Remote sensing image crop drawing method based on multi-classifier domain self-adaption

Through the multi-classifier domain adaptive method, the problem of insufficient field investigation and generalization capabilities in the existing technology is solved, and an unsupervised domain adaptive network model is constructed, and the differences in target domain classification results are calculated using knowledge distillation loss. It improves the accuracy and migration of crop classification and is suitable for areas where field samples are difficult to obtain.

CN119068353AActive Publication Date: 2024-12-03INST OF AGRI RESOURCES & REGIONAL PLANNING CHINESE ACADEMY OF AGRI SCI
View PDF 7 Cites 0 Cited by

Patent Information

Application Number
CN202411105302.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-13
Publication Date
2024-12-03
Estimated Expiration
2044-08-13

AI Technical Summary

Technical Problem

The prior art requires a large amount of field investigations in crop classification, and the generalization ability and migration ability of the model are insufficient, making it difficult to meet the needs of large-scale crop classification, especially in areas where field samples are difficult to obtain.

Method used

Using the remote sensing image crop mapping method based on multi-classifier domain adaptation, the time series curve is extracted by reconstructing satellite images, an unsupervised domain adaptive network model is constructed, and the source domain and target domain data are classified using feature extractors and classifiers. , , calculate the difference in target domain classification results using knowledge distillation losses, and maximize the decision boundary of the classifier.

Benefits of technology

It improves the accuracy and migration of crop classification, provides feasible research solutions for areas with difficulty in obtaining field samples, ensures that the classification process of source domain data is not affected by the target domain, and enhances the accuracy of feature extraction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119068353A_ABST
    Figure CN119068353A_ABST
Patent Text Reader

Abstract

The invention discloses a remote sensing image crop drawing method based on multi-classifier domain self-adaption, and the method comprises the steps: 1, reconstructing a Sentinel-2 satellite image, and extracting a time sequence curve according to the random distribution of scattered points of a crop distribution diagram; 2, extracting training set data features of a source domain and a target domain by using a feature extractor, classifying source domain data by using a classifier, and aligning the data features of the source domain and the target domain by using a domain aligner; 3, randomly generating a series of hyper-parameter training models, calculating risk function values according to the source domain training set, the source domain test set and the target domain training set, and selecting an optimal model; and 4, extracting a target domain time sequence curve in sequence, and applying the optimal model to crop spatial distribution drawing. According to the method, crop distribution drawing is carried out on the unlabeled target domain by utilizing the thought of transfer learning, the features of the source domain data are better extracted by utilizing the feature extractor and the source domain classifier, the classification decision boundary is maximized by utilizing the plurality of classifiers, and the accuracy of crop classification is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a remote sensing image crop mapping method based on multi-classifier domain adaptation, and belongs to the field of remote sensing crop classification. Background Art

[0002] Agricultural production activities play a vital role in the national social economy. Clarifying the spatial and temporal distribution of crops and the structure of crop planting is the basis for the development of modern smart agriculture. Crop classification is the process of automatically identifying and classifying vegetation cover in farmland using remote sensing technology and machine learning methods. In the early days, crop classification mainly relied on manual visual interpretation or the identification of vegetation with the help of some simple image processing techniques. Although traditional classification methods are simple and easy to use, they are limited by the subjectivity of human eye recognition and the limitations of image processing technology, and cannot meet the needs of large-scale crop classification. With the development of remote sensing technology, the application of remote sensing platforms such as satellites and drones has provided a richer source of data for crop classification. By obtaining remote sensing data in different bands, vegetation information in farmland can be accurately extracted, providing reliable data support for crop classification.

[0003] With the development of deep learning technology, crop classification based on supervised learning using spatial and temporal characteristics of crops has become a mainstream research strategy. Although these methods have achieved great success, they require a lot of field surveys to collect ground crop samples, which is time-consuming and laborious; secondly, due to economic, administrative or other reasons, it is difficult for many regions to conduct large-scale field surveys and make classification results public; finally, although supervised network models can achieve good performance in their research fields, the generalization ability of the models is still unclear, and the transferability of research results is greatly limited. Summary of the invention

[0004] In order to solve the shortcomings of the above technologies, the present invention provides a remote sensing image crop mapping method based on multi-classifier domain adaptation.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] A remote sensing image crop mapping method based on multi-classifier domain adaptation includes the following steps:

[0007] Step 1: Reconstruct satellite images to extract time series curves and create deep learning datasets. The specific steps are as follows:

[0008] Step 101: Reconstruct the time series Sentinel-2 satellite images during the crop growth period based on the discrete cosine transform-penalized least squares regression method;

[0009] Step 102: randomly distribute scattered points in the source domain and the target domain according to the open source crop distribution map, and extract the time series vegetation index curve with crop categories;

[0010] Step 103: Use the SG filter to smooth the time series vegetation index curve to obtain the source domain training set Source domain test set Target domain training set and the target domain test set

[0011] Step 2: Build an unsupervised domain adaptation network model, use the feature extractor to extract the data features of the source domain training set and the target domain training set, use the classifier to classify the source domain training data, and use the domain aligner to align the data features of the source domain and target domain training sets. The specific steps are as follows:

[0012] Step 201: Train the feature extractor Γ and the source domain classifier C S ; Use feature extractor Γ to extract data features of source domain training set, and pass through source domain classifier C S Classify the features and compare the classification results with the labels of the source domain training set Calculate the source domain cross entropy loss The formula is as follows:

[0013]

[0014] Where k represents different categories, p S represents the source domain classifier, f represents the feature extractor, is the indicator function; the calculation formula of the optimization objective is as follows:

[0015]

[0016] Step 202: fix the feature extractor Γ and train the classifiers C1 and C2; use the feature extractor Γ to extract the source domain data features, and obtain different classification results of the source domain data through the classifiers C1 and C2 respectively, and calculate the two source domain cross entropy losses with the labels of the source domain training set and The feature extractor Γ is used to extract the features of the target domain data, and different prediction results of the target domain data are obtained through two classifiers C1 and C2. The knowledge distillation loss of the two prediction results of the target domain is calculated. The formula is as follows:

[0017]

[0018] in represents KL divergence, p represents different classifiers; the calculation formula of the optimization target is as follows:

[0019]

[0020] Step 203: Fix classifiers C1 and C2, train feature extractor Γ; use feature extractor Γ to extract features of target domain data, obtain different prediction results of target domain data through classifiers C1 and C2, and calculate the knowledge distillation loss of the two prediction results The calculation formula of the optimization target is as follows:

[0021]

[0022] Step 3: Randomly generate a series of hyperparameters to train the model. Calculate the risk assessment function of the model based on the source domain training set, source domain test set, and target domain training set. Use deep embedding verification risk to evaluate the performance of the model. Select the hyperparameter with the smallest risk assessment function value to train and obtain the optimal model.

[0023] Step 4: Extract the time series curve of each pixel point in the target domain in sequence on the Sentinel-2 satellite image and perform SG filtering, and use the optimal model to predict the crop spatial distribution map of the target domain.

[0024] The present invention constructs an unsupervised domain adaptation network model, and draws on the idea of ​​transfer learning to map crop distribution in an unlabeled target domain. The multi-classifier domain adaptation model uses convolution to extract data features from the source domain and the target domain, uses multiple classifiers to classify the extracted features, and uses knowledge distillation loss to measure the difference in classification results in the target domain.

[0025] The specific advantages of the present invention are as follows:

[0026] 1) The present invention provides a feasible research solution for crop distribution mapping in areas where field samples are difficult to obtain;

[0027] 2) The present invention uses convolution to extract source domain data features, and uses an independent classifier to classify the extracted source domain features, which can ensure that the classification process of the source domain data is not affected by the target domain data, which is conducive to the feature extractor to better extract key information of the source domain data;

[0028] 3) The present invention fixes the feature extractor, trains two classifiers, uses knowledge distillation loss to calculate the difference in classification results in the target domain, maximizes the decision boundary of the classifier, and effectively improves the accuracy of classification of different crops. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 Flow chart of the implementation of the present invention.

[0030] Figure 2 It is a framework diagram of the multi-classifier domain adaptation model of the present invention.

[0031] Figure 3 This is the first step of the model training of the present invention; (a) the framework structure diagram of the first step of model training; (b) the training result of the first step.

[0032] Figure 4 This is the second step of the model training of the present invention; (a) the framework structure diagram of the second step of model training; (b) the training result of the second step.

[0033] Figure 5 This is the third step of the model training of the present invention; (a) the framework structure diagram of the third step of model training; (b) the training result of the third step.

[0034] Figure 6 Part of the classification results of the implementation cases of the present invention are demonstrated; (a) reconstructed Sentinel-2 satellite image; (b) labels of corn and winter wheat corresponding to the Czech Republic; (c) classification results of the multi-classifier domain adaptation model. DETAILED DESCRIPTION

[0035] The present invention is described in detail below in conjunction with specific embodiments.

[0036] refer to Figure 1 A remote sensing image crop mapping method based on multi-classifier domain adaptation, the method of the present invention comprises the following steps:

[0037] Step 1: Reconstruct the Sentinel-2 satellite image, extract the time series curve according to the randomly distributed scattered points of the crop distribution map, and use the SG filter to smooth the time series curve to produce a deep learning dataset. The specific steps are as follows:

[0038] Step 101: Reconstruct the blue, green, red, near-infrared and short-wave infrared images of the Sentinel-2 satellite with a resolution of 10 meters and a cloud coverage rate of less than 30% during the crop growth period based on a discrete cosine transform-penalized least squares regression method at a time interval of 10 days;

[0039] Step 102: randomly distribute 10,000 scattered points in the source domain and the target domain according to the open source crop distribution map, and extract the time series vegetation index curve with crop categories;

[0040] Step 103: Use the SG filter to smooth the time series vegetation index curve to obtain the source domain training set Source domain test set Target domain training set and the target domain test set

[0041] Step 2: Build an unsupervised domain adaptation network model, Figure 2For the structure of the multi-classifier domain adaptation model, the feature extractor is used to extract the data features of the source domain training set and the target domain training set, the classifier is used to classify the source domain training data, and the domain aligner is used to align the data features of the source domain and target domain training sets. The specific steps are as follows:

[0042] Step 201: Train the feature extractor Γ and the source domain classifier C S ,like Figure 3 shown in Figure 3 In (a), the feature extractor Γ is used to extract the data features of the source domain training set, and then the source domain classifier C S Classify the features and compare the classification results with the labels of the source domain training set Calculate the source domain cross entropy loss The formula is as follows:

[0043]

[0044] Where k represents different categories, p S represents the source domain classifier, f represents the feature extractor, is the indicator function; the calculation formula of the optimization objective is as follows:

[0045]

[0046] After continuous training of the model, Figure 3 (b) shows the final training result of this step. The feature extractor Γ can well extract the source domain data features, and the source domain classifier C S Can accurately distinguish the source crop categories;

[0047] Step 202: Fix the feature extractor Γ and train the classifiers C1 and C2, as Figure 4 shown in Figure 4 In (a), the feature extractor Γ is used to extract the features of the source domain data, and the classifiers C1 and C2 respectively obtain different classification results of the source domain data, which are different from the labels of the source domain training set. Calculate the cross entropy loss of the two source domains and The feature extractor Γ is used to extract the features of the target domain data, and different prediction results of the target domain data are obtained through two classifiers C1 and C2. The knowledge distillation loss of the two prediction results of the target domain is calculated. The formula is as follows:

[0048]

[0049] in represents KL divergence, p represents different classifiers; the calculation formula of the optimization target is as follows:

[0050]

[0051] After continuous training of the model, Figure 4 (b) shows the final training result of this step. When the target domain does not have labeled crop categories, classifiers C1 and C2 can well distinguish the source domain crop categories and accurately find the optimal decision boundary for misclassification of target domain data.

[0052] Step 203: Fix classifiers C1 and C2, and train feature extractor Γ, such as Figure 5 shown in Figure 5 In (a), the feature extractor Γ is used to extract the features of the target domain data, and different prediction results of the target domain data are obtained through classifiers C1 and C2. The knowledge distillation loss of the two prediction results is calculated. The calculation formula of the optimization target is as follows:

[0053]

[0054] After continuous training of the model, Figure 5 (b) shows the final training result of this step. The feature extractor Γ maps the source domain and target domain data features to a better feature space. When the target domain does not have annotated crop categories, the classifiers C1 and C2 can accurately distinguish different crop categories in the source domain and the target domain.

[0055] Step 3: Randomly generate a series of hyperparameters to train the model. Calculate the risk assessment function of the model based on the source domain training set, source domain test set, and target domain training set. Use deep embedding verification risk to evaluate the performance of the model. Select the hyperparameter with the smallest risk assessment function value to train and obtain the optimal model.

[0056] Step 4: Extract the time series curve of each pixel point in the target domain in sequence on the Sentinel-2 satellite image and perform SG filtering, and use the optimal model to predict the crop spatial distribution map of the target domain.

[0057] Table 1 shows the experimental results of the existing model and the model multi-classifier domain adaptation (MCDA) proposed in this paper in three groups of study areas. The accuracy rate refers to the ratio of the number of correct predictions to the total number of prediction results, and the F1 score is the harmonic mean of precision and recall. In the experiment in Table 1, Kansas is used as the source domain, the Czech Republic is used as the target domain, and corn and winter wheat are used as the main research objects; the lower Mississippi River basin M R As the source domain, HLJ in the northeastern region of Heilongjiang Province R As the target domain, rice was used as the main research object; C S As the source area, the lower Mississippi River basin M SAs the target domain, soybean is taken as the main research object. From Table 1, it can be seen that the multi-classifier domain adaptation model achieves the best accuracy and F1 score in the three groups of experiments, indicating the effectiveness of the present invention.

[0058] Table 1

[0059] source domain Target domain source domain Target domain source domain Target domain Kansas Czech Republic <![CDATA[M R ]]> <![CDATA[HLJ R ]]> <![CDATA[C S ]]> <![CDATA[M S ]]> Evaluation indicators Accuracy <![CDATA[F1 score]]> Accuracy <![CDATA[F1 score]]> Accuracy <![CDATA[F1 score]]> 1DCNN 83.60% 84.00% 82.44% 82.84% 81.02% 79.17% LSTM 79.46% 80.14% 77.62% 78.85% 80.44% 80.46% DANN 82.86% 83.42% 82.54% 82.14% 81.40% 81.23% MMDA 84.14% 84.81% 86.66% 86.34% 80.94% 81.29% AdvSKM 84.94% 81.47% 89.40% 89.78% 80.20% 79.18% MCDA 86.36% 86.42% 92.74% 92.79% 82.34% 82.40%

[0060] Figure 6 It shows some classification results of Czech. Figure 6 (a) is a partially reconstructed Sentinel-2 satellite image. Figure 6 (b) are the corresponding labels for corn and winter wheat, Figure 6 (c) is the classification result of the multi-classifier domain adaptation model, which shows that the present invention can better extract the crop type of the target domain when there is no labeled crop category in the target domain, thereby effectively completing the classification task.

[0061] The present invention has the following beneficial effects: The present invention constructs an unsupervised domain adaptive network model based on the idea of ​​transfer learning, maps crop distribution in an unlabeled target domain, and provides a feasible research scheme for crop classification tasks in areas where field samples are difficult to obtain. The multi-classifier domain adaptive model uses a feature extractor to extract data features of the source domain and target domain training sets, and uses multiple classifiers to maximize the classification decision boundary to improve the accuracy of crop classification. The present invention uses convolution to extract source domain data features, and uses an independent classifier to classify the extracted source domain features, which can ensure that the classification process of the source domain data is not affected by the target domain data, which is conducive to the feature extractor to better extract key information of the source domain data. The present invention fixes the feature extractor, trains two classifiers, uses knowledge distillation loss to calculate the difference in the classification results of the target domain, maximizes the decision boundary of the classifier, and effectively improves the accuracy of different crop classifications. In addition, the present invention can also be further used to study crop classification tasks with irregular spatial distribution, complex planting structure, and uneven proportion of classification objects, and provide data and technical support.

[0062] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all such improvements and changes should fall within the scope of protection of the appended claims of the present invention.

Claims

1. A remote sensing image crop mapping method based on multi-classifier domain adaptation, It is characterized in that The following steps are involved: Step 1: Reconstruct satellite images to extract time series curves and create data sets; Step 2: Build an unsupervised domain adaptation network model, use the feature extractor to extract the data features of the source domain training set and the target domain training set, use the classifier to classify the source domain training data, and use the domain aligner to align the data features of the source domain and target domain training sets; Step 3: Randomly generate a series of hyperparameters to train the model, calculate the risk assessment function of the model based on the source domain training set, source domain test set, and target domain training set, and select the optimal hyperparameters to train and obtain the optimal model; Step 4: Use the optimal model to predict the crop spatial distribution map of the target domain.

2. The remote sensing image crop mapping method based on multi-classifier domain adaptation according to claim 1, It is characterized in that The specific steps of step 1, making data are as follows: Step 101: Reconstruct the time series Sentinel-2 satellite images during the crop growth period based on the discrete cosine transform-penalized least squares regression method; Step 102: randomly distribute scattered points in the source domain and the target domain according to the open source crop distribution map, and extract the time series vegetation index curve with crop categories; Step 103: Use the SG filter to smooth the time series vegetation index curve to obtain the source domain training set Source domain test set Target domain training set and the target domain test set 3. The remote sensing image crop mapping method based on multi-classifier domain adaptation according to claim 1, It is characterized in that The specific steps of step 2, constructing an unsupervised domain adaptive network model are as follows: Step 201: Train the feature extractor Γ and the source domain classifier C S ; Use feature extractor Γ to extract data features of source domain training set, and pass through source domain classifier C S Classify the features and compare the classification results with the labels of the source domain training set Calculate the source domain cross entropy loss The formula is as follows: Where k represents different categories, p S represents the source domain classifier, f represents the feature extractor, is the indicator function; the calculation formula of the optimization objective is as follows: Step 202: Fix feature extractor Γ and train classifier C 1 and C 2 ; Use feature extractor Γ to extract source domain data features, and pass through classifier C 1 and C 2 Different classification results of the source domain data are obtained respectively, and two source domain cross entropy losses are calculated with the labels of the source domain training set. and The feature extractor Γ is used to extract the features of the target domain data, and then passes through two classifiers C 1 and C 2 Get different prediction results of the target domain data and calculate the knowledge distillation loss of the two prediction results of the target domain The formula is as follows: in represents KL divergence, p represents different classifiers; the calculation formula of the optimization target is as follows: Step 203: Fixing the classifier C 1 and C 2 , train the feature extractor Γ; The feature extractor Γ is used to extract the features of the target domain data, and then the features are extracted by the classifier C 1 and C 2 Get different prediction results of the target domain data and calculate the knowledge distillation loss of the two prediction results The calculation formula of the optimization target is as follows:

4. The remote sensing image crop mapping method based on multi-classifier domain adaptation according to claim 1, It is characterized in that In step 3, deep embedding verification risk is used to evaluate the performance of the model, and the hyperparameter with the smallest risk assessment function value is selected to train the optimal model.

5. The remote sensing image crop mapping method based on multi-classifier domain adaptation according to claim 1, It is characterized in that In step 4, the time series curve of each pixel point in the target domain is sequentially extracted from the Sentinel-2 satellite image and SG filtering is performed, and the crop spatial distribution map of the target domain is predicted using the optimal model.

Citation Information

Patent Citations

  • Multi-target unsupervised domain adaptive method for large-range ground feature segmentation

    CN114220003A

  • Remote sensing image classification method based on multi-classifier adversarial active transfer learning

    CN115620160A

  • Cross-domain remote sensing image confrontation migration method based on weighted discrimination and multiple classifiers

    CN116030304A

  • Passive unsupervised domain adaptive image classification method

    CN116824216A

  • Cross-domain multi-modal remote sensing image classification method based on comparative learning

    CN116912595A