A farmland water, fertilizer and pesticide collaborative intelligent prescription map generation method based on multi-source data fusion

Abstract

The application discloses a farmland water, fertilizer and pesticide collaborative intelligent prescription map generation method based on multi-source data fusion, constructs an inter-class confusion matrix through K-means clustering and an SVM multi-classifier, accurately locates a high-risk area of the same spectrum of foreign matters, and clearly targets a specific object for subsequent processing; a band weight self-adaptive optimizer is designed by fusing a genetic algorithm, a classification accuracy and an inter-class separation degree are taken as targets, and the band weight is self-adaptively adjusted; a teacher-student double-branch feature distillation neural network with heterogeneous parameters is constructed, a composite loss function of knowledge distillation and intra-class covariance loss is designed, robust features are extracted, and spectral confusion is inhibited from the source; a node evaluation function of comprehensive sample consistency, information gain and execution cost is designed, a decision tree is simplified through a branch and bound pruning algorithm, and a decision path is dynamically and efficiently selected; Mahalanobis distance is introduced, low-confidence decisions are corrected in combination with K-nearest neighbor weighted median, a prescription map with confidence annotations is generated, and the robustness and credibility of the decisions are improved.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture, specifically to a method for generating intelligent prescription maps for the coordinated use of water, fertilizer, and pesticides in farmland based on multi-source data fusion. Background Technology

[0002] The development of smart agriculture is driving agricultural production towards digitalization, precision, and intelligence. Intelligent prescription mapping technology for coordinated water, fertilizer, and pesticide application in farmland, based on multi-source data fusion, has become a core means to achieve variable application of water, fertilizer, and pesticides and efficient resource utilization. However, multi-source sensing data in farmland commonly exhibits phenomena such as similar spectral characteristics for different crops and different spectral characteristics for the same crop. Different types of crop stress have similar spectral characteristics, while the same stress may exhibit different spectral characteristics due to environmental interference, severely interfering with the accuracy of crop stress diagnosis and becoming a core obstacle to collaborative decision-making based on multi-source information.

[0003] In existing technologies, some solutions improve decision-making efficiency by optimizing decision tree pruning algorithms, but this only optimizes the tree structure and does not address the problem of underlying feature confusion, which can easily lead to errors in agricultural operation decisions. Other solutions use Mahalanobis distance to build classifiers for feature discrimination, but the class covariance matrix they rely on can be estimated incorrectly due to spectral confusion, leading to erroneous judgments of sample confidence and masking decision-making risks. Furthermore, general multi-source fusion models struggle to remove confused features, and subsequent decision optimization and confidence testing, based on flawed features, cannot guarantee the accuracy and reliability of prescription decisions, easily leading to mismatched application of water, fertilizer, and pesticides, exacerbating crop stress, causing economic losses, and environmental pollution. Therefore, it is essential to design a method for generating intelligent prescription maps for coordinated water, fertilizer, and pesticide application in farmland based on multi-source data fusion to analyze and suppress multi-source data feature confusion from the source, ensuring the accuracy and robustness of the generated coordinated prescription maps. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a method for generating intelligent prescription maps for synergistic effects of water, fertilizer, and pesticides in farmland based on multi-source data fusion. This method primarily solves the problem that different crop stress types have similar spectral characteristics, while the same stress can have different spectral characteristics due to environmental interference, which seriously interferes with the accuracy of crop stress diagnosis.

[0005] The technical solution to achieve the purpose of this invention is as follows: A method for generating intelligent prescription maps for synergistic irrigation, fertilizer, and pesticide application in farmland based on multi-source data fusion includes the following steps: Multispectral remote sensing images of farmland were acquired and preprocessed. Cross-validation of K-means clustering and support vector machine multi-classifiers was used to construct an inter-class confusion matrix, locate hyperspectral confusion risk areas, and generate spatial distribution binary mask maps. Spectral data were extracted based on the binary mask maps of confusion risk areas. The results of local spatial neighborhood method noise assessment, LASSO regression sparse feature selection, and principal component analysis were combined to determine the set of redundant bands and output a subset of effective band spectral data. A teacher-student dual-branch neural network was constructed. By integrating the knowledge distillation loss and intra-class covariance loss into a composite loss function, the student network was trained to extract robust features with high discriminative power and low intra-class variance from effective band data. An adaptive band weight optimizer is constructed, with classification accuracy and inter-class separation as optimization objectives. A band weight optimization mechanism is designed by integrating a genetic algorithm to solve for the optimal band weight vector, which adaptively strengthens discriminative bands and suppresses confusing bands. An initial decision tree is constructed based on agronomic knowledge and historical data. The structure is simplified through a branch-and-bound pruning algorithm to achieve dynamic decision path selection for different farmland units. Mahalanobis distance test is introduced to evaluate the decision confidence. For farmland units with low confidence, a nearest neighbor weighted median correction decision scheme is adopted. Finally, a water, fertilizer and pesticide synergistic variable operation prescription map with confidence label is generated.

[0006] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention constructs an inter-class confusion matrix using K-means clustering and SVM multi-classifiers to accurately locate high-risk regions of different objects sharing the same spectrum, providing a clear target for subsequent processing; it integrates genetic algorithms to design a band weight adaptive optimizer, adaptively adjusting band weights with classification accuracy and inter-class separation as objectives; it constructs a parameter heterogeneous teacher-student dual-branch feature distillation neural network, designs a composite loss function of knowledge distillation and intra-class covariance loss to extract robust features and suppress spectral confusion at its source; it designs a node evaluation function that integrates sample consistency, information gain, and execution cost, and simplifies the decision tree through a branch-and-bound pruning algorithm to dynamically and efficiently select decision paths; it introduces Mahalanobis distance and combines it with K-nearest neighbor weighted median to correct low-confidence decisions, generating prescription maps with confidence labels to improve the robustness and credibility of decisions. Attached Figure Description

[0007] Figure 1 This is a flowchart of a method for generating intelligent prescription maps for farmland water, fertilizer and pesticide synergy based on multi-source data fusion, as described in an embodiment of this application. Figure 2 This is a flowchart illustrating the multi-source data fusion and feature decoupling process in an embodiment of this application; Figure 3 This is a flowchart of the band adaptive optimization and decision tree construction process in the embodiments of this application; Figure 4 This is a flowchart of prescription diagram generation and reliability verification in an embodiment of this application. Detailed Implementation

[0008] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0009] like Figure 1As shown in the figure, a specific embodiment of the present invention discloses a method for generating intelligent prescription maps for synergistic application of water, fertilizer and pesticides in farmland based on multi-source data fusion, comprising the following steps: Step S101: Acquire and preprocess multispectral remote sensing images of farmland. Construct an inter-class confusion matrix through cross-validation of K-means clustering and support vector machine multi-classifiers to locate hyperspectral confusion risk areas and generate a spatial distribution binary mask map. Extract spectral data based on the binary mask map of confusion risk areas. Determine the redundant band set by combining the results of local spatial neighborhood method noise assessment, LASSO regression sparse feature selection, and principal component analysis, and output a subset of effective band spectral data. Step S102: Construct a teacher-student dual-branch neural network. By integrating a composite loss function that combines knowledge distillation loss and intra-class covariance loss, train the student network to extract robust features with high discriminative power and low intra-class variance from effective band data. Step S103: Construct a band weight adaptive optimizer with classification accuracy and inter-class separation as optimization objectives. Integrate genetic algorithm to design a band weight optimization mechanism, solve for the optimal band weight vector, and adaptively strengthen discriminative bands and suppress confusing bands. Step S104: Construct an initial decision tree based on agronomic knowledge and historical data, simplify the structure through a branch-and-bound pruning algorithm, and realize dynamic decision path selection for different farmland units; introduce Mahalanobis distance test to evaluate decision confidence, and adopt a nearest neighbor weighted median correction decision scheme for farmland units with low confidence, and finally generate a water, fertilizer and pesticide synergistic variable operation prescription map with confidence label.

[0010] This embodiment provides a method for generating intelligent prescription maps for coordinated irrigation, fertilization, and pesticide application in farmland based on multi-source data fusion. It constructs an inter-class confusion matrix using K-means clustering and an SVM multi-classifier to accurately locate high-risk regions with similar spectra of different objects, providing a clear target for subsequent processing. A genetic algorithm is integrated to design an adaptive band weight optimizer, which adaptively adjusts band weights with the goals of classification accuracy and inter-class separation. A teacher-student dual-branch feature distillation neural network with heterogeneous parameters is constructed, and a composite loss function combining knowledge distillation and intra-class covariance loss is designed to extract robust features and suppress spectral confusion at its source. A node evaluation function integrating sample consistency, information gain, and execution cost is designed, and a branch-and-bound pruning algorithm is used to simplify the decision tree, dynamically and efficiently selecting the decision path. Mahalanobis distance is introduced, combined with K-nearest neighbor weighted median to correct low-confidence decisions, generating prescription maps with confidence annotations to improve the robustness and reliability of the decisions.

[0011] Step S101: Acquire and preprocess multispectral remote sensing images of farmland. Construct an inter-class confusion matrix through cross-validation of K-means clustering and support vector machine multi-classifiers to locate hyperspectral confusion risk areas and generate a spatial distribution binary mask map. Extract spectral data based on the binary mask map of confusion risk areas. Determine the redundant band set by combining the results of local spatial neighborhood method noise assessment, LASSO regression sparse feature selection, and principal component analysis, and output a subset of effective band spectral data. In some preferred embodiments, multispectral remote sensing images of the target farmland area are acquired and a standardized preprocessing procedure is performed on the images. First, radiometric calibration is performed to convert the raw digital quantization values ​​into apparent radiance, eliminating sensor response differences. Then, the FLAASH atmospheric correction model is used to eliminate the effects of atmospheric scattering and absorption by aerosols and water vapor, obtaining the true surface reflectance. Next, using the farmland GPS control points as a reference, a quadratic polynomial transformation method is used for geometric fine correction to ensure that the image spatial coordinates are consistent with the geographic coordinate system. After data standardization preprocessing, the reflectance of each band is normalized. The K-means clustering algorithm is used to perform unsupervised classification on the preprocessed images. The number of clusters K is set according to the actual farmland scene. For example, in this embodiment, the value of K is set to 5, and the images are divided into healthy vegetation area, water stress area, nitrogen stress area, disease stress area, and bare soil area. During the clustering process, spectral reflectance is used as a feature, and Euclidean distance is used as a similarity measure. For each cluster category, a typical region with stable spectral features is selected in the image to avoid interference from edge and abnormal pixels. The reflectance values ​​of all pixels within the selected typical region are extracted in each band to construct a multi-dimensional spectral feature vector sample library for each category. Each sample in the sample library is a row vector with the same dimension as the number of image bands. The spatial coordinates of the samples are also recorded to provide a basis for subsequent risk area location. Based on the multi-dimensional spectral feature vector sample library, an SVM multi-classifier is trained using a one-to-many training strategy. Each category is sequentially set as a positive sample, and all other categories are merged into negative samples. K binary SVM classification sub-models are trained, and finally, all binary sub-models are integrated to form a complete SVM multi-classifier. The SVM multi-classifier is used to predict the entire image, and the predicted category results are compared with the initial cluster labels. The number of pixels with the true category i predicted as category j is counted, forming a confusion matrix C, where the matrix elements C... ij The number of pixels representing true class i that are predicted as class j, where N is the total number of classes; calculate each off-diagonal element C in the confusion matrix C. ij (Where i≠j) represents the proportion of total pixels, and this proportion exceeds a preset threshold. Category The image was determined to have significant foreign object homospectral confusion; then all pixels belonging to these high confusion risk categories were located in the image, and a binary mask map corresponding to the hyperspectral confusion risk area was generated, where the mask value of 1 indicates the hyperspectral confusion risk area. In some preferred embodiments, a binary mask image of the hyperspectral confusion risk region is loaded, and this binary mask image is used to perform a masking operation on the original multispectral image to extract all pixels in the original multispectral image with a mask value of 1, thereby constructing a risk pixel set S. R For the risk pixel set S R For high-dimensional spectral data, the noise level is first quantitatively assessed using the local spatial neighborhood method, which defines a neighborhood centered on each pixel. For a given window, calculate the standard deviation of the reflectance values ​​of all pixels within that window in band b. After traversing all pixels and all bands, the average noise standard deviation of the entire scene is plotted. The curve that varies with the band number b. The calculation formula is as follows: , Where b ranges from 1 to M, M is the total number of bands, and p represents the set of risk pixels S. R The pixel number in the middle; Bands with values ​​significantly higher than the average of their preceding and following bands are marked as high-noise candidate bands. ; The LASSO regression model is used for sparse feature selection to obtain information-redundant candidate bands. The LASSO regression model uses crop leaf area index y as the response variable and reflectance X of all bands as the feature variable, and sets the objective function as follows: Regularization parameters are determined through cross-validation. The value of is obtained after regression, resulting in a coefficient vector. ,Will The bands corresponding to coefficients with an absolute value of zero or close to zero are identified as candidate bands for information redundancy. Simultaneously, principal component analysis is performed in parallel to calculate the correlation coefficient matrix of reflectance data for all bands and solve for its eigenvalues. And sort them in descending order, calculate the cumulative contribution rate:

[0012] Will make for the first time The smallest All original bands following the principal component corresponding to the value, i.e., those bands with lower contribution rates, are marked as candidate bands for secondary information. ; For high-noise candidate bands Information redundancy candidate bands Candidate bands for secondary information Find the intersection to determine the final set of redundant spectral bands. Redundant spectral bands are filtered out from the original data, and a subset of effective spectral data is output and standardized. like Figure 2 As shown, in step S102, a teacher-student dual-branch neural network is constructed. By integrating the knowledge distillation loss and the intra-class covariance loss into a composite loss function, the student network is trained to extract robust features with high discriminative power and low intra-class variance from effective band data. In some preferred embodiments, a two-branch deep feature self-distillation neural network model is constructed, wherein the teacher network T and the student network S have isomorphic convolutional skeletons but different depths. The input to the teacher network is complete multispectral data of M bands with dimensions of [missing information]. The input to the student network is a subset of valid band spectral data with dimensions of [missing information]. Adding a feature projection head after the terminal feature extraction layers of the teacher and student networks maps high-dimensional features to a unified low-dimensional embedding space. The loss function of the teacher-student dual-branch neural network is a composite loss function. It includes knowledge distillation loss. The loss uses KL divergence to align the probability distributions of the teacher network and the student network in the embedding space, and the calculation formula is as follows:

[0013] in The training batch size is C, and the total number of clusters N obtained from the clustering in step one is C. and Let represent the normalized probabilities of the teacher network and student network for the i-th sample belonging to class c, respectively; secondly, introduce within-class covariance loss. To directly minimize the dispersion of features among similar samples, for all samples belonging to the same true class c in a batch, low-dimensional feature vectors extracted via the student network are collected. Calculate the in-class feature covariance matrix:

[0014] Among them B c This represents the set of sample indices for category c in the batch. If is the mean vector of the features of this class of samples, then the within-class covariance loss is defined as the sum of the Frobenius norms of the covariance matrices of all classes, calculated as follows:

[0015] The composite loss function is the knowledge distillation loss. Intra-class covariance loss Weighted sum of two items ,in It is a hyperparameter that controls the intensity of intra-class compactness penalty; knowledge is transferred from the teacher network to the student network through knowledge distillation, and the student network can extract and output low-dimensional feature vectors with high discriminative power and high intra-class consistency from noisy and redundant data; like Figure 3 As shown, in step S103, an adaptive band weight optimizer is constructed, with classification accuracy and inter-class separation as optimization objectives. A band weight optimization mechanism is designed by integrating a genetic algorithm to solve for the optimal band weight vector, which adaptively strengthens discriminative bands and suppresses confusing bands. In some preferred embodiments, the band weight optimization problem is transformed into a multi-objective global optimization problem. An adaptive band weight optimizer is constructed, with classification accuracy and inter-class separation as optimization objectives. A genetic algorithm is integrated to design the band weight optimization mechanism and solve for the optimal band weight vector. For example, each solution is denoted as a solution with dimension 1. real weight vector ,in This vector represents the number of elements in the effective band spectral data subset and will be used for element-wise weighting of the input spectral data; the optimization objective is to maximize the fitness function. The function is defined as classification accuracy. Inter-class separation The weighted sum is calculated using the following formula:

[0016] in, The results are obtained by inputting the weighted spectral data into the student network S and statistically analyzing the consistency between its output category and the true label. The Fisher discriminant ratio is used to quantify the overall separation degree of all categories in the feature space. It is calculated based on the weighted features extracted by the student network, and the calculation method is as follows:

[0017] in, Represents the inter-class scatter matrix. Represents the within-class scatter matrix. Represents the trace of a matrix. It is a preset weighting coefficient used to balance the importance of accuracy and separation. A genetic algorithm is used to solve for the optimal band weight vector. A population of P individuals is randomly initialized, with each individual representing a weight vector w. In each generation, the fitness of each individual is calculated. The selection process employs a roulette wheel selection method based on fitness ratios to choose parent individuals; simulated binary crossover and polynomial mutation operations are applied to the selected parents to generate offspring, with crossover probabilities... and mutation probability Based on the standard deviation of population fitness Dynamic adjustment, the adjustment rule is as follows: Increase when below the threshold To enhance population diversity; optimize the process by iteratively executing until the maximum number of generations is reached. Alternatively, the optimal fitness may not improve for multiple generations, and the final output will be the optimal band weight vector corresponding to the individual with the highest fitness. ; like Figure 4 As shown, in step S104, an initial decision tree is constructed based on agronomic knowledge and historical data. The structure is simplified through a branch-and-bound pruning algorithm to achieve dynamic decision path selection for different farmland units. Mahalanobis distance test is introduced to evaluate the decision confidence. For farmland units with low confidence, a nearest neighbor weighted median correction decision scheme is adopted. Finally, a water, fertilizer and pesticide synergistic variable operation prescription map with confidence label is generated. In some preferred embodiments, an initial decision tree is constructed based on an agronomic knowledge base and historical stress cases. Each internal node of the tree corresponds to a binary judgment rule based on multi-source features, which include low-dimensional feature vectors extracted by the student network. and the weight vector obtained by the genetic algorithm. The weighted spectral features associate each leaf node with a specific water, fertilizer, and pesticide synergistic intervention scheme. A branch-and-bound pruning algorithm is introduced to simplify the decision structure; its core is defining a node value evaluation function. ;

[0018] in, Represents a node in a decision tree. The consistency of the training samples covered by this node in terms of stress category labels is quantified using Gini impurity calculation. This represents the information gain resulting from feature determination at that node. This represents the computational and data acquisition costs required to execute the decision at this node. Preset weighting coefficients are used to balance the importance of the three indicators; the pruning process starts from the initial decision tree. Starting from the leaf nodes, backtrack upwards, for each internal node... Calculate the current value of the subtree rooted at it. And the value if it were replaced with a leaf node. ,like If established, pruning will be performed, among which A small constant tolerance threshold is used to prevent over-pruning; this comparison and pruning operation is performed recursively until the tree structure is stable, ultimately resulting in a streamlined decision tree. In real-time decision-making, for any farmland unit, the system starts from... Starting from the root node, based on the specific eigenvalues ​​of that unit... Based on weighted spectral features, branches are dynamically selected and traversed downwards according to node judgment rules until a leaf node is reached and the corresponding collaborative intervention scheme is obtained. By eliminating redundant judgment paths based on confusing features, it is ensured that the decision directly points to the optimal scheme determined by the dominant stress factor. In some preferred embodiments, for the simplified decision tree obtained by the branch-bound pruning algorithm Each leaf node represents a decision category Collect the feature vectors of all training samples belonging to this category. This vector integrates deep features extracted from the student network. And the optimal weights for morality are solved using a genetic algorithm. The key spectral features are weighted, and then the mean vector of the feature distribution for that category is calculated. With covariance matrix ;

[0019] For each farmland unit to be generated as the prescription map, it is based on its own characteristics Classified into a specific category A preliminary water, fertilizer, and pesticide recommendation scheme was obtained. To evaluate the reliability of this scheme for this unit, the feature vector of this unit was then calculated. Go to its category center The Mahalanobis distance is calculated using the following formula: ;

[0020] This distance measures the degree of deviation between the unit's features and the typical features of the category in a multidimensional related feature space; a preset distance is based on a set of degrees of freedom. Confidence threshold of chi-square distribution ,in For feature vectors The dimension; if Then the decision of this unit is determined to have a high confidence level and the original solution is adopted. If identified as a low-confidence unit, the alternative decision-making mechanism is triggered; the system searches for the unit's satisfaction within the feature space. The K nearest neighbor units of the confidence test are determined by Mahalanobis distance. The weighted median is calculated using the inverse of the spatial distance of the decision schemes that can be approached the nearest neighbor units as the weight, and is taken as the final scheme of the low confidence unit. The final decision scheme, which integrates all tested farmland units, is then applied to each unit with a... The decision confidence labels derived from the value conversion generate a water, fertilizer and pesticide synergistic variable operation prescription map with quantitative reliability indicators, including the spatial location and crop type of farmland unit, the application parameters of water, fertilizer and pesticide for each farmland unit, and the decision confidence of each farmland unit. Preferably, in this embodiment, the dataset used to train each model in the method for generating intelligent prescription maps for coordinated irrigation, fertilization, and pesticide application in farmland based on multi-source data fusion is a multispectral image library containing two crops, wheat and corn. The images were collected by the Sentinel-2 satellite and its accompanying UAV, covering three key growth stages: jointing, heading, and grain filling, and including four state samples: healthy, water-stressed, nitrogen-stressed, and disease-stressed. A total of 1200 image blocks of 512*512 pixels are randomly divided into training, validation, and test sets at a ratio of 70%, 15%, and 15%, respectively. Each image is accompanied by the true leaf area index obtained through ground sampling. In the data preprocessing stage, after radiometric calibration, atmospheric correction, and geometric correction, the reflectance of each band is normalized using the formula: ,in and The minimum and maximum values ​​for each band in the training set are given. Gaussian noise with a mean of 0 and a standard deviation of 0.02 is randomly added online to the training data input to the teacher network for data augmentation. In terms of network structure, the teacher network uses an 8-layer convolutional neural network as its backbone, specifically configured with convolutional layers having channels of [64, 128, 256, 512], followed by batch normalization and ReLU activation functions, and downsampling using 2*2 max pooling. The student network uses a 4-layer convolutional backbone with channels of [32, 64, 128, 256]. Both networks are connected to the same feature projection head, which uses a fully connected layer to reduce the feature dimension to 128, followed by a batch normalization layer. The training hyperparameters are set as follows: the optimizer is Adam, and its parameters are... The initial learning rate was set to 0.001, and a cosine annealing strategy was used for adjustment, with the following formula: ;

[0021] in The total number of training rounds is 150, the batch size is set to 32, and the total number of training rounds is 150; composite loss function. Weighting coefficients in After verification using grid search, the value was set to 0.5. Training was implemented using the PyTorch framework with an early stopping mechanism. If the validation set loss did not decrease for 15 consecutive rounds, training was terminated.

[0022] In this embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of a method for generating intelligent prescription maps for coordinated irrigation, fertilizer, and pesticide application in farmland based on multi-source data fusion.

[0023] These computer programs may also be loaded onto a computer or other programmable data processing equipment, causing a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes can be implemented by different modules for different steps.

[0024] The aforementioned program can run on a processor or be stored in memory (or computer-readable medium). Computer-readable medium includes both permanent and non-permanent, removable and non-removable media, and information storage can be achieved by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable medium does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0025] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for generating intelligent prescription maps for synergistic effects of water, fertilizer, and pesticides in farmland based on multi-source data fusion, characterized in that, Includes the following steps: A spectral feature space is obtained by acquiring and preprocessing multispectral remote sensing images of farmland. By cross-comparing unsupervised clustering and supervised classification results, pixel sets in the spectral feature space that are prone to class confusion are identified, and corresponding spatial distribution masks are generated. Spectral data of the corresponding region is extracted based on the spatial distribution mask, redundant bands are identified and removed, and a subset of effective band spectral data is output. A heterogeneous teacher-student network architecture is constructed, and a composite loss function integrating knowledge distillation and intra-class covariance constraint is designed to train the student network to extract robust features from the effective band spectral data subset. A band weight optimization problem is constructed with classification accuracy and inter-class separation as comprehensive objectives. The optimal band weight vector is solved using a genetic algorithm. The crossover and mutation probabilities are dynamically adjusted based on population diversity to adaptively adjust the weights of each spectral band. A decision tree is constructed based on agronomic knowledge and historical data. The robust features and the optimal band weight vector are used as the node features of the decision tree to complete the decision path planning. The decision results are corrected through confidence test and a water, fertilizer and pesticide synergistic variable operation prescription map is generated.

2. The method for generating intelligent prescription maps for synergistic irrigation, fertilization, and pesticide application in farmland based on multi-source data fusion as described in claim 1, characterized in that, The construction of a heterogeneous teacher-student network architecture includes: Construct a teacher network and a student network with isomorphic convolutional skeletons but different depths. Input complete multispectral data into the teacher network and a subset of effective band spectral data into the student network. Add feature projection heads after the feature extraction layers at the end of the two networks to map high-dimensional features to a unified low-dimensional embedding space. Complete network training based on a composite loss function, output a robust feature extraction model, and extract robust features from the subset of effective band spectral data.

3. The method for generating intelligent prescription maps for synergistic irrigation, fertilization, and pesticide application in farmland based on multi-source data fusion as described in claim 2, characterized in that... The method for solving the optimal band weight vector includes: Define the solution space for the band weight optimization problem, construct a fitness function with classification accuracy and inter-class separation as the core; randomly initialize the population of the genetic algorithm, calculate the fitness of individuals within the population, select parent individuals to perform crossover and mutation operations and adjust the crossover and mutation probabilities according to the population diversity, and iteratively solve for the optimal band weight vector.

4. The method for generating intelligent prescription maps for synergistic irrigation, fertilization, and pesticide application in farmland based on multi-source data fusion as described in claim 3, characterized in that... The method for generating the spatially distributed mask includes the following steps: An unsupervised clustering algorithm is used to classify farmland areas into categories. Spectral feature vectors of typical areas of each category are extracted from the spectral feature space to construct a multidimensional spectral feature vector sample library. A multi-classifier is trained based on the multidimensional spectral feature vector sample library. The classification results are compared with the clustering labels to construct an inter-class confusion matrix and locate hyperspectral confusion risk areas. After locating the hyperspectral confusion risk areas, a corresponding spatial distribution mask is generated.

5. The method for generating intelligent prescription maps for farmland water, fertilizer, and pesticide synergy based on multi-source data fusion as described in claim 4, characterized in that, The method for obtaining the effective band spectral data subset includes: Based on the spatial distribution mask, spectral data of the corresponding area is extracted from the multispectral remote sensing image of farmland. The noise level of the spectral data is evaluated by the local spatial neighborhood method to identify high-noise bands. The sparse feature selection method is used to identify information-redundant bands. The principal component analysis method is used to identify secondary information bands. The intersection of the three types of bands is used to determine the set of redundant spectral bands and output the effective band spectral data subset.

6. The method for generating intelligent prescription maps for synergistic irrigation, fertilization, and pesticide application in farmland based on multi-source data fusion as described in claim 5, characterized in that... The method for obtaining the decision path plan includes the following steps: An initial decision tree is constructed based on agronomic knowledge and historical stress cases. The robust features and optimal band weight vector are used to construct the node features of the decision tree. A node value evaluation function is defined. The tree is backtracked from the leaf nodes to the top. Branch and bound pruning operations are performed to simplify the structure of the decision tree and construct the decision tree based on the features of the farmland unit. The dynamic decision path selection is completed.

7. The method for generating intelligent prescription maps for farmland water, fertilizer, and pesticide synergy based on multi-source data fusion as described in claim 6, characterized in that, The process of correcting the decision result through confidence testing includes the following steps: Calculate the mean vector and covariance matrix of the feature distribution corresponding to each leaf node of the decision tree; calculate the Mahalanobis distance from the feature vector of the farmland unit to be decided to the center of its class; determine the confidence level of the decision result based on a preset threshold; for farmland units with low confidence, find the nearest neighbor farmland unit in the feature space by Mahalanobis distance; filter the nearest neighbor units that meet the confidence test; and correct the decision result of the low confidence unit based on the decision scheme of the nearest neighbor units.

8. A computer device, characterized in that: It includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that: The device stores a computer program that, when executed by a processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 8.

Images