A neural network-based intelligent crop recommendation system

Abstract

The application discloses a kind of crop intelligent recommendation system based on neural network, including following module: agricultural multi-source data acquisition module, and agricultural data in different geographical regions and seasons are collected and preprocessed;Cross-environment data division module, form multiple independent environment data sets;Deep feature extraction network module, extract high-dimensional semantic features;Invariant constraint optimization module, construct unified stable feature model satisfying the generalization requirement across regions and seasons;Generalization performance verification module, quantitatively evaluate model generalization accuracy;Personalized crop recommendation decision module, generate personalized crop recommendation list;Recommended result visualization interaction module, show recommended result to user.The application significantly improves the cross-environment generalization ability and personalized recommendation accuracy of crop recommendation system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent agriculture, and in particular to a crop intelligent recommendation system based on a neural network. BACKGROUND

[0002] Crop intelligent recommendation technology is one of the important branches in the field of modern intelligent agriculture, and the technical core lies in using data-driven methods to provide scientific and reliable crop planting recommendations for farmers according to different regional and seasonal agricultural environmental conditions. The current main technical route relies on traditional machine learning and deep learning methods. The traditional machine learning method usually extracts soil, meteorological, and historical agricultural operation knowledge features by artificial means, and uses a classification model to predict the recommendation task, which has the advantages of strong interpretability and low computational cost. However, due to its feature dependence on expert experience, the model generalization ability is obviously insufficient when facing complex environmental changes or unknown features.

[0003] With the maturity of deep learning technology, in recent years, crop recommendation systems based on multi-layer neural network structures have appeared, which extract high-dimensional feature representations from raw data and use the powerful nonlinear modeling capability of neural networks to improve recommendation accuracy and generalization performance. It has become one of the mainstream technologies in the field of crop intelligent recommendation. However, the existing neural network methods generally have a large difference in data distribution in actual cross-regional and cross-seasonal applications. The climate, soil properties, and planting habits of different regions differ significantly, resulting in poor generalization performance of traditional deep learning recommendation models under cross-environmental conditions, significantly reducing the accuracy and reliability of the recommendation results, and directly restricting the widespread application of intelligent recommendation technology in agricultural production practice. In addition, the existing technology lacks sufficient interactive verification mechanism between the recommendation results and the actual needs of users, and cannot fully utilize user feedback to dynamically optimize the quality of the model recommendation.

[0004] Therefore, how to provide a crop intelligent recommendation system based on a neural network is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide a crop intelligent recommendation system based on a neural network. In view of the problem that the generalization performance of existing crop recommendation methods significantly decreases under cross-regional and cross-seasonal conditions, an improved invariant risk minimization algorithm and a deep feature extraction network are fused to obtain a unified stable feature representation model through a feature invariance constraint optimization module, thereby realizing individualized and accurate crop recommendation. The present application has the advantages of high generalization performance, high recommendation accuracy, and strong reliability in practical application.

[0006] According to the crop intelligent recommendation system based on a neural network of the present application, the system comprises:

[0007] An agricultural multi-source data acquisition module is configured to collect multi-source data in different geographical regions and different seasons to form a standardized agricultural multi-source data set.

[0008] A cross-environment data division module is configured to receive the standardized agricultural multi-source data set and divide the data set according to geographical region characteristics and seasonal time characteristics to form a plurality of independent environment data sets.

[0009] A deep feature extraction network module is configured to receive the plurality of independent environment data sets and perform feature embedding and feature extraction processing on each independent environment data set to obtain corresponding high-dimensional semantic feature vectors.

[0010] A feature invariance constraint optimization module is configured to perform optimization constraint processing on differences between high-dimensional semantic feature vectors of different independent environment data sets by using an improved invariance risk minimization algorithm to obtain a unified stable feature representation model.

[0011] A cross-environment generalization performance verification module is configured to receive the unified stable feature representation model and perform generalization capability testing by using pre-divided cross-region and cross-season verification data sets.

[0012] An individualized crop recommendation decision module is configured to receive the unified stable feature representation model that passes the generalization capability testing, perform inference calculation on real-time agricultural data of a target region and a target season, and obtain an individualized crop recommendation list.

[0013] A recommendation result visualization interaction module is configured to receive the individualized crop recommendation list, perform data visualization processing and human-computer interaction interface processing, and form a recommendation result visualization page.

[0014] Optionally, the agricultural multi-source data acquisition module comprises:

[0015] Soil property data, meteorological condition data, crop planting history data, and farmland management measure data in different geographical regions and different seasons are collected, and after missing value detection, abnormal value detection, and data format conversion are performed on all collected data, a standardized agricultural multi-source data set is obtained.

[0016] Optionally, the cross-environment data division module comprises:

[0017] A region-season combination matrix is established according to geographical region characteristics and seasonal time characteristics in the standardized agricultural multi-source data set, and cross mapping is performed.

[0018] Data sample amounts corresponding to each environment combination in the region-season combination matrix after cross mapping are analyzed, and environment combinations with data sample amounts lower than a preset sample threshold are removed.

[0019] For the environmental combination whose sample size meets the preset sample threshold, the statistical distribution characteristics of key features in the soil property data, meteorological condition data and crop planting history data in the environmental combination are calculated, and the consistency of the data in the environmental combination is evaluated by taking the statistical distribution characteristic similarity as an index;

[0020] The environmental combination whose statistical distribution characteristic similarity does not meet the consistency requirement is segmented and adjusted until the data consistency in each environmental combination meets the preset statistical distribution similarity threshold;

[0021] Each environmental combination whose statistical distribution characteristic similarity meets the data consistency requirement is respectively given a region mark and a season mark, and a plurality of independent environmental data sets with region and season identification are formed.

[0022] Optionally, the deep feature extraction network module comprises:

[0023] The soil property data, meteorological condition data, crop planting history data and farmland management measure data in each independent environmental data set are respectively subjected to numerical normalization processing;

[0024] Independent feature embedding networks are respectively constructed for the normalized soil property data, meteorological condition data, crop planting history data and farmland management measure data, so as to obtain low-dimensional embedding feature representations corresponding to each data type;

[0025] The low-dimensional embedding feature representations of each data type are respectively calculated through an attention mechanism to obtain corresponding fusion weights, and feature splicing is performed according to the fusion weights to form a fusion feature vector of each independent environmental data set;

[0026] A deep feature extraction network with alternately arranged convolutional neural network layers and fully connected neural network layers is built for the fusion feature vector, wherein the convolutional neural network layers are used to extract local feature correlations in the fusion feature vector, and the fully connected neural network layers are used to realize global nonlinear mapping of the local features;

[0027] A dynamic feature calibration structure is arranged between the convolutional neural network layers and the fully connected neural network layers, and dynamic standardization processing is implemented by calculating the statistical mean and variance of the output features of each network layer;

[0028] A global adaptive pooling operation is performed on the feature vector processed by the dynamic feature calibration structure to obtain a high-dimensional semantic feature vector.

[0029] Optionally, the independent feature embedding networks are respectively constructed as follows:

[0030] The feature embedding network constructed for the soil attribute data comprises an input layer and three full-connection hidden layers with dimension decreasing layer by layer, the number of neurons of the input layer is consistent with the feature dimension of the soil attribute data, the number of neurons of the three full-connection hidden layers decreases by half, quarter and eighth respectively, and the ReLU activation function is used to perform nonlinear mapping after each hidden layer, and finally a low-dimensional embedding feature representation with eighth of the input layer is obtained;

[0031] The feature embedding network constructed for the weather condition data comprises an input layer, two convolutional neural network layers and a full-connection neural network layer, the number of neurons of the input layer is consistent with the feature dimension of the weather condition data, the two convolutional neural network layers adopt one-dimensional convolution structure, the convolution kernel size of each convolutional neural network layer is 3 and the step is 1, and the ReLU activation function is used to process the convolution output, a global maximum pooling layer is connected after the second convolutional neural network layer to reduce the dimension, and the full-connection neural network layer is used to reduce the dimension of the pooled output to obtain a low-dimensional embedding feature representation with fixed dimension;

[0032] The feature embedding network constructed for the crop planting history data comprises an input layer, two long short-term memory network hidden layers and a full-connection neural network layer, the number of neurons of the input layer is consistent with the feature dimension of the planting history data, the number of neurons of the two long short-term memory network hidden layers decreases layer by layer, the output feature dimension of the first long short-term memory network hidden layer is half of the input dimension, the output feature dimension of the second long short-term memory network hidden layer is quarter of the input dimension, and the long short-term memory network output feature is further reduced in dimension by the full-connection neural network layer to obtain a low-dimensional embedding feature representation with fixed dimension;

[0033] The feature embedding network constructed for the farmland management measure data comprises an input layer, two multi-head self-attention mechanism network layers and a full-connection neural network layer, the number of neurons of the input layer is consistent with the feature dimension of the farmland management measure data, the number of attention heads of each multi-head self-attention mechanism network layer is 4, the query matrix, key matrix and value matrix of the feature are obtained by linear transformation, the attention weight of each feature is calculated by scaled dot-product attention, and the feature representation is obtained by weighted summation, after the two layers of multi-head self-attention mechanism processing, the dimension is reduced by the full-connection neural network layer to obtain a low-dimensional embedding feature representation with fixed dimension.

[0034] Optionally, the feature invariance constraint optimization module comprises:

[0035] performing data normalization processing on the high-dimensional semantic feature vectors from multiple independent environment data sets;

[0036] align the feature space of the high-dimensional semantic feature vectors of the plurality of normalized independent environment data sets, and calculate corresponding distribution center vectors;

[0037] construct a unified feature optimization constraint loss function by using the improved invariant risk minimization algorithm, the unified feature optimization constraint loss function including a feature distribution difference loss term and a cross-environment consistency loss term;

[0038] perform iterative optimization processing on the unified feature optimization constraint loss function by using a gradient descent algorithm to obtain an optimized feature representation;

[0039] perform consistency evaluation on the feature distribution stability of the optimized feature representation, determine whether the optimized feature representation reaches a preset feature distribution stability threshold by calculating the distance change amplitude between the feature distribution center vectors of each environment data set before and after optimization;

[0040] when the optimized feature representation does not reach the preset feature distribution stability threshold, adjust the weight parameters of the feature distribution difference loss term and the cross-environment consistency loss term in the unified feature optimization constraint loss function, and repeat the gradient descent algorithm optimization process until a unified stable feature representation model that meets the preset feature distribution stability threshold is obtained.

[0041] Optionally, the unified feature optimization constraint loss function constructed by using the improved invariant risk minimization algorithm specifically includes:

[0042] calculate the robust center vector, the covariance matrix and the probability distribution distance of the high-dimensional semantic feature vector of each independent environment data set respectively;

[0043] construct a robust center alignment loss term according to the Euclidean distance between the robust center vectors of each independent environment data set, construct a covariance alignment loss term according to the covariance matrix difference between each independent environment data set, and construct a distribution distance alignment loss term according to the probability distribution distance between each independent environment data set, linearly combine the robust center alignment loss term, the covariance alignment loss term and the distribution distance alignment loss term with a preset first weight parameter to form the feature distribution difference loss term;

[0044] robust class center vectors corresponding to each class in each independent environment dataset, and construct an intra-class radius contraction loss term according to the Euclidean distance of each data sample to the corresponding environment class center vector, construct a class conditional center alignment loss term according to the Euclidean distance between the robust class center vectors of the same class in each independent environment dataset, and construct a gradient consistency loss term according to the loss gradient of the shared classifier parameters on each independent environment dataset, linearly combine the intra-class radius contraction loss term, the class conditional center alignment loss term and the gradient consistency loss term with a preset second weight parameter to form a cross-environment consistency loss term;

[0045] linearly combine the feature distribution difference loss term and the cross-environment consistency loss term with a preset third weight parameter to form a unified feature optimization constraint loss function.

[0046] Optionally, the cross-environment generalization performance verification module comprises:

[0047] The pre-divided cross-region and cross-season verification data sets are subjected to the same data normalization processing and standardized feature conversion as the training data set;

[0048] Feature embedding and deep feature extraction are performed on all samples of the verification data set to obtain a high-dimensional semantic feature vector corresponding to each verification sample;

[0049] Inference calculation is performed on the high-dimensional semantic feature vector of each verification sample to obtain a crop recommendation prediction result corresponding to each verification sample;

[0050] The crop real planting result corresponding to each verification sample in the cross-region and cross-season verification data set is compared with the crop recommendation prediction result one by one, and the recommendation accuracy is calculated as a quantitative indicator of the generalization ability;

[0051] The recommendation accuracy of different geographical regions in the cross-region verification data set and the recommendation accuracy under different seasonal conditions in the cross-season verification data set are calculated respectively, and the recommendation accuracy difference of the unified stable feature representation model under different regions and different seasonal conditions is analyzed;

[0052] A cross-environment generalization performance index matrix is constructed according to the calculated cross-region and cross-season recommendation accuracy;

[0053] When the average recommendation accuracy in the cross-environment generalization performance index matrix does not reach a preset generalization performance threshold, the region or seasonal data in the verification data set that causes low recommendation accuracy is returned to the feature invariance constraint optimization module for iterative optimization until the average recommendation accuracy in the cross-environment generalization performance index matrix meets the preset generalization performance threshold requirement.

[0054] Optionally, the personalized crop recommendation decision module comprises:

[0055] Performing data preprocessing on real-time agricultural data collected in the target region and the target season;

[0056] Using the unified stable feature representation model that passes the generalization ability test, performing feature embedding and deep feature extraction processing on the preprocessed real-time agricultural data to obtain a high-dimensional semantic feature vector corresponding to the real-time agricultural data of the target region and the target season;

[0057] According to the high-dimensional semantic feature vector, using the shared classifier corresponding to the unified stable feature representation model to perform inference calculation to obtain an initial recommendation score corresponding to the candidate crop species;

[0058] Taking the actual planting data of crops in the target region and the target season in the historical crop planting data as auxiliary reference, calculating the historical planting frequency and planting area proportion of each candidate crop, and fusing the historical planting frequency and planting area proportion with the initial recommendation score by weighted calculation to form a comprehensive recommendation score;

[0059] The comprehensive recommendation score of the candidate crop species is sorted again, and according to the preset recommendation quantity threshold, the top several crops with the highest comprehensive recommendation score are selected from the sorting result to form a personalized crop recommendation list, and the recommendation priority is determined according to the sorting order;

[0060] Based on the personalized crop recommendation list, combining the corresponding soil property, meteorological condition and farmland management measure data in the real-time agricultural data of the target region and the target season, an evaluation index of planting suitability of the recommended crop is constructed;

[0061] When the planting suitability evaluation index does not meet the preset suitability threshold, the unified stable feature representation model is called again for re-inference calculation until the planting suitability evaluation index of all crop species in the recommendation list meets the preset suitability threshold requirement, and finally the crop personalized recommendation list is output.

[0062] Optionally, the recommendation result visualization interaction module comprises:

[0063] Receiving the crop personalized recommendation list output from the personalized crop recommendation decision module;

[0064] Using a visualization chart generation method to construct a visualization recommendation score comparison chart;

[0065] According to the soil property data, meteorological condition data and farmland management measure data in the real-time agricultural data of the target region and the target season, a crop planting suitability radar chart is formed;

[0066] Adopt geographic information system technology, according to the geographical position and geographical space data of target area, build interactive geographical space distribution map;

[0067] A user interaction control interface is provided in the recommendation result visualization page, allowing the user to perform crop type screening, recommendation score threshold setting, crop historical planting area and real-time agricultural data feature screening through the interaction interface, and dynamically adjusting the recommendation result visualization display content in real time;

[0068] When the user adjusts the recommendation screening conditions through the interaction control interface, the personalized crop recommendation decision module is called again for real-time inference calculation, and the recommendation result in the visualization page is dynamically updated;

[0069] The user interaction behavior data of the recommendation result visualization page is recorded and analyzed, and the user's preferred crop type, the user's set recommendation score threshold and the real-time agricultural data feature screening condition are counted.

[0070] The beneficial effects of the present application are:

[0071] (1) The present application realizes the unified and stable feature representation of the crop recommendation model under the conditions of cross-region and cross-season by adopting the improved invariant risk minimization algorithm combined with the deep feature extraction network, effectively improves the generalization performance of the crop intelligent recommendation system, and enhances the reliability and accuracy of the cross-environment recommendation result.

[0072] (2) The present application realizes the accurate quantification and dynamic feedback optimization of the model generalization ability through the design of the cross-environment data division module and the cross-environment generalization performance verification module, significantly improves the scientificity of the recommendation result, and shows better adaptability and robustness in different geographical regions and seasonal scenes.

[0073] (3) The present application improves the accuracy and practicality of cross-environment personalized recommendation through the personalized crop recommendation decision module and the recommendation result visualization interaction module, effectively solves the problem of low recommendation accuracy caused by large data distribution difference in cross-region and cross-season in the prior art, breaks through the bottleneck of poor generalization ability of the existing recommendation system, realizes the significant progress of personalized recommendation accuracy and user interaction experience, and effectively improves the practical application value of the crop intelligent recommendation system. BRIEF DESCRIPTION OF DRAWINGS

[0074] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application. In the drawings:

[0075] Fig. 1 A system overall structure diagram of a crop intelligent recommendation system based on neural network is provided.

[0076] Fig. 2 A unified stable feature representation model construction and optimization flowchart of a neural network-based crop intelligent recommendation system is proposed for the present application. DETAILED DESCRIPTION

[0077] The present application will now be further described in detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams and only illustrate the basic structure of the present application in a schematic manner, and therefore only show the components related to the present application.

[0078] REFERENCE Figs. 1-2 A neural network-based crop intelligent recommendation system, comprising:

[0079] An agricultural multi-source data acquisition module is used to collect soil property data, meteorological condition data, crop planting history data and farmland management measure data in different geographical regions and different seasons, and after preliminary cleaning and preprocessing, a standardized agricultural multi-source data set is formed;

[0080] A cross-environment data division module is connected with the agricultural multi-source data acquisition module, used to receive the standardized agricultural multi-source data set, and divide the standardized agricultural multi-source data set according to geographical region features and seasonal time features, forming multiple independent environment data sets with region labels and season labels;

[0081] A deep feature extraction network module is connected with the cross-environment data division module, used to receive the multiple independent environment data sets, and perform feature embedding and feature extraction processing on each independent environment data set in the multiple independent environment data sets respectively, to obtain a high-dimensional semantic feature vector corresponding to each independent environment data set;

[0082] A feature invariance constraint optimization module is connected with the deep feature extraction network module, used to receive the high-dimensional semantic feature vectors corresponding to each independent environment data set, and through an improved invariance risk minimization algorithm, the differences between the high-dimensional semantic feature vectors of different independent environment data sets are optimized and constrained, to obtain a unified stable feature representation model meeting the cross-region and cross-season invariance requirements;

[0083] A cross-environment generalization performance verification module is connected with the feature invariance constraint optimization module, used to receive the unified stable feature representation model, and use pre-divided cross-region and cross-season verification data sets to test the generalization ability, to obtain a model accuracy quantification result under cross-region and cross-season conditions;

[0084] a personalized crop recommendation decision module, connected with the cross-environment generalization performance verification module, configured to receive the unified stable feature representation model that passes the generalization capability test, and obtain a crop personalized recommendation list for a target region and a target season through inference calculation on real-time agricultural data of the target region and the target season;

[0085] a recommendation result visualization interaction module, connected with the personalized crop recommendation decision module, configured to receive the crop personalized recommendation list, and perform data visualization processing and human-computer interaction interface processing to form a recommendation result visualization page that can be directly displayed to a user.

[0086] In the embodiment, the agricultural multi-source data acquisition module comprises:

[0087] The soil property data, the meteorological condition data, the crop planting history data and the farmland management measure data of different geographical regions and different seasons are collected, wherein the soil property data includes soil type, soil pH, soil fertility level and soil water content data, the meteorological condition data includes environmental temperature, air humidity, precipitation, light intensity and wind speed data, the crop planting history data includes crop variety, planting area, planting density, planting date and harvesting date data, and the farmland management measure data includes fertilizer type and application amount, pesticide application type and application amount, irrigation frequency and irrigation water amount data. After missing value detection, outlier detection and data format conversion are performed on all collected data, a standardized agricultural multi-source data set is obtained.

[0088] In the embodiment, the cross-environment data division module comprises:

[0089] According to the geographical region features and the seasonal time features in the standardized agricultural multi-source data set, a preliminary region-season combination matrix is established, and the standardized agricultural multi-source data set is cross-mapped according to the two dimensions of region and season. The cross-mapping is to divide the data records with the same or similar geographical region features in the standardized agricultural multi-source data set into corresponding region groups and divide the data records with the same or similar seasonal time features into corresponding season groups, respectively, by taking the geographical region features as the first dimension and the seasonal time features as the second dimension, and by cross-corresponding the region groups and the season groups to form region-season combinations and mapping the data in the standardized agricultural multi-source data set into the corresponding region-season combinations to form a cross-mapped region-season combination matrix.

[0090] The data sample amounts corresponding to each environment combination in the cross-mapped region-season combination matrix are analyzed, and the environment combinations with data sample amounts lower than a preset sample threshold are removed.

[0091] For the environment combination whose sample size meets the preset sample threshold, the statistical distribution characteristics of the key features in the soil property data, the meteorological condition data and the crop planting history data in the environment combination are calculated, the soil pH value and the soil fertility level in the soil property data, the temperature and the precipitation in the meteorological condition data, and the planting area and the planting density in the crop planting history data are extracted for each environment combination respectively, the mean, the variance and the distribution interval of each key feature are calculated, the data distribution curve of each key feature is obtained by using the histogram statistical method, and the statistical distribution characteristics are determined by calculating the mean difference, the variance difference and the distribution curve similarity of each key feature in each environment combination. The consistency of the data in each environment combination is evaluated by taking the statistical distribution characteristic similarity as an index.

[0092] The environment combination whose statistical distribution characteristic similarity does not meet the consistency requirement is segmented and adjusted until the data consistency in each environment combination meets the preset statistical distribution similarity threshold.

[0093] Each environment combination whose statistical distribution characteristic similarity meets the data consistency requirement is respectively given a region mark and a season mark, and a plurality of independent environment data sets with clear and exclusive region and season marks are formed.

[0094] In the embodiment, the deep feature extraction network module comprises:

[0095] The numerical normalization processing is respectively performed on the soil property data, the meteorological condition data, the crop planting history data and the farmland management measure data in each independent environment data set, so as to eliminate the dimension and numerical difference between the data.

[0096] The independent feature embedding network is respectively constructed for the normalized soil property data, the meteorological condition data, the crop planting history data and the farmland management measure data, and the low-dimensional embedding feature representation corresponding to each data type is obtained.

[0097] The attention score corresponding to each data type is calculated by the attention mechanism for the low-dimensional embedding feature representation of each data type, the attention score of each data type is normalized to obtain the fusion weight corresponding to each data type, and the low-dimensional embedding feature representation of each data type is weighted according to the fusion weight. The low-dimensional embedding feature representation of each data type is multiplied by the corresponding fusion weight and then concatenated to form the fusion feature vector of each independent environment data set.

[0098] A deep feature extraction network of alternately arranged convolutional neural network layers and fully connected neural network layers is built for the fusion feature vector, wherein the convolutional neural network layers are used to extract local feature correlations in the fusion feature vector, and the fully connected neural network layers are used to realize global nonlinear mapping of the local features;

[0099] The fusion feature vector is a one-dimensional feature sequence, and the initial dimension is set to 128 dimensions;

[0100] A fully connected neural network layer is used to map the feature dimension of the fusion feature vector from 128 dimensions to 256 dimensions, and a ReLU activation function is used to perform nonlinear transformation on the mapped features to obtain an initial feature representation;

[0101] The initial feature representation is input into a first one-dimensional convolutional neural network layer, wherein the convolution kernel size is 3, the number of convolution kernels is 128, the step is 1, a ReLU activation function is used to obtain nonlinear mapping of local features, and the feature dimension after convolution operation is 128 dimensions;

[0102] The output feature vector of the first convolutional neural network layer is input into a second fully connected neural network layer, the number of neurons of the second fully connected neural network layer is fixedly set to 64, a ReLU activation function is used to perform nonlinear transformation to realize reduction of the feature dimension;

[0103] The 64-dimensional feature vector output by the second fully connected neural network layer is input into a second one-dimensional convolutional neural network layer, wherein the convolution kernel size is 3, the number of convolution kernels is 64, the step is 1, and a ReLU activation function is used to realize further nonlinear mapping of local features, and the output dimension is maintained at 64 dimensions;

[0104] The 64-dimensional feature vector output by the second convolutional neural network layer is input into a third fully connected neural network layer, the number of neurons of the third fully connected neural network layer is fixedly set to 32, and a ReLU activation function is used for feature dimension mapping and nonlinear transformation;

[0105] Through the above-mentioned alternately arranged convolutional neural network layers and fully connected neural network layers, a total of not less than 6 layers are obtained, and finally a high-dimensional semantic feature vector with a fixed dimension of 32 is obtained;

[0106] A dynamic feature calibration structure is arranged between the convolutional neural network layers and the fully connected neural network layers, and dynamic standardization processing is implemented by calculating the statistical mean and variance of the output features of each network layer to eliminate the instability caused by the change of feature scale in the network training process;

[0107] The global adaptive pooling operation is performed on the feature vector processed by the dynamic feature calibration structure to obtain a high-dimensional semantic feature vector with fixed dimensions and environment difference expression capability.

[0108] In the embodiment, the independent feature embedding network is constructed respectively, specifically:

[0109] The feature embedding network constructed for the soil attribute data includes an input layer and three full connection hidden layers with dimension decreasing layer by layer, the number of neurons of the input layer is consistent with the feature dimension of the soil attribute data, the number of neurons of the three full connection hidden layers decreases by half, quarter and eighth respectively according to the number of neurons of the input layer, and the ReLU activation function is used to perform nonlinear mapping after each hidden layer, and finally a low-dimensional embedding feature representation with eighth of the input layer is obtained;

[0110] The feature embedding network constructed for the meteorological condition data includes an input layer, two convolutional neural network layers and a full connection neural network layer, the number of neurons of the input layer is consistent with the feature dimension of the meteorological condition data, the two convolutional neural network layers adopt one-dimensional convolution structure, the convolution kernel size of each convolutional neural network layer is 3 and the step is 1, and the ReLU activation function is used to process the convolution output, a global maximum pooling layer is connected after the second convolutional neural network layer to reduce the dimension, and the full connection neural network layer is used to reduce the dimension of the pooled output to obtain a low-dimensional embedding feature representation with fixed dimension;

[0111] The feature embedding network constructed for the crop planting history data includes an input layer, two long short-term memory network hidden layers and a full connection neural network layer, the number of neurons of the input layer is consistent with the feature dimension of the planting history data, the number of neurons of the two long short-term memory network hidden layers decreases layer by layer, the output feature dimension of the first long short-term memory network hidden layer is half of the input dimension, the output feature dimension of the second long short-term memory network hidden layer is quarter of the input dimension, and the long short-term memory network output feature is reduced in dimension by the full connection neural network layer to obtain a low-dimensional embedding feature representation with fixed dimension;

[0112] The feature embedding network constructed for the farmland management measure data includes an input layer, two multi-head self-attention mechanism network layers and a full connection neural network layer, the number of neurons of the input layer is consistent with the feature dimension of the farmland management measure data, the number of attention heads of each multi-head self-attention mechanism network layer is 4, the query matrix, key matrix and value matrix of the feature are obtained by linear transformation, the attention weight of each feature is calculated by scaled dot-product attention, and the feature representation is obtained by weighted summation, after the two layers of multi-head self-attention mechanism processing, the dimension reduction processing is performed by the full connection neural network layer to obtain a low-dimensional embedding feature representation with fixed dimension.

[0113] In this embodiment, the feature invariance constraint optimization module comprises:

[0114] Data normalization is performed on the high-dimensional semantic feature vectors from the multiple independent environment data sets to eliminate data distribution deviation caused by scale differences between different environment feature vectors;

[0115] The high-dimensional semantic feature vectors of the multiple independent environment data sets after normalization are aligned in feature space, and the distribution center vectors corresponding to the high-dimensional semantic feature vectors of each independent environment data set are calculated, and the Euclidean distance between the distribution center vectors of each independent environment data set is used as an initial measure of feature distribution difference between environments;

[0116] A unified feature optimization constraint loss function is constructed using an improved invariance risk minimization algorithm, which includes a feature distribution difference loss term and a cross-environment consistency loss term, wherein the feature distribution difference loss term is used to constrain the distance between the distribution center vectors corresponding to the high-dimensional semantic feature vectors of each independent environment data set, and the cross-environment consistency loss term is used to constrain the distance between the high-dimensional semantic feature vectors and the corresponding distribution center vectors within each independent environment data set;

[0117] An iterative optimization process is performed on the unified feature optimization constraint loss function using a gradient descent algorithm to reduce the distance between the distribution center vectors corresponding to the high-dimensional semantic feature vectors of each independent environment data set, and the distance between the high-dimensional semantic feature vectors and the corresponding distribution center vectors within each independent environment data set, to obtain an optimized feature representation that meets the cross-region and cross-season invariance requirements;

[0118] The feature distribution stability of the optimized feature representation is evaluated for consistency by calculating the distance change amplitude between the feature distribution center vectors of each environment data set before and after optimization to determine whether the optimized feature representation meets a preset feature distribution stability threshold;

[0119] When the optimized feature representation does not meet the preset feature distribution stability threshold, the weight parameters of the feature distribution difference loss term and the cross-environment consistency loss term in the unified feature optimization constraint loss function are adjusted, and the gradient descent algorithm optimization process is repeated until a unified stable feature representation model that meets the preset feature distribution stability threshold is obtained.

[0120] In this embodiment, the unified feature optimization constraint loss function constructed using the improved invariance risk minimization algorithm specifically comprises:

[0121] The robust center vector is calculated by using a robust mean estimation method, the covariance matrix difference is calculated by using a symmetric Log-Euclidean distance method, and the probability distribution distance is calculated by using a sliced Wasserstein distance method.

[0122] The robust center alignment loss term is constructed according to the Euclidean distances between the robust center vectors of the calculated independent environment data sets, the covariance alignment loss term is constructed according to the covariance matrix differences between the calculated independent environment data sets, and the distribution distance alignment loss term is constructed according to the probability distribution distances between the calculated independent environment data sets. The robust center alignment loss term, the covariance alignment loss term and the distribution distance alignment loss term are linearly combined with a preset first weight parameter to form a feature distribution difference loss term.

[0123] The robust class center vector corresponding to each class in each independent environment data set is calculated, the intra-class radius contraction loss term is constructed according to the Euclidean distance between each data sample and the corresponding environment class center vector, the class conditional center alignment loss term is constructed according to the Euclidean distance between the robust class center vectors of the same class in each independent environment data set, and the gradient consistency loss term is constructed according to the loss gradient of the shared classifier parameters on each independent environment data set. The intra-class radius contraction loss term, the class conditional center alignment loss term and the gradient consistency loss term are linearly combined with a preset second weight parameter to form a cross-environment consistency loss term.

[0124] The feature distribution difference loss term and the cross-environment consistency loss term are linearly combined with a preset third weight parameter to form a unified feature optimization constraint loss function.

[0125] In the embodiment, the cross-environment generalization performance verification module comprises:

[0126] The pre-divided cross-region and cross-season verification data sets are respectively subjected to the same data normalization processing and standardized feature conversion as the training data set, so as to ensure that the verification data and the training data are in the same feature distribution space.

[0127] The feature embedding and deep feature extraction are performed on all samples of the verification data set to obtain a high-dimensional semantic feature vector corresponding to each verification sample.

[0128] Based on the shared classifier corresponding to the unified stable feature representation model, the high-dimensional semantic feature vector of each verification sample is subjected to inference calculation to obtain a crop recommendation prediction result corresponding to each verification sample.

[0129] According to the corresponding crop true planting results and crop recommendation prediction results of each verification sample in the cross-region and cross-season verification data set, the number of samples in each verification data set whose recommendation prediction results completely match the true planting results is counted, and the recommendation accuracy is calculated as a quantitative indicator of generalization ability;

[0130] The recommendation accuracy of different geographical regions in the cross-region verification data set and the recommendation accuracy under different seasonal conditions in the cross-season verification data set are calculated respectively, and the difference in recommendation accuracy of the unified stable feature representation model under different regions and different seasonal conditions is analyzed;

[0131] According to the calculated cross-region and cross-season recommendation accuracy, a cross-environment generalization performance index matrix is constructed, and the mean and variance of all recommendation accuracies in the matrix are used as overall evaluation indicators of the cross-environment generalization performance of the unified stable feature representation model;

[0132] When the mean of the recommendation accuracy in the cross-environment generalization performance index matrix does not reach the preset generalization performance threshold, the region or seasonal data in the verification data set that causes low recommendation accuracy is returned to the feature invariance constraint optimization module for iterative optimization until the mean of the recommendation accuracy in the cross-environment generalization performance index matrix meets the preset generalization performance threshold requirement.

[0133] In the embodiment, the personalized crop recommendation decision module comprises:

[0134] The real-time agricultural data collected in the target region and target season is subjected to data preprocessing, and the data preprocessing method is consistent with that in the agricultural multi-source data acquisition module, so as to ensure the consistency of the data format and feature distribution of the real-time agricultural data and the standardized agricultural multi-source data set;

[0135] The high-dimensional semantic feature vector corresponding to the real-time agricultural data in the target region and target season is obtained by performing feature embedding and deep feature extraction processing on the preprocessed real-time agricultural data using the unified stable feature representation model confirmed by the cross-environment generalization performance verification module to pass the generalization ability test;

[0136] According to the high-dimensional semantic feature vector, inference calculation is performed on the shared classifier corresponding to the unified stable feature representation model to obtain the initial recommendation score corresponding to the candidate crop species;

[0137] The actual planting situation data of the target region and target season crops in the historical crop planting data is taken as auxiliary reference, the historical planting frequency and planting area proportion of each candidate crop are calculated, and the historical planting frequency and planting area proportion are fused with the initial recommendation score by weighted calculation to form a comprehensive recommendation score;

[0138] The comprehensive recommendation scores of the candidate crop types are re-ranked, and according to a preset recommendation quantity threshold, a number of crop types with the highest comprehensive recommendation scores are selected from the ranking results to form a personalized crop recommendation list, and the recommendation priorities are determined according to the ranking order;

[0139] Based on the personalized crop recommendation list, soil properties, meteorological conditions and farmland management measures data corresponding to real-time agricultural data of the target region and the target season are combined to construct a planting suitability evaluation index of the recommended crops, which is used to further evaluate the matching degree of the recommended crops and the real-time agricultural data characteristics to verify the scientificity and rationality of the recommendation list;

[0140] When the planting suitability evaluation index does not meet the preset suitability threshold, the unified stable feature representation model is called again for re-reasoning calculation for the candidate crop types that do not meet the standard, and the comprehensive recommendation scores are re-ranked again in combination with the corrected comprehensive recommendation scores, until the planting suitability evaluation index of all crop types in the recommendation list meets the preset suitability threshold requirement, and finally the personalized crop recommendation list suitable for the target region and the target season is output.

[0141] In the embodiment, the recommendation result visualization interaction module comprises:

[0142] The personalized crop recommendation list output by the personalized crop recommendation decision module is received, and the personalized crop recommendation list comprises crop types, recommendation priorities, recommendation scores and planting suitability evaluation indexes;

[0143] Using a visualization chart generation method, the recommendation scores and the planting suitability evaluation indexes of the crops are intuitively displayed in the form of interactive bar charts or column charts according to the ranking of the crop recommendation priorities and the recommendation scores, and a visualization recommendation score comparison chart is constructed;

[0144] According to the soil property data, the meteorological condition data and the farmland management measure data in the real-time agricultural data of the target region and the target season, the matching degree between the recommended crop types and the real-time agricultural data characteristics is displayed in the form of a radar chart or a spider chart to form a crop planting suitability radar chart;

[0145] Using geographic information system technology, the historical planting positions and area distribution of the recommended crop types in the target region are dynamically marked on an electronic map according to the geographic position and geographic spatial data of the target region to construct an interactive geographic spatial distribution chart;

[0146] A user interaction control interface is provided in the recommendation result visualization page to allow the user to perform crop type screening, recommendation score threshold setting, historical crop planting area and real-time agricultural data characteristic screening through the interaction interface to dynamically adjust the visualization display content of the recommendation results in real time.

[0147] When the user adjusts the recommendation filtering conditions through the interactive control interface, the personalized crop recommendation decision module is re-invoked for real-time inference calculation, and the recommendation results in the visual page are dynamically updated to respond to the user's interactive operation;

[0148] The user interaction behavior data of the recommendation result visualization page are recorded and analyzed, the user preferred crop types, the user set recommendation score threshold and real-time agricultural data feature filtering conditions are counted, and the analysis results are stored and returned to the feature invariance constraint optimization module for subsequent recommendation model generalization performance optimization and iterative improvement of recommendation effect.

[0149] Embodiment 1:

[0150] In order to verify the feasibility of the application in implementation, the application is applied to the crop intelligent planting recommendation task of a certain intelligent agricultural demonstration base, and the actual application background is the personalized and accurate recommendation decision of crops in different geographical regions and seasonal environments. The traditional method usually uses expert knowledge and conventional statistical learning models, such as decision tree or support vector machine algorithm, to recommend crops, but this kind of method causes the problem of data distribution drift caused by regional differences and seasonal changes, which leads to significant decline in generalization performance across regions and seasons, insufficient recommendation accuracy, and serious impact on crop yield and economic benefits, which needs to be solved urgently.

[0151] In the actual implementation process, first, the agricultural multi-source data acquisition module collects agricultural basic data in multiple seasons in multiple regions, the collected data includes soil type, soil pH, soil fertility level, soil moisture content, environmental temperature, air humidity, precipitation, light intensity, wind speed, and crop planting history information and farmland management measure information, and the data is processed for missing value and abnormal value, and the standardized agricultural multi-source data set is formed. Then, through the cross-environment data division module, the standardized agricultural multi-source data set is divided into independent environment data sets in the form of region and season feature cross mapping; then, the deep feature extraction network module is used to extract features and embed features of each environment data set to obtain high-dimensional semantic feature vectors; further, the feature invariance constraint optimization module is used to construct a unified stable feature representation model, specifically, an improved invariance risk minimization algorithm is used, a robust center vector, a covariance matrix and a probability distribution distance are defined to construct a unified feature optimization constraint loss function, and after 32 iterations of gradient descent algorithm, the feature distribution stability loss value converges to 0.043, and finally a unified stable feature representation model is obtained.

[0152] After the model is constructed, the cross-environment generalization performance verification module is used to quantitatively test the model generalization capability of the cross-region and cross-season verification data set, and the recommendation accuracy is taken as the evaluation index. The recommendation accuracy of the five typical regions in the cross-region verification data set is all above 85%, and the recommendation accuracy of the spring, summer, autumn and winter in the cross-season verification data set is 87.6%, 89.4%, 86.3% and 84.7% respectively, which shows obvious generalization advantage and robustness.

[0153] In the actual recommendation decision stage, the individualized crop recommendation decision module is used to obtain a comprehensive recommendation score by model inference calculation and combination of historical planting frequency and planting area proportion, and the final high-precision individualized crop recommendation list is generated through secondary sorting and suitability evaluation index feedback optimization.

[0154] The recommendation result is displayed to the user in the form of an interactive bar chart, a radar chart and a geographical spatial distribution chart through the recommendation result visualization interaction module. The user can dynamically adjust the recommendation score threshold and the filtering condition, and the system responds to the user interaction in real time, thereby improving the user experience and decision efficiency.

[0155] In the implementation process, five crop planting samples in different seasons of typical regions are tracked and compared, and the specific data is shown in Table 1.

[0156] Table 1 Comparison table of crop recommendation prediction and actual planting result

[0157]

[0158] As can be seen from the data in Table 1, the crop intelligent recommendation system based on neural network proposed by the present application has high prediction accuracy and actual application effect. In the way of double verification of recommendation score and planting suitability index, the reliability of crop recommendation decision is significantly improved. Taking A-01 spring corn as an example, the predicted yield is 9.4 t / ha, and the actual yield is 9.6 t / ha, with an error of only 2.1%; the predicted yield of A-02 summer soybean is 3.1 t / ha, and the actual yield is 3.2 t / ha, with an error of 3.1%, which reflects the outstanding performance advantage of the present application in crop recommendation task. In addition, the actual generalization ability of the system for different regions and seasonal environments is high, which reflects the significant progress over the traditional technical scheme.

[0159] Through the specific verification and analysis of the above embodiments, the crop intelligent recommendation system of the present application combines feature invariance constraint optimization with deep neural network, effectively solves the problem of insufficient recommendation accuracy of traditional methods under cross-region and cross-season conditions, and significantly improves the generalization performance and recommendation accuracy of the system. The technical advantage is obvious, and the practical application potential is huge.

[0160] The above description is only the preferred embodiment of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can make equivalent substitutions or changes according to the technical solution and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered in the protection scope of the present application.

Claims

1. A neural network based intelligent recommendation system for crops characterized in that, The utility model relates to an agricultural multi-source data acquisition module for collecting multi-source data in different geographical regions and different seasons to form a standardized agricultural multi-source data set. A cross-environment data division module is used to receive the standardized agricultural multi-source data set and divide it according to geographical region characteristics and seasonal time characteristics to form multiple independent environment data sets. A deep feature extraction network module is used to receive the multiple independent environment data sets and perform feature embedding and feature extraction processing on each independent environment data set to obtain corresponding high-dimensional semantic feature vectors. A feature invariance constraint optimization module is used to optimize and constrain the differences between the high-dimensional semantic feature vectors of different independent environment data sets by using an improved invariance risk minimization algorithm to obtain a unified stable feature representation model. A cross-environment generalization performance verification module is used to receive the unified stable feature representation model and test its generalization ability using pre-divided cross-region and cross-season verification data sets. An individualized crop recommendation decision module is used to receive the unified stable feature representation model that has passed the generalization ability test, infer and calculate real-time agricultural data for a target region and a target season, and obtain an individualized crop recommendation list. A recommendation result visualization interaction module is used to receive the individualized crop recommendation list, perform data visualization processing and human-computer interaction interface processing, and form a recommendation result visualization page. The improved invariance risk minimization algorithm is used to construct a unified feature optimization constraint loss function, which specifically includes the following steps: Robust center vectors, covariance matrices, and probability distribution distances are calculated for the high-dimensional semantic feature vectors of each independent environment data set. A robust center alignment loss term is constructed based on the Euclidean distances between the robust center vectors of each independent environment data set, a covariance alignment loss term is constructed based on the differences in covariance matrices between each independent environment data set, and a distribution distance alignment loss term is constructed based on the probability distribution distances between each independent environment data set. Robust class center vectors corresponding to each class in each independent environment data set are calculated, an intra-class radius contraction loss term is constructed based on the Euclidean distances between each data sample and the corresponding environment class center vector, a class conditional center alignment loss term is constructed based on the Euclidean distances between the robust class center vectors of the same class in each independent environment data set, and a gradient consistency loss term is constructed based on the loss gradients of the shared classifier parameters on each independent environment data set. The intra-class radius contraction loss term, the class conditional center alignment loss term, and the gradient consistency loss term are linearly combined with a preset second weight parameter to form a cross-environment consistency loss term. The feature distribution difference loss term and the cross-environment consistency loss term are linearly combined with a preset third weight parameter to form a unified feature optimization constraint loss function. 2.The neural network-based crop intelligent recommendation system according to claim 1, wherein, The agricultural multi-source data acquisition module includes Collect soil property data, meteorological condition data, crop planting history data and farmland management measure data in different geographical regions and different seasons, and after detecting missing values, detecting outliers and converting data formats, a standardized agricultural multi-source data set is obtained. 3.The neural network-based crop intelligent recommendation system according to claim 1, wherein, The cross-environment data division module comprises: According to the geographical region features and seasonal time features in the standardized agricultural multi-source data set, a region-season combination matrix is established, and cross mapping is performed; The data sample size corresponding to each environment combination in the cross-mapped region-season combination matrix is analyzed, and environment combinations with a data sample size below a preset sample threshold are removed; For environment combinations with a sample size meeting the preset sample threshold, the statistical distribution characteristics of key features in the soil property data, meteorological condition data and crop planting history data within the environment combination are calculated, and the consistency of the data within the environment combination is evaluated using the statistical distribution characteristic similarity as an index; The environment combinations whose statistical distribution characteristic similarity does not meet the consistency requirement are segmented and adjusted until the data consistency within each environment combination meets the preset statistical distribution similarity threshold; Each environment combination whose statistical distribution characteristic similarity meets the data consistency requirement is respectively given a region label and a season label, forming a plurality of independent environment data sets with region and season identification. 4.The neural network-based crop intelligent recommendation system according to claim 1, wherein, The deep feature extraction network module comprises: Numerical normalization is performed on the soil property data, meteorological condition data, crop planting history data and farmland management measure data in each independent environment data set; An independent feature embedding network is constructed for the normalized soil property data, meteorological condition data, crop planting history data and farmland management measure data to obtain low-dimensional embedding feature representations corresponding to each data type; The low-dimensional embedding feature representations of each data type are respectively calculated through an attention mechanism to obtain corresponding fusion weights, and the features are spliced according to the fusion weights to form a fusion feature vector for each independent environment data set; A deep feature extraction network comprising alternating convolutional neural network layers and fully connected neural network layers is built for the fusion feature vector, wherein the convolutional neural network layers are used to extract local feature correlations in the fusion feature vector, and the fully connected neural network layers are used to realize global nonlinear mapping of local features; A dynamic feature calibration structure is set between the convolutional neural network layers and the fully connected neural network layers to implement dynamic standardization processing by calculating the statistical mean and variance of the output features of each network layer; Global adaptive pooling is performed on the feature vectors processed by the dynamic feature calibration structure to obtain high-dimensional semantic feature vectors. 5.The neural network-based crop intelligent recommendation system according to claim 4, characterized in that, The independent feature embedding networks are respectively constructed as follows: The feature embedding network constructed for the normalized soil attribute data comprises an input layer and three fully connected hidden layers with dimension reduction layer by layer, the number of neurons of the input layer is consistent with the feature dimension of the soil attribute data, the number of neurons of the three fully connected hidden layers decreases by half, quarter and eighth of the number of neurons of the input layer in turn, and the ReLU activation function is used to perform nonlinear mapping after each hidden layer, and finally a low-dimensional embedding feature representation with an eighth of the input layer is obtained; The feature embedding network constructed for the normalized weather condition data comprises an input layer, two convolutional neural network layers and a fully connected neural network layer, the number of neurons of the input layer is consistent with the feature dimension of the weather condition data, the two convolutional neural network layers adopt one-dimensional convolution structure, the convolution kernel size of each convolutional neural network layer is 3 and the step is 1, and the ReLU activation function is used to process the convolution output, a global maximum pooling layer is connected after the second convolutional neural network layer to reduce the dimension, and the fully connected neural network layer reduces the dimension of the pooled output to obtain a low-dimensional embedding feature representation with a fixed dimension; The feature embedding network constructed for the normalized crop planting history data comprises an input layer, two long short-term memory network hidden layers and a fully connected neural network layer, the number of neurons of the input layer is consistent with the feature dimension of the planting history data, the number of neurons of the two long short-term memory network hidden layers decreases layer by layer, the feature dimension of the output of the first long short-term memory network hidden layer is half of the input dimension, the feature dimension of the output of the second long short-term memory network hidden layer is quarter of the input dimension, and the long short-term memory network output feature is reduced in dimension by the fully connected neural network layer to obtain a low-dimensional embedding feature representation with a fixed dimension; The feature embedding network constructed for the normalized farmland management measure data comprises an input layer, two multi-head self-attention mechanism network layers and a fully connected neural network layer, the number of neurons of the input layer is consistent with the feature dimension of the farmland management measure data, the number of attention heads of each multi-head self-attention mechanism network layer is 4, the query matrix, key matrix and value matrix of the feature are obtained by linear transformation, the attention weight of each feature is calculated by scaled dot-product attention, and the feature representation is obtained by weighted summation, after two layers of multi-head self-attention mechanism processing, the dimension is reduced by the fully connected neural network layer to obtain a low-dimensional embedding feature representation with a fixed dimension. 6.The neural network-based crop intelligent recommendation system according to claim 1, wherein, The feature invariance constraint optimization module comprises: performing data normalization processing on the high-dimensional semantic feature vectors from multiple independent environment data sets; aligning the high-dimensional semantic feature vectors of the multiple independent environment data sets after normalization processing in the feature space, and calculating the corresponding distribution center vectors; constructing a unified feature optimization constraint loss function by using an improved invariant risk minimization algorithm, the unified feature optimization constraint loss function comprises a feature distribution difference loss term and a cross-environment consistency loss term; performing iterative optimization processing on the unified feature optimization constraint loss function by using a gradient descent algorithm to obtain an optimized feature representation; The consistency of the feature distribution stability of the optimized feature representation is evaluated, and whether the optimized feature representation reaches a preset feature distribution stability threshold is determined by calculating the distance change amplitude between the feature distribution center vectors of each environment data set before and after optimization. When the optimized feature representation does not reach the preset feature distribution stability threshold, the weight parameters of the feature distribution difference loss term and the cross-environment consistency loss term in the unified feature optimization constraint loss function are adjusted, and the gradient descent algorithm optimization process is repeatedly executed until a unified stable feature representation model that meets the preset feature distribution stability threshold is obtained. 7.The neural network-based crop intelligent recommendation system according to claim 1, wherein, The cross-environment generalization performance verification module comprises: The pre-divided cross-regional and cross-seasonal verification data sets are subjected to the same data normalization processing and standardized feature conversion as the training data set; Feature embedding and deep feature extraction are performed on all samples of the verification data set to obtain high-dimensional semantic feature vectors corresponding to each verification sample; Inference calculation is performed on the high-dimensional semantic feature vectors of each verification sample to obtain crop recommendation prediction results corresponding to each verification sample; The crop real planting results and crop recommendation prediction results corresponding to each verification sample in the cross-regional and cross-seasonal verification data set are compared one by one, and the recommendation accuracy is calculated as a quantitative indicator of generalization ability; The recommendation accuracy of different geographical regions in the cross-regional verification data set and the recommendation accuracy under different seasonal conditions in the cross-seasonal verification data set are calculated respectively to analyze the recommendation accuracy difference of the unified stable feature representation model under different regional and seasonal conditions; According to the calculated cross-regional and cross-seasonal recommendation accuracy, a cross-environment generalization performance index matrix is constructed; When the average recommendation accuracy in the cross-environment generalization performance index matrix does not reach the preset generalization performance threshold, the regional or seasonal data in the verification data set that causes low recommendation accuracy is returned to the feature invariance constraint optimization module for iterative optimization until the average recommendation accuracy in the cross-environment generalization performance index matrix meets the preset generalization performance threshold requirement. 8.The neural network-based crop intelligent recommendation system according to claim 1, wherein, The individualized crop recommendation decision module comprises: Performing data preprocessing on real-time agricultural data collected in the target region and target season; Using the unified stable feature representation model that passes the generalization ability test, performing feature embedding and deep feature extraction on the preprocessed real-time agricultural data to obtain high-dimensional semantic feature vectors corresponding to the target region and target season real-time agricultural data; According to the high-dimensional semantic feature vectors, using the shared classifier corresponding to the unified stable feature representation model to perform inference calculation to obtain initial recommendation scores corresponding to the candidate crop species; Taking the actual planting data of crops in the target region and target season in the historical crop planting data as auxiliary reference, calculating the historical planting frequency and planting area proportion of each candidate crop, and fusing the historical planting frequency and planting area proportion with the initial recommendation score by weighted calculation to form a comprehensive recommendation score. The individualized crop recommendation decision module comprises: Performing data preprocessing on real-time agricultural data collected in the target region and target season; Using the unified stable feature representation model that passes the generalization ability test, performing feature embedding and deep feature extraction on the preprocessed real-time agricultural data to obtain high-dimensional semantic feature vectors corresponding to the target region and target season real-time agricultural data; According to the high-dimensional semantic feature vectors, using the shared classifier corresponding to the unified stable feature representation model to perform inference calculation to obtain initial recommendation scores corresponding to the candidate crop species; Taking the actual planting data of crops in the target region and target season in the historical crop planting data as auxiliary reference, calculating the historical planting frequency and planting area proportion of each candidate crop, and fusing the historical planting frequency and planting area proportion with the initial recommendation score by weighted calculation to form a comprehensive recommendation score. The comprehensive recommendation scores of the candidate crop types are re-ranked, and according to a preset recommendation quantity threshold, a number of crop types with the highest comprehensive recommendation scores are selected from the ranking results to form a personalized crop recommendation list, and the recommendation priorities are determined according to the ranking orders; Based on the personalized crop recommendation list, soil properties, meteorological conditions and farmland management measures data corresponding to the real-time agricultural data of the target region and the target season are combined to construct a planting suitability evaluation index of the recommended crops; When the planting suitability evaluation index does not meet the preset suitability threshold, the unified stable feature representation model is called again for re-reasoning calculation until the planting suitability evaluation indexes of all crop types in the recommendation list meet the preset suitability threshold requirements, and finally the personalized crop recommendation list is output. 9.The neural network-based crop intelligent recommendation system according to claim 1, wherein, The recommendation result visualization interaction module comprises: Receiving the personalized crop recommendation list output by the personalized crop recommendation decision module; Using a visualization chart generation method, a visualization recommendation score comparison chart is constructed; According to the soil property data, meteorological condition data and farmland management measure data in the real-time agricultural data of the target region and the target season, a crop planting suitability radar chart is formed; Using geographic information system technology, an interactive geographic spatial distribution chart is constructed according to the geographic location and geographic spatial data of the target region; A user interaction control interface is provided in the recommendation result visualization page to allow users to perform crop type screening, recommendation score threshold setting, crop historical planting area and real-time agricultural data feature screening through the interaction interface, and to dynamically adjust the visualization display content of the recommendation result; When the user adjusts the recommendation screening conditions through the interaction control interface, the personalized crop recommendation decision module is called again for real-time reasoning calculation, and the recommendation result in the visualization page is dynamically updated; The user interaction behavior data of the recommendation result visualization page is recorded and analyzed, and the user preferred crop types, the user set recommendation score threshold and the real-time agricultural data feature screening conditions are counted.

Images