Crop distribution drawing sample preparation method and device based on multi-source knowledge migration

Abstract

The invention provides a crop distribution drawing sample preparation method and device based on multi-source knowledge migration. The method comprises the following steps: respectively obtaining multiple groups of source domain representative sample sets and target domain representative sample sets based on clustering analysis; based on the multi-dimensional combination features of the multiple groups of source domain representative sample sets and the target domain representative sample set, multiple groups of source domain optimal sample sets are determined, and the multiple groups of source domain optimal sample sets are labeled and supplemented based on historical training sample points and a crop classification map to obtain multiple groups of source domain to-be-migrated sample sets; and through a domain adaptive model, mapping the multiple groups of source domain to-be-migrated sample sets and the target domain representative sample set to a shared feature space, training a classifier on the shared feature space based on the multiple groups of source domain to-be-migrated sample sets, and carrying out category prediction on the target domain representative sample set by using the trained classifier, and determining a plurality of groups of target domain labeling sample sets based on a category prediction result. The demand of on-site sampling can be reduced, and the efficiency and accuracy of sample preparation are improved.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing monitoring technology, and in particular to a method and device for preparing crop distribution mapping samples based on multi-source knowledge migration. Background Art

[0002] In recent years, crop classification and mapping using remote sensing images has become one of the important means of agricultural monitoring, playing a key role in crop planting area statistics, yield estimation and agricultural policy formulation. However, the current remote sensing image crop classification technology still faces the following major challenges: Sample dependence: High-precision crop classification and mapping rely heavily on high-quality and large numbers of training samples. These samples usually need to be obtained through field surveys, which is a time-consuming, laborious and costly process.

[0003] Time limitation: Since factors such as crop planting patterns, climate conditions, and land use patterns may change every year, the training samples collected in previous years are often not directly applicable to the classification tasks of the current year. This means that large-scale sample collection needs to be carried out again every year, which greatly increases the workload and cost.

[0004] Spatial heterogeneity: Crop growth characteristics, planting methods, and environmental conditions in different regions may vary significantly, which makes it difficult to directly apply samples collected in one region to classification tasks in other regions, limiting the generalization ability of the model.

[0005] Data imbalance: It may be difficult to obtain sufficient training samples for some crop types due to small planting areas and scattered distribution, resulting in poor recognition ability of the classification model for these categories, affecting the overall classification accuracy.

[0006] Insufficient utilization of temporal information: The growth process of crops has obvious temporal characteristics, but traditional classification methods often cannot fully utilize the rich temporal information contained in multi-temporal remote sensing images, resulting in limited classification accuracy. Summary of the invention

[0007] The present invention provides a method and device for preparing crop distribution mapping samples based on multi-source knowledge migration, which is used to solve the defects of the prior art such as high sample dependence, strong time limitation, and large spatial heterogeneity, reduce the demand for field sampling, improve the efficiency and accuracy of sample preparation, and make full use of historical data accumulated over many years, thus overcoming the limitations of traditional methods in processing time series data and spatial heterogeneity.

[0008] The present invention provides a method for preparing crop distribution mapping samples based on multi-source knowledge migration, comprising the following steps: Based on cluster analysis, a plurality of source domain representative sample sets and a target domain representative sample set are obtained respectively; wherein the plurality of source domain representative sample sets and the target domain representative sample set are obtained respectively based on a plurality of source domain images corresponding to a plurality of historical years in a remote sensing image database and a target domain image of a target year; Based on the multi-dimensional combined features of the multiple groups of source domain representative sample sets and the multi-dimensional combined features of the target domain representative sample sets, multiple groups of source domain preferred sample sets are determined from the multiple groups of source domain representative sample sets, and the multiple groups of source domain preferred sample sets are annotated and supplemented based on historical training sample points and crop classification maps to obtain multiple groups of source domain sample sets to be migrated; Based on the domain adaptation model, the multiple groups of source domain sample sets to be migrated and the representative sample sets of the target domain are mapped to a shared feature space, and classifiers are trained on the shared feature space based on the multiple groups of source domain sample sets to be migrated. The trained classifiers are used to perform category prediction on the representative sample sets of the target domain, and based on the results of the category prediction, multiple groups of target domain labeled sample sets are determined.

[0009] Optionally, the method further comprises: Performing confidence evaluation on the multiple groups of target domain labeled sample sets; Based on the confidence evaluation results of the multiple groups of target domain labeled sample sets, weighted fusion is performed on the multiple groups of target domain labeled sample sets; A consistency check is performed on the weighted fused target domain labeled sample set, and based on the result of the consistency check, the weighted fused target domain labeled sample set is optimized, and the optimized target domain labeled sample set is used as the target domain labeled sample set for model training.

[0010] Optionally, based on the multi-dimensional combined features of the multiple groups of source domain representative sample sets and the multi-dimensional combined features of the target domain representative sample sets, determining multiple groups of source domain preferred sample sets from the multiple groups of source domain representative sample sets specifically includes: Based on the multi-dimensional combined features of the multiple groups of source domain representative sample sets and the multi-dimensional combined features of the target domain representative sample set, calculating the maximum mean difference value between each group of source domain representative sample sets and the target domain representative sample set; Based on the maximum mean difference value, a heuristic search strategy is adopted to determine the source domain preferred sample set corresponding to each group of source domain representative sample sets.

[0011] Optionally, based on the domain adaptation model, mapping the multiple groups of source domain sample sets to be migrated and the target domain representative sample set to a shared feature space specifically includes: The sample features of any source domain sample set to be migrated are represented as a source domain matrix, and the sample features of the representative sample set of the target domain are represented as a target domain matrix; Inputting the source domain matrix and the target domain matrix into a transfer component analysis model to obtain a subspace mapping matrix; Based on the subspace mapping matrix, the multiple groups of source domain sample sets to be migrated and the target domain representative sample set are mapped to a shared feature space.

[0012] Optionally, after determining multiple groups of source domain preferred sample sets, the method further includes: For any source domain preferred sample set, based on a category balance threshold, checking the balance of each crop category in the any source domain preferred sample set; When any crop category in the arbitrary source domain preferred sample set is lower than the category balance threshold, samples are supplemented from the source domain representative sample set corresponding to the arbitrary source domain preferred sample set to the arbitrary source domain preferred sample set.

[0013] Optionally, the multi-dimensional combined features include one or more of spectral features, texture features, phenological features, temporal features and spatial features.

[0014] The present invention also provides a crop distribution mapping sample preparation device based on multi-source knowledge migration, comprising the following modules: An acquisition module is used to acquire multiple groups of source domain representative sample sets and target domain representative sample sets based on cluster analysis; wherein the multiple groups of source domain representative sample sets and the target domain representative sample sets are respectively acquired based on multiple groups of source domain images corresponding to multiple historical years and target domain images of target years in a remote sensing image database; A screening module is used to determine a plurality of source domain preferred sample sets from the plurality of source domain representative sample sets based on the multi-dimensional combined features of the plurality of source domain representative sample sets and the multi-dimensional combined features of the target domain representative sample sets, and to annotate and supplement the plurality of source domain preferred sample sets based on historical training sample points and crop classification maps to obtain a plurality of source domain sample sets to be migrated; A determination module is used to map the multiple groups of source domain sample sets to be migrated and the representative sample sets of the target domain to a shared feature space based on a domain adaptation model, and train classifiers on the shared feature space based on the multiple groups of source domain sample sets to be migrated, use the trained classifiers to perform category prediction on the representative sample sets of the target domain, and determine multiple groups of target domain labeled sample sets based on the results of the category prediction.

[0015] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for preparing crop distribution mapping samples based on multi-source knowledge migration as described in any one of the above is implemented.

[0016] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-described methods for preparing crop distribution mapping samples based on multi-source knowledge migration.

[0017] The present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements any of the above-mentioned methods for preparing crop distribution mapping samples based on multi-source knowledge migration.

[0018] The method and device for preparing crop distribution mapping samples based on multi-source knowledge migration provided by the present invention determine multiple groups of source domain preferred sample sets based on the multi-dimensional combined features of multiple groups of source domain representative sample sets and target domain representative sample sets, and then annotate and supplement them according to historical training sample points and crop classification maps to obtain multiple groups of source domain sample sets to be migrated, and then map the multiple groups of source domain sample sets to be migrated and the target domain representative sample sets to a shared feature space according to a domain adaptive model, train classifiers on the shared feature space based on the multiple groups of source domain sample sets to be migrated, use the trained classifiers to predict the categories of the target domain representative sample sets, and determine multiple groups of target domain labeled sample sets based on the category prediction results, which can reduce the demand for field sampling, improve the efficiency and accuracy of sample preparation, and make full use of historical data accumulated over many years to overcome the limitations of traditional methods in processing time series data and spatial heterogeneity problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0020] Figure 1 A schematic flow chart of the method for preparing crop distribution mapping samples based on multi-source knowledge migration provided by the present invention.

[0021] Figure 2 The present invention provides an overall flow chart of the embodiments.

[0022] Figure 3 A schematic diagram of the process flow of the multi-time series knowledge distillation method provided by the present invention.

[0023] Figure 4 A schematic diagram of the structure of a crop distribution mapping sample preparation device based on multi-source knowledge migration provided by the present invention.

[0024] Figure 5 This is a schematic structural diagram of an electronic device provided by the present invention. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] Figure 1 A schematic diagram of the process of preparing crop distribution mapping samples based on multi-source knowledge migration provided by the present invention is shown in FIG. Figure 1 As shown, the method comprises the following steps: Step 100: Based on cluster analysis, multiple groups of source domain representative sample sets and target domain representative sample sets are obtained respectively; wherein the multiple groups of source domain representative sample sets and target domain representative sample sets are obtained respectively based on multiple groups of source domain images corresponding to multiple historical years in a remote sensing image database and target domain images of target years.

[0027] Specifically, the remote sensing image database can be a database provided by an existing monitoring remote sensing system, such as CropWatch, etc. It is understandable that the data in the remote sensing impact database has undergone necessary preprocessing, such as radiation calibration, atmospheric correction, geometric correction, etc., and is comparable and ready for use.

[0028] In the embodiment of the present invention, if it is necessary to obtain crop distribution mapping samples for a target year, source domain images corresponding to a plurality of historical years and target domain images for the target year may be obtained from a remote sensing image database.

[0029] Optionally, in the process of obtaining source domain images corresponding to multiple historical years from the remote sensing image database, for any historical year, it is possible to first determine whether ground survey samples of crop types and crop type distribution maps of that historical year can be obtained. If ground survey samples of crop types and crop type distribution maps of that historical year can be obtained, then the source domain images corresponding to that historical year can be obtained for preparation of crop classification samples for the subsequent target year.

[0030] After obtaining multiple groups of source domain images corresponding to multiple historical years and a group of target domain images for the target year, a representative sample set of each group of images (source domain images or target domain images) can be obtained through cluster analysis, thereby obtaining multiple groups of source domain representative sample sets and target domain representative sample sets.

[0031] Optionally, the cluster analysis is performed as follows: (1) Setting the category balance threshold and sample size requirements ; (2) Perform cluster analysis on the images within the group to obtain a cluster structure that is compact within the class and separated between classes; (3) Count the number of samples in each cluster and calculate the category balance metric; (4) Calculate the neighborhood density of each sample and the distance to the cluster center; (5) Sort the samples in each cluster from small to large, and select samples from each cluster in turn, while giving priority to samples with small neighborhood density and large distance from cluster center; (6) Repeat the steps in (5) until the preset category balance threshold and total sample size requirements are reached, and a representative sample set is obtained.

[0032] Step 101: Based on the multi-dimensional combined features of the multiple groups of source domain representative sample sets and the multi-dimensional combined features of the target domain representative sample sets, multiple groups of source domain preferred sample sets are determined from the multiple groups of source domain representative sample sets, and the multiple groups of source domain preferred sample sets are annotated and supplemented based on historical training sample points and crop classification maps to obtain multiple groups of source domain sample sets to be migrated.

[0033] Specifically, first, for each sample in multiple groups of source domain representative sample sets and target domain representative sample sets, a multi-dimensional combined feature of each sample may be determined.

[0034] Optionally, the multi-dimensional combined feature includes one or more of the following features: (1) Spectral characteristics: Utilize the information of multiple spectral bands of images to calculate various vegetation indices, such as Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Soil-Adjusted Vegetation Index (SAVI), etc. These indices can effectively reflect the growth status, leaf area index, chlorophyll content and other physiological characteristics of crops.

[0035] (2) Texture features: The texture features of crops in images, such as homogeneity, contrast, and entropy, are extracted using methods such as the Gray Level Co-occurrence Matrix (GLCM) and wavelet transform. These texture features can describe the spatial structure and variation of crop canopies and play an important role in distinguishing different crop types.

[0036] (3) Phenological characteristics: Multi-temporal images are used to invert crop phenological parameters, such as sowing period, harvesting period, and length of the growing period. These phenological characteristics can reflect the growth and development process of crops and are of great value for identifying different crop varieties and estimating yields.

[0037] (4) Time series characteristics: Time series analysis methods such as Fourier transform and wavelet analysis are used to extract the time series characteristics of crops during the growing season, such as peak time, curve slope, amplitude, etc. These time series characteristics can describe the dynamic changes of crops at different growth stages and help distinguish crop types with similar spectral characteristics but different growth rhythms.

[0038] (5) Spatial characteristics: Using spatial analysis methods of images, such as texture matrix and shape index, the spatial distribution characteristics of crops in images are extracted, such as patch size, shape complexity, spatial autocorrelation, etc. These spatial characteristics can reflect the planting method and spatial layout characteristics of crops, and are of great significance for distinguishing field crops from cash crops.

[0039] Optionally, after extracting the above features, the extracted features need to be standardized to eliminate the dimensional differences between different features and improve the comparability and robustness of the features. Through image feature extraction, the original multi-temporal remote sensing image data can be converted into a high-dimensional feature space, and each pixel is represented as a multi-dimensional feature vector, which can more comprehensively and meticulously characterize the characteristics of crops.

[0040] After determining the multi-dimensional features of each sample, for any group of source domain representative sample sets, source domain samples that are consistent with the target domain feature distribution can be determined from the group of source domain representative sample sets based on the similarity between the multi-dimensional combined features of the source domain representative sample sets and the multi-dimensional combined features of the target domain representative sample sets, thereby obtaining the source domain preferred sample set.

[0041] After obtaining the source domain preferred sample set corresponding to each group of source domain representative sample sets, multiple groups of source domain preferred sample sets can be annotated and supplemented according to historical training sample points and crop classification maps to obtain multiple groups of source domain sample sets to be migrated that are both representative and of high annotation quality.

[0042] Optionally, you can first check whether there are historical training sample points and crop classification maps in the time period corresponding to the source domain preferred sample set, and handle them in the following three cases: ① If there are only historical training sample points: The historical training sample points are merged with the source domain preferred sample set to form the source domain sample set to be migrated. During the merging process, the original labels of the sample points in the source domain preferred sample set are retained; for the unlabeled sample points, no category assignment is performed. In this way, the existing high-quality labeling information can be maximized while retaining the source domain preferred sample set.

[0043] ②If there is only a crop classification map: Based on the crop classification map corresponding to the source domain image, the sample points in the source domain preferred sample set are assigned categories. Specifically, the location coordinates of each sample point are spatially matched with the crop classification map, the crop category label of the location is extracted, and it is assigned to the sample point. After this process, the source domain preferred sample set becomes the source domain sample set to be migrated with category annotation information.

[0044] ③If both historical training sample points and crop classification maps exist: The historical training sample points are merged with the source domain preferred sample set to form the initial source domain image sample set to be migrated. For the sample points that have been labeled in the source domain preferred sample set, their original labels are retained; for the unlabeled sample points, the category is assigned with reference to the crop classification map. Specifically, the position coordinates of each unlabeled sample point are spatially matched with the crop classification map, the crop category label at that position is extracted, and it is assigned to the sample point. Through this process, all sample points in the source domain preferred sample set will be labeled with categories, taking into account the information of historical sample points and crop classification maps, forming a comprehensive and high-quality source domain sample set to be migrated.

[0045] Step 102: Based on the domain adaptation model, multiple groups of source domain sample sets to be migrated and representative sample sets of the target domain are mapped to a shared feature space, and classifiers are trained on the shared feature space based on the multiple groups of source domain sample sets to be migrated. The trained classifiers are used to perform category prediction on the representative sample sets of the target domain, and multiple groups of target domain labeled sample sets are determined based on the results of the category prediction.

[0046] Specifically, in the embodiment of the present invention, a domain adaptation model is used to solve the distribution difference between the source domain and the target domain, and multiple groups of source domain sample sets to be migrated and target domain representative sample sets are mapped to a shared feature space.

[0047] Then, a classifier can be trained based on the source domain sample set to be migrated and its annotated category labels to provide a basis for category prediction of target domain samples. Appropriate classifiers can be selected according to task requirements, such as support vector machines (SVM), random forests (RF), neural networks (NN), etc., which are not limited in the embodiments of the present invention.

[0048] After obtaining the trained classifier, the classifier can be used to predict the category of the representative sample set of the target domain and obtain the category prediction results, such as the predicted category label and confidence score. The confidence score here reflects the degree of confidence of the classifier in the prediction result and can be used for subsequent sample screening.

[0049] Based on the predicted category label and confidence score, high-confidence samples can be extracted from the representative sample set of the target domain as the target domain labeled samples. High-confidence samples refer to those samples that are predicted by the classifier to be a certain category with a high confidence, and they are likely to be correct prediction results. By setting a confidence threshold, samples with scores higher than the threshold can be regarded as high-confidence samples.

[0050] Optionally, high confidence samples can be further screened and optimized, such as: (1) Category balance: Ensure that the number of extracted high-confidence samples is distributed as evenly as possible across categories to avoid having too many or too few samples in certain categories.

[0051] (2) Spatial distribution: Consider the spatial distribution characteristics of the samples and try to select samples with uniform spatial distribution and wide coverage to improve the representativeness of the samples.

[0052] (3) Uncertainty measurement: In addition to confidence scores, other uncertainty measurement indicators can be introduced, such as entropy and marginal probability, to comprehensively evaluate the reliability of samples and select samples with the smallest uncertainty.

[0053] Different source domain sample sets to be transferred can be trained to obtain different classifiers, so multiple target domain labeled sample sets can be determined based on the results of different classifier category predictions. The samples in multiple target domain labeled sample sets come from the target domain, but the category labels are obtained with the help of source domain knowledge, so they can be regarded as the result of weakly supervised learning, which can provide a large amount of training data for the target domain, and is of great significance for improving the performance of the model in the target domain.

[0054] The method for preparing crop distribution mapping samples based on multi-source knowledge migration provided by the present invention determines multiple groups of source domain preferred sample sets according to the multi-dimensional combined features of multiple groups of source domain representative sample sets and target domain representative sample sets, and then labels and supplements them according to historical training sample points and crop classification maps to obtain multiple groups of source domain sample sets to be migrated, and then maps the multiple groups of source domain sample sets to be migrated and the target domain representative sample sets to a shared feature space according to a domain adaptive model, trains classifiers on the shared feature space according to the multiple groups of source domain sample sets to be migrated, uses the trained classifiers to perform category prediction on the target domain representative sample sets, and determines multiple groups of target domain labeled sample sets based on the category prediction results, which can reduce the demand for field sampling, improve the efficiency and accuracy of sample preparation, and make full use of historical data accumulated over many years, thus overcoming the limitations of traditional methods in processing time series data and spatial heterogeneity problems.

[0055] Optionally, the method further comprises: Conduct confidence evaluation on multiple sets of target domain labeled sample sets; Based on the confidence evaluation results of multiple sets of target domain labeled sample sets, weighted fusion is performed on multiple sets of target domain labeled sample sets; A consistency check is performed on the target domain labeled sample set after weighted fusion, and based on the result of the consistency check, the target domain labeled sample set after weighted fusion is optimized, and the optimized target domain labeled sample set is used as the target domain labeled sample set for model training.

[0056] Specifically, since the quality and relevance of source domain data in different historical years may vary, directly using these multiple sets of target domain annotated sample sets for training may introduce noise and inconsistency.

[0057] Therefore, after obtaining multiple groups of target domain labeled sample sets, the embodiments of the present invention can also perform multi-sequence knowledge distillation on the target domain labeled sample sets. This step aims to make full use of the knowledge redundancy and complementarity of multiple time series, and automatically optimize the target domain labeled sample sets through weighted fusion and consistency screening to improve the reliability and consistency of the samples.

[0058] The specific process of multi-order knowledge distillation is as follows: (1) Confidence evaluation: For each set of target domain labeled samples, evaluate its confidence level. The confidence of sample annotation can be measured by using indicators such as sample prediction probability and marginal probability. The higher the confidence, the greater the reliability of sample annotation.

[0059] (2) Weighted fusion: Based on the confidence evaluation results, multiple sets of target domain labeled samples are weightedly fused. Specifically, for each target domain sample, its weighted average label in multiple sets of labeled results is calculated, and the weight is determined by the confidence of the corresponding labeled sample. In this way, samples with high confidence will be given greater weights during the fusion process, thereby dominating the fusion results.

[0060] (3) Consistency check: After obtaining the fused target domain labeled samples, a consistency check is performed. Specifically, for each sample, the consistency between its fused label and multiple sets of original annotation results is compared. If the consistency is higher than a certain threshold, the fused annotation of the sample is considered reliable; conversely, if the consistency is lower, it indicates that there are large differences in the annotation results of the sample at different time series, which may be a noise or difficult sample.

[0061] (4) Iterative optimization: According to the consistency test results, the fused target domain annotated samples are iteratively optimized. Specifically, for samples with low consistency, you can consider reducing their confidence weights or even eliminating these samples, and then repeat steps (2) and (3) until the consistency level of all samples meets the requirements. In this way, unreliable or controversial samples can be gradually filtered out, retaining the most reliable and consistent sample subset.

[0062] (5) Annotation confirmation: After multiple rounds of iterative optimization, a set of target domain annotated samples with high confidence and high consistency is finally obtained. This set of samples integrates the knowledge of multiple time series and has undergone weighted fusion and consistency screening. Therefore, it has high reliability and stability and can serve as the basis for subsequent model training and prediction.

[0063] Optionally, based on the multi-dimensional combination features of the multiple groups of source domain representative sample sets and the multi-dimensional combination features of the target domain representative sample sets, determining multiple groups of source domain preferred sample sets from the multiple groups of source domain representative sample sets specifically includes: Based on the multi-dimensional combined features of multiple groups of source domain representative sample sets and the multi-dimensional combined features of target domain representative sample sets, the maximum mean difference value between each group of source domain representative sample sets and the target domain representative sample sets is calculated; Based on the maximum mean difference value, a heuristic search strategy is adopted to determine the preferred source domain sample set corresponding to each group of source domain representative sample sets.

[0064] Specifically, in the process of determining multiple sets of source domain preferred sample sets from multiple sets of source domain representative sample sets, the Maximum Mean Discrepancy (MMD) criterion can be used to measure the distribution difference between the source domain and the target domain in the feature space. The smaller the MMD value, the closer the data distribution of the two domains is, and the greater the potential for knowledge transfer. Therefore, the ultimate goal is to select a sample subset that minimizes the MMD value between the subset and the representative sample set of the target domain while taking into account the class balance.

[0065] First, the MMD can be calculated. Based on the extracted multi-dimensional combined features, the MMD value between the representative sample set of the source domain and the representative sample set of the target domain can be calculated. MMD can be estimated using the following formula: in, and They represent the representative sample set of the source domain and the representative sample set of the target domain respectively, and They represent the representative sample sets of the source domain. samples, and the representative sample set of the target domain samples, represents the feature mapping function, and They represent the number of samples in the representative sample set of the source domain and the representative sample set of the target domain respectively.

[0066] Then, a heuristic search strategy can be used to gradually select samples from the representative sample set of the source domain and add them to the preferred sample set of the source domain. Each time when selecting, samples that can minimize the MMD value can be given priority.

[0067] Optionally, the following strategies may be adopted in the embodiments of the present invention: Initialize the source domain optimal sample set to an empty set; For each sample in the representative sample set of the source domain, calculate the change in the MMD value after adding it to the preferred sample set of the source domain; Select the sample with the smallest change in MMD value and add it to the optimal sample set; The above process is repeated until the source domain optimal sample set reaches a preset size, or the minimum value of the MMD value change is greater than a certain threshold.

[0068] Optionally, based on the domain adaptation model, multiple groups of source domain sample sets to be migrated and target domain representative sample sets are mapped to a shared feature space, specifically including: The sample features of any source domain sample set to be migrated are represented as a source domain matrix, and the sample features of the representative sample set of the target domain are represented as a target domain matrix; Input the source domain matrix and the target domain matrix into the transfer component analysis model to obtain the subspace mapping matrix; Based on the subspace mapping matrix, multiple groups of source domain sample sets to be migrated and target domain representative sample sets are mapped to the shared feature space.

[0069] Specifically, the domain adaptation model in the embodiment of the present invention can be a Transfer Component Analysis (TCA) model. TCA is a domain adaptation method based on matrix decomposition. By learning a subspace mapping, samples of the source domain and the target domain are projected into a shared feature space, so that the distribution difference between the two domains in the space is minimized.

[0070] The specific steps are as follows: (1) Data preparation: Represent the sample features of the source domain as a matrix , the sample features of the target domain are represented as a matrix . Each row of these two matrices corresponds to a sample, and each column corresponds to a feature dimension.

[0071] (2) TCA model construction: The goal of the TCA model is to learn a subspace mapping matrix , the original feature space is mapped to a shared feature space so that the MMD values ​​of the source domain and the target domain in this space are minimized.

[0072] (3) Feature conversion: using the learned mapping matrix , map the source domain sample set to be migrated and the target domain representative sample set to the shared feature space. That is, the sample feature matrix and and Multiply them together to get their representation in the shared feature space and : Through the above steps, we can obtain sample representations of the two domains in the shared feature space, and their distribution differences have been minimized, which is beneficial to subsequent knowledge transfer and classification tasks.

[0073] Optionally, after determining multiple groups of source domain preferred sample sets, the method further includes: For any source domain preferred sample set, based on the category balance threshold, check the balance of each crop category in any source domain preferred sample set; When any crop category in any source domain preferred sample set is lower than the category balance threshold, samples are added to the source domain preferred sample set from the source domain representative sample set corresponding to the any source domain preferred sample set.

[0074] Specifically, for any preferred sample set of the source domain, its samples should contain multiple crop categories. If the number of samples of some categories is obviously too small, it may affect the effect of transfer learning.

[0075] Therefore, after determining multiple groups of source domain preferred sample sets, the category balance of each group of source domain preferred sample sets can also be checked. In the embodiment of the present invention, a category balance threshold can be set, and the balance of each crop category needs to be checked according to the category balance threshold.

[0076] For crop categories whose sample numbers are lower than the category balance threshold, a certain number of samples can be added to the source domain preferred sample set from the source domain representative sample set corresponding to the source domain preferred sample set to ensure the category balance of the source domain preferred sample set.

[0077] It is understandable that different category balance thresholds can be set according to different crop categories to meet the actual needs of the sample.

[0078] The crop distribution mapping sample preparation method provided by the present invention is further explained below through examples in specific application scenarios.

[0079] The core idea of ​​the intelligent sample preparation method for crop distribution mapping based on multi-source historical knowledge migration proposed in this embodiment is to use multi-source migration learning technology to migrate the remote sensing image data and annotation information of historical years to the target year, thereby generating high-quality training samples for the crop classification mapping task of the target year. This method effectively solves the problem of requiring a large number of field samples every year in traditional methods, and significantly improves the efficiency and quality of sample preparation.

[0080] Multi-source transfer learning is an advanced machine learning paradigm that can use data and knowledge from multiple source fields to assist learning tasks in the target field. In this embodiment, the source field data includes crop distribution maps and crop training sample points in historical years, while the target field data is remote sensing images of the current year. Through multi-source transfer learning, useful information in historical data can be effectively transferred to the target year, thereby generating high-quality crop classification training samples.

[0081] The core technologies of this method include field adaptive algorithms such as Transfer Component Analysis (TCA). These algorithms can align the features of historical sample points and target images to achieve effective migration and generation of sample points. At the same time, this method also takes into account the balance of categories and the maximization of information to ensure that the generated samples are representative and balanced.

[0082] Specifically, this method first extracts multi-temporal image data of the target year and historical years from the remote sensing image database. Then, through a series of steps, including image feature extraction, representative sample selection, source domain optimal sample set generation, feature alignment, etc., the crop classification training samples of the target year are finally generated. This process not only makes full use of the information of historical data, but also improves the quality and reliability of samples through multi-source knowledge fusion.

[0083] The innovation of this method is that it can automatically complete the migration from historical data to samples of the target year, greatly reducing the need for manual sampling while ensuring the quality and representativeness of the samples. This method provides a new idea and effective way to solve the problem of sample acquisition in remote sensing image crop classification.

[0084] Figure 2 The overall flow diagram of the embodiment provided by the present invention is as follows: Figure 2 As shown, the specific steps of implementing this method are as follows: 1. Image acquisition From the CropWatch remote sensing image database, we extracted multi-temporal image data of the target year and historical years of the study area. These image data have undergone necessary preprocessing, such as radiometric calibration, atmospheric correction, geometric correction, etc., and are comparable and ready for use.

[0085] For each historical year, determine whether it meets one of the following conditions: (1) There are ground survey samples of the crop type in that year in the CropWatch ground survey database; (2) In CropWatch’s crop classification product database, there is a spatial distribution map of crop types for that year.

[0086] The image data of historical years that meet the above conditions are multi-temporal images of qualified historical years, which can be used for the preparation of crop classification samples in the subsequent target year. The image data of historical years that do not meet the conditions are not used. In this embodiment, the multi-temporal images of qualified historical years are called source domain images ( ), the multi-temporal image of the target year is called the target domain image ( ). Through this step, the available source domain images and target domain images can be quickly screened out from the massive remote sensing image database to prepare the required image data for subsequent sample migration.

[0087] After step 1 is completed, we get Group source domain image ( ) and a set of target domain images ( ). The following processing is divided into two levels: (1) Outer loop (steps 2-7): For each set of source domain images ( ), and compare it with the target domain image Input together, and execute steps 2 to 7 in sequence. After each cycle, a set of The generated target domain annotation sample points (set as ).

[0088] (2) Overall processing (step 8): Complete all After processing the source domain of the group, we can get Group target domain labeled sample points .this The sample points will be integrated and optimized in step 8 to finally obtain a high-quality target domain labeled sample set. .

[0089] 2. Image feature extraction The purpose of this step is to and target domain image In the process, multi-dimensional combined features are extracted, which can effectively characterize crop characteristics and facilitate crop classification, providing a more comprehensive and effective feature representation for subsequent sample selection and knowledge transfer.

[0090] Specifically, this method extracts the following five types of features: (1) Spectral characteristics: Utilize the information of multiple spectral bands of images to calculate various vegetation indices, such as NDVI, EVI, SAVI, etc. These indices can effectively reflect the growth status, leaf area index, chlorophyll content and other physiological characteristics of crops.

[0091] (2) Texture features: Using methods such as GLCM and wavelet transform of images, we can extract the texture features of crops on images, such as homogeneity, contrast, entropy, etc. These texture features can describe the spatial structure and variation characteristics of crop canopies, and play an important role in distinguishing different crop types.

[0092] (3) Phenological characteristics: Multi-temporal images are used to invert crop phenological parameters, such as sowing period, harvesting period, and length of the growing period. These phenological characteristics can reflect the growth and development process of crops and are of great value for identifying different crop varieties and estimating yields.

[0093] (4) Time series characteristics: Time series analysis methods such as Fourier transform and wavelet analysis are used to extract the time series characteristics of crops during the growing season, such as peak time, curve slope, amplitude, etc. These time series characteristics can describe the dynamic changes of crops at different growth stages and help distinguish crop types with similar spectral characteristics but different growth rhythms.

[0094] (5) Spatial characteristics: Using spatial analysis methods of images, such as texture matrix and shape index, the spatial distribution characteristics of crops in images are extracted, such as patch size, shape complexity, spatial autocorrelation, etc. These spatial characteristics can reflect the planting method and spatial layout characteristics of crops, and are of great significance for distinguishing field crops from cash crops.

[0095] After extracting the above features, the extracted features need to be standardized to eliminate the dimensional differences between different features and improve the comparability and robustness of the features. Through image feature extraction, the original multi-temporal remote sensing image data can be converted into a high-dimensional feature space, and each pixel is represented as a multi-dimensional feature vector. This feature representation can more comprehensively and meticulously characterize the characteristics of crops, laying a solid foundation for subsequent sample selection and knowledge transfer.

[0096] 3. Selection of representative sample sets of source domain images and target domain images For source domain images and target domain image , respectively execute the following processes to obtain the representative sample set of source domain images (set as ) and the representative sample set of target domain images (set as ): (1) Setting the category balance threshold and sample size requirements ; (2) Perform cluster analysis on the images to obtain a cluster structure with compact clusters within the cluster and separated clusters between clusters; (3) Count the number of samples in each cluster and calculate the category balance metric; (4) Calculate the neighborhood density of each sample and the distance to the cluster center; (5) Sort the samples in each cluster from small to large, and select samples from each cluster in turn, while giving priority to samples with small neighborhood density and large distance from cluster center; (6) Repeat the steps in (5) until the preset category balance threshold and total sample size requirements are reached, and a representative sample set is obtained.

[0097] 4. Generation of optimal sample set of source domain images Step 3 extracts a representative sample set of source domain images (set as ), only the information of the source domain is extracted. In this step, the information of the target domain needs to be further considered to select the sample subset that is most conducive to knowledge transfer and form the optimal sample set of the source domain image (set as The purpose of this step is to prevent negative transfer, that is, to avoid introducing some samples in the source domain that are not conducive to transfer into the target domain, which will affect the effect of transfer learning.

[0098] Specifically, the MMD criterion is used to measure the distribution difference between the source domain and the target domain in the feature space. The smaller the MMD value, the closer the data distribution of the two domains is, and the greater the potential for knowledge transfer. Therefore, the ultimate goal is to select a sample subset that minimizes the MMD value between the subset and the representative sample set of the target domain while taking into account the class balance.

[0099] The specific process is as follows: ①MMD calculation: Based on the features extracted in the above image feature extraction step, calculate the representative sample set of the source domain image Representative sample set of target domain images MMD values ​​between . MMD can be estimated by the following formula: in, and They represent the representative sample set of source domain images and the representative sample set of target domain images respectively. and They represent the representative sample sets of source domain images. samples, and the representative sample set of the target domain image samples, represents the feature mapping function, and Represent the number of samples in the representative sample set of source domain images and the representative sample set of target domain images respectively.

[0100] ② Sample selection: Adopt a heuristic search strategy to gradually select samples from the representative sample set of source domain images and add them to the optimal sample set. Each time when selecting, give priority to those samples that can minimize the MMD value. Specifically, the following strategies can be adopted: Initialize the optimal sample set to an empty set; For each sample in the representative sample set of the source domain image, calculate the change in the MMD value after adding it to the optimal sample set; Select the sample with the smallest change in MMD value and add it to the optimal sample set; The above process is repeated until the optimal sample set reaches a preset size, or the minimum value of the MMD value change is greater than a certain threshold.

[0101] ③Balance check: After obtaining the optimal sample set, you also need to check its category balance. If the number of samples in some categories is significantly small, it may affect the effect of transfer learning. Therefore, set a category balance threshold. For categories with a sample number lower than the threshold, add a certain number of samples from the representative sample set of the source domain image to ensure the category balance of the optimal sample set.

[0102] 5. Generation of source domain image sample sets to be migrated This step will adopt different strategies to optimize the source domain sample set according to the availability of historical training sample points and crop classification maps. Perform category annotation and supplementation to finally obtain a sample set of source domain images to be migrated that is both representative and of high annotation quality .

[0103] The specific process is to first check whether there are historical training sample points and crop classification maps in the time period corresponding to the source domain image, and then handle them in the following three cases: ① If there are only historical training sample points: The historical training sample points are merged with the optimal sample set of the source domain image to form the sample set of the source domain image to be migrated. During the merging process, the original labels of the sample points that have been labeled in the optimal sample set are retained; for the unlabeled sample points, no category assignment is performed. In this way, the optimal sample set can be retained while maximizing the use of the existing high-quality labeling information.

[0104] ②If there is only a crop classification map: Based on the crop classification map corresponding to the source domain image, the sample points in the optimal sample set of the source domain image are assigned categories. Specifically, the position coordinates of each sample point are spatially matched with the crop classification map, the crop category label of the position is extracted, and it is assigned to the sample point. After this process, the optimal sample set of the source domain image becomes the sample set of the source domain image to be migrated with category annotation information.

[0105] ③If both historical training sample points and crop classification maps exist: The historical training sample points are merged with the optimal sample set of the source domain image to form the initial sample set of the source domain image to be migrated. For the sample points that have been labeled in the optimal sample set, their original labels are retained; for the unlabeled sample points, the category is assigned with reference to the crop classification map. Specifically, the position coordinates of each unlabeled sample point are spatially matched with the crop classification map, the crop category label at that position is extracted, and it is assigned to the sample point. Through this process, all sample points in the optimal sample set of the source domain image will be labeled with categories, taking into account the information of the historical sample points and the crop classification map, forming a comprehensive and high-quality sample set of the source domain image to be migrated.

[0106] 6. Feature alignment based on domain adaptation After obtaining the source domain image sample set to be migrated ( ) and a representative sample set of target domain images ( ), and then input them into the domain adaptation model for feature alignment. The specific domain adaptation model used in this method is the TCA model. TCA is a domain adaptation method based on matrix decomposition. By learning a subspace mapping, the samples of the source domain and the target domain are projected into a shared feature space, so that the distribution difference between the two domains in the space is minimized.

[0107] Specifically, the steps are as follows: (1) Data preparation: Represent the sample features of the source domain as a matrix , the sample features of the target domain are represented as a matrix . Each row of these two matrices corresponds to a sample, and each column corresponds to a feature dimension.

[0108] (2) TCA model construction: The goal of the TCA model is to learn a subspace mapping matrix , the original feature space is mapped to a shared feature space so that the MMD values ​​of the source domain and the target domain in this space are minimized.

[0109] (3) Feature conversion: using the learned mapping matrix , the source domain image sample set to be migrated and the target domain image candidate sample set are mapped to the shared feature space. Specifically, this embodiment maps the sample feature matrix and and Multiply them together to get their representation in the shared feature space and : Through the above steps, we obtain sample representations of the two domains in the shared feature space, and their distribution differences have been minimized, which is beneficial to subsequent knowledge transfer and classification tasks.

[0110] 7. Generation of target domain image annotation samples After completing feature alignment in the shared feature space, we need to use the source domain sample set to be migrated ( ) has category-labeled samples, and representative samples of the target domain ( ) to perform category prediction and screening to obtain labeled samples of the target domain ( The purpose of this step is to make full use of the knowledge of the source domain to generate high-quality labeled samples for the target domain, in preparation for subsequent model training and prediction.

[0111] Specifically, the steps are as follows: ① Classifier training: Use the labeled samples in the sample set of the source domain image to be migrated to train a machine learning classifier in the shared feature space. The classifier can be any supervised learning model, such as support vector machine (SVM), random forest, neural network, etc. The shared feature representation of the source domain sample is used as input, and its corresponding category label is used as output, and the classifier is trained to minimize the classification error.

[0112] ② Target domain sample prediction: Use the trained classifier to predict the category of the candidate sample set of the target domain image. Specifically, the shared feature representation of the target domain samples is input into the classifier to obtain their predicted category labels and confidence scores. The confidence score here reflects the degree of confidence of the classifier in the prediction result and can be used for subsequent sample screening.

[0113] ③ Extraction of high-confidence samples: Based on the prediction results and confidence scores, high-confidence samples are extracted from the candidate sample set of the target domain image as the target domain image annotation samples. The high-confidence samples here refer to those samples that are predicted by the classifier as a certain category with a high confidence, and they are likely to be correct prediction results. By setting a confidence threshold, samples with scores higher than the threshold are regarded as high-confidence samples. In addition, high-confidence samples can be further screened and optimized, such as: Category balance: Ensure that the number of extracted high-confidence samples is distributed as evenly as possible across categories to avoid too many or too few samples in certain categories.

[0114] Spatial distribution: Consider the spatial distribution characteristics of samples, and try to select samples with uniform spatial distribution and wide coverage to improve the representativeness of samples.

[0115] Uncertainty measurement: In addition to the confidence score, this embodiment can also introduce other uncertainty measurement indicators, such as entropy, edge probability, etc., to comprehensively evaluate the reliability of the samples and select the samples with the smallest uncertainty.

[0116] Through the above steps, we get the target domain image annotation sample set ( ). These samples come from the target domain, but they are labeled with the help of source domain knowledge, so they can be regarded as the result of weakly supervised learning. Although these annotations may not be as accurate as manually collected samples, they can provide a large amount of training data for the target domain, which is of great significance for improving the performance of the model in the target domain.

[0117] It should be noted that when extracting high-confidence samples, the confidence threshold and screening strategy should be flexibly selected in combination with the specific task and data characteristics. Setting the threshold too high may lead to insufficient sample quantity, while setting it too low may introduce more noise annotations. Therefore, this embodiment needs to balance between sample quality and quantity and select the optimal threshold and strategy combination.

[0118] 8. Multi-temporal knowledge distillation This step aims to make full use of the knowledge redundancy and complementarity of multiple time series, automatically optimize the target domain labeled sample set through weighted fusion and consistency screening, and improve the reliability and consistency of the samples. In the previous step, this method uses TCA for feature alignment and annotates the target domain samples based on the source domain knowledge, and obtains Group target domain labeled samples ( However, since the quality and relevance of source domain data from different historical years may vary, it is not feasible to directly use this Training with a set of samples may bring noise and inconsistency. Therefore, this embodiment designs a multi-time series knowledge distillation method to The sample sets are optimized and merged to obtain a set of high-quality and consistent target domain labeled sample sets ( ).

[0119] Figure 3 A schematic diagram of the multi-time series knowledge distillation method provided by the present invention is shown in FIG. Figure 3 Specifically, the following steps are adopted: (1) Confidence evaluation: For each set of target domain labeled samples, evaluate its confidence level. The confidence of the sample can be measured by using indicators such as the sample’s prediction probability and marginal probability. The higher the confidence, the greater the reliability of the sample’s labeling.

[0120] (2) Weighted fusion: Based on the confidence evaluation results, Specifically, for each target domain sample, calculate its The weighted average label in the group annotation result is determined by the confidence of the corresponding labeled sample. In this way, samples with high confidence will be given greater weight in the fusion process, thus dominating the fusion result.

[0121] (3) Consistency check: After obtaining the fused target domain labeled samples, a consistency check is performed. Specifically, for each sample, its fused label is compared with The consistency between the original annotation results of the group is measured. If the consistency is higher than a certain threshold, the fusion annotation of the sample is considered to be reliable. On the contrary, if the consistency is low, it indicates that there are large differences in the annotation results of the sample at different time series, which may be noise or difficult samples.

[0122] (4) Iterative optimization: According to the consistency test results, the fused target domain annotated samples are iteratively optimized. Specifically, for samples with low consistency, you can consider reducing their confidence weights or even eliminating these samples, and then repeat steps (2) and (3) until the consistency level of all samples meets the requirements. In this way, unreliable or controversial samples can be gradually filtered out, retaining the most reliable and consistent sample subset.

[0123] (5) Annotation confirmation: After multiple rounds of iterative optimization, a set of target domain annotated samples with high confidence and high consistency is finally obtained. This set of samples combines The time series knowledge has been weighted fused and screened for consistency, so it has high reliability and stability and can serve as the basis for subsequent model training and prediction.

[0124] It is worth noting that this multi-time series knowledge distillation method is an unsupervised optimization strategy that does not require human participation in labeling or verification, and relies entirely on the confidence assessment and consistency analysis of the sample itself. The advantages of this method are that it has a high degree of automation, can process large-scale sample data, saves labor costs, can significantly improve the quality of target domain samples, and lays a solid foundation for model training and prediction in downstream applications. But at the same time, its effectiveness also depends on the accuracy of confidence assessment and consistency verification, and must be adjusted and optimized according to specific tasks and data characteristics.

[0125] The crop distribution mapping sample preparation device provided by the present invention is described below. The crop distribution mapping sample preparation device described below and the crop distribution mapping sample preparation method described above can be referred to each other.

[0126] Figure 4 The schematic diagram of the structure of the crop distribution mapping sample preparation device based on multi-source knowledge migration provided by the present invention is as follows: Figure 4 As shown, the device includes the following modules: The acquisition module 400 is used to respectively acquire multiple groups of source domain representative sample sets and target domain representative sample sets based on cluster analysis; wherein the multiple groups of source domain representative sample sets and target domain representative sample sets are respectively acquired based on multiple groups of source domain images corresponding to multiple historical years and target domain images of target years in a remote sensing image database; A screening module 410 is used to determine multiple groups of source domain preferred sample sets from multiple groups of source domain representative sample sets based on the multi-dimensional combined features of the multiple groups of source domain representative sample sets and the multi-dimensional combined features of the target domain representative sample sets, and to annotate and supplement the multiple groups of source domain preferred sample sets based on historical training sample points and crop classification maps to obtain multiple groups of source domain sample sets to be migrated; Determination module 420 is used to map multiple groups of source domain sample sets to be migrated and representative sample sets of the target domain to a shared feature space based on a domain adaptation model, train classifiers on the shared feature space based on the multiple groups of source domain sample sets to be migrated, use the trained classifiers to perform category prediction on the representative sample sets of the target domain, and determine multiple groups of target domain labeled sample sets based on the results of the category prediction.

[0127] Optionally, the device further comprises an iterative optimization module, configured to: Conduct confidence evaluation on multiple sets of target domain labeled sample sets; Based on the confidence evaluation results of multiple sets of target domain labeled sample sets, weighted fusion is performed on multiple sets of target domain labeled sample sets; A consistency check is performed on the target domain labeled sample set after weighted fusion, and based on the result of the consistency check, the target domain labeled sample set after weighted fusion is optimized, and the optimized target domain labeled sample set is used as the target domain labeled sample set for model training.

[0128] Optionally, based on the multi-dimensional combination features of the multiple groups of source domain representative sample sets and the multi-dimensional combination features of the target domain representative sample sets, determining multiple groups of source domain preferred sample sets from the multiple groups of source domain representative sample sets specifically includes: Based on the multi-dimensional combined features of multiple groups of source domain representative sample sets and the multi-dimensional combined features of target domain representative sample sets, the maximum mean difference value between each group of source domain representative sample sets and the target domain representative sample sets is calculated; Based on the maximum mean difference value, a heuristic search strategy is adopted to determine the preferred source domain sample set corresponding to each group of source domain representative sample sets.

[0129] Optionally, based on the domain adaptation model, multiple groups of source domain sample sets to be migrated and target domain representative sample sets are mapped to a shared feature space, specifically including: The sample features of any source domain sample set to be migrated are represented as a source domain matrix, and the sample features of the representative sample set of the target domain are represented as a target domain matrix; Input the source domain matrix and the target domain matrix into the transfer component analysis model to obtain the subspace mapping matrix; Based on the subspace mapping matrix, multiple groups of source domain sample sets to be migrated and target domain representative sample sets are mapped to the shared feature space.

[0130] Optionally, the screening module is further configured to: After determining multiple groups of source domain preferred sample sets, for any source domain preferred sample set, based on a category balance threshold, checking the balance of each crop category in any source domain preferred sample set; When any crop category in any source domain preferred sample set is lower than the category balance threshold, samples are added to the source domain preferred sample set from the source domain representative sample set corresponding to the any source domain preferred sample set.

[0131] Optionally, the multi-dimensional combined features include one or more of spectral features, texture features, phenological features, temporal features and spatial features.

[0132] Figure 5 A schematic diagram of the structure of an electronic device provided by the present invention, such as Figure 5 As shown, the electronic device may include: a processor 510, a communications interface 520, a memory 530 and a communication bus 540, wherein the processor 510, the communications interface 520 and the memory 530 communicate with each other via the communication bus 540. The processor 510 may call the logic instructions in the memory 530 to execute the crop distribution mapping sample preparation method, which includes: Based on cluster analysis, multiple groups of source domain representative sample sets and target domain representative sample sets are obtained respectively; wherein the multiple groups of source domain representative sample sets and target domain representative sample sets are obtained based on multiple groups of source domain images corresponding to multiple historical years and target domain images of target years in the remote sensing image database respectively; Based on the multi-dimensional combined features of multiple source domain representative sample sets and the multi-dimensional combined features of target domain representative sample sets, multiple source domain preferred sample sets are determined from the multiple source domain representative sample sets, and the multiple source domain preferred sample sets are annotated and supplemented based on historical training sample points and crop classification maps to obtain multiple source domain sample sets to be migrated; Based on the domain adaptation model, multiple groups of source domain sample sets to be migrated and representative sample sets of the target domain are mapped to a shared feature space, and classifiers are trained on the shared feature space based on multiple groups of source domain sample sets to be migrated. The trained classifiers are used to predict the categories of the representative sample sets of the target domain. Based on the results of the category prediction, multiple groups of target domain labeled sample sets are determined.

[0133] In addition, the logic instructions in the above-mentioned memory 530 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0134] On the other hand, the present invention further provides a computer program product, the computer program product includes a computer program, the computer program can be stored in a non-transitory computer-readable storage medium, when the computer program is executed by a processor, the computer can execute the crop distribution mapping sample preparation method provided by the above methods, the method includes: Based on cluster analysis, multiple groups of source domain representative sample sets and target domain representative sample sets are obtained respectively; wherein the multiple groups of source domain representative sample sets and target domain representative sample sets are obtained based on multiple groups of source domain images corresponding to multiple historical years and target domain images of target years in the remote sensing image database respectively; Based on the multi-dimensional combined features of multiple source domain representative sample sets and the multi-dimensional combined features of target domain representative sample sets, multiple source domain preferred sample sets are determined from the multiple source domain representative sample sets, and the multiple source domain preferred sample sets are annotated and supplemented based on historical training sample points and crop classification maps to obtain multiple source domain sample sets to be migrated; Based on the domain adaptation model, multiple groups of source domain sample sets to be migrated and representative sample sets of the target domain are mapped to a shared feature space, and classifiers are trained on the shared feature space based on multiple groups of source domain sample sets to be migrated. The trained classifiers are used to predict the categories of the representative sample sets of the target domain. Based on the results of the category prediction, multiple groups of target domain labeled sample sets are determined.

[0135] In another aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which is implemented when the computer program is executed by a processor to execute the crop distribution mapping sample preparation method provided by the above methods, the method comprising: Based on cluster analysis, multiple groups of source domain representative sample sets and target domain representative sample sets are obtained respectively; wherein the multiple groups of source domain representative sample sets and target domain representative sample sets are obtained based on multiple groups of source domain images corresponding to multiple historical years and target domain images of target years in the remote sensing image database respectively; Based on the multi-dimensional combined features of multiple source domain representative sample sets and the multi-dimensional combined features of target domain representative sample sets, multiple source domain preferred sample sets are determined from the multiple source domain representative sample sets, and the multiple source domain preferred sample sets are annotated and supplemented based on historical training sample points and crop classification maps to obtain multiple source domain sample sets to be migrated; Based on the domain adaptation model, multiple groups of source domain sample sets to be migrated and representative sample sets of the target domain are mapped to a shared feature space, and classifiers are trained on the shared feature space based on multiple groups of source domain sample sets to be migrated. The trained classifiers are used to predict the categories of the representative sample sets of the target domain. Based on the results of the category prediction, multiple groups of target domain labeled sample sets are determined.

[0136] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, i.e., they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative effort.

[0137] Through the description of the above implementation modes, those skilled in the art can clearly understand that each implementation mode can be implemented by means of software plus a necessary general hardware platform, or of course by hardware. Based on such an understanding, the above technical solution can essentially or in other words be embodied in the form of a software product that contributes to the prior art. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each embodiment or some parts of the embodiment.

[0138] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for preparing crop distribution mapping samples based on multi-source knowledge migration, characterized in that: include: Based on cluster analysis, a plurality of source domain representative sample sets and a target domain representative sample set are obtained respectively; wherein the plurality of source domain representative sample sets and the target domain representative sample set are obtained respectively based on a plurality of source domain images corresponding to a plurality of historical years in a remote sensing image database and a target domain image of a target year; Based on the multi-dimensional combined features of the multiple groups of source domain representative sample sets and the multi-dimensional combined features of the target domain representative sample sets, multiple groups of source domain preferred sample sets are determined from the multiple groups of source domain representative sample sets, and the multiple groups of source domain preferred sample sets are annotated and supplemented based on historical training sample points and crop classification maps to obtain multiple groups of source domain sample sets to be migrated; Based on the domain adaptation model, the multiple groups of source domain sample sets to be migrated and the representative sample sets of the target domain are mapped to a shared feature space, and classifiers are trained on the shared feature space based on the multiple groups of source domain sample sets to be migrated. The trained classifiers are used to perform category prediction on the representative sample sets of the target domain, and based on the results of the category prediction, multiple groups of target domain labeled sample sets are determined.

2. The method for preparing crop distribution mapping samples based on multi-source knowledge migration according to claim 1, characterized in that: The method further comprises: Performing confidence evaluation on the multiple groups of target domain labeled sample sets; Based on the confidence evaluation results of the multiple groups of target domain labeled sample sets, weighted fusion is performed on the multiple groups of target domain labeled sample sets; A consistency check is performed on the weighted fused target domain labeled sample set, and based on the result of the consistency check, the weighted fused target domain labeled sample set is optimized, and the optimized target domain labeled sample set is used as the target domain labeled sample set for model training.

3. The method for preparing crop distribution mapping samples based on multi-source knowledge migration according to claim 1, characterized in that: Based on the multi-dimensional combined features of the multiple groups of source domain representative sample sets and the multi-dimensional combined features of the target domain representative sample sets, determining multiple groups of source domain preferred sample sets from the multiple groups of source domain representative sample sets specifically includes: Based on the multi-dimensional combined features of the multiple groups of source domain representative sample sets and the multi-dimensional combined features of the target domain representative sample set, calculating the maximum mean difference value between each group of source domain representative sample sets and the target domain representative sample set; Based on the maximum mean difference value, a heuristic search strategy is adopted to determine the source domain preferred sample set corresponding to each group of source domain representative sample sets.

4. The method for preparing crop distribution mapping samples based on multi-source knowledge migration according to claim 1, characterized in that: Based on the domain adaptation model, the multiple source domain sample sets to be migrated and the target domain representative sample set are mapped to a shared feature space, specifically including: The sample features of any source domain sample set to be migrated are represented as a source domain matrix, and the sample features of the representative sample set of the target domain are represented as a target domain matrix; Inputting the source domain matrix and the target domain matrix into a transfer component analysis model to obtain a subspace mapping matrix; Based on the subspace mapping matrix, the multiple groups of source domain sample sets to be migrated and the target domain representative sample set are mapped to a shared feature space.

5. The method for preparing crop distribution mapping samples based on multi-source knowledge migration according to claim 1 or 3, characterized in that: After determining multiple groups of source domain preferred sample sets, the method further includes: For any source domain preferred sample set, based on a category balance threshold, checking the balance of each crop category in the any source domain preferred sample set; When any crop category in the arbitrary source domain preferred sample set is lower than the category balance threshold, samples are supplemented from the source domain representative sample set corresponding to the arbitrary source domain preferred sample set to the arbitrary source domain preferred sample set.

6. The method for preparing crop distribution mapping samples based on multi-source knowledge migration according to claim 1 or 3, characterized in that: The multi-dimensional combined features include one or more of spectral features, texture features, phenological features, temporal features and spatial features.

7. A sample preparation device for crop distribution mapping based on multi-source knowledge migration, characterized in that: include: An acquisition module is used to acquire multiple groups of source domain representative sample sets and target domain representative sample sets based on cluster analysis; wherein the multiple groups of source domain representative sample sets and the target domain representative sample sets are respectively acquired based on multiple groups of source domain images corresponding to multiple historical years and target domain images of target years in a remote sensing image database; A screening module is used to determine a plurality of source domain preferred sample sets from the plurality of source domain representative sample sets based on the multi-dimensional combined features of the plurality of source domain representative sample sets and the multi-dimensional combined features of the target domain representative sample sets, and to annotate and supplement the plurality of source domain preferred sample sets based on historical training sample points and crop classification maps to obtain a plurality of source domain sample sets to be migrated; A determination module is used to map the multiple groups of source domain sample sets to be migrated and the representative sample sets of the target domain to a shared feature space based on a domain adaptation model, and train classifiers on the shared feature space based on the multiple groups of source domain sample sets to be migrated, use the trained classifiers to perform category prediction on the representative sample sets of the target domain, and determine multiple groups of target domain labeled sample sets based on the results of the category prediction.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the method for preparing crop distribution mapping samples based on multi-source knowledge migration as described in any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the method for preparing crop distribution mapping samples based on multi-source knowledge migration as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, it implements the method for preparing crop distribution mapping samples based on multi-source knowledge migration as described in any one of claims 1 to 6.

Images