Entity alignment method for multi-modal crop knowledge graph

Abstract

The invention discloses an entity alignment method for a multi-modal crop knowledge graph, and belongs to the technical field of knowledge graphs and agricultural intelligent analysis, and the method comprises the steps: obtaining a data feature item set of an agricultural field, and carrying out the time sequence compensation, and generating a growth time sequence feature parameter; performing feature extraction on the data feature item set, performing fusion to generate a multi-modal fusion feature mapping graph, constructing a dynamic feature matching network based on growth time sequence feature parameters, and performing entity node traversal on a reference knowledge graph to generate a candidate alignment set and a corresponding difference unit set; filtering conflict nodes in the candidate alignment set according to a three-dimensional confidence evaluation model to generate an effective alignment chain; and generating a graph updating instruction based on the difference unit set, and reconstructing a knowledge graph topological structure in combination with the effective alignment chain. According to the method, multi-modal feature bridging, dynamic weight optimization of growth stage perception and an incremental conflict resolution mechanism are adopted, so that cross-domain agricultural entity accurate alignment and real-time adaptive updating of the knowledge graph can be realized.

Description

Technical Field

[0001] The present invention relates to the field of knowledge graphs and agricultural intelligent analysis technology, and in particular to an entity alignment method for multimodal crop knowledge graphs. Background Art

[0002] Modern agricultural knowledge graphs must integrate multimodal information such as drone imagery, environmental sensor time-series data, and agricultural technical documentation to comprehensively model crop growth status. Existing technologies, entity alignment, as a core component of knowledge graph construction, aims to establish a unified agricultural knowledge system through feature mapping across data sources to support precision agriculture decision-making. However, crop growth is influenced by complex factors such as regional climate and planting practices, resulting in significant nonlinear correlations and temporal dynamics in multimodal features, posing significant challenges to traditional alignment methods.

[0003] Current entity alignment technologies mainly rely on static semantic similarity calculations, such as text alignment based on word embeddings (e.g., the TransE model) or image alignment based on visual feature matching (e.g., the Siamese network). Existing solutions, such as the cross-modal alignment framework proposed in [1] based on the attention mechanism, achieve joint mapping of images and texts through a shared feature space; and [2] use graph neural networks to fuse sensor data with knowledge graph nodes and align them using node embedding similarity. Some methods introduce time series analysis, but are mostly limited to simple sliding window processing of single-dimensional time series features.

[0004] However, static weight allocation strategies cannot adapt to the stage characteristics of crop growth. For example, the difference in temperature sensitivity during the seedling stage and light sensitivity during the flowering stage is not dynamically modeled. The heterogeneity of multimodal features leads to a semantic gap in the alignment process. For example, it is difficult to effectively associate the text term "leaf blight" with the lesion morphology in image features. Existing incremental update methods lack fine-grained control over local feature conflicts, and graphs are easily invalidated due to node redundancy or the accumulation of outdated information. These problems severely limit the practical application of agricultural knowledge graphs in field management scenarios. Summary of the Invention

[0005] To solve the above problems, the present invention provides an entity alignment method for multimodal crop knowledge graphs, which adopts multimodal feature bridging, dynamic weight optimization of growth stage perception and incremental conflict resolution mechanism, and can achieve precise alignment of cross-domain agricultural entities and real-time adaptive updating of knowledge graphs.

[0006] The above objectives can be achieved through the following solutions:

[0007] The invention discloses an entity alignment method for a multimodal crop knowledge graph, comprising acquiring multimodal data from an agricultural site, extracting image features, sensor features, and text features from the multimodal data to generate a data feature item set containing a timestamp parameter; performing time series compensation on the data feature item set according to a preset growth stage division strategy to generate growth time series feature parameters; constructing a three-level feature calibration channel to extract visual feature vectors, physical feature matrices, and concept feature sets from the data feature item set, and fusing them to generate a multimodal fusion feature map; adjusting the spatiotemporal weights of the multimodal fusion feature map according to the growth time series feature parameters to construct a dynamic feature matching network; utilizing the dynamic feature matching network to traverse entity nodes of a preset benchmark knowledge graph to generate a candidate alignment set and a corresponding difference unit set; filtering conflicting nodes in the candidate alignment set according to a preset three-dimensional confidence evaluation model to generate a valid alignment chain; triggering an incremental feature alignment process based on the difference unit set to generate a graph update instruction; and reconstructing a knowledge graph topology structure according to the valid alignment chain and the graph update instruction.

[0008] Optionally, the acquiring of multimodal data from an agricultural site, extracting image features, sensor features, and text features from the multimodal data to generate a data feature item set containing a timestamp parameter includes: acquiring drone aerial image stream data, extracting plant morphological features through a preset leaf texture segmentation algorithm, and obtaining image features; collecting multi-source time series data from a sensor group, calculating soil-atmosphere coupling features through a preset environmental factor association model, and obtaining sensor features containing a timestamp parameter; parsing heterogeneous text data from an agricultural technology document library, generating standardized agricultural concepts through a preset dialect term conversion table, and obtaining text features; and aggregating the image features, the sensor features, and the text features to generate a data feature item set containing a timestamp parameter.

[0009] Optionally, the data feature item set is time-series compensated according to a preset growth stage division strategy to generate growth time series characteristic parameters, including: dividing the growth stage into five stages of sowing period, seedling period, flowering period, maturity period and harvest period according to a preset phenological cycle template of the target crop; establishing a corresponding feature compensation coefficient matrix for each growth stage, wherein the temperature feature compensation weights of the seedling period and flowering period are increased to a preset weight threshold; matching the timestamp parameters in the data feature item set with the phenological cycle template to generate a growth stage identifier; activating the corresponding feature compensation coefficient matrix based on the growth stage identifier to compensate and correct the data feature item set to obtain growth time series characteristic parameters.

[0010] Optionally, the construction of a three-level feature calibration channel, extracting visual feature vectors, physical feature matrices and conceptual feature sets from the data feature item set, and fusing them to generate a multimodal fusion feature map includes: deploying an image texture recognizer, performing leaf vein direction analysis and spot morphology detection on the image features to generate a visual feature vector with topological constraints; activating a preset sensor correction ring, correcting abnormal temperature and humidity data in the sensor features according to a preset soil type comparison table to generate a physical feature matrix; starting a preset text semantic disambiguator, using a preset crop alias dictionary to perform synonym merging on the text features to generate a conceptual feature set; synchronizing the time axis of the visual feature vector, the physical feature matrix and the conceptual feature set, and fusing them to generate a multimodal fusion feature map.

[0011] Optionally, adjusting the spatiotemporal weights of the multimodal fusion feature map according to the growth time series characteristic parameters to construct a dynamic feature matching network includes: setting a spatiotemporal weight allocation strategy according to the growth stage identifier in the growth time series characteristic parameters; adjusting the spatial correlation strength in the multimodal fusion feature map based on a preset regional planting calendar; constructing a topological structure based on the visual feature vector, the physical feature matrix and the conceptual feature set; and constructing a dynamic feature matching network according to the spatiotemporal weight allocation strategy, the spatial correlation strength and the topological structure.

[0012] Optionally, filtering the conflicting nodes in the candidate alignment set according to a preset three-dimensional confidence assessment model to generate a valid alignment chain includes: establishing a first confidence layer based on the node similarity score output by the dynamic feature matching network; establishing a second confidence layer in combination with the conflict intensity parameter in the difference unit set; fusing the first confidence layer and the second confidence layer to generate a three-dimensional confidence assessment model, and determining it as a valid alignment chain when the total score output by the three-dimensional confidence assessment model exceeds a preset threshold.

[0013] Optionally, filtering the conflicting nodes in the candidate alignment set according to a preset three-dimensional confidence assessment model to generate a valid alignment chain also includes: collecting entity association feedback information in the farmer's operation log to generate an artificial verification factor; injecting the artificial verification factor into the three-dimensional confidence assessment model as a third confidence layer; when the deviation between the output value of the first confidence layer and the output value of the third confidence layer exceeds a preset tolerance, restoring the current multimodal fusion feature map to a previous preset version.

[0014] Optionally, the generation of graph update instructions by triggering the incremental feature alignment process based on the difference unit set includes: extracting new variety features and disappeared features from the difference unit set to generate a difference feature package; performing local graph mapping calculation on the difference feature package to generate an incremental update vector; and using a preset regional knowledge exchange protocol to distribute the incremental update vector to knowledge graph nodes in adjacent regions.

[0015] Optionally, the local graph mapping calculation for the difference feature package to generate an incremental update vector includes: locking the parent node associated with the difference feature package to generate a feature influence domain; executing a lightweight graph traversal algorithm within the feature influence domain to generate a local matching result; performing topological consistency verification on the local matching result and the global knowledge graph; and generating an incremental update vector based on the topological verification result.

[0016] Optionally, reconstructing the knowledge graph topology structure according to the valid alignment chain and the graph update instruction includes: establishing cross-modal feature connection edges according to the entity relationship in the valid alignment chain; inserting new feature nodes and marking invalid nodes based on the graph update instruction; eliminating redundant connections and strengthening core feature links through a preset topology optimization algorithm; generating an updated knowledge graph with a version identification and synchronizing it to the agricultural decision-making terminal.

[0017] Compared with the prior art, the present invention has the following advantages:

[0018] 1. Through the deep fusion of multimodal data and the dynamic allocation of spatiotemporal weights, the accuracy of crop entity alignment is significantly improved. A three-level calibration channel based on visual features, sensor timing parameters, and agronomic concepts effectively eliminates the bias of a single data source. The dynamic feature matching network adjusts spatiotemporal weights to adapt to the stage-specificity of crop growth, thereby improving the accuracy of alignment tasks such as pest and disease identification and planting pattern association.

[0019] 2. A technical solution based on incremental learning and local graph updates enables adaptive iteration of the knowledge graph. Through an incremental feature alignment process for differential unit sets, combined with a three-dimensional confidence assessment model to filter conflicting nodes, dynamic information such as new varieties and new agronomic techniques can be integrated in real time, avoiding the high computational overhead of traditional global update methods while ensuring the topological consistency of the knowledge graph and improving maintenance efficiency.

[0020] 3. The combination of cross-modal association modeling and manual feedback synergy enhances robustness to complex agricultural scenarios. The introduction of manual verification factors as the third level of confidence assessment automatically triggers the feature rollback mechanism when there is a significant deviation between farmer experience and algorithm results. This not only retains the explanatory power of domain knowledge, but also avoids the spread of misalignment caused by data noise, thereby comprehensively improving the reliability of agricultural decision-making.

[0021] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0023] Figure 1 It is a flowchart of the entity alignment method of the multimodal crop knowledge graph according to an embodiment of the present invention.

[0024] Figure 2 1 is a schematic diagram of the multimodal data collection and processing flow of an embodiment of the present invention. DETAILED DESCRIPTION

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0026] Reference Figure 1 One embodiment of the present invention proposes an entity alignment method for a multimodal crop knowledge graph, which adopts multimodal feature bridging, dynamic weight optimization based on growth stage perception, and incremental conflict resolution mechanism to achieve precise alignment of cross-domain agricultural entities and real-time adaptive update of the knowledge graph.

[0027] The method of this embodiment specifically includes:

[0028] Acquiring multimodal data from an agricultural site, extracting image features, sensor features, and text features from the multimodal data to generate a data feature item set including a timestamp parameter;

[0029] According to a preset growth stage division strategy, time series compensation is performed on the data feature item set to generate growth time series feature parameters;

[0030] Constructing a three-level feature calibration channel to extract visual feature vectors, physical feature matrices, and conceptual feature sets from the data feature item set, and fusing them to generate a multimodal fusion feature map;

[0031] Adjusting the spatiotemporal weights of the multimodal fusion feature map according to the growth time series feature parameters to construct a dynamic feature matching network; using the dynamic feature matching network to traverse the entity nodes of a preset benchmark knowledge graph to generate a candidate alignment set and a corresponding difference unit set;

[0032] Filtering conflicting nodes in the candidate alignment set according to a preset three-dimensional confidence evaluation model to generate a valid alignment chain; triggering an incremental feature alignment process based on the difference unit set to generate a graph update instruction;

[0033] Reconstruct the knowledge graph topology structure according to the valid alignment chain and the graph update instruction.

[0034] This method, based on multimodal feature fusion and a dynamic weight allocation mechanism, enhances the expressive power of crop entity features through spatiotemporal compensation and three-level calibration. It combines a confidence gradient model with incremental learning to achieve dynamic alignment and updating of the knowledge graph. Cross-modal bridging addresses the semantic gap inherent in single data sources, adapting to the complex and variable nature of the agricultural environment. A hierarchical compensation mechanism based on growth stages improves the accuracy of capturing changes in crop physiological states. Dynamic topology optimization reduces redundant information interference and enhances the decision-making support capabilities of the knowledge graph.

[0035] Optionally, acquiring multimodal data from an agricultural site and extracting image features, sensor features, and text features from the multimodal data to generate a data feature item set containing a timestamp parameter includes:

[0036] Obtain drone aerial image stream data, extract plant morphological features through a preset leaf texture segmentation algorithm, and obtain image features;

[0037] Specifically, the drone cruises along a preset route and altitude to capture farmland images, generating a continuous frame stream from a single aerial shot. This image stream is then processed using a leaf texture segmentation algorithm. This algorithm, based on a U-Net network architecture, uses an encoder to extract leaf edge features and a decoder to restore pixel-level segmentation masks. The output is a set of morphological features, including plant height, canopy cover, and lesion distribution index. For example, the segmented leaf regions are quantitatively represented using RGB mean and texture entropy values.

[0038] The leaf texture segmentation algorithm uses a U-Net-based semantic segmentation model. The input image is resized to 512×512, and the output is a binary mask and texture feature vector. Plant morphological features include plant height calculated using stereo vision ranging, canopy cover as the percentage of leaf pixels, and the lesion distribution index as the ratio of lesion area to total leaf area.

[0039] Collect multi-source time series data from the sensor group, calculate the soil-atmosphere coupling characteristics through the preset environmental factor correlation model, and obtain sensor characteristics including timestamp parameters;

[0040] Specifically, the sensor group deployed in the field (temperature, humidity, light, soil conductivity) collects time series data at 5-minute intervals. The environmental factor association model is a mathematical model based on partial least squares regression (PLS), and the formula is:

[0041]

[0042] Among them, Y is soil moisture, X is the matrix of atmospheric temperature, air humidity and light intensity, is the regression coefficient matrix, and ∈ is the error term. After model fitting, the soil-atmosphere coupling characteristics are represented as the joint vector of regression residuals and predicted values, and the timestamp is recorded as the Unix time code of the acquisition moment.

[0043] Multi-source time series data represents a temporal sequence of multidimensional environmental parameters, while soil-atmosphere coupling characteristics are quantitative indicators reflecting the dynamic relationship between the two. The environmental factor correlation model is trained and generated using historical data, supporting the modeling of nonlinear relationships between temperature and humidity.

[0044] Parse the heterogeneous text data in the agricultural technology document library, generate standardized agricultural concepts through a preset dialect term conversion table, and obtain text features;

[0045] Specifically, agricultural technical documents, including PDF reports, handwritten field notes, and speech transcriptions, are converted into structured text using optical character recognition (OCR) and automated speech recognition (ASR) technologies. The dialect term conversion table is a predefined key-value mapping table, for example, mapping the dialect term "rice fever" to the standard term "rice blast." Text features are generated using TF-IDF vectorization.

[0046] The heterogeneous text data consists of agricultural technical documents in multiple formats and languages, and the dialect terminology conversion table is a mapping dictionary between regional agricultural terminology and standardized terminology. Text features are bag-of-words models or BERT embedding vectors after stop word removal and stemming.

[0047] The image features, the sensor features, and the text features are aggregated to generate a data feature item set including a timestamp parameter.

[0048] Specifically, image features (such as canopy cover at 08:00), sensor features (such as soil moisture forecast at 08:00), and text features (such as "leaf yellowing" recorded at 08:00) are aligned by timestamp to generate multimodal data items indexed by timestamp. For example, three types of features at the same moment are concatenated into a high-dimensional tensor and stored in a time series database. The data feature item set is a multimodal feature set organized by time series, with timestamp parameters in the ISO 8601 standard time format. Feature aggregation is achieved through time window sliding matching.

[0049] For example, in the rice-growing area of Jiangnan, a drone captured an aerial image of a field at 09:00. The leaf texture segmentation algorithm detected a canopy coverage of 65% and a diseased area of 2%. The sensor recorded an atmospheric temperature of 28°C and a soil moisture of 45% at this time. The environmental factor association model calculated the coupled feature residual to be 0.3. At the same time, the farmer's handwritten log recorded "early rice leaf tips scorched and withered", which was mapped to "early leaf blight" using the dialect conversion table. These three types of features were aggregated according to the 09:00 timestamp to generate a data item set. Through time alignment and feature fusion of multi-source data, the co-occurrence relationship between environmental anomalies (high temperature and low humidity) and crop diseases (leaf blight) can be accurately associated, avoiding the one-sided interpretation of single sensor data.

[0050] Optionally, performing time series compensation on the data feature item set according to a preset growth stage division strategy to generate growth time series feature parameters includes:

[0051] The target crop is divided into five growth stages: sowing period, seedling period, flowering period, maturity period and harvest period according to the preset phenological cycle template;

[0052] Specifically, a phenological cycle template is established based on the historical growth data of the target crop (such as wheat), defining the sowing period as the time from seed placement to the unfolding of the first true leaf, the seedling period as the time from plant height reaching 10 cm to the completion of tillering, the flowering period as the time from heading to the end of flowering, the maturity period as the time from grain filling to the hardness reaching the standard, and the harvest period as the time from plant yellowing to the completion of harvesting. The time window of each stage is dynamically adjusted according to the longitude and latitude and the climate zone, and the template accuracy is ±3 days. Among them, the phenological cycle template is a standardized time interval table of crop physiological stages, which is generated based on the accumulated temperature model and historical agricultural records.

[0053] A corresponding feature compensation coefficient matrix is established for each growth stage, where the temperature feature compensation weights of the seedling and flowering stages are increased to the preset weight threshold;

[0054] Specifically, the compensation coefficient matrix is a diagonal matrix with the same dimension as the number of features. The temperature feature weight in the seedling stage is set to 0.9 (threshold), and the humidity weight is 0.6; the temperature weight in the flowering stage is 0.8, and the light weight is 0.7; the weight of each feature in the other stages is 0.5 by default. For example, if the night temperature is lower than 15°C during the flowering stage, the influence of the temperature feature in the fusion mapping is enhanced by a weight of 0.8. For the compensation coefficient matrix W stage ,have:

[0055] W stage =diag(w1,w2,…,w n ),

[0056] Among them, w n is the compensation weight of the nth feature, the diagonal elements are set according to the characteristics of the growth stage, and diag is the diagonal matrix.

[0057] Matching and calculating the timestamp parameter in the data feature item set with the phenological cycle template to generate a growth stage identifier;

[0058] Specifically, the data timestamp is converted to Julian days and compared with the start and end dates of each stage in the template. If the timestamp falls within a certain stage's date range, the stage identifier is output, such as "M2" for the seedling stage. For data that spans multiple stages, such as the boundary between flowering and maturity, a sliding window algorithm is used to determine the dominant stage. The growth stage identifier is a unique coding label used to index the compensation coefficient matrix.

[0059] The data feature item set is compensated and corrected based on the feature compensation coefficient matrix corresponding to the activation of the growth stage identifier to obtain growth time series feature parameters.

[0060] Specifically, the seedling compensation matrix is selected based on the identifier "M2." The temperature feature value in the data item is multiplied by 0.9, the humidity feature value is multiplied by 0.6, and the other features remain at their original values. For example, if the original temperature data is 25°C, it will be 22.5°C after compensation. This strengthens the seedling's sensitivity to temperature.

[0061] The compensation correction is a weighted adjustment operation of the characteristic value, and the growth time series characteristic parameter is the compensated feature set.

[0062] For example, the seedling stage of winter wheat in North China runs from November 1st to December 10th. On a particular day, the temperature was 8°C and the humidity was 70%. The system determined the period to be the seedling stage based on the timestamp and activated a temperature weight of 0.9. After compensation, the temperature characteristic value increased to 7.2°C, and the humidity remained at 70% × 0.6 = 42%. The compensated low-temperature feature is more pronounced, aligning with the "frost damage vulnerability period" entity in the baseline map. By amplifying the sensitive characteristics of key growth stages through staged compensation, we can avoid misjudgments caused by low absolute values of environmental data and improve the accuracy of low-temperature frost damage warnings.

[0063] Alternatively, as Figure 2 As shown, the three-level feature calibration channel is constructed, and the visual feature vector, physical feature matrix and concept feature set are extracted from the data feature item set, and fused to generate a multimodal fusion feature map, including:

[0064] Deploy an image texture recognizer to perform leaf vein direction analysis and spot morphology detection on the image features to generate a visual feature vector with topological constraints;

[0065] Specifically, the image texture recognition function is based on a modified VGG16 network. It extracts features such as the branching angle of the primary vein and the density of secondary veins through a max pooling layer, outputting a 256-dimensional vector. Topological constraints are implemented by modeling the connectivity of leaf veins using a graph convolutional network (GCN). Lesion detection uses Mask R-CNN to locate lesion boundaries and calculate their area percentage. Topological constraints are spatial association rules based on graph structures, and the visual feature vector includes quantitative morphological and pathological indicators.

[0066] Activate a preset sensor correction ring and correct abnormal temperature and humidity data in the sensor characteristics according to a preset soil type comparison table to generate a physical characteristic matrix;

[0067] Specifically, the soil type comparison table defines the water holding capacity and thermal conductivity of sandy soil, clay, and other soil types. The sensor calibration loop uses the correction factor corresponding to the soil type (for example, multiplying sandy soil moisture data by 1.2) and filters outliers that exceed the three standard deviations (for example, temperatures > 50°C are marked as missing). The physical feature matrix is the corrected 20×20 spatiotemporal distribution data. The sensor calibration loop is a data cleaning and calibration module based on the physical characteristics of the environment.

[0068] Starting a preset text semantic disambiguator, and using a preset crop alias dictionary to perform synonym merging on the text features to generate a concept feature set;

[0069] Specifically, the text semantic disambiguator uses the BERT model combined with an agricultural dictionary to unify "seedling rot" and "rice damping-off" into "rice damping-off," generating standardized concept codes (such as ICD-AC001). The concept feature set includes normalized pathological terms and their TF-IDF weights. The crop alias dictionary is an authoritative lexicon that maps regional aliases to scientific names.

[0070] The visual feature vector, the physical feature matrix and the concept feature set are synchronized on a time axis and fused to generate a multimodal fusion feature map.

[0071] Specifically, a cross-modal feature bridging channel is established. Using timestamps as key values, the channel aligns visual vectors, physical matrices, and concept sets at the millisecond level, then uses tensor concatenation and dimensionality reduction (PCA) to generate a fusion map. For example, a visual vector (256D) + a physical matrix (400D) + a concept set (50D) is compressed into a 512-dimensional fused feature. The cross-modal feature bridging channel is an integrated module for spatiotemporal alignment of multi-source data and feature concatenation.

[0072] For example, at a certain moment, image features detected an abnormal angle of 23° between the main veins of a leaf, sensors indicated a 20% drop in soil moisture, and a textual record indicated "leaf tip wilting." After synchronizing these three features, the bridging channel fused the mapping to highlight the "water stress-structural deformation" association, matching it with the "drought response" entity in the baseline atlas. By topologically correlating and fusing multimodal features, it is possible to decouple the co-occurring factors of complex symptoms, such as distinguishing leaf deformation caused by drought from that caused by disease, thereby improving diagnostic specificity.

[0073] Optionally, adjusting the spatiotemporal weights of the multimodal fusion feature map according to the growth time series feature parameters to construct a dynamic feature matching network includes:

[0074] Setting a spatiotemporal weight allocation strategy according to a growth stage identifier in the growth time series characteristic parameter;

[0075] Specifically, the spatiotemporal weight includes spatial weight and time weight. The spatial weight is the weighted intensity of the spatial neighborhood of the feature map, which is used to adjust the contribution of spatial features at different growth stages; the time weight is the weighted intensity of the time window of the feature map, which is used to adjust the importance of the time dimension at different growth stages. For the spatial weight W spa and time weight W tem ,have:

[0076] W spa =α1*S stage ,

[0077]

[0078] Among them, α1 is the spatial sensitivity factor (default 0.7), S stage is the spatial extension of the growth stage (e.g., flowering stage = 1.2), β1 is the time attenuation coefficient (default 0.5), T stage is the stage duration (days). The weights are ultimately applied to the spatial neighborhood and time window of the feature map.

[0079] adjusting the spatial correlation strength in the multimodal fusion feature map based on a preset regional planting calendar;

[0080] Specifically, the regional planting calendar records the crop rotation plans and variety distribution in different regions. If the current region is a double-season rice area, the association edge weight between early rice and late rice varieties is strengthened, such as increasing it from 0.6 to 0.8. spa ,have:

[0081] C spa =γ1*log(1+D region ),

[0082] Among them, γ1 is the regional weight factor, D region For the planting similarity of adjacent regions, by adjusting the value of spatial association intensity, the correlation between different varieties or regions can be strengthened or weakened, making the dynamic feature matching network more consistent with the actual planting situation.

[0083] A topological structure is constructed based on the visual feature vector, the physical feature matrix and the conceptual feature set.

[0084] Specifically, a multi-head attention mechanism is used to calculate cross-attention scores between image, sensor, and text features, and based on these scores, node connections and feature weights are reconstructed. For example, during the flowering period, because physiological activities such as flower opening and pollination are more sensitive to environmental factors such as light and temperature, the weights of image and sensor features in the fused feature map can be increased accordingly. For the attention score Attention(Q,K,V), we have:

[0085]

[0086] Where Q is the query matrix, K is the key matrix, V is the value matrix, and K T is the transpose of the key matrix, B is the dimension, and softmax is the attention calculation function. The query matrix represents the features or information of interest. In entity alignment within a multimodal crop knowledge graph, it represents features from a modality, such as an image, and is used to query related features from other modalities, such as sensors or text. The key matrix performs a dot product operation with the query matrix to calculate similarity between the two. The key matrix can be considered a "key" storing feature information from other modalities, used for matching with the query matrix. The value matrix contains the actual values or information. In the attention mechanism, once the similarity between the query and key matrices is calculated, the value matrix performs a weighted sum based on these similarities to generate the final attention output. The dimension represents the dimensions of the query and key matrices. Node connections are reconstructed based on the attention scores to construct the topology of the dynamic network.

[0087] A dynamic feature matching network is constructed according to the spatiotemporal weight allocation strategy, the spatial association strength, and the topological structure.

[0088] Specifically, assume that image features, sensor features, and text features of rice at different growth stages are acquired, and a data feature item set containing a timestamp parameter is generated. Next, the spatiotemporal weights of the multimodal fusion feature map are adjusted based on the growth time series feature parameters. Node connections and feature weights are reconstructed based on attention scores to construct a dynamic feature matching network.

[0089] For example, during the flowering period in a double-crop rice region, increased spatial correlation strength led to a strong correlation between pest and disease characteristics of different rice crops in the same plot. The attention mechanism assigned a weight of 0.75 to image lesion features and 0.6 to sensor data, ultimately accurately aligning the entities for "early rice sheath blight" and "late rice susceptible period." By allocating attention guided by regional knowledge, the ability to discover entity associations within localized cropping patterns is enhanced.

[0090] Optionally, filtering conflicting nodes in the candidate alignment set according to a preset three-dimensional confidence assessment model to generate a valid alignment chain includes:

[0091] Establishing a first confidence layer based on the conflicting node similarity scores of the candidate alignment set;

[0092] Specifically, the three-dimensional confidence assessment model adopts a two-layer assessment mechanism, including at least the first confidence layer and the second confidence layer. Based on the node similarity score output by the dynamic feature matching network, the nodes in the candidate alignment set are preliminarily screened. The node similarity score output by the dynamic feature matching network (usually a value between 0 and 1) is converted into a confidence value. This is usually achieved through a probability conversion function such as a S-type function, so that nodes with high similarity scores obtain higher confidence. For the output value C1 of the first confidence layer, there is:

[0093]

[0094] Where s is the similarity score, τ is the slope factor (default 5), and θ is the threshold (default 0.7). When s = 0.85, C1 ≈ 0.83. The node similarity score is a quantitative measure of the matching between entity attributes and relationships, and the first confidence layer is a probabilistic conversion module based on similarity.

[0095] establishing a second confidence layer in combination with the conflict strength parameter in the difference unit set;

[0096] Specifically, the conflict intensity parameter includes the number of attribute conflicts (such as inconsistent names) and the number of relationship conflicts (such as inconsistent parent classes). For the second confidence level output value C2, there are:

[0097]

[0098] Among them, G i is the count of type i conflicts, Li is the weight of the count of type i conflicts, and max(G) represents the maximum possible conflict value, which is used to normalize the conflict intensity. For example, a name conflict has a weight of 0.5, and a relationship conflict has a weight of 0.8. If the total conflict value is 2 and the maximum possible conflict value is 5, then C2 = 0.6. The difference unit set is the set of unmatched entities and their conflicting features, and the conflict intensity parameter is a weighted statistic of the type and number of conflicts.

[0099] The first confidence layer and the second confidence layer are fused to generate a three-dimensional confidence assessment model. When the total score output by the three-dimensional confidence assessment model exceeds a preset threshold, the alignment chain is determined to be a valid alignment chain.

[0100] Specifically, the total score is calculated using a three-dimensional confidence assessment model with a preset threshold of 0.75. If the total score is 0.72, it is filtered out if it is below the threshold. The three-dimensional confidence assessment model is a multi-dimensional scoring fusion framework, and the valid alignment chain is the entity matching relationship that has passed the confidence screening. The three-dimensional confidence assessment model is a multi-dimensional scoring fusion framework, and the valid alignment chain is the entity matching relationship that has passed the confidence screening.

[0101] Optionally, filtering conflicting nodes in the candidate alignment set according to a preset three-dimensional confidence assessment model to generate a valid alignment chain further includes:

[0102] Collect entity-related feedback information from farmers' operation logs to generate artificial verification factors;

[0103] Specifically, farmers use the terminal to correct the automatic alignment results (for example, marking a misclassified "rice blast" as "insect pest"). The system records the timestamp and content of the correction, forming a manual verification factor. Each correction is assigned a weight coefficient (for example, 0.9 for expert users and 0.7 for general users). The entity association feedback information is the correct entity relationship annotation after manual intervention, and the manual verification factor is the normalized credibility score.

[0104] injecting the artificial verification factor as a third confidence layer into the three-dimensional confidence assessment model;

[0105] Specifically, for the total score C total ,have:

[0106] C total =α2*C1+β2*C2+γ2*C3,

[0107] Here, α2 is the adjustment coefficient for the first confidence layer, β2 is the adjustment coefficient for the second confidence layer, γ2 is the adjustment coefficient for the third confidence layer, and C3 is the manual verification factor. For example, after expert correction, C3 = 0.9. Assuming the adjustment coefficient for the first confidence layer is 0.3, the adjustment coefficient for the second confidence layer is 0.2, the output value of the first confidence layer is 0.73, and the output value of the second confidence layer is 0.64, then the total score is 0.3 × 0.73 + 0.2 × 0.64 + 0.3 × 0.9 = 0.649. If the threshold is 0.75, the alignment chain is considered valid. The third confidence layer is a quantitative integration of the manual verification results, and the weight is dynamically adjusted based on user permissions.

[0108] When the deviation between the output value of the first confidence layer and the output value of the third confidence layer exceeds a preset tolerance, the current multimodal fusion feature map is restored to the previous preset version.

[0109] Specifically, the preset tolerance is 0.2. If the system score C1 = 0.8 and the manual score C3 = 0.4, and the deviation 0.4 > 0.2, the rollback mechanism restores the current feature map to the last stable version and re-extracts the alignment features. The feature rollback mechanism is a data recovery process under version control to prevent the cascading propagation of misalignments.

[0110] For example, a "leaf spot" disease was aligned to a "fungal disease" with a C1 of 0.78, and the farmer corrected it to "bacterial leaf streak" with a C3 of 0.82. Because the deviation, 0.82-0.78 = 0.04, was within the tolerance, the system retained the correction. However, if the farmer corrected it to an "insect pest" with a C3 of 0.3, and the deviation, 0.78-0.3 = 0.48, was greater than 0.2, a rollback to the previous version would be triggered. This dynamic balance between manual feedback and automatic alignment respects domain experience while preventing malicious or erroneous interference.

[0111] Optionally, triggering the incremental feature alignment process based on the difference unit set to generate a graph update instruction includes:

[0112] Extracting new variety features and missing features from the difference unit set to generate a difference feature package;

[0113] Specifically, the differential feature package contains two types of data: newly added features (such as disease-resistance gene markers and new pesticide names) and disappeared features (such as discontinued pesticide identifiers). Differential encoding technology is used for compressed storage, for example, in JSON format to record feature key-value pairs and operation types (ADD / DELETE). Newly added variety features are entity attributes that are not registered in the knowledge graph, while disappeared features are entity identifiers that have expired or been eliminated.

[0114] Performing local atlas mapping calculation on the difference feature package to generate an incremental update vector;

[0115] Specifically, a lightweight graph traversal is performed in the local subgraph, such as a pruning optimized version of breadth-first search. For the incremental update vector ΔF, we have:

[0116]

[0117] Among them, d i is the number of hops between the i-th newly added feature and the parent node, λ is the attenuation factor, which can be 0.8, f i is the feature weight of the i-th newly added feature. For example, if the newly added "drought-resistant rice varieties" is associated with the parent node "rice variety library", the hop count is 1 and the weight is 0.9, then ΔF = 0.8 1 ×0.9=0.72. The incremental update vector is a numerical vector that represents the influence of the newly added features and is used to sort the update priorities of distributed nodes.

[0118] The incremental update vector is distributed to the knowledge graph nodes in adjacent regions using a preset regional knowledge exchange protocol.

[0119] Specifically, the protocol defines geographic-based knowledge dissemination rules, such as requiring nodes in adjacent counties to synchronize incremental updates within 24 hours. The gossip protocol is used for distribution to ensure eventual consistency of update vectors. The regional knowledge exchange protocol serves as a data synchronization mechanism in distributed systems, supporting cross-regional knowledge sharing.

[0120] For example, a new lodging-resistant corn variety is added to a region. Local mapping calculates its association weight with the corn breeding library as 0.8, generating an incremental vector of 0.64. After distribution to three adjacent county nodes, each node decides whether to incorporate this feature based on its local planting strategy. This enables regionally adaptive updates and prevents the global proliferation of invalid knowledge, such as automatically filtering tropical varieties for updates in high-altitude and cold regions.

[0121] Optionally, performing local atlas mapping calculation on the difference feature package to generate an incremental update vector includes:

[0122] Locking the parent node associated with the difference feature package to generate a feature influence domain;

[0123] Specifically, the SPARQL query language is used to search for parent nodes that have a "belongs to" or "related to" relationship with the newly added feature, forming an influence domain radius (the default is 3 hops). For example, the influence domain of the newly added "drone pesticide application technology" includes the parent nodes "precision agriculture equipment" and "pesticide spraying methods." The feature influence domain is the set of entities in the knowledge graph that are directly or indirectly associated with the differential feature.

[0124] Executing a lightweight graph traversal algorithm within the feature influence domain to generate a local matching result;

[0125] Specifically, the lightweight algorithm limits the traversal depth to ≤ 3 hops and prunes irrelevant branches (such as edges with an association weight < 0.3). The output is a list of candidate nodes that the new feature may connect to. The lightweight graph traversal algorithm is an optimized subgraph search method for resource-constrained environments.

[0126] The local matching result is topologically verified with the global knowledge graph to obtain a topological verification result.

[0127] Specifically, verification criteria include the absence of circular dependencies and core attribute conflicts. For example, if a newly added "Organic Fertilizer A" conflicts with "Chemical Fertilizer B" in the global graph, the node will be marked as pending review. Topological consistency verification is a logical conflict checking mechanism based on graph constraints.

[0128] An incremental update vector is generated according to the topology verification result.

[0129] Specifically, the incremental update vector typically contains information such as the identifiers of new or missing features, the identifiers of associated nodes, and the association weights. New and missing features are extracted from the entity alignment results to form a differential feature package. A graph database query language is used to retrieve entities with specific relationships to the differential features and determine the feature impact domain. A lightweight graph traversal algorithm is executed within the feature impact domain to search for candidate nodes that may be associated with the new or missing features. Topological consistency verification is performed on the candidate nodes to ensure that the update operation does not disrupt the structural integrity and logical consistency of the knowledge graph. An incremental update vector is generated based on the verified candidate nodes and their association weights with the differential features.

[0130] For example, suppose a new disease-resistant gene marker "GeneX" is discovered during the entity alignment process, and this marker is associated with the rice variety "RiceA". The differential feature package contains the new disease-resistant gene marker "GeneX"; the SPARQL query language is used to retrieve entities that have a "belongs to" relationship with "GeneX", and it is found that "RiceA" is its associated rice variety. Therefore, "RiceA" and its related nodes (such as planting areas, climatic conditions, etc.) constitute the feature influence domain; a lightweight graph traversal algorithm is executed within the feature influence domain to search for candidate nodes that may be associated with "GeneX". Suppose that during the traversal process, it is found that the planting area "AreaY" of "RiceA" is indirectly associated with "GeneX" (such as introduced through some genetic improvement technology); check whether the association between "GeneX", "RiceA" and "AreaY" is consistent with genetic principles and common sense of agricultural practice. The hypothesis is verified; based on the verification results, an incremental update vector is generated. For example, an incremental update vector might contain the following information: a newly added disease resistance gene marker "GeneX," associated with the rice variety "RiceA" with an association weight of 0.9 (indicating a strong association), and an indirect association with the growing region "AreaY" with an association weight of 0.6 (indicating a moderate association). This incremental update vector will be used to update the knowledge graph, adding "GeneX" and its associations with "RiceA" and "AreaY" to the graph.

[0131] Optionally, reconstructing the knowledge graph topology structure according to the valid alignment chain and the graph update instruction includes:

[0132] Establishing cross-modal feature connection edges according to the entity relationships in the valid alignment chain;

[0133] Specifically, based on the entity relationships in the effective alignment chain, cross-modal feature connection edges are established in the knowledge graph. These connection edges represent the association between different modal features, such as the association between image features and sensor features, the association between text features and image features, etc. By establishing these connection edges, the fusion and association analysis of multimodal data can be achieved. The connection edge attributes include the data source modality (such as image, sensor, text), confidence score and time series identifier. These attributes help to evaluate the reliability and timeliness of the connection edges, and provide an important reference for subsequent knowledge reasoning and application. For example, the causal edge established between the image feature "leaf spot area" and the sensor feature "humidity anomaly" has a weight of 0.75, indicating that there is a strong association between the two.

[0134] Insert new feature nodes and mark invalid nodes based on the graph update instruction;

[0135] Specifically, new feature nodes are inserted into the knowledge graph based on information such as the newly added variety characteristics in the graph update instructions. These newly added feature nodes represent new varieties, technologies, or methods emerging in the agricultural field, injecting new information into the knowledge graph. To distinguish newly added feature nodes from existing nodes, they are typically marked with a special color, such as green. Furthermore, historical version information for newly added feature nodes is retained for easy tracing and rollback when necessary.

[0136] Specifically, invalid nodes are marked in the knowledge graph based on information such as missing features in the graph update instructions. These invalid nodes may no longer be applicable to the current agricultural landscape due to factors such as outdated technology, cultivar phasing out, or management strategy adjustments. To visually demonstrate the status of invalid nodes, they are typically marked in gray. Furthermore, historical version information for invalid nodes is retained for analysis and recovery when needed.

[0137] Eliminate redundant connections and strengthen core feature links through a preset topology optimization algorithm;

[0138] Specifically, in order to optimize the topological structure of the knowledge graph and improve its simplicity and accuracy, a topology optimization algorithm, such as the Louvain community discovery algorithm, can be used. By calculating the similarity between nodes, nodes with similarity above a threshold are merged into one node. This helps reduce redundant information in the knowledge graph and improve its query efficiency. Edges with weights below a threshold are deleted to remove noise and irrelevant information in the knowledge graph. This helps improve the simplicity and readability of the knowledge graph. For core feature links in the knowledge graph (such as the association between yield and fertilizer application), their weights are increased to highlight their importance. This helps users quickly locate key information during queries and analysis.

[0139] Generate an updated knowledge graph with version identification and synchronize it to the agricultural decision-making terminal.

[0140] Specifically, based on the entity relationships in the effective alignment chain, cross-modal feature connection edges are established in the knowledge graph to achieve the fusion and association analysis of multimodal data; based on the information such as new variety features and disappeared features in the graph update instructions, new feature nodes are inserted into the knowledge graph, and invalid nodes are marked to ensure that the knowledge graph can be updated in a timely manner; by merging nodes with high similarity, pruning low-weight edges, strengthening core feature links, etc., the topological structure of the knowledge graph is optimized to improve its simplicity and accuracy; after completing the above steps, an updated knowledge graph with version identification is generated to track the change history of the knowledge graph and synchronize it to the agricultural decision-making terminal for use by farmers and agricultural experts.

[0141] For example, assume that multimodal data from a rice-growing region is acquired through drone aerial photography and sensor monitoring. After processing using entity alignment methods, a valid alignment chain and graph update instructions are generated. A dynamic feature matching network generates a valid alignment chain, indicating that the image feature "leaf spot" corresponds to the text feature "rice blast" with high confidence. Based on the difference unit set, an incremental feature alignment process is triggered, generating a graph update instruction indicating the need to add a new feature node "disease resistance gene marker" and mark an invalid node "old pesticide identifier." A connecting edge is established in the knowledge graph, connecting the image feature "leaf spot" with the text feature "rice blast." The data source modality is set to image and text, the confidence score is high, and the time series identifier is set to the current time. A new feature node "disease resistance gene marker" is inserted into the knowledge graph and marked green. The invalid node "old pesticide identifier" is marked gray, and its historical version information is retained. The Louvain community discovery algorithm is applied to merge nodes with high similarity, such as merging multiple similar rice variety nodes into a single node. Low-weight edges are pruned, such as removing unimportant connections. Core feature links are strengthened, such as increasing the weight of the connection between yield and fertilizer application. An updated knowledge graph is generated and a version identifier (such as V2.0) is added to distinguish it from previous versions. The updated knowledge graph is synchronized to agricultural decision-making terminals for use by farmers and agricultural experts. Through these steps, the topology of the knowledge graph is reconstructed and optimized, more accurately reflecting the actual conditions in rice-growing areas and providing strong support for agricultural decision-making.

[0142] It should be noted that any embodiment of the present invention can be applied as long as the purpose of the present invention is achieved. The above description is only an exemplary embodiment of the present invention and should not be used to limit the scope of the present invention.

[0143] That is, any equivalent changes and modifications made according to the teachings of the present invention are still within the scope of the present invention. Those skilled in the art will readily conceive of other embodiments of the present invention after considering the disclosure of the specification and practical truths. This application is intended to cover any variations, uses, or adaptations of the present invention that follow the general principles of the present invention and include common knowledge or customary technical means in the art not described herein.

Claims

1. The entity alignment method of multimodal crop knowledge graph is characterized by: The method comprises: Acquiring multimodal data from an agricultural site, extracting image features, sensor features, and text features from the multimodal data to generate a data feature item set including a timestamp parameter; According to a preset growth stage division strategy, time series compensation is performed on the data feature item set to generate growth time series feature parameters; Constructing a three-level feature calibration channel to extract visual feature vectors, physical feature matrices, and conceptual feature sets from the data feature item set, and fusing them to generate a multimodal fusion feature map; Adjusting the spatiotemporal weights of the multimodal fusion feature map according to the growth time series feature parameters to construct a dynamic feature matching network; using the dynamic feature matching network to traverse the entity nodes of a preset benchmark knowledge graph to generate a candidate alignment set and a corresponding difference unit set; Filtering conflicting nodes in the candidate alignment set according to a preset three-dimensional confidence evaluation model to generate a valid alignment chain; triggering an incremental feature alignment process based on the difference unit set to generate a graph update instruction; Reconstruct the knowledge graph topology structure according to the valid alignment chain and the graph update instruction.

2. The entity alignment method for a multimodal crop knowledge graph according to claim 1, characterized in that: The step of acquiring multimodal data from an agricultural site and extracting image features, sensor features, and text features from the multimodal data to generate a data feature item set containing a timestamp parameter includes: Obtain drone aerial image stream data, extract plant morphological features through a preset leaf texture segmentation algorithm, and obtain image features; Collect multi-source time series data from the sensor group, calculate the soil-atmosphere coupling characteristics through the preset environmental factor correlation model, and obtain sensor characteristics including timestamp parameters; Parse the heterogeneous text data in the agricultural technology document library, generate standardized agricultural concepts through a preset dialect term conversion table, and obtain text features; The image features, the sensor features, and the text features are aggregated to generate a data feature item set including a timestamp parameter.

3. The entity alignment method for a multimodal crop knowledge graph according to claim 2 is characterized in that: The step of performing time series compensation on the data feature item set according to a preset growth stage division strategy to generate growth time series feature parameters includes: The target crop is divided into five growth stages: sowing period, seedling period, flowering period, maturity period and harvest period according to the preset phenological cycle template; A corresponding feature compensation coefficient matrix is established for each growth stage, where the temperature feature compensation weights of the seedling and flowering stages are increased to the preset weight threshold; Matching and calculating the timestamp parameter in the data feature item set with the phenological cycle template to generate a growth stage identifier; The data feature item set is compensated and corrected based on the feature compensation coefficient matrix corresponding to the activation of the growth stage identifier to obtain growth time series feature parameters.

4. The entity alignment method for a multimodal crop knowledge graph according to claim 3 is characterized in that: The constructing of the three-level feature calibration channel, extracting the visual feature vector, the physical feature matrix and the concept feature set from the data feature item set, and fusing them to generate a multimodal fusion feature map includes: Deploy an image texture recognizer to perform leaf vein direction analysis and spot morphology detection on the image features to generate a visual feature vector with topological constraints; Activate a preset sensor correction ring and correct abnormal temperature and humidity data in the sensor characteristics according to a preset soil type comparison table to generate a physical characteristic matrix; Starting a preset text semantic disambiguator, and using a preset crop alias dictionary to perform synonym merging on the text features to generate a concept feature set; The visual feature vector, the physical feature matrix and the concept feature set are synchronized on a time axis and fused to generate a multimodal fusion feature map.

5. The entity alignment method for a multimodal crop knowledge graph according to claim 4 is characterized in that: The step of adjusting the spatiotemporal weights of the multimodal fusion feature map according to the growth time series feature parameters to construct a dynamic feature matching network includes: Setting a spatiotemporal weight allocation strategy according to a growth stage identifier in the growth time series characteristic parameter; adjusting the spatial correlation strength in the multimodal fusion feature map based on a preset regional planting calendar; Constructing a topological structure based on the visual feature vector, the physical feature matrix and the concept feature set; A dynamic feature matching network is constructed according to the spatiotemporal weight allocation strategy, the spatial association strength, and the topological structure.

6. The entity alignment method for a multimodal crop knowledge graph according to claim 1, characterized in that: The filtering of conflicting nodes in the candidate alignment set according to the preset three-dimensional confidence evaluation model to generate a valid alignment chain includes: Establishing a first confidence layer according to the node similarity scores output by the dynamic feature matching network; establishing a second confidence layer in combination with the conflict strength parameter in the difference unit set; The first confidence layer and the second confidence layer are fused to generate a three-dimensional confidence assessment model. When the total score output by the three-dimensional confidence assessment model exceeds a preset threshold, the alignment chain is determined to be a valid alignment chain.

7. The entity alignment method for a multimodal crop knowledge graph according to claim 6, characterized in that: The filtering of conflicting nodes in the candidate alignment set according to the preset three-dimensional confidence evaluation model to generate a valid alignment chain further includes: Collect entity-related feedback information from farmers' operation logs to generate artificial verification factors; injecting the artificial verification factor as a third confidence layer into the three-dimensional confidence assessment model; When the deviation between the output value of the first confidence layer and the output value of the third confidence layer exceeds a preset tolerance, the current multimodal fusion feature map is restored to the previous preset version.

8. The entity alignment method for a multimodal crop knowledge graph according to claim 7, characterized in that: The generating of the atlas update instruction by triggering the incremental feature alignment process based on the difference unit set includes: Extracting new variety features and missing features from the difference unit set to generate a difference feature package; Performing local atlas mapping calculation on the difference feature package to generate an incremental update vector; The incremental update vector is distributed to the knowledge graph nodes in adjacent regions using a preset regional knowledge exchange protocol.

9. The entity alignment method for a multimodal crop knowledge graph according to claim 8, characterized in that: The performing local atlas mapping calculation on the difference feature package to generate an incremental update vector includes: Locking the parent node associated with the difference feature package to generate a feature influence domain; Executing a lightweight graph traversal algorithm within the feature influence domain to generate a local matching result; Verify the topological consistency of the local matching result with the global knowledge graph; An incremental update vector is generated according to the topology verification result.

10. The entity alignment method for a multimodal crop knowledge graph according to claim 1, characterized in that: Reconstructing the knowledge graph topology structure according to the valid alignment chain and the graph update instruction includes: Establishing cross-modal feature connection edges according to the entity relationships in the valid alignment chain; Insert new feature nodes and mark invalid nodes based on the graph update instruction; Eliminate redundant connections and strengthen core feature links through a preset topology optimization algorithm; Generate an updated knowledge graph with version identification and synchronize it to the agricultural decision-making terminal.

Images