Intelligent evaluation method and system for variety regional suitability

Abstract

The invention provides a variety region suitability intelligent evaluation method and system, and relates to the technical field of digital agriculture. The method comprises the following steps: synchronously collecting registration information of an environment sensor, an imaging device and a variety registration database; performing space-time registration and cleaning on the acquired data to obtain a cleaned data set; extracting normalization, time window statistics and domain derivation features from the data, and constructing a unified feature data set; inputting the data into a pre-trained cross-residual attention network model, and outputting adaptability evaluation of each variety in the target area; sorting the evaluation results into a suitability catalogue and pushing the suitability catalogue to a user side; and collecting feedback of actually measured yield, growth period and the like, and iteratively updating model parameters. According to the invention, through multi-source data fusion, dynamic optimization and closed-loop learning, efficient, scientific and real-time selection of regional varieties is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of digital agriculture technology, in particular to a method and system for intelligent evaluation of regional suitability of varieties. BACKGROUND

[0002] With the rapid development of precision agriculture and digital agriculture, agricultural producers increasingly rely on multi-source data to guide variety selection and field management. Environmental information can be obtained in real time through IoT sensor networks, and crop phenotypes can be obtained quickly by high-resolution imaging systems, while variety registration databases provide basic information on genetic characteristics of varieties.

[0003] Artificial intelligence technology is accelerating its penetration into the agricultural sector, and deep learning and multi-modal fusion models can extract complex implicit relationships from large-scale, heterogeneous data to accurately predict regional adaptability and potential yield of varieties. In addition, the rise of open data sharing and agricultural large models makes cross-regional variety evaluation possible.

[0004] Existing evaluations rely heavily on single-dimensional indicators or expert experience, making it difficult to consider real-time environmental fluctuations, dynamic changes in phenotypes, and differences in variety genotypes simultaneously. Model updates lag, leading to delayed adaptability determinations. Results are often presented in scattered reports, lacking standardized variety catalog output mechanisms for decision-making. SUMMARY

[0005] To overcome the shortcomings of the prior art, the present application aims to provide a method and system for intelligent evaluation of regional suitability of varieties, which realizes intelligent evaluation and dynamic optimization of variety adaptability for regional environments, and improves the scientificity, real-time performance and application value of variety selection.

[0006] To achieve the above-mentioned purpose, the present application provides the following solutions:

[0007] A method for intelligent evaluation of regional suitability of varieties, comprising:

[0008] synchronously acquiring environmental data collected by environmental sensors, crop phenotype data collected by imaging devices, and registration information in a variety registration database in a target region;

[0009] performing spatio-temporal registration and data cleaning on the environmental data, the crop phenotype data, and the registration information to obtain a cleaned data set;

[0010] performing feature extraction on the cleaned data set to form a unified feature data set;

[0011] inputting the feature data set into a pre-trained machine learning model to obtain an adaptability evaluation result of the variety in the target region;

[0012] The adaptability evaluation result is arranged as a suitability catalog and pushed to a user end;

[0013] Feedback data of actual planting is collected, and the machine learning model is iteratively updated based on the feedback data.

[0014] Preferably, the environment data collected by the environment sensors in the target area, the crop phenotype data collected by the imaging device, and the registration information in the variety registration database are synchronously acquired, including:

[0015] A plurality of types of environment sensors including temperature and humidity sensors, soil moisture sensors, and light sensors are deployed at preset sampling nodes in the target area, and the environment data is collected by the environment sensors at a set time interval;

[0016] The crop surface chlorophyll fluorescence image and the visible light image are obtained based on a mobile imaging device, and the images are time-stamped to obtain the crop phenotype data;

[0017] The remote variety registration database is accessed through a network interface to retrieve the registration information of the variety code, breeding unit, growth period, and suitable planting area corresponding to the target crop;

[0018] The environment data, crop phenotype data, and registration information are synchronously matched based on a unified time index.

[0019] Preferably, the environment data, the crop phenotype data, and the registration information are subjected to space-time registration and data cleaning to obtain a cleaned data set, including:

[0020] The time stamps of the environment data and the crop phenotype data are corrected at a set time interval as a unified sampling period, and the data sequence is obtained by thinning and resampling to the unified sampling period;

[0021] The global positioning coordinates recorded by the environment sensors and the imaging device are converted to a local coordinate system of a field plot, and a one-to-one spatial mapping table is established based on the field plot number and the variety code to obtain a spatially aligned data sequence; the spatial mapping table is used to associate the time-aligned data sequence with the registration information of the corresponding variety;

[0022] The 3σ principle is applied to the continuous numerical variable in the spatially aligned data sequence to identify an over-threshold record, and the over-threshold record and the sensing or imaging missing record are jointly marked as a missing record;

[0023] repairing the missing records by spline interpolation for continuous numerical variables and by forward fill for discrete categorical variables in the missing records, and deleting the records of a time period corresponding to a plot when the proportion of continuous missing of any variable in the plot exceeds a preset threshold, to obtain the repaired data records;

[0024] complementing the registration information of the counterpart varieties to the repaired data records by using the space mapping table;

[0025] merging the repaired data records according to a unified time index to generate the cleaned data set. Preferably, feature extraction is performed on the cleaned data set to form a unified feature data set, including:

[0026] performing minimum-maximum normalization on continuous numerical variables in the cleaned data set and performing one-hot encoding on discrete categorical variables to obtain data records with unified scales;

[0027] calculating statistics of mean, standard deviation, extreme value and linear trend coefficient for each preset sliding time window based on the data records with unified scales to generate time window statistical features;

[0028] calculating domain features of evapotranspiration stress index, vegetation index and phenotype change rate from the data records with unified scales according to empirical formulas in the agricultural field;

[0029] applying mutual information screening and low-variance elimination rules to the and the domain features to retain features with mutual information scores higher than a set threshold and variances greater than zero;

[0030] concatenating the retained features with feature fields of the registration information of the corresponding varieties according to a unified time index and plot number to form a unified feature vector;

[0031] forming a unified feature data set according to the feature vector.

[0032] Preferably, the feature data set is input into a pre-trained machine learning model to obtain an adaptability evaluation result of the varieties in the target region, including:

[0033] calling a weight file and a hyperparameter configuration of a cross-residual attention network that is offline-trained and version-solidified in a model inference node;

[0034] mapping the unified feature data set into a fixed-length feature vector according to a predetermined field order, and inputting the fixed-length feature vector into a network structure of the cross-residual attention network to obtain the adaptability evaluation result.

[0035] Preferably, the workflow of the cross-residual attention network is:

[0036] performing linear transformation twice and Hadamard product on the fixed-length feature vector to generate an explicit feature interaction term;

[0037] concatenating the explicit feature interaction term with the fixed-length feature vector to obtain an extended vector;

[0038] inputting the extended vector into three fully connected layers in sequence to obtain a nonlinear transformation vector; each of the fully connected layers is followed by ReLU activation and Dropout of 0.2; a residual connection is provided between the output and the input of each of the fully connected layers to realize identity mapping;

[0039] performing multi-head scaled dot-product attention operation on the nonlinear transformation vector to generate a weight vector;

[0040] weighting corresponding feature channels with the weight vector to strengthen the response to high-contribution features to obtain a weighted vector;

[0041] inputting the weighted vector into a fully connected layer with a hidden unit number of 32 and a Sigmoid activation function in sequence to output an adaptive score;

[0042] comparing the adaptive score with a preset suitability threshold to map the suitability evaluation result into four levels of high suitability, moderate suitability, general suitability, or unsuitability.

[0043] Preferably, the suitability evaluation result is sorted into a suitability directory and pushed to the user end, including:

[0044] annotating each variety according to the adaptive score result to obtain an annotated variety result, and associating the variety identification, target region position, and predicted yield information;

[0045] sorting the annotated variety result according to the suitability level to construct a suitability directory containing variety number, suitability level, evaluation time, and predicted yield interval;

[0046] pushing the suitability directory to the user end in a structured data format for the user to browse, download, or further filter and use.

[0047] Preferably, feedback data of actual planting is collected, and the machine learning model is iteratively updated through the feedback data, including:

[0048] After the planting period of the variety ends, the actual yield, growth period, occurrence of diseases and pests, and management measures of the corresponding field are recorded to form a sample of the feedback data;

[0049] Data cleaning and structuring are performed on the feedback data, and the adaptive evaluation result is matched to construct an updated training set containing the input features and actual output control relationship;

[0050] Based on the updated training set, incremental learning is performed on the original machine learning model to update the model parameters, so as to improve the prediction accuracy of the model for similar environment-variety combinations.

[0051] A variety regional suitability intelligent evaluation system comprises:

[0052] A data acquisition unit is configured to synchronously acquire environment data collected by an environment sensor, crop phenotype data collected by an imaging device, and registration information in a variety registration database in a target region;

[0053] A registration and cleaning unit is configured to perform spatio-temporal registration and data cleaning on the environment data, the crop phenotype data, and the registration information to obtain a cleaned data set;

[0054] A feature construction unit is configured to perform feature extraction on the cleaned data set to form a unified feature data set;

[0055] A model evaluation unit is configured to input the feature data set into a pre-trained machine learning model to obtain an adaptive evaluation result of a variety in the target region;

[0056] A catalog generation unit is configured to arrange the adaptive evaluation result into a suitability catalog and push the catalog to a user end;

[0057] A feedback updating unit is configured to collect feedback data of actual planting and iteratively update the machine learning model based on the feedback data.

[0058] The present application discloses the following technical effects:

[0059] The present application realizes synchronous acquisition and unified management of multi-source data by fusing environment data, crop phenotype data, and variety registration information, significantly improving the matching efficiency between regional planting conditions and variety adaptability; by introducing spatio-temporal registration and data cleaning mechanism, the data quality and feature consistency are ensured, providing a reliable foundation for model training; the pre-trained machine learning model is used for adaptability prediction, which not only improves the scientificity and intelligence of variety selection, but also continuously updates the model according to user feedback data, realizes dynamic learning and adaptability optimization under regional differences; finally, the standardized suitability catalog is output, which helps agricultural decision-makers to quickly identify high adaptability varieties, improves the success rate of planting and the stability of yield, and has high practicality and popularization value. BRIEF DESCRIPTION OF DRAWINGS

[0060] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings described below only constitute some of the embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0061] Figure 1 The method flowchart provided for the embodiments of the present application is as shown in

[0062] Figure 2 The system structure schematic diagram provided for the embodiments of the present application is as shown in DETAILED DESCRIPTION

[0063] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments only constitute some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0064] The purpose of the present application is to provide a variety of regional suitability intelligent evaluation method and system, which realizes the intelligent evaluation and dynamic optimization of variety adaptability for regional environment, and improves the scientificity, real-time performance and application value of variety selection.

[0065] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0066] Figure 1 The method flowchart provided for the embodiments of the present application is as shown in Figure 1 The present application provides a variety of regional suitability intelligent evaluation method, which comprises:

[0067] Step 100: synchronously acquiring environmental data collected by an environmental sensor in a target region, crop phenotype data collected by an imaging device, and registration information in a variety registration database;

[0068] Step 200: performing space-time registration and data cleaning on the environmental data, the crop phenotype data and the registration information to obtain a cleaned data set;

[0069] Step 300: performing feature extraction on the cleaned data set to form a unified feature data set;

[0070] Step 400: inputting the feature data set into a pre-trained machine learning model to obtain an adaptability evaluation result of the variety in the target region;

[0071] Step 500: collate the adaptive evaluation results into a suitability catalog and push to the user end;

[0072] Step 600: collect feedback data of actual planting, and update the machine learning model through the feedback data iteration.

[0073] Preferably, the environment data collected by the environment sensor in the target area, the crop phenotype data collected by the imaging device and the registration information in the variety registration database are synchronously acquired, including:

[0074] A plurality of types of environment sensors including temperature and humidity sensors, soil moisture sensors and light sensors are deployed at preset sampling nodes in the target area, and the environment data is collected by the environment sensors at a set time interval;

[0075] The mobile imaging device is used to obtain crop surface chlorophyll fluorescence images and visible light images, and the images are time-stamped to obtain the crop phenotype data;

[0076] The remote variety registration database is accessed through a network interface to retrieve the registration information of the target crop corresponding to the variety code, the breeding unit, the growth period and the suitable planting area;

[0077] The environment data, crop phenotype data and registration information are synchronously matched according to a unified time index.

[0078] Specifically, compared with the traditional single sampling device, the multi-type environment sensors deployed in the field of the present application have their own independent sampling clock. In order to ensure that the data has a unified time reference system in the model input stage, a time anchor point must be established through a master clock issuing mechanism. Specifically, the present application introduces a synchronization mechanism with an edge computing gateway as the center node: the node periodically broadcasts a time synchronization signal (which can be based on NTP or GPS time signal) to each sensor terminal, and the sensor collects data and synchronously timestamps, and uploads it through a local bus or a low-power wide-area network. The sampling frequency and the model time window are consistent (for example, 10 minutes), which avoids subsequent interpolation error diffusion. This synchronous collection framework provides a guarantee for the time feature stability of the machine learning model of the present application.

[0079] Optionally, for the chlorophyll fluorescence image and the visible light image taken by the imaging device, the application emphasizes the "time-field-crop" ternary binding mechanism, which is the important basis for ensuring that the image can correctly correspond to the target crop. During the shooting process, the mobile imaging device (such as a ground unmanned platform or a multi-rotor unmanned aerial vehicle carrying an image module) needs to record the following parameters in real time and embed them in the image metadata: imaging timestamp (ms precision); imaging location information (GPS coordinates + operation trajectory index); crop row index (combined with operation trajectory, camera field of view width, and field path rules calculation); imaging device ID and parameter configuration identifier (to ensure the traceability of image acquisition settings). Through the above binding strategy, it can be ensured that the subsequent image data and sensing data are aligned under the condition of the same field and the same time, and are positioned to the crop row or even the single plant level. This mechanism can prevent the risk of imaging misallocation and misallocation, and is a prerequisite for image phenotype data to be effectively utilized in the machine learning pipeline.

[0080] Optionally, when adjusting the variety registration information, the application does not only extract the variety name through manual input or annotation, but also relies on field code and sowing records, and combines image auxiliary identification mechanism to realize automatic traceability of variety code. The specific process is as follows:

[0081] Based on the planting task list provided by the field management platform, the crop variety code sown in each field is obtained;

[0082] If it is missing or not clearly marked, the variety category is automatically identified through an image recognition model (using sowing label cards or plant characteristics);

[0083] Once the variety code is confirmed, only the following key fields are called from the remote variety registration database: growth period, suitable planting area, resistance level, and breeding unit;

[0084] The fields are mapped to a unified data structure and bound to the corresponding field and sampling period as one of the machine learning input features.

[0085] Preferably, the environmental data, the crop phenotype data and the registration information are subjected to spatio-temporal registration and data cleaning to obtain a cleaned data set, comprising:

[0086] The time stamps of the environmental data and the crop phenotype data are corrected at a set time interval as a unified sampling period, and the data sequence is obtained by thinning and resampling to the unified sampling period;

[0087] The global positioning coordinates recorded by the environmental sensor and the imaging device are converted into a field local coordinate system, and a one-to-one spatial mapping table is established according to the field number and the variety code to obtain a spatially aligned data sequence; the spatial mapping table is used to associate the time-aligned data sequence with the registration information of the corresponding variety.

[0088] applying 3σ principle to continuous numerical variables in the spatial alignment data sequence to identify super-threshold records, and jointly marking the super-threshold records and sensing or imaging missing records as missing records;

[0089] repairing the missing records by using spline interpolation for continuous numerical variables in the missing records and by using forward filling for discrete classification variables in the missing records, and deleting the records of the time period corresponding to the field for each field when the proportion of continuous missing of any variable in the field exceeds a preset threshold, to obtain the data records after missing repair;

[0090] complementing the registration information of the matching varieties to the data records after missing repair by using the spatial mapping table;

[0091] merging the data records after missing repair according to a unified time index to generate the cleaned data set.

[0092] Optionally, in order to eliminate the respective clock error of the heterogeneous collection terminal, the application deploys a master clock synchronization service on the edge computing gateway, corrects the local clock of each sensor and imaging device by GPS time service or NTP protocol regularly. The original environmental data and crop phenotype image data collected are reported after carrying the local time stamp. Subsequently, the system takes the above-mentioned "set time interval" as a unified sampling period, and converts the original time series data into a time alignment data sequence strictly corresponding to the unified sampling period by a thinning resampling algorithm (for example, taking an integer minute or linear interpolation). This process ensures that the subsequent model input has comparable multi-source data support at each time node, which is a necessary link to realize fine time series analysis and cannot be replaced by experience alignment or manual annotation.

[0093] In order to accurately map the global positioning coordinates to the specific field, the application adopts a two-level coordinate conversion mechanism: first, project the GPS coordinates of the sensor and the imaging device into the local geodetic coordinate system (such as WGS84→CGCS2000), and then complete the conversion from the geodetic coordinate to the "field number" label by using the ray method to determine the point falling into the polygon boundary library generated by measuring the boundary of the field in advance. According to the sowing account or label information, the system maintains a relationship table (spatial mapping table). Through this table, the application binds the generated time alignment data sequence with the registration information of the corresponding varieties one by one to form a spatial alignment data sequence, which ensures that each record is clearly associated with the specific field and variety attribute, which is a key technical guarantee to realize the variety level analysis.

[0094] ​The present application introduces a 3σ algorithm based on dynamic threshold for continuous numerical indicators (such as temperature, humidity, light intensity, fluorescence intensity, etc.) in the spatial alignment data sequence, calculates the local mean and standard deviation through windowing, automatically identifies the deviation points that exceed the threshold, and marks them as missing records together with the missing records caused by communication failure or obstruction. Unlike traditional global threshold, the dynamic threshold of the present application can adapt to seasonal and variety differences, improving the identification accuracy. For the marked missing records, the continuous numerical variables are restored by cubic spline interpolation, maintaining the smoothness of the data curve; the discrete categorical variables (such as growth stage labels) are restored based on the time series forward filling strategy to avoid unreasonable stage jumps; if the length of continuous missing of a variable in the same field exceeds the preset (such as 10% of the sampling period), all data in that time period are automatically removed to prevent over-interpolation affecting model training.

[0095] After obtaining the missing repaired data records, the present embodiment again calls the spatial mapping table to accurately supplement the corresponding variety registration information (growth period, suitable area, etc.) to each spatio-temporal alignment record. For the possible field missing or version conflict in the registration information, the present application designs a "minimum update priority" strategy: preferentially using the latest registration version to fill in the missing, if there is a version inconsistency, the conflicting entry is recorded in the verification log for manual review and inclusion in the historical case library. Finally, all environmental, phenotypic and registration fields are merged according to the unified time index to generate a cleaned data set for feature extraction and model training. This fusion step ensures that multi-source data is presented completely in the same record, which is the only legal input for subsequent intelligent evaluation.

[0096] Preferably, feature extraction is performed on the cleaned data set to form a unified feature data set, including:

[0097] The continuous numerical variables in the cleaned data set are normalized by minimum-maximum, and the discrete categorical variables are executed by one-hot encoding to obtain scale-unified data records;

[0098] Based on the scale-unified data records, statistical quantities of mean, standard deviation, extreme value and linear trend coefficient are calculated for each preset sliding time window to generate time window statistical features;

[0099] According to the empirical formula in the agricultural field, the scale-unified data records are used to calculate the evapotranspiration stress index, the vegetation index and the field features of the phenotypic change rate;

[0100] Mutual information screening and low variance rejection rules are applied to the and the field features to retain features with mutual information scores higher than a set threshold and variances greater than zero;

[0101] The reserved features are spliced with the feature fields of the registration information of the corresponding variety according to unified time indexing and field block numbering, to form a unified feature vector;

[0102] A unified feature dataset is formed according to the feature vector.

[0103] Preferably, the feature dataset is input into a pre-trained machine learning model to obtain an adaptability evaluation result of the variety in the target region, including:

[0104] The weight file and the hyperparameter configuration of the cross-residual attention network which is completed offline and version solidified are called in the model inference node;

[0105] The unified feature dataset is mapped into a fixed-length feature vector according to a predetermined field order, and the fixed-length feature vector is input into the network structure of the cross-residual attention network to obtain the adaptability evaluation result.

[0106] Specifically, the minimum-maximum normalization algorithm is called for each continuous numerical variable (such as temperature, humidity, chlorophyll fluorescence intensity, etc.) in the cleaning dataset, which is linearly mapped to the interval [0, 1]; and one-hot encoding is performed on each discrete classification variable (such as growth stage, pest and disease grade), which is mapped into a binary vector. This module manages the mapping upper and lower limits through a unified parameter file to ensure that variables from different sources have comparability in the same numerical scale, which is the basis for subsequent fusion statistics and domain feature calculation.

[0107] For each time series data of each field, the embodiment adopts a double-ended queue structure to maintain a preset sliding window (for example, 24 hours or 7 days). During the window sliding process, the mean, standard deviation, maximum value, minimum value and linear regression trend coefficient of each normalized variable in the window are calculated in real time, and these statistics are output in time sequence as a "time window statistical feature" array. This dynamic calculation mechanism significantly improves the ability to capture short-term environmental fluctuations and crop response relationships compared to one-time statistics, which is a key link to realize complex dynamic adaptability analysis. The embodiment further integrates a formula library in the agricultural field to call calculation functions such as evapotranspiration stress index (ET0 / crop evapotranspiration model), vegetation index (such as NDVI) and phenotype change rate (based on the difference between two consecutive frames of fluorescence intensity and the ratio of time difference) for scale-unified data records, to generate "field features" representing crop physiological conditions and environmental stress. This link converts raw numerical values into high semantic value indicators, making up for the lack of understanding of crop growth mechanisms with only statistical features. The embodiment sorts the mutual information scores between each feature and the target adaptability score based on the mutual information evaluator for the full set of candidate features (i.e., field features) generated by the above steps, and removes redundant features with scores below a preset threshold or zero variance. The remaining features are then spliced with the variety registration information (fields such as growth period length and suitable area label after encoding) according to consistent time index and field number to form a final fixed-length feature vector set, i.e., a "unified feature dataset".

[0108] Preferably, the workflow of the cross-residual attention network is as follows:

[0109] Performing linear transformation twice on the fixed-length feature vector and taking Hadamard product to generate explicit feature interaction items;

[0110] Concatenating the explicit feature interaction items with the fixed-length feature vector to obtain an extended vector;

[0111] Inputting the extended vector into three fully connected layers in turn to obtain a nonlinear transformation vector; each fully connected layer is followed by ReLU activation and Dropout of 0.2; a residual connection is set between the output and the input of each fully connected layer to realize identity mapping;

[0112] Performing multi-head scaled dot-product attention operation on the nonlinear transformation vector to generate a weight vector;

[0113] Weighting the corresponding feature channels with the weight vector to strengthen the response to high-contribution features to obtain a weighted vector;

[0114] Inputting the weighted vector into a fully connected layer with a hidden unit number of 32 and a Sigmoid activation function in turn to output an adaptability score;

[0115] The adaptability score is compared with a preset suitability threshold, and the adaptability evaluation result is mapped into four levels of high suitability, relatively suitable, general or unsuitable.

[0116] Optionally, the network of the embodiment first performs two independent linear mappings (weight matrices W1 and W2) on the input fixed-length feature vector X respectively to obtain intermediate representations X1 = W1X + b1 and X2 = W2X + b2. b1 and b2 are trainable bias parameters matched with the corresponding weight matrices.

[0117] Subsequently, the Hadamard product X cross = X1 ⊙ X2 is used to explicitly construct feature interaction items, thereby capturing second-order cross information between original features; this step replaces the traditional implicit cross layer and can realize high-dimensional feature fusion with low complexity, which is a key innovative point to improve the expression ability of the model. Then, X cross is concatenated with the original vector X in the feature dimension to form an extended vector X' = [X; X cross ], which provides more abundant input dimensions for subsequent deep nonlinear transformation.

[0118] The extended vector X' sequentially passes through three groups of substructures composed of fully connected layers, ReLU activation functions and Dropout layers. The output of each group of fully connected layers is added to the corresponding input through a skip connection to realize an identity mapping. This "multi-layer residual" design not only avoids gradient disappearance of deep networks, but also retains shallow feature information, ensuring that original feature expression will not be lost when deepening the network depth. Compared with simply stacking nonlinear layers, the residual structure of the present application can significantly enhance the fitting ability for complex feature patterns while maintaining computational efficiency.

[0119] On the basis of the nonlinearly transformed vector, the present application introduces a multi-head scaled dot-product attention mechanism: first, the query, key and value projections of multiple attention heads are calculated, and the attention weight matrix is generated according to the scaled dot-product formula. After Softmax normalization, the final weight vector W att is obtained. Subsequently, W att is used to weight the aforementioned nonlinearly transformed vector in the feature channel to form a weighted vector X att , thereby strengthening the response to key contribution features. Finally, X att is input into a fully connected layer with 32 hidden units and is activated by Sigmoid to output an adaptability score S ∈ [0, 1]; by comparing with a preset threshold, the adaptability score is mapped into a four-level evaluation result of high suitability, relatively suitable, general or unsuitable. The combination design of the above attention weighting and small output layer realizes dynamic emphasis on the importance of different features while maintaining lightweight computation.

[0120] In the present embodiment, to map the adaptability score value S output by the model to the four-level classification of high suitability, moderate suitability, general suitability, or unsuitability, the system sets preset threshold values based on quantile rules of historical regional planting data: specifically, taking the average yield of the same variety in the region in the past three years and the model score as the reference, the statistical distribution of the score S in the training set is divided into four segments, where the variety with a score greater than 0.85 is defined as "high suitability", the score between 0.70-0.85 is "moderate suitability", the score between 0.50-0.70 is "general", and the score below 0.50 is "unsuitable"; the threshold values can be dynamically adjusted according to the characteristics of the region or user demand, and the default values are set to [0.50, 0.70, 0.85] after fitting with historical data. The above scoring classification criteria ensure that the model evaluation results have statistical consistency with the actual yield performance, and have strong interpretability, which is convenient for agricultural management personnel to make decisions.

[0121] Preferably, the adaptability evaluation results are sorted into a suitability directory and pushed to the user end, including:

[0122] According to the adaptability score results, each variety is labeled with the corresponding adaptability level, and the variety identification, target region location, and predicted yield information are associated to obtain the labeled variety results;

[0123] The labeled variety results are sorted by adaptability level to construct a suitability directory containing variety number, adaptability level, evaluation time, and predicted yield interval;

[0124] The suitability directory is pushed to the user end in a structured data format for the user to browse, download, or further filter and use.

[0125] In the embodiment, the catalog generation unit first labels each variety as "highly suitable", "relatively suitable", "average" or "unsuitable" according to the dynamic mapping rule of the adaptability score S output by the model evaluation unit and the preset grading threshold, and automatically associates the internal unique identifier of the variety, the corresponding field or region coordinates, and the upper and lower limits of the predicted yield calculated based on the model confidence interval, to generate the labeled variety data record; then, the catalog generation unit calls a lightweight sorting service to sort in batches with the adaptability level as the primary key ("highly suitable" highest, "unsuitable" lowest), and arrange the varieties in the same level in descending order of the upper limit of the predicted yield, and finally output an ordered list containing fields such as "variety number", "adaptability level", "evaluation timestamp", "predicted yield lower-upper limit", etc.; finally, the catalog generation unit serializes the ordered list into standardized structured data according to the predefined JSON template, and pushes it to the user-side display layer through the system internal message bus or RESTful interface at one time, ensuring that the catalog data can be loaded, browsed, downloaded or further filtered in the Web interface, mobile App or third-party API call.

[0126] Preferably, feedback data of actual planting is collected, and the machine learning model is iteratively updated based on the feedback data, including:

[0127] After the planting period of the variety ends, the actual yield, growth period, occurrence of diseases and pests, and management measures of the corresponding field are recorded to form a sample of the feedback data;

[0128] The feedback data is cleaned and structured, and matched with the adaptability evaluation results to build an updated training set containing input features and actual output control relationships;

[0129] Based on the updated training set, incremental learning is performed on the original machine learning model to update the model parameters, so as to improve the prediction accuracy of the model for similar environment-variety combinations.

[0130] In the embodiment, for feedback update after the planting period of the variety ends, the data recovery process is automatically triggered through the interface with the field operation management system to accurately obtain the actual yield, growth period, occurrence of diseases and pests, and corresponding control records of each field, and the unique field number and harvest timestamp are attached to realize automatic and seamless collection of feedback data. This triggering mechanism eliminates the time lag and omission of traditional manual registration, which is a key technical link to ensure the timeliness and accuracy of feedback samples.

[0131] Subsequently, the embodiment uses the aforementioned spatio-temporal registration framework to align the feedback data to the original evaluation records in the field block through the master clock synchronization and spatial mapping table, and adopts an intelligent anomaly detection strategy based on the model residual distribution: for continuous fields such as yield and growth period, the local distribution of the residual between the model predicted value and the measured value is calculated to identify outliers beyond the dynamic upper and lower limits; for outliers and missing items, a residual-driven local regression completion method is used to fill in, keeping the data trend consistent; for classification fields such as pest and management records, forward inference repair based on similar field blocks or adjacent time periods is introduced; this intelligent completion process breaks through the simple interpolation or lookup table method, significantly improves the quality of feedback data, and ensures the integrity and reliability of the updated training set.

[0132] Finally, the embodiment performs incremental fine-tuning based on sample importance on the original cross-residual attention network: first, the environment-variety combinations with large prediction errors in the updated training set are oversampled to enhance the model's learning of weak intervals; then only the parameters of the attention sublayer and the output mapping layer in the network are updated in small batches online, and the learning rate and dropout rate are controlled to ensure that the generalization performance of the original model is not affected; after each fine-tuning, a new model version is generated and verified in the inference environment to achieve continuous and adaptive accurate prediction capability.

[0133] In a preferred embodiment of the present application, the machine learning model adopts a cross-residual attention network structure for high-order interaction modeling and dynamic adjustment of feature channel weights on the input multi-dimensional feature vector, thereby outputting the adaptability score of the variety in the regional environment. However, those skilled in the art should understand that, under the premise of meeting the multi-source heterogeneous data expression and regression prediction task, the machine learning model can also use other known supervised learning structures, such as integrated learning models based on Gradient Boosting Decision Tree (GBDT), XGBoost or LightGBM, to complete adaptability score prediction using their non-linear feature and feature importance evaluation capabilities, especially for regional analysis in small sample or feature sparse scenarios.

[0134] Further, in the deep learning modeling scenario, the machine learning model can also be replaced by a multi-layer perception (MLP), a gated recurrent unit network (GRU), or a long short-term memory network (LSTM) structure, among others, where LSTM is particularly suitable for modeling time series features such as environmental change trends within a sliding window; for tasks that require simultaneous extraction of spatial and temporal patterns, a hybrid neural network structure that combines one-dimensional convolution and attention mechanism can also be used. The above-mentioned various network structures can be flexibly selected according to the data feature dimension, sample quantity, and model precision requirement, and all constitute feasible technical implementation ways of the present application.

[0135] Corresponding to the above method, as shown in the embodiment, a variety of regional suitability intelligent evaluation system is also provided, comprising: Figure 2 A data acquisition unit is configured to synchronously acquire environmental data collected by an environmental sensor, crop phenotype data collected by an imaging device, and registration information in a variety of registration database in a target region;

[0136] A registration and cleaning unit is configured to perform spatio-temporal registration and data cleaning on the environmental data, the crop phenotype data, and the registration information to obtain a cleaned data set;

[0137] A feature construction unit is configured to perform feature extraction on the cleaned data set to form a unified feature data set;

[0138] A model evaluation unit is configured to input the feature data set into a pre-trained machine learning model to obtain an adaptability evaluation result of the variety in the target region;

[0139] A catalog generation unit is configured to arrange the adaptability evaluation result into a suitability catalog and push the catalog to a user end;

[0140] A feedback updating unit is configured to collect feedback data of actual planting and iteratively update the machine learning model based on the feedback data.

[0141]

[0142] The beneficial effects of the present application are as follows:

[0143] (1) The present application synchronously collects environmental sensor data, crop phenotype imaging data, and variety registration information, realizes integrated management and fusion of multi-source heterogeneous information, and provides comprehensive, real-time, and high-quality input data sources for regional variety suitability evaluation. By using master clock synchronization and spatial mapping table mechanisms, the accurate alignment of various data in time and space is ensured, and the model input noise caused by different data synchronization or position error is greatly reduced, thereby laying a solid data foundation for subsequent analysis and prediction.

[0144] (2) In the cleaning and registration link, the present application adopts dynamic threshold abnormality identification, spline interpolation repair, and residual-driven intelligent completion technology, designs special repair strategies for continuous and discrete variables respectively, and combines with preset rules to remove serious missing sections, which not only retains the key fluctuation characteristics of real data, but also effectively removes abnormality and noise, significantly improves the integrity and reliability of the cleaned data set, and provides an irreplaceable high-quality data set for feature extraction and model training.

[0145] ​(3) For feature construction, the present application innovatively integrates minimum-maximum normalization, sliding time window statistics, agricultural field derived indicators and mutual information screening mechanism, which not only retains the core dynamic information of environment and phenotype data, but also generates physiological stress and growth state features with high semantic value through field formula, and finally inputs the cross-residual attention network with a simplified and efficient feature vector, realizes the deep fusion and expression of complex multi-dimensional information, and significantly improves the generalization ability and interpretability of the model under variable environment.

[0146] (4) The cross-residual attention network used in the present application combines feature interaction, deep residual connection and multi-head attention mechanism, which not only improves the ability to capture and strengthen high-dimensional cross information, but also realizes the rapid adaptation and continuous optimization of the model to the latest feedback data through online incremental fine-tuning and resampling strategy based on sample importance, so that the variety suitability evaluation can be updated iteratively with planting practice, and finally outputs the standardized suitability directory, providing scientific, real-time and operable variety selection guidance for agricultural decision makers, and helping efficient and stable yield.

[0147] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments, and the same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the related parts can be referred to the method part.

[0148] In this paper, specific examples are used to illustrate the principles and implementation methods of the present application, and the above examples are only used to help understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation method and application range will be changed. In summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A method for intelligent assessment of regional suitability of a variety, characterized in that, include: Simultaneously acquire environmental data collected by environmental sensors in the target area, crop phenotypic data collected by imaging devices, and registration information from the variety registration database; Spatiotemporal registration and data cleaning are performed on the environmental data, crop phenotypic data, and registration information to obtain a cleaned dataset; Feature extraction is performed on the cleaned dataset to form a unified feature dataset; The feature dataset is input into a pre-trained machine learning model to obtain the adaptability evaluation results of the variety in the target region; The results of the adaptive evaluation are compiled into a suitability catalog and pushed to the user terminal; Collect feedback data from actual planting, and iteratively update the machine learning model using the feedback data.

2. The intelligent assessment method for regional suitability of varieties according to claim 1, characterized in that, Simultaneously acquire environmental data collected by environmental sensors in the target area, crop phenotypic data collected by imaging devices, and registration information from the variety registration database, including: Sampling nodes are preset within the target area, and multiple types of environmental sensors, including temperature and humidity sensors, soil moisture sensors, and light sensors, are deployed. Environmental data is collected at set time intervals based on the environmental sensors. The crop surface chlorophyll fluorescence image and visible light image are acquired using a mobile imaging device, and the images are timestamped to obtain the crop phenotypic data. Access the remote variety registration database via network interface to retrieve the registration information corresponding to the target crop, including variety code, breeding unit, growth period, and suitable planting area. The environmental data, crop phenotypic data, and registration information are synchronized and matched using a unified time index.

3. The intelligent assessment method for regional suitability of varieties according to claim 1, characterized in that, Spatiotemporal registration and data cleaning are performed on the environmental data, the crop phenotypic data, and the registration information to obtain a cleaned dataset, including: Using a set time interval as a unified sampling period, the timestamps of the environmental data and the crop phenotypic data are corrected, and time-aligned data sequences are obtained by sparse resampling to the unified sampling period. The global positioning coordinates recorded by the environmental sensor and the imaging device are converted into the local coordinate system of the field, and a one-to-one spatial mapping table is established based on the field number and variety code to obtain a spatially aligned data sequence; the spatial mapping table is used to associate the time-aligned data sequence with the registration information of the corresponding variety. The 3σ principle is applied to the continuous numerical variables in the spatially aligned data sequence to identify records exceeding the threshold, and the records exceeding the threshold and the missing records of sensing or imaging are jointly marked as missing records; Spline interpolation is used to repair continuous numerical variables in the missing records, and forward filling is used to repair discrete categorical variables in the missing records. When the proportion of continuous missing variables in any field exceeds a preset threshold, the records of the corresponding time period of the field are deleted to obtain the missing data records. The registration information of the corresponding varieties is supplemented into the missing data record using the spatial mapping table; The missing and repaired data records are merged according to a unified time index to generate the cleaned dataset.

4. The intelligent assessment method for regional suitability of varieties according to claim 1, characterized in that, Feature extraction is performed on the cleaned dataset to form a unified feature dataset, including: Min-max normalization is applied to continuous numerical variables in the cleaned dataset, and one-hot encoding is performed on discrete categorical variables to obtain data records with uniform scale. Based on the data records with the uniform scale, the statistical properties of mean, standard deviation, extreme values ​​and linear trend coefficient are calculated for each preset sliding time window to generate time window statistical features; Based on empirical formulas in the agricultural field, the domain characteristics of evapotranspiration stress index, vegetation index, and phenotypic change rate are calculated using data records of the aforementioned uniform scale. Mutual information filtering and low variance elimination rules are applied to the aforementioned and the aforementioned domain features to retain features with mutual information scores higher than a set threshold and variance greater than zero. The retained features are concatenated with the feature fields of the corresponding variety's registration information according to the unified time index and field number to form a unified feature vector; A unified feature dataset is constructed based on the feature vectors.

5. The intelligent assessment method for regional suitability of varieties according to claim 1, characterized in that, The feature dataset is input into a pre-trained machine learning model to obtain the adaptability evaluation results of the variety in the target region, including: The weight file and hyperparameter configuration of the cross-residual attention network, which has been trained offline and is version-fixed, are called in the model inference node; The unified feature dataset is mapped into a fixed-length feature vector according to a predetermined field order, and the fixed-length feature vector is fed into the network structure of the cross-residual attention network to obtain the adaptive evaluation result.

6. The intelligent assessment method for regional suitability of varieties according to claim 5, characterized in that, The workflow of the cross-residual attention network is as follows: The fixed-length feature vector is subjected to two linear transformations and a Hadamard product to generate explicit feature interaction terms; The explicit feature interaction term is concatenated with the fixed-length feature vector to obtain an extended vector; The extended vector is sequentially input into three fully connected layers to obtain a nonlinear transformation vector; each fully connected layer is followed by ReLU activation and Dropout of 0.2; a residual connection is set between the output and input of each fully connected layer to achieve an identity mapping; Perform multi-head scaling dot product attention operation on the nonlinear transformation vector to generate a weight vector; The corresponding feature channels are weighted using the weight vector to enhance the response to high-contribution features, thereby obtaining a weighted vector. The weighted vector is sequentially input into a fully connected layer with 32 hidden units and a Sigmoid activation function, and an adaptive score is output. The adaptability score is compared with a preset adaptability threshold and mapped to an adaptability evaluation result of four levels: high adaptability, relatively suitable, average, or unsuitable.

7. The intelligent assessment method for regional suitability of varieties according to claim 1, characterized in that, The adaptive evaluation results are compiled into a suitability catalog and pushed to the user terminal, including: Based on the adaptability score results, each variety is labeled with a corresponding adaptability level, and associated with the variety identifier, target area location, and predicted yield information to obtain the labeled variety results; The labeled variety results are sorted according to their adaptability level, and a suitability catalog containing variety number, adaptability level, evaluation time, and predicted yield range is constructed. The suitability catalog is pushed to the user terminal in a structured data format for the user to browse, download, or further filter.

8. The intelligent assessment method for regional suitability of varieties according to claim 1, characterized in that, Collecting feedback data from actual planting and iteratively updating the machine learning model using the feedback data includes: After the planting cycle of the variety ends, the actual yield, growth period, occurrence of diseases and pests and management measures of the corresponding field are collected to form a sample of the feedback data; The feedback data is cleaned and structured, and then matched with the adaptive evaluation results to construct an updated training set containing the relationship between input features and actual output. Incremental learning is performed on the original machine learning model based on the updated training set to update the model parameters and improve the model's prediction accuracy for similar environment-variety combinations.

9. A variety regional suitability intelligent assessment system, characterized in that, include: The data acquisition unit is used to simultaneously acquire environmental data collected by environmental sensors in the target area, crop phenotypic data collected by imaging devices, and registration information in the variety registration database; The registration and cleaning unit is used to perform spatiotemporal registration and data cleaning on the environmental data, the crop phenotypic data and the registration information to obtain a cleaned dataset. The feature construction unit is used to extract features from the cleaned dataset to form a unified feature dataset; The model evaluation unit is used to input the feature dataset into a pre-trained machine learning model to obtain the adaptability evaluation results of the variety in the target region. The catalog generation unit is used to organize the adaptive evaluation results into a suitability catalog and push it to the user terminal; The feedback update unit is used to collect feedback data from actual planting and iteratively update the machine learning model using the feedback data.

Images