CN115965867A - Remote sensing image earth surface coverage classification method based on pseudo label and category dictionary learning - Google Patents
Remote sensing image earth surface coverage classification method based on pseudo label and category dictionary learning Download PDFInfo
- Publication number
- CN115965867A CN115965867A CN202211727361.7A CN202211727361A CN115965867A CN 115965867 A CN115965867 A CN 115965867A CN 202211727361 A CN202211727361 A CN 202211727361A CN 115965867 A CN115965867 A CN 115965867A
- Authority
- CN
- China
- Prior art keywords
- domain
- target domain
- dictionary
- category
- surface coverage
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Landscapes
- Image Analysis (AREA)
Abstract
The invention relates to a remote sensing image earth surface coverage classification method based on pseudo label and category dictionary learning, which comprises the following steps: extracting deep features for source domain images and target domain images by utilizing a pre-trained convolutional neural network, and screening reliable pseudo label data of a target domain; combining the source domain label data and the target domain reliable pseudo label data to form a combined training sample, and obtaining a source domain category dictionary by combining a category dictionary learning method; combining the reliable pseudo label data of the target domain to obtain a transition type dictionary and a final target domain type dictionary; combining the source domain category dictionary, the transition category dictionary and the target domain category dictionary to obtain domain invariant feature expressions of the source domain and the target domain; and (3) expressing and training the SVM classifier by using the domain invariant features of the joint training sample, and predicting the earth surface coverage category of the target domain image. The method effectively improves the classification precision and reliability of the remote sensing image ground surface coverage under the condition that the training sample is lost.
Description
Technical Field
The invention relates to the field of image processing technology, in particular to a remote sensing image earth surface coverage classification method based on pseudo labels and class dictionary learning.
Background
The ground surface covering refers to a complex of elements of the ground surface covered by natural buildings and artificial buildings, and comprises ground surface vegetation, soil, lakes, marsh wetlands and various buildings such as roads, houses and the like. Surface coverage is an important compelling factor for global environmental changes and has received widespread attention in recent years.
With the development of remote sensing technology and sensors, the resolution of remote sensing images is higher and higher, which provides feasibility for mapping of earth surface coverage. The earth surface coverage classification is an important link in earth surface coverage mapping and determines the quality of the earth surface coverage mapping. Currently, methods for classifying remote sensing images are mainly classified into two categories: pixel-oriented methods, object-oriented methods, and scene-oriented methods.
The pixels are basic units of the remote sensing images, and the method for classifying the remote sensing images by using the statistical information of the pixels is the simplest and most effective method. However, in the high-resolution remote sensing image, the ground object area corresponding to a single pixel is small. Therefore, the remote sensing image ground surface coverage classification by using the pixel-based method generates a large amount of Hu Jiaoyan noise, which reduces the precision of the ground surface coverage classification.
To overcome the disadvantages of the pixel-based approach, the object-oriented classification approach has quietly emerged. The object-oriented method firstly segments the remote sensing image into independent objects with homogeneous spectrum and continuous space, and then classifies the objects. However, the size of the object obtained by image segmentation depends on the segmentation scale parameter of the image. A larger segmentation scale may result in the feature being under-segmented, while a smaller segmentation scale may result in the feature being over-segmented.
The remote sensing image scene classification can make the most intuitive understanding of the whole remote sensing image and is greatly convenient for workers in other fields, such as urban construction planners, to make correct decisions or plans, so the remote sensing image scene classification becomes an important task in the field of intelligent remote sensing information processing. However, the task of scene classification for remote sensing images has the following problems:
1) Insufficient training sample number can lead to under-fitting of the scene classification model, and the generalization capability of the scene classification model is insufficient.
2) High intra-class difference and inter-class similarity exist in the high-resolution remote sensing image scene, and the difficulty of remote sensing image scene classification can be improved.
Disclosure of Invention
In order to solve the prior technical problem, the invention provides a remote sensing image earth surface coverage classification method based on pseudo labels and class dictionary learning.
The invention comprises the following concrete contents: the remote sensing image earth surface coverage classification method based on the pseudo label and the class dictionary learning comprises the following steps:
step (10), supposing that a source domain is the existing sample data and a target domain is the remote sensing image to be classified, extracting deep features for the source domain and the target domain images by using a pre-trained convolutional neural network, and screening reliable pseudo tag data of the target domain by calculating the distance between the target domain and the source domain images in a feature space;
step (20), the source domain label data and the target domain reliable pseudo label data form a joint training sample, and a source domain category dictionary is obtained by utilizing the joint training sample and combining a category dictionary learning method;
step (30), combining the reliable pseudo label data of the target domain, and obtaining a transition type dictionary and a final target domain type dictionary by reducing the reconstruction error of the target domain and avoiding the sudden change of the dictionary in the iterative process;
step (40), combining the source domain category dictionary, the transition category dictionary and the target domain category dictionary to obtain domain invariant feature expressions of the source domain and the target domain;
and (50) expressing and training the SVM classifier by using the domain invariant features of the joint training samples, and predicting the earth surface coverage category of the target domain image by using the trained SVM classification model.
Further, the steps of (10) obtaining the reliable pseudo tag specifically include:
(10a) Collecting existing sample data aiming at the earth surface coverage categories existing in the remote sensing images to be classified;
(10b) Assuming that a source domain is the existing sample data and a target domain is the remote sensing image to be classified; extracting deep features for source domain images and target domain images by utilizing a pre-training network;
(10c) The center of each class of the source domain in the feature space is calculated according to the following formula:
wherein
Represents a deep feature corresponding to the ith image of the source field>
The number of source field samples corresponding to the category k, <' >>
And &>
Respectively representing the ith image of the source domain and the corresponding earth surface coverage type;
(10d) Calculating the Euclidean distance between the t-th feature of the target domain and the k-th class center of the source domain in a feature space according to the following formula:
(10e) The Euclidean distance is converted into the probability that the tth feature of the target domain belongs to the kth class according to the following formula:
wherein C is the number of earth surface coverage categories;
(10f) Calculating the pseudo label of the t-th feature of the target domain according to the following formula
(10g) And setting a threshold value, and selecting the target domain image with the classification probability higher than a certain threshold value as a target domain reliable pseudo label.
Further, the source domain category dictionary calculation method in step (20) is as follows:
(20a) Combining all sample data of a source domain with the screened reliable pseudo label data of the target domain to obtain a combined training sample set;
(20b) Based on the joint training sample set, learning a dictionary of each category of the source domain according to the following formula:
wherein, Y i All features of the ith class, D i Dictionary representing ith category to be learned, X i,i Represents from D i Alpha is a parameter balancing the recognition and reconstruction capabilities of the feature, X j,i Represents from D j The characteristic expression obtained in (1), T 0 Representing an upper limit of a non-zero number in the representation of the feature;
(20c) The dictionaries of each category are connected in series to obtain a final source domain category dictionary D s ＝{D 1 ,D 2 ,...,D C }。
Further, the learning method of the transition category visual dictionary in step (30) comprises the following steps:
(30a) Updating the target domain reliable pseudo label according to the source domain category dictionary in combination with the steps (10 b) - (10 g);
(30b) Calculating a transition category visual dictionary by combining the reliable pseudo labels of the target domain according to the following formula
Y t The characteristics of the target domain are represented,
represents Y t Is at>
The characteristic expression obtained above, k representing the number of iterations, is then evaluated>
Representing a dictionary obtained by the kth iteration of the ith category;
(30c) And repeating iteration until convergence, wherein the final transition category visual dictionary is the target domain category visual dictionary.
Further, the step (50) comprises:
(50a) Training an SVM classifier by combining domain invariant feature expression of a source domain sample label and a target domain reliable pseudo label to obtain a ground surface coverage classification model;
(50b) And predicting the target domain image by using the earth surface coverage classification model to obtain an earth surface coverage classification prediction result.
Aiming at the defects of the prior art, the remote sensing image migration ground surface coverage classification method based on the pseudo label and the class dictionary learning restrains the negative influence of insufficient training samples, large intra-class difference and high inter-class similarity on the high-resolution remote sensing image ground surface coverage classification, and effectively improves the precision and reliability of the remote sensing image ground surface coverage classification under the condition of the missing training samples.
Drawings
The following further explains embodiments of the present invention with reference to the drawings.
FIG. 1 is a flow chart of a remote sensing image migration earth surface coverage classification method based on pseudo label and class dictionary learning;
FIG. 2 is source domain sample data;
FIG. 3 is a diagram of an image to be classified in a target domain;
FIG. 4 is a target domain reliable pseudo label result obtained by a remote sensing image migration earth surface coverage classification method based on pseudo label and category dictionary learning;
FIG. 5 is a classification result of a remote sensing image migration earth surface coverage classification method on a target domain image based on pseudo labels and class dictionary learning.
Detailed Description
The invention is explained in further detail below with reference to the drawings. With reference to the attached drawing 1, the method comprises the following specific steps:
(10) Assuming that a source domain is the existing sample data, a target domain is the remote sensing image to be classified, extracting deep features for the source domain and the target domain image by utilizing a pre-trained convolutional neural network, and screening reliable pseudo label data of the target domain by calculating the distance between the target domain and the source domain image in a feature space.
(10a) And collecting existing sample data aiming at the earth surface coverage categories existing in the remote sensing images to be classified.
(10b) The method includes the steps that a source domain is assumed to be the existing sample data, and a target domain is the remote sensing image to be classified. And extracting deep features for the source domain image and the target domain image by utilizing a pre-training network.
Fig. 2 shows source domain sample data of 8 different earth surface coverage categories, and fig. 3 shows a target domain image to be classified.
(10c) The center of each class of the source domain in the feature space is computed as follows.
Wherein
Represents the deep feature corresponding to the ith image of the source domain>
The number of source field samples corresponding to the category k, <' >>
And &>
Respectively representing the ith image of the source domain and the corresponding earth surface coverage type.
(10d) And according to the following formula, calculating the Euclidean distance between the t-th feature of the target domain and the k-th class center of the source domain in a feature space.
(10e) The Euclidean distance is converted into the probability that the tth feature of the target domain belongs to the kth class according to the following formula.
Where C is the number of surface coverage categories
(10f) Calculating the pseudo label of the t-th feature of the target domain according to the following formula
(10g) And setting a threshold value, and selecting the target domain image with the classification probability higher than a certain threshold value as a target domain reliable pseudo label.
FIG. 4 shows the reliable pseudo labels of the target domain obtained by using the remote sensing image migration earth surface coverage classification method based on the pseudo labels and class dictionary learning.
(20) And the source domain label data and the target domain reliable pseudo label data form a joint training sample, and a source domain category dictionary is obtained by utilizing the joint training sample and combining a category dictionary learning method.
(20a) All sample data of the source domain and the screened reliable pseudo label data of the target domain are combined to obtain a combined training sample set
(20b) Based on the joint training sample set, the dictionary of each category of the source domain is learned according to the following formula.
Wherein, Y i All features of the ith class, D i Dictionary representing ith category to be learned, X i,i Represents from D i Alpha is a parameter balancing the recognition and reconstruction capabilities of the feature, X j,i Represents from D j The characteristic expression obtained in (1), T 0 Representing an upper limit for a non-zero number in the representation of the feature.
(20c) The dictionaries of each category are connected in series to obtain a final source domain category dictionary D s ＝{D 1 ,D 2 ,...,D C }。
(30) And combining the reliable pseudo label data of the target domain, and obtaining a transition type dictionary and a final target domain type dictionary by reducing the reconstruction error of the target domain and avoiding the sudden change of the dictionary in the iteration process.
(30a) And (5) updating the reliable pseudo labels of the target domain according to the source domain category dictionary in combination with the steps (10 b) - (10 g).
(30b) Calculating a transition category visual dictionary by combining the reliable pseudo labels of the target domain according to the following formula
Y t A feature of the target domain is represented,
represents Y t Is at>
The characteristic expression obtained above, k representing the number of iterations, is then evaluated>
Representing the dictionary obtained by the kth iteration of the ith category.
(30c) And repeating iteration until convergence, wherein the final transition category visual dictionary is the target domain category visual dictionary.
(40) And combining the source domain category dictionary, the transition category dictionary and the target domain category dictionary to obtain the domain invariant feature expression of the source domain and the target domain.
(40a) And combining the source domain type dictionary, the transition type dictionary and the target domain type dictionary according to the following formula to obtain the domain invariant feature expression of the source domain image.
(40b) And combining the source domain type dictionary, the transition type dictionary and the target domain type dictionary according to the following formula to obtain the domain invariant feature expression of the target domain image.
(50) And (3) expressing and training the SVM classifier by using the domain invariant features of the joint training samples, and predicting the earth surface coverage category of the target domain image by using the trained SVM classification model.
(50a) And training the SVM classifier by combining the domain invariant feature expression of the source domain sample label and the reliable pseudo label of the target domain to obtain a surface coverage classification model.
(50b) And predicting the target domain image by using the earth surface coverage classification model to obtain an earth surface coverage classification result.
Fig. 5 shows the result of the earth surface coverage classification of the remote sensing image migration earth surface coverage classification method on the target domain image based on the label and the class dictionary learning.
The method realizes the remote sensing image migration earth surface coverage classification method based on the pseudo label and the class dictionary learning, and can provide core technical support for the application fields of national land resource investigation, environment monitoring, agriculture and forestry monitoring general survey, disaster early warning assessment and the like. Compared with the existing remote sensing image migration earth surface coverage classification method, the method has the obvious advantages that the negative influence of insufficient training samples, large intra-class difference and high inter-class similarity on high-resolution remote sensing image earth surface coverage classification can be effectively inhibited, and the accuracy and reliability of remote sensing image earth surface coverage classification can be effectively improved under the condition that the training samples are lost.
In the previous description, numerous specific details were set forth in order to provide a thorough understanding of the present invention. The foregoing description is that of the preferred embodiment of the invention only, and the invention can be practiced in many ways other than as described herein, so that the invention is not limited to the specific implementations disclosed above. And that those skilled in the art may, using the methods and techniques disclosed above, make numerous possible variations and modifications to the disclosed embodiments, or modify equivalents thereof, without departing from the scope of the claimed embodiments. Any simple modification, equivalent change and modification of the above embodiments according to the technical essence of the present invention are within the scope of the technical solution of the present invention.
Claims (5)
1. The remote sensing image earth surface coverage classification method based on the pseudo label and the class dictionary learning is characterized in that: the method comprises the following steps:
step (10), supposing that a source domain is the existing sample data and a target domain is the remote sensing image to be classified, extracting deep features for the source domain and the target domain images by utilizing a pre-trained convolutional neural network, and screening reliable pseudo label data of the target domain by calculating the distance between the target domain and the source domain images in a feature space;
step (20), the source domain label data and the target domain reliable pseudo label data form a joint training sample, and a source domain category dictionary is obtained by utilizing the joint training sample and combining a category dictionary learning method;
step (30), combining the reliable pseudo label data of the target domain, and obtaining a transition type dictionary and a final target domain type dictionary by reducing the reconstruction error of the target domain and avoiding the sudden change of the dictionary in the iterative process;
step (40), combining the source domain category dictionary, the transition category dictionary and the target domain category dictionary to obtain domain invariant feature expressions of the source domain and the target domain;
and (50) utilizing the domain invariant feature expression of the joint training sample to train an SVM classifier, and utilizing the trained SVM classification model to predict the earth surface coverage category of the target domain image.
2. The remote sensing image surface coverage classification method based on the pseudo label and the class dictionary learning as claimed in claim 1, wherein the reliable pseudo label obtaining in step (10) comprises the following specific steps:
(10a) Collecting existing sample data aiming at the earth surface coverage categories existing in the remote sensing images to be classified;
(10b) Assuming that a source domain is the existing sample data and a target domain is the remote sensing image to be classified; extracting deep features for the source domain image and the target domain image by utilizing a pre-training network;
(10c) The center of each class of the source domain in the feature space is calculated according to the following formula:
wherein
Represents a deep feature corresponding to the ith image of the source field>
The number of source field samples corresponding to the category k, <' >>
And &>
Respectively representing the ith image of the source domain and the corresponding earth surface coverage classRespectively;
(10d) Calculating the Euclidean distance between the t-th feature of the target domain and the k-th class center of the source domain in a feature space according to the following formula:
(10e) The Euclidean distance is converted into the probability that the tth feature of the target domain belongs to the kth class according to the following formula:
wherein C is the number of earth surface coverage categories;
(10f) Calculating the pseudo label of the t-th feature of the target domain according to the following formula
(10g) And setting a threshold value, and selecting the target domain image with the classification probability higher than a certain threshold value as a target domain reliable pseudo label.
3. The remote sensing image surface coverage classification method based on the pseudo label and the class dictionary learning as claimed in claim 2, characterized in that the source domain class dictionary calculation method in the step (20) is as follows:
(20a) Combining all sample data of a source domain with the screened reliable pseudo label data of the target domain to obtain a combined training sample set;
(20b) Based on the joint training sample set, learning a dictionary of each category of the source domain according to the following formula:
wherein, Y i All features representing the ith category, D i Dictionary representing ith category to be learned, X i,i Represents from D i Alpha is a parameter balancing the recognition and reconstruction capabilities of the feature, X j,i Represents from D j The characteristic expression obtained in (1), T 0 Representing an upper limit of a non-zero number in the representation of the feature;
(20c) The dictionaries of each category are connected in series to obtain a final source domain category dictionary D s ＝{D 1 ,D 2 ,...,D C }。
4. The remote sensing image surface coverage classification method based on the pseudo label and class dictionary learning of the claim 3 is characterized in that the transition class visual dictionary learning method of the step (30) comprises the following steps:
(30a) Updating the target domain reliable pseudo label according to the source domain category dictionary in combination with the steps (10 b) - (10 g);
(30b) Calculating a transition category visual dictionary by combining the reliable pseudo labels of the target domain according to the following formula
Y t The characteristics of the target domain are represented,
represents Y t Is at>
The characteristic expression obtained above, k representing the number of iterations, is then evaluated>
Representing a dictionary obtained by the kth iteration of the ith category;
(30c) And repeating iteration until convergence, wherein the final transition category visual dictionary is the target domain category visual dictionary.
5. The remote sensing image surface coverage classification method based on pseudo label and class dictionary learning as claimed in claim 1, characterized in that the step (50) comprises:
(50a) Training an SVM classifier by combining domain invariant feature expression of a source domain sample label and a target domain reliable pseudo label to obtain a ground surface coverage classification model;
(50b) And predicting the target domain image by using the earth surface coverage classification model to obtain an earth surface coverage classification prediction result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211727361.7A CN115965867A (en) | 2022-12-30 | 2022-12-30 | Remote sensing image earth surface coverage classification method based on pseudo label and category dictionary learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211727361.7A CN115965867A (en) | 2022-12-30 | 2022-12-30 | Remote sensing image earth surface coverage classification method based on pseudo label and category dictionary learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115965867A true CN115965867A (en) | 2023-04-14 |
Family
ID=87359844
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211727361.7A Pending CN115965867A (en) | 2022-12-30 | 2022-12-30 | Remote sensing image earth surface coverage classification method based on pseudo label and category dictionary learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115965867A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116704363A (en) * | 2023-05-22 | 2023-09-05 | 中国地质大学（武汉） | Deep learning model, land coverage classification method and device |
-
2022
- 2022-12-30 CN CN202211727361.7A patent/CN115965867A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116704363A (en) * | 2023-05-22 | 2023-09-05 | 中国地质大学（武汉） | Deep learning model, land coverage classification method and device |
| CN116704363B (en) * | 2023-05-22 | 2024-01-26 | 中国地质大学（武汉） | Land coverage classification method and device based on deep learning model |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Zhang et al. | Scale Sequence Joint Deep Learning (SS-JDL) for land use and land cover classification | |
| US10984532B2 (en) | Joint deep learning for land cover and land use classification | |
| EP3614308B1 (en) | Joint deep learning for land cover and land use classification | |
| CN110969088B (en) | Remote sensing image change detection method based on significance detection and deep twin neural network | |
| CN109359623B (en) | Hyperspectral image migration classification method based on depth joint distribution adaptive network | |
| CN106991382A (en) | A kind of remote sensing scene classification method | |
| CN109871875B (en) | Building change detection method based on deep learning | |
| CN110309868A (en) | In conjunction with the hyperspectral image classification method of unsupervised learning | |
| CN112232371B (en) | American license plate recognition method based on YOLOv3 and text recognition | |
| Tao et al. | Scene context-driven vehicle detection in high-resolution aerial images | |
| CN110598564B (en) | OpenStreetMap-based high-spatial-resolution remote sensing image transfer learning classification method | |
| CN109741340B (en) | Ice cover radar image ice layer refined segmentation method based on FCN-ASPP network | |
| CN109063754A (en) | A kind of remote sensing image multiple features combining classification method based on OpenStreetMap | |
| CN114694038A (en) | High-resolution remote sensing image classification method and system based on deep learning | |
| Beucher et al. | Interpretation of convolutional neural networks for acid sulfate soil classification | |
| Yuan et al. | Fine-grained classification of urban functional zones and landscape pattern analysis using hyperspectral satellite imagery: A case study of Wuhan | |
| CN110414616A (en) | A kind of remote sensing images dictionary learning classification method using spatial relationship | |
| Baeta et al. | Learning deep features on multiple scales for coffee crop recognition | |
| CN115471739A (en) | Cross-domain remote sensing scene classification and retrieval method based on self-supervision contrast learning | |
| Genitha et al. | Classification of satellite images using new fuzzy cluster centroid for unsupervised classification algorithm | |
| CN115965867A (en) | Remote sensing image earth surface coverage classification method based on pseudo label and category dictionary learning | |
| CN114549909A (en) | Pseudo label remote sensing image scene classification method based on self-adaptive threshold | |
| CN114973019A (en) | Deep learning-based geospatial information change detection classification method and system | |
| CN110766696A (en) | Satellite image segmentation method based on improved rough set clustering algorithm | |
| CN109886303A (en) | A kind of TrAdaboost sample migration aviation image classification method based on particle group optimizing |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |