge needs to be constructed between the nodes corresponding to these two agricultural concepts.

[0063] After obtaining the agricultural knowledge graph, "fertilization efficiency" can be used as the target node for knowledge subgraph mining to determine other nodes related to the target node of "fertilization efficiency" in the agricultural knowledge graph and their graph structure relationships, so as to obtain the domain knowledge subgraph corresponding to fertilization efficiency. In some embodiments, traversal can be performed along the connection edge starting from the target node, and three screening conditions can be applied synchronously:

[0064] Semantic influence degree screening: Calculate the influence degree of the currently traversed node on the target node, and screen nodes with an influence degree higher than the first preset threshold;

[0065] In some embodiments, in order to calculate the influence degree of the currently traversed node on the target node, a sequence of passing nodes included in the path between the currently traversed node and the target node can be obtained, and based on the co-occurrence times of any two adjacent nodes in the sequence of passing nodes in each agricultural text and the total number of agricultural texts, the influence degree between any two adjacent nodes can be determined. Among them, the ratio of the co-occurrence times of any two adjacent nodes in each agricultural text to the total number of agricultural texts can be used as the influence degree between the two adjacent nodes. Subsequently, based on the influence degrees between any two adjacent nodes in the sequence of passing nodes, the influence degree of the currently traversed node on the target node can be determined. For example, the product of the influence degrees between any two adjacent nodes in the sequence of passing nodes can be used as the influence degree of the currently traversed node on the target node. In some other embodiments, in order to further improve the calculation accuracy of the influence degree between any two adjacent nodes and more accurately quantify the correlation between two nodes in the agricultural field, for any two adjacent nodes in the above sequence of passing nodes, the starting positions of the agricultural concepts corresponding to the two adjacent nodes in the corresponding agricultural text can be determined (the method for determining the starting positions is the same as that in the above embodiments and will not be elaborated here), the target text segment between the starting positions of the two agricultural concepts can be obtained, the entity mentions in the target text segment can be sequentially extracted to obtain an entity mention sequence, and the entity mention sequence can be compressed based on the co-reference set to obtain a compressed mention sequence. Based on the length of the compressed mention sequence, in combination with the co-occurrence times of the two nodes in each agricultural text and the total number of agricultural texts, the influence degree between the two nodes can be determined. For example, the first ratio of the co-occurrence times of the two nodes in each agricultural text to the total number of agricultural texts and the second ratio of the preset value to the length of the compressed mention sequence can be calculated, and the product of the first ratio and the second ratio can be used as the influence degree between the two nodes.

[0066] Topological distance constraint: Filter nodes with the number of hops (hop number) less than the second preset threshold from the target node, and limit the number of hops between the subgraph nodes and the target node not to exceed the preset number of layers to ensure the tightness of knowledge association.

[0067] Cross-document breadth control: Filter nodes with the number of agricultural texts crossed by the path to the target node (if the agricultural text identifiers corresponding to multiple edges of the path are the same, the number of agricultural texts crossed by these edges is recorded as 1) less than the third preset threshold, so that the number of documents crossed by the association path between any node in the subgraph and the target node does not exceed the preset number of articles, avoiding knowledge fragmentation.

[0068] Only when the currently traversed node simultaneously meets the above three conditions, the currently traversed node is added to the domain knowledge subgraph, and at the same time, the connection relationship between the node and other nodes in the domain knowledge subgraph is constructed according to the graph structure relationship of the node in the agricultural knowledge graph, and then continue to traverse along the connection edge of the node.

[0069] Through the above process, a domain knowledge sub-graph covering influencing factors related to fertilization efficiency, such as soil physical structure (e.g., bulk density, permeability coefficient), crop root architecture (e.g., lateral root density, main root elongation rate), and fertilizer kinetic characteristics (e.g., nitrogen release curve, phosphorus fixation rate), can be finally generated, forming a knowledge framework for subsequent feature engineering.

[0070] Based on this domain knowledge sub-graph, an initial feature set for predicting fertilization efficiency is determined. Among them, the initial features in the initial feature set correspond to the nodes in the domain knowledge sub-graph. Subsequently, the feature values of each initial feature in the initial feature set corresponding to various deep fertilization methods for different cultivated lands and different crops are obtained, thereby constructing a feature selection sample for feature selection. Based on the feature values of each initial feature in the initial feature set corresponding to the above various deep fertilization methods, efficiency prediction features that can effectively predict fertilization efficiency are selected from the initial feature set, and the efficiency of the current deep fertilization method for a specific cultivated land and a specific crop is predicted based on the above efficiency prediction features.

[0071] Such as Figure 2 shown, selecting efficiency prediction features from the initial feature set based on the feature values of each initial feature in the initial feature set corresponding to various deep fertilization methods includes:

[0072] Step 210, based on the feature values of each initial feature in the initial feature set corresponding to various deep fertilization methods, construct multiple feature selection samples;

[0073] Step 220, train the encoder and decoder in the autoencoder based on the feature selection samples to obtain a trained first autoencoder;

[0074] Step 230, construct a feature selection model based on the encoder and weight configuration layer in the trained first autoencoder; wherein, the parameters of the weight configuration layer are used to adjust the weights corresponding to each initial feature;

[0075] Step 240, train the feature selection model based on the feature selection samples to obtain the weights corresponding to each initial feature configured by the weight configuration layer;

[0076] Step 250, based on the weights corresponding to each initial feature configured by the weight configuration layer, select the initial features with non-zero weights from the initial feature set as the efficiency prediction features.

[0077] Here, based on the feature values of each initial feature in the initial feature set corresponding to different deep fertilization methods for different cultivated lands and different crops, multiple different feature selection samples are constructed. The encoder and decoder in the autoencoder are trained based on the feature selection samples to obtain a trained first autoencoder.

[0078] Among them, an autoencoder with sparse constraints can be constructed to extract the non-linear relationships between initial features. The encoder part includes an input layer (for receiving the feature values of each initial feature in any feature selection sample), an encoder, and a decoder. To enable it to effectively capture the complex relationships between initial features, when training the autoencoder, the feature selection sample can be input into the autoencoder to obtain the feature encoding vector corresponding to the feature selection sample output by the encoder in the autoencoder and the feature reconstruction result reconstructed by the decoder in the autoencoder based on this feature encoding vector. Subsequently, the Laplacian matrix of the feature selection sample and the Laplacian matrix of the feature reconstruction result are obtained, so as to adjust the model parameters of the encoder and decoder in the autoencoder based on the similarity between the feature selection sample and the feature reconstruction result and the similarity between the Laplacian matrix of the feature selection sample and the Laplacian matrix of the feature reconstruction result. It can be seen that when training the autoencoder, the training objective is to ensure that the encoder can accurately extract the semantic features of the input data. In addition, it is also necessary to retain the spatial structure relationship between features while reducing the dimensionality of features, so as to ensure that the encoder in the autoencoder can fully consider the structural relationship between different features in the knowledge graph when extracting features.

[0079] Based on the trained autoencoder, a feature selection model can be constructed by combining an encoder in the trained first autoencoder with a weight configuration layer according to the transfer learning idea. Among them, the parameters of the weight configuration layer are used to assign trainable weights to each initial feature. The parameters of the weight configuration layer can be initialized by the Xavier normal distribution and ensured to be non - negative through the projected gradient descent method during the training process. The feature selection model is trained based on the above - mentioned feature selection samples to achieve precise filtering of redundant features. Specifically, based on the current parameters of the weight configuration layer of the feature selection model and any feature selection sample, the input feature sample can be determined. Here, the current parameters of the weight configuration layer can be multiplied point - by - point with the feature values of each initial feature in the feature selection sample to obtain the input feature sample. Subsequently, the input feature sample is encoded by the encoder of the feature selection model to obtain the feature encoding vector corresponding to the input feature sample. On the other hand, the feature selection sample is also encoded by the encoder of the feature selection model to obtain the feature encoding vector corresponding to the feature selection sample. By determining the L1 norm of the current parameters of the encoder of the feature selection model and the current parameters of the weight configuration layer, the parameters of the weight configuration layer and the encoder in the feature selection model are adjusted based on the difference between the feature encoding vector corresponding to the above - mentioned input feature sample and the feature encoding vector corresponding to the above - mentioned feature selection sample, as well as the L1 norm of the current parameters of the encoder and the current parameters of the weight configuration layer. After the feature selection model is trained, based on the parameters of its weight configuration layer, that is, the weights corresponding to each initial feature configured by the weight configuration layer, the initial features with non - zero weights are selected from the initial feature set as the efficiency prediction features.

[0080] Therefore, based on the selected efficiency prediction features, the efficiency of the current deep fertilization method for a specific cultivated land and a specific crop can be predicted. Among them, a fertilization efficiency prediction model can be constructed by combining the encoder of the trained feature selection model with a regression layer. Here, the input of the encoder of the feature selection model includes the feature values of the current deep fertilization method for a specific cultivated land and a specific crop corresponding to the above - mentioned efficiency prediction features, and the regression layer is used to perform regression calculations based on the feature encoding vector output by the encoder of the feature selection model to obtain the fertilization efficiency of the current deep fertilization method.

[0081] In summary, the method provided by the embodiments of the present invention dynamically extracts a domain knowledge subgraph related to fertilization efficiency through the semantic association characteristics of the agricultural knowledge graph, where nodes related to fertilization efficiency can be screened out through triple screening conditions; subsequently, feature screening is performed based on the characteristics of the domain knowledge subgraph. By constructing a feature selection sample formed by the feature values of each initial feature in the initial feature set corresponding to multiple deep fertilization methods, and combining the feature compression ability of the autoencoder for feature screening, redundant features are screened out, and efficiency prediction features highly related to fertilization efficiency are selected from the initial feature set, thereby constructing an intelligent prediction model for fertilization efficiency prediction, overcoming the problem of the feature dimension disaster caused by the difficulty in determining effective features for fertilization prediction and the problem of the decline in prediction accuracy caused by a large number of redundant features.

[0082] The deep fertilization efficiency prediction device provided by the present invention will be described below. The deep fertilization efficiency prediction device described below can be mutually corresponded and referred to the deep fertilization efficiency prediction method described above.

[0083] Based on any of the above embodiments, Figure 3 is a schematic structural diagram of the deep fertilization efficiency prediction device provided by the present invention. As Figure 3 shown, the device includes:

[0084] A knowledge extraction unit 310, configured to obtain a domain knowledge subgraph corresponding to fertilization efficiency based on the agricultural knowledge graph; each node in the agricultural knowledge graph corresponds to an agricultural concept in the agricultural text, and the edge attributes connecting each node include the agricultural text identifier to which the nodes connected by the corresponding edge belong;

[0085] An initial feature extraction unit 320, configured to determine an initial feature set for predicting fertilization efficiency based on the domain knowledge subgraph, and obtain the feature values of each initial feature in the initial feature set corresponding to multiple deep fertilization methods for different cultivated lands and different crops;

[0086] A feature selection and prediction unit 330, configured to select efficiency prediction features from the initial feature set based on the feature values of each initial feature in the initial feature set corresponding to the multiple deep fertilization methods, and perform efficiency prediction on the current deep fertilization method for a specific cultivated land and a specific crop based on the efficiency prediction features;

[0087] Among them, the obtaining of the domain knowledge subgraph corresponding to fertilization efficiency based on the agricultural knowledge graph includes:

[0088] Starting from the target node corresponding to the fertilization efficiency, traverse along the connecting edges. If the influence degree of the currently traversed node on the target node is higher than the first preset threshold, the number of hops between the currently traversed node and the target node is less than the second preset threshold, and the number of agricultural texts spanned by the currently traversed node and the target node is less than the third preset threshold, then add the currently traversed node to the domain knowledge subgraph and continue traversing along the connecting edges of the currently traversed node.

[0089] The device provided by the embodiments of the present invention dynamically extracts a domain knowledge subgraph related to fertilization efficiency through the semantic association characteristics of the agricultural knowledge graph, and nodes related to fertilization efficiency can be screened out through triple screening conditions; subsequently, feature screening is performed based on the characteristics of the domain knowledge subgraph. By constructing a feature selection sample formed by the feature values of each initial feature in the initial feature set corresponding to multiple deep fertilization methods, combined with the feature compression ability of the autoencoder, redundant features are screened out, and efficiency prediction features highly related to fertilization efficiency are selected from the initial feature set, thereby constructing an intelligent prediction model for fertilization efficiency prediction, overcoming the problem of feature dimension disaster caused by the difficulty in determining effective features for fertilization prediction and the problem of decreased prediction accuracy caused by a large number of redundant features.

[0090] Based on any of the above embodiments, the influence degree of the currently traversed node on the target node is calculated based on the following method:

[0091] Obtain the sequence of passing nodes included in the path between the currently traversed node and the target node;

[0092] Based on the co-occurrence times of any adjacent nodes in the sequence of passing nodes in each agricultural text and the total number of agricultural texts, determine the influence degree between the any adjacent nodes;

[0093] Based on the influence degrees between any adjacent nodes in the sequence of passing nodes, determine the influence degree of the currently traversed node on the target node.

[0094] Based on any of the above embodiments, the selection of efficiency prediction features from the initial feature set based on the feature values of each initial feature in the initial feature set corresponding to the multiple deep fertilization methods includes:

[0095] Based on the feature values of each initial feature in the initial feature set corresponding to multiple deep fertilization methods, construct multiple feature selection samples;

[0096] Train the encoder and decoder in the autoencoder based on the feature selection samples to obtain a trained first autoencoder;

[0097] Construct a feature selection model based on the encoder and weight configuration layer in the trained first autoencoder; wherein, the parameters of the weight configuration layer are used to adjust the weights corresponding to each initial feature;

[0098] Train the feature selection model based on the feature selection samples to obtain the weights corresponding to each initial feature configured by the weight configuration layer;

[0099] Based on the weights corresponding to each initial feature configured by the weight configuration layer, select the initial features with non-zero weights from the initial feature set as the efficiency prediction features.

[0100] Based on any of the above embodiments, training the encoder and decoder in the autoencoder based on the feature selection samples includes:

[0101] Input any feature selection sample into the autoencoder to obtain the feature encoding vector corresponding to the any feature selection sample output by the encoder in the autoencoder and the feature reconstruction result reconstructed by the decoder in the autoencoder based on the feature encoding vector;

[0102] Obtain the Laplacian matrix of the any feature selection sample and the Laplacian matrix of the feature reconstruction result;

[0103] Based on the similarity between the any feature selection sample and the feature reconstruction result and the similarity between the Laplacian matrix of the any feature selection sample and the Laplacian matrix of the feature reconstruction result, adjust the model parameters of the encoder and decoder in the autoencoder.

[0104] Based on any of the above embodiments, training the feature selection model based on the feature selection samples includes:

[0105] Determine the input feature sample based on the current parameters of the weight configuration layer of the feature selection model and any feature selection sample;

[0106] Encode the input feature sample based on the encoder of the feature selection model to obtain the feature encoding vector corresponding to the input feature sample;

[0107] Encode the any feature selection sample based on the encoder of the feature selection model to obtain the feature encoding vector corresponding to the any feature selection sample;

[0108] Determine the current parameters of the encoder of the feature selection model and the L1 norm of the current parameters of the weight configuration layer of the feature selection model;

[0109] Adjust the parameters of the weight configuration layer and the encoder in the feature selection model based on the difference between the feature encoding vector corresponding to the input feature sample and the feature encoding vector corresponding to any one of the feature selection samples, as well as the L1 norm of the current parameters of the encoder of the feature selection model and the current parameters of the weight configuration layer of the feature selection model.

[0110] Based on any of the above embodiments, the efficiency prediction of the current deep fertilization method for a specific cultivated land and a specific crop based on the efficiency prediction features includes:

[0111] Construct a fertilization efficiency prediction model based on the encoder and the regression layer of the feature selection model;

[0112] Among them, the input of the encoder of the feature selection model includes the feature values of the efficiency prediction features corresponding to the current deep fertilization method for a specific cultivated land and a specific crop, and the regression layer is used to perform regression calculations based on the feature encoding vector output by the encoder of the feature selection model to obtain the fertilization efficiency of the current deep fertilization method.

[0113] Figure 4 is a schematic structural diagram of the electronic device provided by the present invention, as Figure 4As shown in the figure, the electronic device may include: a processor 410, a memory 420, a communications interface 430, and a communication bus 440. Among them, the processor 410, the memory 420, and the communication interface 430 complete communication with each other through the communication bus 440. The processor 410 may call the logical instructions in the memory 420 to execute the deep fertilization efficiency prediction method, which includes: based on the agricultural knowledge graph, obtaining a domain knowledge subgraph corresponding to the fertilization efficiency; each node in the agricultural knowledge graph corresponds to an agricultural concept in the agricultural text, and the edge attributes connecting each node include the agricultural text identifier to which the nodes connected by the corresponding edge belong; based on the domain knowledge subgraph, determining an initial feature set for performing fertilization efficiency prediction, and obtaining the feature values of each initial feature in the initial feature set corresponding to various deep fertilization methods for different cultivated lands and different crops; selecting efficiency prediction features from the initial feature set based on the feature values of each initial feature in the initial feature set corresponding to the various deep fertilization methods, and predicting the efficiency of the current deep fertilization method for a specific cultivated land and a specific crop based on the efficiency prediction features; where, the obtaining the domain knowledge subgraph corresponding to the fertilization efficiency based on the agricultural knowledge graph includes: starting from the target node corresponding to the fertilization efficiency and traversing along the connecting edge, if the influence degree of the currently traversed node on the target node is higher than the first preset threshold, the number of hops between the currently traversed node and the target node is less than the second preset threshold, and the number of agricultural texts spanned by the currently traversed node and the target node is less than the third preset threshold, then add the currently traversed node to the domain knowledge subgraph, and continue to traverse along the connecting edge of the currently traversed node.

[0114] In addition, when the logical instructions in the above-mentioned memory 420 can be implemented in the form of software function units and sold or used as an independent product, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0115] On the other hand, the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the deep fertilization efficiency prediction method provided by each of the above methods. The method includes: obtaining a domain knowledge subgraph corresponding to fertilization efficiency based on an agricultural knowledge graph; each node in the agricultural knowledge graph corresponds to an agricultural concept in agricultural texts, and the edge attributes connecting each node include the agricultural text identifiers to which the nodes connected by the corresponding edges belong; determining an initial feature set for predicting fertilization efficiency based on the domain knowledge subgraph, and obtaining the feature values of each initial feature in the initial feature set corresponding to multiple deep fertilization methods for different cultivated lands and different crops; selecting efficiency prediction features from the initial feature set based on the feature values of each initial feature in the initial feature set corresponding to the multiple deep fertilization methods, and predicting the efficiency of the current deep fertilization method for a specific cultivated land and a specific crop based on the efficiency prediction features; wherein, the obtaining a domain knowledge subgraph corresponding to fertilization efficiency based on the agricultural knowledge graph includes: starting from the target node corresponding to fertilization efficiency and traversing along the connecting edges. If the influence degree of the currently traversed node on the target node is higher than a first preset threshold, the number of hops between the currently traversed node and the target node is less than a second preset threshold, and the number of agricultural texts spanned by the currently traversed node and the target node is less than a third preset threshold, then add the currently traversed node to the domain knowledge subgraph and continue to traverse along the connecting edges of the currently traversed node.

[0116] In another aspect, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it is used to implement the deep fertilization efficiency prediction method provided above. The method includes: based on an agricultural knowledge graph, obtaining a domain knowledge sub-graph corresponding to fertilization efficiency; each node in the agricultural knowledge graph corresponds to an agricultural concept in agricultural texts, and the edge attributes connecting each node include the agricultural text identifiers to which the nodes connected by the corresponding edges belong; based on the domain knowledge sub-graph, determining an initial feature set for predicting fertilization efficiency, and obtaining the feature values of each initial feature in the initial feature set corresponding to various deep fertilization methods for different cultivated lands and different crops; based on the feature values of each initial feature in the initial feature set corresponding to the various deep fertilization methods, selecting efficiency prediction features from the initial feature set, and predicting the efficiency of the current deep fertilization method for a specific cultivated land and a specific crop based on the efficiency prediction features; wherein, the step of obtaining a domain knowledge sub-graph corresponding to fertilization efficiency based on the agricultural knowledge graph includes: starting from a target node corresponding to fertilization efficiency and traversing along the connecting edges. If the influence degree of the currently traversed node on the target node is higher than a first preset threshold, the number of hops between the currently traversed node and the target node is less than a second preset threshold, and the number of agricultural texts spanned by the currently traversed node and the target node is less than a third preset threshold, then add the currently traversed node to the domain knowledge sub-graph and continue to traverse along the connecting edges of the currently traversed node.

[0117] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0118] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, also by hardware. Based on such an understanding, the above technical solutions, in essence, or the part that contributes to the prior art can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0119] 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 foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the deep fertilization efficiency, characterized in that, Including: Based on an agricultural knowledge graph, obtain a domain knowledge subgraph corresponding to fertilization efficiency; each node in the agricultural knowledge graph corresponds to an agricultural concept in agricultural texts, and the edge attributes connecting each node include the agricultural text identifiers to which the nodes connected by the corresponding edges belong. Based on the domain knowledge subgraph, determine an initial feature set for predicting fertilization efficiency, and obtain the feature values of various deep fertilization methods for different cultivated lands and different crops corresponding to each initial feature in the initial feature set. Based on the feature values of various deep fertilization methods corresponding to each initial feature in the initial feature set, select efficiency prediction features from the initial feature set, and based on the efficiency prediction features, predict the efficiency of the current deep fertilization method for a specific cultivated land and a specific crop. Among them, the obtaining of the domain knowledge subgraph corresponding to fertilization efficiency based on the agricultural knowledge graph includes: Starting from the target node corresponding to fertilization efficiency, traverse along the connecting edges. If the influence degree of the currently traversed node on the target node is higher than a first preset threshold, the number of hops between the currently traversed node and the target node is less than a second preset threshold, and the number of agricultural texts spanned by the currently traversed node and the target node is less than a third preset threshold, then add the currently traversed node to the domain knowledge subgraph and continue to traverse along the connecting edges of the currently traversed node.

2. The deep fertilization efficiency prediction method according to claim 1, wherein The influence degree of the currently traversed node on the target node is calculated based on the following method: Obtain the sequence of passing nodes included in the path between the currently traversed node and the target node. Based on the co-occurrence times of any adjacent nodes in the sequence of passing nodes in each agricultural text and the total number of agricultural texts, determine the influence degree between any adjacent nodes. Based on the influence degrees between any adjacent nodes in the sequence of passing nodes, determine the influence degree of the currently traversed node on the target node.

3. The deep fertilization efficiency prediction method according to claim 1, characterized in that The selecting of the efficiency prediction features from the initial feature set based on the feature values of various deep fertilization methods corresponding to each initial feature in the initial feature set includes: Based on the feature values of various deep fertilization methods corresponding to each initial feature in the initial feature set, construct multiple feature selection samples. Based on the feature selection samples, train the encoder and decoder in the autoencoder to obtain a trained first autoencoder. Based on the encoder and the weight configuration layer in the trained first autoencoder, construct a feature selection model; among them, the parameters of the weight configuration layer are used to adjust the weights corresponding to each initial feature. Based on the feature selection samples, train the feature selection model to obtain the weights corresponding to each initial feature configured by the weight configuration layer. Based on the weights corresponding to each initial feature configured by the weight configuration layer, select the initial features with non-zero weights from the initial feature set as the efficiency prediction features.

4. The deep fertilization efficiency prediction method according to claim 3, wherein The training of the encoder and decoder in the autoencoder based on the feature selection samples includes: Input any feature selection sample into the autoencoder, to obtain the feature encoding vector corresponding to the any feature selection sample output by the encoder in the autoencoder, and the feature reconstruction result reconstructed by the decoder in the autoencoder based on the feature encoding vector; Obtain the Laplacian matrix of the any feature selection sample and the Laplacian matrix of the feature reconstruction result; Based on the similarity between the any feature selection sample and the feature reconstruction result, and the similarity between the Laplacian matrix of the any feature selection sample and the Laplacian matrix of the feature reconstruction result, adjust the model parameters of the encoder and decoder in the autoencoder.

5. The deep fertilization efficiency prediction method according to claim 3, characterized in that The training of the feature selection model based on the feature selection samples includes: Based on the current parameters of the weight configuration layer of the feature selection model and any feature selection sample, determine the input feature sample; Based on the encoder of the feature selection model, encode the input feature sample to obtain the feature encoding vector corresponding to the input feature sample; Based on the encoder of the feature selection model, encode the any feature selection sample to obtain the feature encoding vector corresponding to the any feature selection sample; Determine the current parameters of the encoder of the feature selection model and the L1 norm of the current parameters of the weight configuration layer of the feature selection model; Based on the difference between the feature encoding vector corresponding to the input feature sample and the feature encoding vector corresponding to the any feature selection sample, and the L1 norm of the current parameters of the encoder of the feature selection model and the current parameters of the weight configuration layer of the feature selection model, adjust the parameters of the weight configuration layer and the encoder in the feature selection model.

6. The deep fertilization efficiency prediction method according to claim 3, wherein The efficiency prediction of the current deep fertilization method for a specific cultivated land and a specific crop based on the efficiency prediction features includes: Construct a fertilization efficiency prediction model based on the encoder and regression layer of the feature selection model; Wherein, the input of the encoder of the feature selection model includes the feature values corresponding to the efficiency prediction features of the current deep fertilization method for a specific cultivated land and a specific crop, and the regression layer is used to perform regression calculation based on the feature encoding vector output by the encoder of the feature selection model to obtain the fertilization efficiency of the current deep fertilization method.

7. A device for predicting deep fertilization efficiency, characterized in that It includes: A knowledge extraction unit, configured to obtain a domain knowledge subgraph corresponding to the fertilization efficiency based on the agricultural knowledge graph; each node in the agricultural knowledge graph corresponds to an agricultural concept in the agricultural text, and the edge attributes connecting each node include the agricultural text identifiers to which the nodes connected by the corresponding edges belong; An initial feature extraction unit, configured to determine an initial feature set for performing fertilization efficiency prediction based on the domain knowledge subgraph, and obtain the feature values of each initial feature in the initial feature set corresponding to multiple deep fertilization methods for different cultivated lands and different crops; A feature selection and prediction unit, configured to select efficiency prediction features from the initial feature set based on the feature values of each initial feature in the initial feature set corresponding to the multiple deep fertilization methods, and predict the efficiency of the current deep fertilization method for a specific cultivated land and a specific crop based on the efficiency prediction features; Among them, the obtaining of the domain knowledge subgraph corresponding to the fertilization efficiency based on the agricultural knowledge graph includes: Traverse along the connection edges starting from the target node corresponding to the fertilization efficiency. If the influence degree of the currently traversed node on the target node is higher than the first preset threshold, the number of hops between the currently traversed node and the target node is less than the second preset threshold, and the number of agricultural texts spanned by the currently traversed node and the target node is less than the third preset threshold, then add the currently traversed node to the domain knowledge subgraph and continue to traverse along the connection edge of the currently traversed node.

8. The deep fertilization efficiency prediction device according to claim 7, characterized in that, The influence degree of the currently traversed node on the target node is calculated based on the following method: Obtain the sequence of passing nodes included in the path between the currently traversed node and the target node; Based on the co-occurrence times of any adjacent nodes in the sequence of passing nodes in each agricultural text and the total number of agricultural texts, determine the influence degree between any adjacent nodes; Based on the influence degrees between any adjacent nodes in the sequence of passing nodes, determine the influence degree of the currently traversed node on the target node.

9. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the deep fertilization efficiency prediction method according to any one of claims 1 to 6.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the deep fertilization efficiency prediction method according to any one of claims 1 to 6.

Images