Quantitative fertilization method for low-nitrogen-resistant corn planting

Abstract

The application discloses a kind of quantitative fertilization methods of low nitrogen resistant corn planting, belong to agricultural big data and intelligent fertilization technical field, the method includes: the field ontology including low nitrogen stress response characteristic is constructed;Setting dynamic weighted entity alignment algorithm of low nitrogen resistant genotype perception, according to entity type self-adaptive adjustment similarity weight;Then propose TransH-Agri agricultural embedding model, introduce growth period timing hyperplane mechanism;Then construct spatio-temporal-diagnosis-genotype three module cross attention collaborative reasoning network, realize feature level deep fusion;Finally, based on low nitrogen stress response characteristic map, genotype is exclusively corrected.The application adopts the quantitative fertilization method of low nitrogen resistant corn planting described above, realizes the precision quantitative decision of low nitrogen resistant corn fertilization by synergistic effect, improves fertilization efficiency.

Description

Technical Field

[0001] This invention belongs to the field of agricultural big data and intelligent fertilization technology, specifically relating to a quantitative fertilization method for planting low-nitrogen-tolerant corn. Background Technology

[0002] Maize is a food crop, and nitrogen is a key nutrient element affecting maize yield. A significant proportion of maize-growing areas have low-nitrogen or medium-low-nitrogen soils, and low-nitrogen stress has become a major limiting factor restricting maize yield improvement. Existing maize fertilization technologies have three major shortcomings: First, fertilization decisions lack precise quantitative basis and have poor temporal and spatial matching. Existing fertilization technologies are mostly based on regional average recommendations or farmers' experience, failing to dynamically adjust according to soil nitrogen supply capacity, variety characteristics, and real-time plant status, resulting in generally low nitrogen fertilizer utilization efficiency.

[0003] Second, the characteristics of low-nitrogen tolerant varieties are disconnected from fertilization techniques, and the genotypic advantages have not been fully utilized. Breeders have bred a number of low-nitrogen tolerant maize varieties, but the existing fertilization techniques still follow the fertilization schemes for ordinary varieties, failing to develop differentiated strategies based on the nitrogen absorption patterns of low-nitrogen tolerant varieties, resulting in "good varieties without good methods".

[0004] Third, knowledge graph fertilization technology lacks agriculture-specific improvements. Existing knowledge graph fertilization methods use general entity alignment and embedding algorithms, without being specifically optimized for the strong temporal characteristics and genotypic features of agriculture. Multi-engine inference still uses a simple parallel weighted mode, leaving room for improvement.

[0005] Therefore, there is an urgent need for a knowledge graph-based precision fertilization technology specifically designed for low-nitrogen-tolerant corn. Summary of the Invention

[0006] The purpose of this invention is to provide a quantitative fertilization method for low-nitrogen-tolerant maize cultivation. This method uses confidence transmission of low-nitrogen-tolerant genotype features and a strong rejection mechanism to ensure the alignment quality of core features. It sets up independent hyperplanes for the five growth stages of maize to achieve agricultural-specific embedding, achieves feature-level deep fusion through three-mode cross-attention, and realizes quantitative mapping of nitrogen-saving amount based on five-dimensional genotype features, thus systematically solving the problem of homogenization in fertilization of low-nitrogen-tolerant varieties.

[0007] To achieve the above objectives, this invention provides a quantitative fertilization method for low-nitrogen-tolerant maize cultivation, comprising the following steps: S1. Collect knowledge and technical specifications from experts in the agricultural field, construct a two-level ontology classification system that includes low nitrogen tolerance attributes, and output the two-level ontology classification system to S2 and S3 simultaneously. S2 receives the two-level ontology classification system output by S1, performs semantic mapping on the structured database, generates standardized structured knowledge triples, and outputs the structured knowledge triples to S4. S3 receives the two-level ontology classification system output by S1, performs entity recognition and relation extraction on unstructured text, generates standardized unstructured knowledge triples, and outputs the unstructured knowledge triples to S4. S4 receives the structured knowledge triples output from S2 and the unstructured knowledge triples output from S3, executes the entity alignment mechanism prioritizing low-nitrogen genotype features, calculates the similarity of nine high-priority genotype attributes, uses a dual mechanism of confidence transmission and strong rejection to carry out entity matching, and outputs a high-quality fused knowledge set that eliminates redundant conflicts to S5. S5 receives the high-quality fusion knowledge set output by S4, performs ontology reasoning and implicit knowledge completion, selects and extracts five-dimensional genotype features from the core attributes of 9 high-priority genotypes, constructs a low-nitrogen tolerance knowledge graph containing a low-nitrogen stress response feature graph subgraph, and outputs the low-nitrogen tolerance knowledge graph to S6 and S8 simultaneously. S6 receives the low-nitrogen tolerance knowledge graph output by S5, performs embedding learning of the independent semantic hyperplane for the reproductive period, sets independent hyperplanes for the five physiological stages of seedling stage, jointing stage, large trumpet stage, silking stage and grain filling stage, generates time-aware vector representations, and outputs the time-aware vector representations to S7. S7 receives the temporal-aware vector representation output by S6, performs spatiotemporal-diagnosis-genotype three-mode cross-attention collaborative reasoning in the same vector space, realizes feature-level deep fusion of the three dimensions, generates fusion reasoning results, and outputs the fusion reasoning results to S8; S8 receives the fusion reasoning results output by S7 and the low-nitrogen tolerance knowledge graph output by S5, calls the low-nitrogen stress response feature graph subgraph to query the five-dimensional genotype feature vector corresponding to the variety, performs quantitative mapping correction of nitrogen saving on the basic nitrogen application amount, allocates fertilizer amount according to each growth stage of maize, and outputs a fertilization plan including fertilization time, fertilizer type, and application amount.

[0008] Preferably, in S1, the 67 attributes are divided into a two-tier architecture of high priority and regular priority; The two-tier attribute system satisfies: ; In the formula, It is a set of core attributes of high-priority genotypes, including nine attributes: low nitrogen tolerance level, nitrate reductase activity, glutamine synthase activity, root nitrogen translocation efficiency, chlorophyll degradation rate, nitrogen redistribution efficiency, root type, late-stage aging resistance, and grain filling rate. For the remaining 58 common attributes; Establish a complete ontology architecture ,in, It is a collection of 7 major categories of entity concepts. It is a set of 23 semantic relations.

[0009] Preferably, in S2, the two-level ontology classification system carrying ontology architecture information output by S1 is received, and the database fields are mapped one by one to knowledge graph triples according to the entity type, relation type, and attribute list defined by the classification system. The generated set of structured knowledge triples is as follows: ; In the formula, For the head entity, For tail entities, Relationships between entities; Database primary keys are mapped to entity identifiers, database foreign keys are mapped to relationships between entities, and field values ​​are mapped to entity attribute values; the categories of triples are all limited to the scope of the two-level ontology classification system definition.

[0010] Preferably, in S3, the two-level ontology classification system carrying ontology architecture information output by S1 is received, and labels are set according to the 7 types of entities and 23 types of relations defined by the classification system. The BERT-BiLSTM-CRF model is used for sequence labeling. The loss function is: ; In the formula, For the sample size, For the first The entity label sequence of a sample, Given a sequence of text as input, For model parameters, is the L2 regularization coefficient.

[0011] Preferably, in S4, the structured knowledge triples output from S2 and the unstructured knowledge triples output from S3 are received, the similarity of nine high-priority genotype attributes is calculated, and a dual mechanism of confidence transmission and strong rejection is executed, including the following sub-steps: S401. Calculate the similarity of 9 high-priority genotype attributes: ; In the formula, For entities With entity In attributes Similarity on; S402, Execution of genotype confidence transmission mechanism: ; In the formula, For the overall similarity of entity pairs, Similarity for 58 common attributes; The threshold for core attribute similarity; These are the weighting coefficients; When the similarity of the 9 core attributes is greater than or equal to At the same time, increase the overall confidence level of the entity pair; S403, Implement a strong veto mechanism for genotype inconsistency: ; In the formula, For entities Its low-nitrogen tolerance rating; If two entities have different low-nitrogen tolerance levels, they are considered mismatched; if the similarity is greater than or equal to the entity merging threshold... Entity merging is performed at that time.

[0012] Preferably, in S5, the high-quality fused knowledge set output by S4 is received, ontology reasoning and implicit knowledge completion are performed based on the high-quality fused knowledge set, genotype features are extracted, and a low nitrogen stress response feature map subgraph is constructed. The spectrum of the low nitrogen stress response characteristics is as follows: ; ; In the formula, A collection of maize varieties. For varieties The five-dimensional genotype feature vector, For low-nitrogen tolerance, the value is [1, 5]. It is a root system type. For nitrogen use efficiency, For later anti-aging properties, For grouting rate; The final knowledge graph is represented as follows: ; In the formula, For a set of entity nodes, Let be the set of relation edges. For a collection of attributes, This is a subplot of the characteristic spectrum of the response to low nitrogen stress.

[0013] Preferably, in S6, the low-nitrogen tolerance knowledge graph output from S5 is received, and an improved TransH-Agri embedding model is used to carry out embedding learning. The improved TransH-Agri embedding model is as follows: ; ; ; In the formula, During the corn growing season, For the head entity vector, For the tail entity vector, For the first The relation vector corresponding to each reproductive period For the first The unit normal vector of the hyperplane corresponding to each reproductive period.

[0014] Preferably, in S7, the temporal-aware vector representation output by S6 is received, and the spatiotemporal matching features, diagnostic linkage features, and genotype adaptation features are all mapped to the same vector space. A multi-head cross-attention mechanism is used to achieve feature-level deep fusion of the three dimensions. The multi-head cross-attention mechanism is as follows: ; ; ; Add residual connections and layer normalization: ; In the formula, Features of the spatiotemporal matching engine To diagnose the characteristics of the linkage engine, For genotype-adapted engine features, The dimension of the key vector. This is a learnable weight matrix.

[0015] Preferably, in S8, the fusion reasoning result output by S7 and the low nitrogen tolerance knowledge graph output by S5 are received, and the genotype feature vector of the variety is queried from the low nitrogen stress response feature graph according to the variety identifier, and the basic nitrogen application rate is specifically corrected. The final nitrogen application rate is calculated using the following formula: ; In the formula, Base nitrogen application rate For low-nitrogen tolerance and nitrogen-saving coefficient; This is the soil nitrogen correction factor.

[0016] This invention also provides an intelligent fertilization decision-making system for quantitative supplemental fertilization of low-nitrogen tolerant maize, comprising: The ontology classification construction module collects expert knowledge and technical specifications in the agricultural field, constructs a two-level ontology classification system that includes low nitrogen tolerance attributes, and outputs the two-level ontology classification system to the knowledge triple extraction module. The knowledge triple extraction module is connected to the ontology classification construction module; it receives the two-level ontology classification system output by the ontology classification construction module, and obtains structured knowledge triples by semantic mapping of the structured database; it uses the BERT-BiLSTM-CRF model to perform entity recognition and relation extraction on unstructured text to obtain unstructured knowledge triples, and outputs the structured knowledge triples and unstructured knowledge triples to the genotype-priority entity alignment and fusion module. The genotype-priority entity alignment and fusion module is connected to the knowledge triple extraction module; it receives structured knowledge triples and unstructured knowledge triples output by the knowledge triple extraction module, calculates the similarity of nine high-priority genotype attributes, completes entity matching through a dual mechanism of confidence transmission and strong rejection, and outputs a high-quality fused knowledge set that eliminates redundant conflicts to the low-nitrogen stress knowledge graph construction module. The low nitrogen stress knowledge graph construction module is connected to the genotype priority entity alignment and fusion module; it receives the high-quality fused knowledge set output by the genotype priority entity alignment and fusion module, performs ontology reasoning and implicit knowledge completion, extracts five-dimensional genotype features, constructs a low nitrogen tolerance knowledge graph containing a low nitrogen stress response feature graph subgraph, and outputs the low nitrogen tolerance knowledge graph to the reproductive period time-series embedding module and the fertilization quantitative correction output terminal, respectively. The reproductive period temporal embedding module is connected to the low nitrogen stress knowledge graph construction module; it receives the low nitrogen tolerance knowledge graph output by the low nitrogen stress knowledge graph construction module, and sets independent semantic hyperplanes for the five reproductive periods (seedling stage, jointing stage, large trumpet stage, silking stage, and grain filling stage) based on the TransH-Agri agricultural embedding model, completes embedding learning, generates temporal awareness vectors, and outputs the temporal awareness vectors to the cross-attention collaborative reasoning module; The cross-attention collaborative reasoning module is connected to the reproductive period time-series embedding module and the low nitrogen stress knowledge graph construction module, respectively. It receives the time-series perception vector output by the reproductive period time-series embedding module, and has built-in spatiotemporal matching engine, diagnostic linkage engine and genotype adaptation engine. It maps the features to the same vector space, and uses a multi-head cross-attention mechanism to achieve deep fusion of spatiotemporal-diagnosis-genotype three-mode features, generates fusion reasoning results, and outputs the fusion reasoning results to the fertilizer quantitative correction output terminal. The fertilization quantitative correction output terminal is connected to the cross-attention collaborative reasoning module and the low nitrogen stress knowledge graph construction module, respectively; it receives the fusion reasoning results output by the cross-attention collaborative reasoning module and the low nitrogen stress knowledge graph construction module; it calls the low nitrogen stress response feature graph subgraph to query the five-dimensional genotype feature vector of the variety, performs nitrogen-saving quantitative mapping correction on the basic nitrogen application amount, allocates the fertilizer application amount according to each growth stage of maize, and outputs a fertilization plan including fertilization time, fertilizer type, and application amount.

[0017] Therefore, the present invention employs the above-mentioned quantitative fertilization method for low-nitrogen-tolerant corn cultivation. Compared with the prior art, the technical solution of the present invention has the following beneficial effects: 1. A dynamic weighted entity alignment technique based on low-nitrogen tolerance genotype perception is adopted to overcome the technical problems of existing entity alignment techniques that use globally fixed weights and do not consider the special characteristics of maize genotypes; the constraint effect of low-nitrogen tolerance genotype attributes in the entity fusion process is strengthened to ensure the matching accuracy of genotype characteristics of maize genotype entities.

[0018] 2. By adopting the TransH-Agri agriculture-specific improved embedding model technology, we can overcome the technical problems that traditional embedding models cannot represent the strong temporal characteristics of agricultural knowledge and cannot reflect the semantic differences of the same entity in different reproductive stages. We can realize the knowledge vectorization representation with reproductive stage time perception and distinguish the semantic expression of the same entity in different reproductive stages.

[0019] 3. It adopts a three-engine cross-attention collaborative reasoning technology to overcome the technical problem of the lack of deep interaction at the feature level caused by the simple parallel reasoning of multiple engines and the weighted fusion in the later stage in the existing technology; it achieves deep feature-level fusion of spatiotemporal, diagnostic and genotypic dimensions in the embedding layer, which is different from the fusion method of simple result weighting.

[0020] 4. By adopting a dedicated correction technique for low nitrogen stress response characteristic maps, we can overcome the technical problems of homogenization of fertilization formulas between low nitrogen tolerant varieties and ordinary varieties, and the disconnect between genotype and cultivation techniques. We can establish a quantitative correlation between low nitrogen tolerance level and nitrogen saving amount to adapt to the fertilization needs of different low nitrogen tolerant maize varieties.

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating an embodiment of a quantitative fertilization method for low-nitrogen-tolerant corn cultivation according to the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used in the present invention should have the ordinary meaning understood by those skilled in the art.

[0024] Example 1 like Figure 1 As shown, this embodiment provides a quantitative fertilization method for planting low-nitrogen-tolerant corn. It should be understood that the specific parameters, models and protocols mentioned in this embodiment are merely examples to help those skilled in the art understand the present invention, and are not intended to limit the present invention.

[0025] The present invention provides a quantitative fertilization method for low-nitrogen-tolerant maize cultivation, comprising the following steps: S1. Collect knowledge and technical specifications from experts in the agricultural field to construct a two-level ontology classification system that includes the exclusive attribute of low nitrogen tolerance; output the classification system to S2 and S3.

[0026] In this step, the 67 attributes are divided into a two-tier architecture of high priority and regular priority; The two-tier attribute system satisfies: ; In the formula, It is a set of core attributes of high-priority genotypes, including nine attributes: low nitrogen tolerance level, nitrate reductase activity, glutamine synthase activity, root nitrogen translocation efficiency, chlorophyll degradation rate, nitrogen redistribution efficiency, root type, late-stage aging resistance, and grain filling rate. For the remaining 58 common attributes; Establish a complete ontology architecture ,in, It is a collection of 7 major categories of entity concepts. It is a set of 23 semantic relations.

[0027] S2. Based on the classification system of S1, semantic mapping is performed on three typical agricultural databases: maize variety database, soil nutrient database, and field meteorological monitoring database, to generate standardized structured knowledge triples; the structured knowledge triples are output to S4.

[0028] In this step, database fields are mapped one by one to knowledge graph triples according to the entity type, relation type, and attribute list defined in S1; The generated set of structured knowledge triples is as follows: ; In the formula, For the head entity, For tail entities, For relationships or attributes between entities; database primary keys are mapped to entity identifiers, database foreign keys are mapped to relationships between entities, and field values ​​are mapped to entity attribute values; the categories of all output triples are strictly limited to the classification system defined in S1.

[0029] S3. Based on the classification system of S1, entity recognition and relation extraction are performed on unstructured texts such as agricultural documents and agricultural technology manuals to generate standardized unstructured knowledge triples; the unstructured knowledge triples are output to S4.

[0030] In this step, labels are set according to the 7 types of entities and 23 types of relationships defined in S1, and the BERT-BiLSTM-CRF model is used for sequence labeling; The loss function is: ; In the formula, For the sample size, For the first The entity label sequence of a sample, Given a sequence of text as input, For model parameters, is the L2 regularization coefficient; all extracted entity and relation categories are strictly limited to the classification system defined in S1.

[0031] S4 receives the structured knowledge triplet from S2 and the unstructured knowledge triplet from S3, executes an entity alignment mechanism prioritizing low-nitrogen-tolerant genotype features, prioritizes calculating the similarity of nine high-priority genotype attributes, uses a dual mechanism of confidence propagation and strong rejection for entity matching, and outputs a high-quality fused knowledge set that eliminates redundant conflicts; the fused knowledge set is then output to S5.

[0032] In this step, the similarity of nine high-priority genotype attributes is calculated, and a dual mechanism of confidence transmission and strong rejection is implemented, including the following sub-steps: S401. Calculate the similarity of 9 high-priority genotype attributes: ; In the formula, For entities With entity In attributes Similarity on; S402, Execution of genotype confidence transmission mechanism: ; In the formula, For the overall similarity of entity pairs, Similarity for 58 common attributes; The threshold for core attribute similarity; These are the weighting coefficients; When the similarity of the 9 core attributes is greater than or equal to At the same time, increase the overall confidence level of the entity pair; S403, Implement a strong veto mechanism for genotype inconsistency: ; In the formula, For entities The low nitrogen tolerance level attribute; when the low nitrogen tolerance levels of two entities are inconsistent, regardless of how high the similarity of other attributes is, they are directly judged as mismatched; entities with a similarity exceeding the threshold of 0.85 are merged.

[0033] S5 receives the high-quality fusion knowledge set from S4, performs ontology reasoning and implicit knowledge completion, specifically extracts the five-dimensional genotype features of maize varieties, and constructs a knowledge graph for low-nitrogen-tolerant maize that includes a subgraph of low-nitrogen stress response feature map; the knowledge graph is then output to S6 and S8.

[0034] In this step, ontology reasoning and implicit knowledge completion are performed based on the high-quality knowledge fused by S4. Five-dimensional core genotypic features of all maize varieties are extracted to construct a sub-graph of low nitrogen stress response feature map. The spectrum of the low nitrogen stress response characteristics is as follows: ; ; In the formula, A collection of maize varieties. For varieties The five-dimensional genotypic feature vectors are: low nitrogen tolerance level (LNT), root type (RT), nitrogen use efficiency (NUE), late-stage aging resistance (SS), and grain filling rate (GFR). The final knowledge graph is represented as follows: ; In the formula, For a set of entity nodes, Let be the set of relation edges. For a collection of attributes, This is a subplot of the low nitrogen stress response characteristic map, containing low nitrogen stress response parameters for 200+ maize varieties.

[0035] S6 receives the knowledge graph from S5 and performs agricultural-specific embedding learning on independent semantic hyperplanes for the five major growth stages of maize. Independent hyperplanes are set for the five physiological stages of seedling, jointing, large trumpet stage, silking stage, and grain filling stage, respectively, and time-aware low-dimensional dense vector representations are generated. The vector representations are output to S7.

[0036] In this step, for the five physiological stages clearly defined in maize agronomy, an independent relational semantic hyperplane is set for each growth stage, and the traditional TransH model is improved specifically for agriculture. The improved TransH-Agri embedding model is as follows: ; ; ; In the formula, During the corn growing season, For the head entity vector, For the tail entity vector, For the first The relation vector corresponding to each reproductive period For the first The relational hyperplane unit normal vectors corresponding to each reproductive period are used for training, and the embedding dimension is set to 128.

[0037] S7 receives the temporal-aware vector representation from S6, performs spatiotemporal-diagnosis-genotype three-mode cross-attention collaborative reasoning in the same vector space, realizes feature-level deep fusion of the three dimensions, and generates fusion reasoning results; the fusion reasoning results are output to S8.

[0038] In this step, the spatiotemporal matching features, diagnostic linkage features, and genotype adaptation features are all mapped to the same vector space output by S6, and an 8-head cross-attention mechanism is used to achieve feature-level deep fusion of the three dimensions. The multi-head cross-attention mechanism is as follows: ; ; ; Add residual connections and layer normalization: ; In the formula, Features of the spatiotemporal matching engine To diagnose the characteristics of the linkage engine, For genotype-adapted engine features, The dimension of the key vector. It is a learnable weight matrix; the spatiotemporal features provide environmental constraints for the genotype features, the genotype features provide a variety baseline for the diagnostic features, and the diagnostic features provide real-time corrections for the spatiotemporal features.

[0039] S8 receives the fusion inference results from S7, calls the low nitrogen stress response feature map constructed by S5, performs quantitative mapping and correction of nitrogen-saving amount on the basic nitrogen application rate based on the five-dimensional genotype features, allocates fertilizer amount according to the growth period, and outputs the final precise fertilization plan.

[0040] In this step, the five-dimensional genotype feature vector of the maize variety is queried from the low nitrogen stress response feature map constructed by S5 based on the maize variety identifier, and the basic nitrogen application rate is specifically corrected. The final nitrogen application rate is calculated using the following formula: ; In the formula, This is the base nitrogen application rate output by S7. This represents the variety's tolerance to low nitrogen levels, with a value range of 1-5. The higher the value, the stronger the variety's tolerance to low nitrogen. For low-nitrogen tolerance and nitrogen-saving coefficient; This is the soil nitrogen correction factor; the higher the soil nitrogen content, the better. The smaller the value, the lower the soil nitrogen content. The larger the value, the more it characterizes the regulatory effect of the plot's background nitrogen supply capacity on the amount of nitrogen applied.

[0041] This embodiment provides an exemplary configuration: for each increase of the low-nitrogen tolerance level, the nitrogen application rate can be reduced by a certain percentage; exemplary nitrogen-saving coefficient. A value of 0.04-0.06 can be selected. This coefficient is only an example parameter for this embodiment. Those skilled in the art can adapt and adjust it in combination with variety breeding data and historical field data of the plot. It does not constitute a necessary limitation of this invention.

[0042] The final output includes a complete fertilization plan that specifies the timing, type, and amount of fertilizer applied during each growth stage.

[0043] This embodiment includes a calculation rule for nitrogen nutrient allocation during the five growth stages of maize, based on the corrected total nitrogen application rate. The allocation is performed in segments based on the base number, using the following formula: ; In the formula, For the first Nitrogen application rate during each growth period This refers to the nitrogen allocation ratio during the corresponding reproductive period; This embodiment provides an example of the nitrogen allocation ratio at each growth stage: seedling stage allocation ratio. Jointing period Big trumpet mouth period Silk-spinning period Grouting period For low-nitrogen-stressed plots, the distribution ratio of each stage can be fine-tuned according to the soil nitrogen level.

[0044] In S402, S403, and S6, the similarity threshold of 0.90, the entity merging threshold of 0.85, and the embedding dimension of 128 are exemplary configuration parameters in this embodiment. In actual engineering applications, those skilled in the art can select appropriate similarity thresholds and model embedding dimensions according to the dataset size and entity matching accuracy requirements.

[0045] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0046] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for quantitative fertilization in low-nitrogen-tolerant maize cultivation, characterized in that, Includes the following steps: S1. Collect knowledge and technical specifications from experts in the agricultural field, construct a two-level ontology classification system that includes low nitrogen tolerance attributes, and output the two-level ontology classification system to S2 and S3 simultaneously. S2 receives the two-level ontology classification system output by S1, performs semantic mapping on the structured database, generates standardized structured knowledge triples, and outputs the structured knowledge triples to S4. S3 receives the two-level ontology classification system output by S1, performs entity recognition and relation extraction on unstructured text, generates standardized unstructured knowledge triples, and outputs the unstructured knowledge triples to S4. S4 receives the structured knowledge triples output from S2 and the unstructured knowledge triples output from S3, executes the entity alignment mechanism prioritizing low-nitrogen genotype features, calculates the similarity of nine high-priority genotype attributes, uses a dual mechanism of confidence transmission and strong rejection to carry out entity matching, and outputs a high-quality fused knowledge set that eliminates redundant conflicts to S5. S5 receives the high-quality fusion knowledge set output by S4, performs ontology reasoning and implicit knowledge completion, selects and extracts five-dimensional genotype features from the core attributes of 9 high-priority genotypes, constructs a low-nitrogen tolerance knowledge graph containing a low-nitrogen stress response feature graph subgraph, and outputs the low-nitrogen tolerance knowledge graph to S6 and S8 simultaneously. S6 receives the low-nitrogen tolerance knowledge graph output by S5, performs embedding learning of the independent semantic hyperplane for the reproductive period, sets independent hyperplanes for the five physiological stages of seedling stage, jointing stage, large trumpet stage, silking stage and grain filling stage, generates time-aware vector representations, and outputs the time-aware vector representations to S7. S7 receives the temporal-aware vector representation output by S6, performs spatiotemporal-diagnosis-genotype three-mode cross-attention collaborative reasoning in the same vector space, realizes feature-level deep fusion of the three dimensions, generates fusion reasoning results, and outputs the fusion reasoning results to S8; S8 receives the fusion reasoning results output by S7 and the low-nitrogen tolerance knowledge graph output by S5, calls the low-nitrogen stress response feature graph subgraph to query the five-dimensional genotype feature vector corresponding to the variety, performs quantitative mapping correction of nitrogen saving on the basic nitrogen application amount, allocates fertilizer amount according to each growth stage of maize, and outputs a fertilization plan including fertilization time, fertilizer type, and application amount.

2. The quantitative fertilization method for low-nitrogen-tolerant maize cultivation according to claim 1, characterized in that, In S1, the 67 attributes are divided into a two-tier architecture of high priority and regular priority; The two-tier attribute system satisfies: ； In the formula, It is a set of core attributes of high-priority genotypes, including nine attributes: low nitrogen tolerance level, nitrate reductase activity, glutamine synthase activity, root nitrogen translocation efficiency, chlorophyll degradation rate, nitrogen redistribution efficiency, root type, late-stage aging resistance, and grain filling rate. For the remaining 58 common attributes; Establish a complete ontology architecture ,in, It is a collection of 7 major categories of entity concepts. It is a set of 23 semantic relations.

3. The quantitative fertilization method for low-nitrogen-tolerant maize cultivation according to claim 2, characterized in that, In S2, the two-level ontology classification system carrying ontology architecture information is received from the output of S1. According to the entity type, relation type, and attribute list defined in the classification system, the database fields are mapped one by one to knowledge graph triples. The generated set of structured knowledge triples is as follows: ； In the formula, For the head entity, For tail entities, Relationships between entities; Database primary keys are mapped to entity identifiers, database foreign keys are mapped to relationships between entities, and field values ​​are mapped to entity attribute values; the categories of triples are all limited to the scope of the two-level ontology classification system definition.

4. The quantitative fertilization method for low-nitrogen-tolerant maize cultivation according to claim 3, characterized in that, In S3, the two-level ontology classification system carrying ontology architecture information is received from the output of S1. Labels are set according to the 7 types of entities and 23 types of relations defined by the classification system, and the BERT-BiLSTM-CRF model is used for sequence labeling. The loss function is: ； In the formula, For the sample size, For the first The entity label sequence of a sample, Given a sequence of text as input, For model parameters, is the L2 regularization coefficient.

5. The quantitative fertilization method for low-nitrogen-tolerant maize cultivation according to claim 4, characterized in that, In S4, the structured knowledge triples output from S2 and the unstructured knowledge triples output from S3 are received. The similarity of the nine high-priority genotype attributes is calculated, and a dual mechanism of confidence transmission and strong rejection is executed, including the following sub-steps: S401. Calculate the similarity of 9 high-priority genotype attributes: ； In the formula, For entities With entity In attributes Similarity on; S402, Execution of genotype confidence transmission mechanism: ； In the formula, For the overall similarity of entity pairs, Similarity for 58 common attributes; The threshold for core attribute similarity; These are the weighting coefficients; When the similarity of the 9 core attributes is greater than or equal to At the same time, increase the overall confidence level of the entity pair; S403, Implement a strong veto mechanism for genotype inconsistency: ； In the formula, For entities Its low-nitrogen tolerance rating; If two entities have different low-nitrogen tolerance levels, they are considered mismatched; if the similarity is greater than or equal to the entity merging threshold... Entity merging is performed at that time.

6. The quantitative fertilization method for low-nitrogen-tolerant maize cultivation according to claim 5, characterized in that, In S5, a high-quality fused knowledge set output from S4 is received. Ontology reasoning and implicit knowledge completion are performed based on the high-quality fused knowledge set, genotype features are extracted, and a low-nitrogen stress response feature map subgraph is constructed. The spectrum of the low nitrogen stress response characteristics is as follows: ； ； In the formula, A collection of maize varieties. For varieties The five-dimensional genotype feature vector, For low-nitrogen tolerance, the value is [1, 5]. It is a root system type. For nitrogen use efficiency, For later anti-aging properties, For grouting rate; The final knowledge graph is represented as follows: ； In the formula, For a set of entity nodes, Let be the set of relation edges. For a collection of attributes, This is a subplot of the characteristic spectrum of the response to low nitrogen stress.

7. A quantitative fertilization method for low-nitrogen-tolerant maize cultivation according to claim 6, characterized in that, In S6, the low-nitrogen tolerance knowledge graph output from S5 is received, and the improved TransH-Agri embedding model is used to carry out embedding learning. The improved TransH-Agri embedding model is as follows: ； ； ； In the formula, During the corn growing season, For the head entity vector, For the tail entity vector, For the first The relation vector corresponding to each reproductive period For the first The unit normal vector of the hyperplane corresponding to each reproductive period.

8. A quantitative fertilization method for low-nitrogen-tolerant maize cultivation according to claim 7, characterized in that, In S7, the temporal-aware vector representation output from S6 is received, and the spatiotemporal matching features, diagnostic linkage features, and genotype adaptation features are all mapped to the same vector space. A multi-head cross-attention mechanism is used to achieve feature-level deep fusion of the three dimensions. The multi-head cross-attention mechanism is as follows: ； ； ； Add residual connections and layer normalization: ； In the formula, Features of the spatiotemporal matching engine To diagnose the characteristics of the linkage engine, For genotype-adapted engine features, The dimension of the key vector. This is a learnable weight matrix.

9. A quantitative fertilization method for low-nitrogen-tolerant maize cultivation according to claim 8, characterized in that, In S8, the fusion reasoning results output by S7 and the low nitrogen tolerance knowledge graph output by S5 are received. Based on the variety identifier, the genotype feature vector of the variety is queried from the low nitrogen stress response feature graph, and the basic nitrogen application rate is specifically corrected. The final nitrogen application rate is calculated using the following formula: ； In the formula, Base nitrogen application rate For low-nitrogen tolerance and nitrogen-saving coefficient; This is the soil nitrogen correction factor.

10. A smart fertilization decision-making system for quantitative supplemental fertilization of low-nitrogen tolerant maize, characterized in that, A quantitative fertilization method for implementing the low-nitrogen-tolerant maize planting method according to any one of claims 1 to 9 includes: The ontology classification construction module collects expert knowledge and technical specifications in the agricultural field, constructs a two-level ontology classification system that includes low nitrogen tolerance attributes, and outputs the two-level ontology classification system to the knowledge triple extraction module. The knowledge triple extraction module is connected to the ontology classification construction module; it receives the two-level ontology classification system output by the ontology classification construction module, and obtains structured knowledge triples by semantic mapping of the structured database; it uses the BERT-BiLSTM-CRF model to perform entity recognition and relation extraction on unstructured text to obtain unstructured knowledge triples, and outputs the structured knowledge triples and unstructured knowledge triples to the genotype-priority entity alignment and fusion module. The genotype-priority entity alignment and fusion module is connected to the knowledge triple extraction module; it receives structured knowledge triples and unstructured knowledge triples output by the knowledge triple extraction module, calculates the similarity of nine high-priority genotype attributes, completes entity matching through a dual mechanism of confidence transmission and strong rejection, and outputs a high-quality fused knowledge set that eliminates redundant conflicts to the low-nitrogen stress knowledge graph construction module. The low nitrogen stress knowledge graph construction module is connected to the genotype priority entity alignment and fusion module; it receives the high-quality fused knowledge set output by the genotype priority entity alignment and fusion module, performs ontology reasoning and implicit knowledge completion, extracts five-dimensional genotype features, constructs a low nitrogen tolerance knowledge graph containing a low nitrogen stress response feature graph subgraph, and outputs the low nitrogen tolerance knowledge graph to the reproductive period time-series embedding module and the fertilization quantitative correction output terminal, respectively. The reproductive period temporal embedding module is connected to the low nitrogen stress knowledge graph construction module; it receives the low nitrogen tolerance knowledge graph output by the low nitrogen stress knowledge graph construction module, and sets independent semantic hyperplanes for the five reproductive periods (seedling stage, jointing stage, large trumpet stage, silking stage, and grain filling stage) based on the TransH-Agri agricultural embedding model, completes embedding learning, generates temporal awareness vectors, and outputs the temporal awareness vectors to the cross-attention collaborative reasoning module; The cross-attention collaborative reasoning module is connected to the reproductive period time-series embedding module and the low nitrogen stress knowledge graph construction module, respectively. It receives the time-series perception vector output by the reproductive period time-series embedding module, and has built-in spatiotemporal matching engine, diagnostic linkage engine and genotype adaptation engine. It maps the features to the same vector space, and uses a multi-head cross-attention mechanism to achieve deep fusion of spatiotemporal-diagnosis-genotype three-mode features, generates fusion reasoning results, and outputs the fusion reasoning results to the fertilizer quantitative correction output terminal. The fertilization quantitative correction output terminal is connected to the cross-attention collaborative reasoning module and the low nitrogen stress knowledge graph construction module, respectively; it receives the fusion reasoning results output by the cross-attention collaborative reasoning module and the low nitrogen stress knowledge graph construction module; it calls the low nitrogen stress response feature graph subgraph to query the five-dimensional genotype feature vector of the variety, performs nitrogen-saving quantitative mapping correction on the basic nitrogen application amount, allocates the fertilizer application amount according to each growth stage of maize, and outputs a fertilization plan including fertilization time, fertilizer type, and application amount.

Images