Ground feature element classification and change detection method, system and device and storage medium

Abstract

The embodiment of the invention provides a ground feature element classification and change detection method, system and device and a storage medium. The method comprises the following steps: acquiring a remote sensing image; and inputting the remote sensing image into a pre-established ground feature element classification and change detection model, and outputting obtained ground feature element classification information and change information. Through fusion of semi-supervision and active learning, a sample is rapidly constructed based on sample enhancement of scene understanding, and adaptability to a complex scene is enhanced; an initial model is quickly constructed based on a smaller amount of labels and non-label column data through self-supervision and few-sample learning, so that the generalization ability of the model is improved; carrying out model training on the initial model based on the rapidly constructed sample through an improved model interpretation algorithm, so as to improve the robustness of the model to noise and improve the model interpretation capability; and finally, through intelligent interpretation result post-processing, efficient and automatic surface feature classification and change detection are realized, and the precision and reliability of surface feature element classification and change detection are greatly improved.

Description

Technical Field

[0001] The embodiments of the present disclosure belong to the technical field of ground feature classification and change detection, and specifically relate to a ground feature classification and change detection method, system, device and storage medium. Background Art

[0002] Remote sensing feature classification is a process of identifying and classifying different types of objects or features on the earth's surface using remote sensing technology. It is widely used in land use and land cover change monitoring, ecological environment monitoring, agricultural resource management, urban planning and other fields. Remote sensing image change detection generally refers to using a remote sensing image of one phase as a reference to detect the difference between another phase and the reference image. It is a theory and method that extracts and describes the characteristics of objects or phenomena of interest that change over time based on observations of remote sensing images and reference data at different phases, and quantitatively analyzes and determines their changes.

[0003] Existing methods for ground feature classification and change detection use manually annotated samples and traditional machine learning methods based on shallow learning. They are inefficient when processing large-scale data sets, have difficulty capturing complex spatial and spectral features, and have limited classification and detection accuracy.

[0004] However, the use of data augmentation, synthetic samples, and deep learning-based interpretation methods may cause the classifier to overfit the enhanced samples because the generated samples may not be completely consistent with the actual data distribution. The synthetic samples may not fully reflect the complexity of real objects, thus affecting the accuracy of classification and detection. The training process of deep learning models requires a large amount of labeled data and is easily affected by sample imbalance and noise. Summary of the invention

[0005] The embodiments of the present disclosure aim to solve at least one of the technical problems existing in the prior art and provide a method, system, device and storage medium for ground feature classification and change detection.

[0006] One aspect of the present disclosure provides a method for ground feature classification and change detection, the method comprising:

[0007] Acquire a remote sensing image; wherein the remote sensing image includes a first phase remote sensing image and a second phase remote sensing image;

[0008] The remote sensing image is input into a pre-established ground feature classification and change detection model, and the ground feature classification information and change information are output; wherein,

[0009] The land feature classification and change detection model is obtained based on the fusion of semi-supervised learning and active learning, sample enhancement based on scene understanding, the combination of self-supervised learning and few-sample learning, and an improved interpretation model.

[0010] Furthermore, the ground feature classification and change detection model is pre-established through the following steps:

[0011] Acquire a remote sensing image sample set; wherein the sample set includes labeled samples and unlabeled samples;

[0012] Using the labeled samples to train a self-supervised model to obtain an initial model, and then using the initial model to screen out high-quality samples from the unlabeled samples;

[0013] Active learning is used to select high-information samples from the unlabeled samples, and the high-information samples are labeled to update the remote sensing image sample set;

[0014] The high-quality samples are enhanced by using a SimCLR network, feature vectors of the enhanced high-quality samples are extracted by using a convolutional neural network, and the initial model is subjected to few-sample learning using the feature vectors and the labeled samples;

[0015] The interpretation algorithm of the initial model is improved by combining a multi-scale convolutional neural network, a two-stream network and a contrastive learning algorithm to obtain a ground feature classification and change detection model.

[0016] Furthermore, the using the initial model to select high-quality samples from the unlabeled samples includes:

[0017] Performing sample enhancement on the unlabeled sample to obtain an enhanced unlabeled sample;

[0018] Inputting the enhanced unlabeled column sample into the initial model to obtain the confidence of each pixel;

[0019] A consistency regularization method is used to compare the confidences of the unlabeled samples and the enhanced unlabeled samples, and samples with confidences greater than a preset threshold are screened out as high-quality samples.

[0020] Furthermore, the adopting active learning to select high-information samples from the unlabeled samples and labeling the high-information samples includes:

[0021] Use the current model to predict the category probability distribution of the unlabeled samples and calculate the entropy value;

[0022] A batch of unlabeled samples with the highest entropy value are determined as high-information samples, and the high-information samples are labeled.

[0023] Furthermore, the entropy value is calculated by the following formula:

[0024]

[0025] Where P iis the predicted probability of the i-th class.

[0026] Furthermore, before using the SimCLR network to enhance the high-quality sample, the method further includes:

[0027] The U-Net network, generative adversarial network and data consistency judgment are used to perform sample enhancement based on scene understanding on the labeled samples.

[0028] Furthermore, after obtaining the classification information and change information of the ground feature elements, the method further includes:

[0029] The classification information and change information of the ground feature elements are subjected to marginalization post-processing.

[0030] Another aspect of the present disclosure provides a system for ground feature classification and change detection, the system comprising:

[0031] An image acquisition module, used to acquire remote sensing images; wherein the remote sensing images include remote sensing images of a first time phase and remote sensing images of a second time phase;

[0032] The classification detection module is used to input the remote sensing image into a pre-established ground feature classification and change detection model, and output ground feature classification information and change information; wherein,

[0033] The land feature classification and change detection model is obtained based on the fusion of semi-supervised learning and active learning, sample enhancement based on scene understanding, the combination of self-supervised learning and few-sample learning, and an improved interpretation model.

[0034] Another aspect of the present disclosure provides an electronic device, comprising:

[0035] At least one processor; and a memory in communication with the at least one processor, for storing one or more programs, which, when executed by the at least one processor, enables the at least one processor to implement the above-mentioned method for land feature classification and change detection.

[0036] Another aspect of the present disclosure provides a computer-readable storage medium storing a computer program, which implements the above-mentioned method for classifying land features and detecting changes when executed by a processor.

[0037] A method, system, device and storage medium for ground feature classification and change detection in an embodiment of the present disclosure rapidly constructs samples by integrating semi-supervised and active learning and sample enhancement based on scene understanding, thereby enhancing adaptability to complex scenes; rapidly constructs an initial model based on a smaller amount of labeled and unlabeled data through self-supervision and few-sample learning, thereby improving the generalization ability of the model; trains the above initial model based on rapidly constructed samples through an improved model interpretation algorithm, thereby improving the model's robustness to noise and the model's interpretation ability; and finally, through intelligent interpretation result post-processing, realizes efficient and automated ground feature classification and change detection, thereby greatly improving the accuracy and reliability of ground feature classification and change detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A schematic diagram of a flow chart of a method for ground feature classification and change detection according to an embodiment of the present disclosure;

[0039] Figure 2 A schematic diagram of before and after intelligent interpretation of a building image according to another embodiment of the present disclosure;

[0040] Figure 3 A schematic diagram of a road image before and after intelligent interpretation according to another embodiment of the present disclosure;

[0041] Figure 4 A schematic diagram of another embodiment of the present disclosure before and after intelligent interpretation of a water body image;

[0042] Figure 5 It is a schematic diagram before and after the intelligent interpretation of general change detection according to another embodiment of the present disclosure;

[0043] Figure 6 It is a structural schematic diagram of a ground feature classification and change detection system according to another embodiment of the present disclosure;

[0044] Figure 7 The figure is a schematic diagram of the structure of an electronic device according to another embodiment of the present disclosure. DETAILED DESCRIPTION

[0045] Remote sensing feature classification is a process of identifying and classifying different types of objects or features on the earth's surface using remote sensing technology. Remote sensing feature classification technology is widely used in land use and land cover change monitoring, ecological environment monitoring, agricultural resource management, urban planning and other fields. Remote sensing image change detection generally refers to using a remote sensing image of one phase as a reference to detect the difference between another phase and the reference image. It is a theory and method that extracts and describes the characteristics of the objects or phenomena of interest that change over time based on observations of remote sensing images and reference data at different phases, and quantitatively analyzes and determines their changes. High-quality sample construction and intelligent interpretation technology for feature classification and change detection are key links in remote sensing applications. High-quality sample data and intelligent interpretation technology can significantly improve the accuracy of feature classification and change detection. Therefore, this paper proposes a high-quality sample construction and intelligent interpretation technology method for feature classification and change detection.

[0046] Among the existing methods for classification and change detection of land features, there are those that use manually annotated samples and traditional machine learning methods based on shallow learning. Manually annotated samples are usually annotated by experts for target tasks based on high-resolution remote sensing images. Shallow learning methods such as support vector machines (SVM), decision trees, and random forests rely on manually extracted features for classification and change detection. Manual labeling is time-consuming and labor-intensive, and the quality of labeling is greatly affected by human factors, especially when dealing with large-scale data sets. In addition, the limited number of annotated samples may lead to sample imbalance problems and affect the generalization ability of the model. Traditional machine learning methods rely on manual feature extraction, which makes it difficult to capture complex spatial and spectral features and has limited classification accuracy. In addition, these methods show poor scalability when dealing with large-scale, multi-dimensional remote sensing data.

[0047] There are also interpretation methods based on deep learning that use data augmentation and synthetic samples. Data augmentation generates more samples by rotating, scaling, flipping, and other operations on existing samples, while synthetic samples use algorithms to generate simulated samples to supplement the data set. Deep learning methods, such as Convolutional Neural Networks (CNN) and spatiotemporal networks, automatically extract features to classify objects and detect changes. However, although data augmentation can increase the diversity of samples, the generated samples may not be completely consistent with the actual data distribution, resulting in overfitting of the classifier to the enhanced samples. In addition, synthetic samples may not fully reflect the complexity of real objects, affecting classification accuracy. Deep learning models have high requirements for data volume and computing resources. The training process requires a large amount of labeled data and is easily affected by sample imbalance and noise.

[0048] In order to make up for the shortcomings of existing methods in sample construction for land feature classification and change detection and to provide more accurate intelligent interpretation, the present invention proposes a technical method for high-quality sample construction and intelligent interpretation for land feature classification and change detection. This method can well deal with various problems of sample construction methods, thereby improving the reliability of remote sensing interpretation.

[0049] The following will be combined with the drawings in the embodiments of the present disclosure to clearly and completely describe the technical solutions in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present disclosure.

[0050] In addition, the described features, structures or characteristics may be combined in one or more embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present disclosure. However, those skilled in the art will appreciate that the technical solutions of the present disclosure may be practiced without one or more of the specific details, or other methods, components, devices, steps, etc. may be adopted. In other cases, known methods, devices, implementations or operations are not shown or described in detail to avoid blurring the various aspects of the present disclosure.

[0051] The flowcharts shown in the accompanying drawings are only exemplary and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps can be decomposed, and some operations / steps can be combined or partially combined, so the actual execution order may change according to actual conditions.

[0052] It should be understood that although the terms first, second, third, etc. may be used in this disclosure to describe various components, these components should not be limited by these terms. These terms are used to distinguish one component from another component. Therefore, the first component discussed below can be referred to as the second component without departing from the teachings of the concepts of the present disclosure. As used in this disclosure, the term "and / or" includes any one of the associated listed items and all combinations of one or more.

[0053] Those skilled in the art will appreciate that the drawings are merely schematic diagrams of example embodiments, and the modules or processes in the drawings are not necessarily necessary for implementing the present disclosure, and therefore cannot be used to limit the protection scope of the present disclosure.

[0054] like Figure 1 As shown, an embodiment of the present disclosure provides a method for ground feature classification and change detection, including:

[0055] Step S1, acquiring remote sensing images; wherein the remote sensing images include remote sensing images of a first time phase and remote sensing images of a second time phase.

[0056] Specifically, color or black-and-white optical remote sensing images of the surface can be obtained by optical sensors installed on satellites, aerial vehicles, etc. The first phase remote sensing image and the second phase remote sensing image are images of the same area obtained at two different times. Remote sensing images of a single phase can be used for classification of ground features, and the combination of remote sensing images of multiple phases can be used for image enhancement or change detection of ground features. Remote sensing images of multiple phases can be obtained at different times based on the same optical sensor, or they can be a fusion of remote sensing images of multiple optical sensors and multiple phases.

[0057] Step S2: input the remote sensing image into a pre-established ground feature classification and change detection model, and output ground feature classification information and change information.

[0058] Specifically, the remote sensing image acquired in the previous step S1 is input into the ground feature classification and change detection model, and the model outputs the classification result and the detection result.

[0059] The ground feature classification and change detection model of the disclosed embodiment realizes high-quality sample construction and efficient sample annotation, including semi-supervised learning and active learning, sample enhancement based on scene understanding, self-supervised learning and few-sample learning, deep learning model and improved model interpretation algorithm. The specific establishment steps are as follows:

[0060] 1. Obtain a remote sensing image sample set, wherein the sample set includes a small number of labeled samples and a large number of unlabeled samples. For example, for a large number of remote sensing image samples obtained, 1,000 sample spots are manually labeled, and the remaining samples are not labeled, and then all samples are cropped to 1024 pixels * 1024 pixels.

[0061] 2. Integration of semi-supervised learning and active learning:

[0062] First, the deep belief network (DBN) is trained using the aforementioned small amount of labeled samples. The learning rate is set to 0.001 and the number of training iterations is 50. The initial model is obtained through training.

[0063] Then, semi-supervised learning is used to further screen high-quality samples, and unlabeled image samples are selected to generate enhanced samples. The enhanced samples may include the original image, the image after arbitrary rotation of 15°, the image after cropping 15% of the area, and the image after the brightness is increased by 20%; the trained initial model is used to predict the above enhanced samples to obtain the category probability prediction confidence of each pixel; then the consistency regularization model is used to compare the prediction results of the samples before and after enhancement, and the confidence similarity is output. Samples with confidence greater than a preset threshold such as 0.9 are screened as high-quality training samples for subsequent model training; the loss function of the above consistency regularization model is shown as follows:

[0064]

[0065] Where P original and P augmented Respectively represent the predicted probability distribution of the image before and after enhancement;

[0066] Then, active learning is used to select the samples with the highest information content from the unlabeled samples for manual labeling, and the sample set and model are updated to improve the labeling efficiency: the current model is used to predict the category probability distribution of the unlabeled samples and calculate the entropy value. The higher the entropy value, the higher the uncertainty of the sample, and therefore the higher the amount of information. The samples with the highest entropy value, such as the top 5%, are selected as the queue to be labeled; the entropy value calculation formula is as follows:

[0067]

[0068] Where P i is the predicted probability of the i-th class.

[0069] 3. Sample enhancement based on scene understanding: In order to increase the extended data set of remote sensing images in different regions, sensors, and resolutions, improve the model's adaptability to complex and unseen scenes, enhance the diversity of samples, and improve the model training effect, it is necessary to expand the number of samples in an automated way.

[0070] (1) Scene segmentation and analysis: The labeled samples are input as training samples, and the U-Net network is used to train the segmentation model. The convolution kernel size is 3×3, the activation function is ReLU, the optimizer is Adam, the learning rate is set to 0.001, the number of iterations (epochs) is 50-100, and the batch size is set to 16.

[0071] (2) Sample enhancement: Based on the trained U-Net model, image data with a data volume that is 8 times larger than the original sample is classified into categories such as buildings, roads, and vegetation. The segmentation results and object category information are input to train the generative adversarial network (GAN) model so that the model can generate a generator and a discriminator. The optimizer is Adam, the learning rate is set to 0.0002, the batch size is 132, and the number of training iterations ranges from 1,000 to 10,000. Based on the trained generative adversarial network, geometric transformations and illumination changes are applied to generate more diverse enhanced remote sensing images that are similar to the actual scene features.

[0072] (3) Data consistency judgment: Input the original image and the image enhanced by the generative adversarial network, and use the structural similarity index (SSIM) to calculate the similarity of the image's structural information. When the similarity exceeds a preset threshold of 0.8, the image data generated by the generative adversarial network is retained and used as high-quality enhanced image data, otherwise the data is discarded. The retained image data can increase the number and diversity of samples and alleviate the problem of insufficient samples. On the other hand, it can enhance the generalization ability, that is, improve the model's classification and change detection capabilities in complex scenes; at the same time, samples are screened through consistency verification to ensure data reliability.

[0073] 4. Combination of self-supervised learning and few-sample learning:

[0074] (1) Self-supervised learning: A large number of unlabeled remote sensing images are input, and the SimCLR network based on contrastive learning is used to generate different enhanced versions of each image (such as rotation, cropping, color jitter, and horizontal flipping); a convolutional neural network such as ResNet50 is used to extract the feature vector of the image, so that the feature vectors of different enhanced versions of the same image are closer, while the feature vectors of different images are farther apart; the SimCLR model is optimized through a contrast loss function such as InfoNCE to enable it to have the ability to distinguish image features; the learning rate during the model training process is set to 0.0001, the batch size is 128, the data augmentation methods are random cropping, color jittering, and horizontal flipping, and the number of iterations is 20,000.

[0075] (2) Few-shot learning: In the case of limited labeled data, few-shot learning techniques such as ProtoNet prototype network can be used to achieve efficient object classification (feature representation obtained by self-supervised learning can be used). Through training with a small number of labeled samples, classification accuracy and generalization ability can be improved; 5 remote sensing images and labeled samples of each category are input, combined with unlabeled remote sensing images, and feature extraction is performed using the SimCLR model. The features of the two are compared to predict the category of the unlabeled image. The training parameters of the ProtoNet network are as follows: learning rate 0.001, number of tasks per batch 32, number of samples per category K = 5, number of iterations 5,000, and optimizer Adam.

[0076] 5. Improve model interpretation:

[0077] The model is interpreted through multi-scale feature extraction and multi-temporal data fusion, and contrastive learning technology is used to enhance the model's change recognition ability and reduce false alarms caused by noise. Multi-scale CNN, two-stream network (Siamese Network) and other methods are used for feature extraction and fusion, combined with contrastive learning algorithm (SimSiam) to enhance the robustness of model interpretation.

[0078] (1) Multi-scale feature extraction: Single-phase remote sensing images and annotated samples, as well as two-phase remote sensing images and change patch samples are input. Multi-scale convolutional neural networks are used to simultaneously capture ground object details and global semantic features. The backbone network is FPN (Feature Pyramid Network), with a learning rate of 0.0001, a batch size of 16, and 50,000 iterations.

[0079] (2) Multi-temporal feature fusion: For change detection, the feature maps F extracted from two phases of remote sensing images are input respectively. t1 and F t2 Based on the two-stream network (Siamese Network), the difference fusion and splicing fusion methods are used to extract and compare the features of the two time points. The loss function in the training process adopts the binary cross entropy loss function to judge the changed and unchanged areas. The learning rate is set to 0.001, the batch size is set to 32, and the number of iterations is set to 30,000.

[0080] (3) Contrastive Learning (SimSiam): Input positive and negative sample pairs, such as features of changed areas and unchanged areas, use two branches to calculate the feature representation of the sample pairs, optimize the contrast loss function of the positive sample pair loss, and output the optimized feature representation for classification and change detection, thereby obtaining the ground feature classification and change detection model of this embodiment.

[0081] Step S3: perform marginalization post-processing on the classification information and change information of the ground feature elements.

[0082] Specifically, the prediction results of the model intelligent interpretation are prone to a large number of fragmentation, small holes, boundary jaggedness and other problems. A method of optimizing the edge of the patch based on the spatial constraints of the edge pixel distribution is adopted. By extracting the semantic features and edge features of the image, post-processing rules are designed, including simplification (such as the Douglas-Peucker algorithm), deletion of small patches according to area, island hole filling (i.e., filling the hole area within the polygon), closing operation (i.e., corroding the patch first and then expanding it), opening operation (i.e., expanding the patch first and then corroding it), building regularization, etc. The vector map patches in the ground feature classification information and change information obtained in the previous step S2 are post-processed by edgeization to obtain the final ground feature classification and change detection results, thereby improving the accuracy of intelligent interpretation of remote sensing images.

[0083] In some embodiments, schematic diagrams of the building image before and after intelligent interpretation are shown in FIG. Figure 2 As shown in the figure, the left side is the building image before interpretation, and the right side is the building image after interpretation; the schematic diagram of the road image before and after intelligent interpretation is shown in Figure 3 As shown in the figure, the left side is the road image before interpretation, and the right side is the road image after interpretation; the schematic diagram of the water body image before and after intelligent interpretation is shown in Figure 4 As shown in the figure, the left side is the water body image before interpretation, and the right side is the water body image after interpretation; the schematic diagram of the general change detection before and after intelligent interpretation is shown in Figure 5 As shown, the left side is the universal change detection image before interpretation, and the right side is the universal change detection image after interpretation.

[0084] A method for ground feature classification and change detection in an embodiment of the present disclosure rapidly constructs samples by integrating semi-supervised and active learning and sample enhancement based on scene understanding, thereby enhancing adaptability to complex scenes; rapidly constructs an initial model based on a smaller amount of labeled and unlabeled data through self-supervision and few-sample learning, thereby improving the generalization ability of the model; trains the above initial model based on rapidly constructed samples through an improved model interpretation algorithm, thereby improving the model's robustness to noise and the model's interpretation ability; and finally, through intelligent interpretation result post-processing, realizes efficient and automated ground feature classification and change detection, thereby greatly improving the accuracy and reliability of ground feature classification and change detection.

[0085] like Figure 6 As shown, another embodiment of the present disclosure provides a ground feature classification and change detection system, the system comprising:

[0086] The image acquisition module 610 is used to acquire remote sensing images; wherein the remote sensing images include remote sensing images of a first time phase and remote sensing images of a second time phase;

[0087] The classification detection module 620 is used to input the remote sensing image into a pre-established ground feature classification and change detection model, and output ground feature classification information and change information; wherein,

[0088] The land feature classification and change detection model is obtained based on the fusion of semi-supervised learning and active learning, sample enhancement based on scene understanding, the combination of self-supervised learning and few-sample learning, and an improved interpretation model.

[0089] Specifically, a ground feature classification and change detection system of an embodiment of the present disclosure is used to implement the ground feature classification and change detection method described in the above embodiments. The specific implementation process has been described in detail in the above embodiments and will not be repeated here.

[0090] A land feature classification and change detection system of an embodiment of the present disclosure rapidly constructs samples by integrating semi-supervised and active learning and sample enhancement based on scene understanding, thereby enhancing adaptability to complex scenes; rapidly constructs an initial model based on a smaller amount of labeled and unlabeled data through self-supervision and few-sample learning, thereby improving model generalization ability; trains the above initial model based on rapidly constructed samples through an improved model interpretation algorithm, thereby improving the model's robustness to noise and improving the model's interpretation ability; and finally, through intelligent interpretation result post-processing, realizes efficient and automated land feature classification and change detection, thereby greatly improving the accuracy and reliability of land feature classification and change detection.

[0091] like Figure 7 As shown, another embodiment of the present disclosure provides an electronic device, including:

[0092] At least one processor 701; and a memory 702 communicatively connected to the at least one processor 701, for storing one or more programs, which, when executed by the at least one processor 701, enable the at least one processor 701 to implement the above-mentioned method for ground feature classification and change detection.

[0093] The memory 702 and the processor 701 are connected in a bus manner, and the bus may include any number of interconnected buses and bridges, and the bus connects various circuits of one or more processors 701 and the memory 702 together. The bus can also connect various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and are therefore not further described herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be one element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices on a transmission medium. The data processed by the processor 701 is transmitted on a wireless medium via an antenna, and further, the antenna also receives data and transmits the data to the processor 701.

[0094] The processor 701 is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interfaces, voltage regulation, power management and other control functions. The memory 702 can be used to store data used by the processor 701 when performing operations.

[0095] Yet another embodiment of the present disclosure provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the above-mentioned method for classifying land features and detecting changes.

[0096] The computer-readable storage medium may be included in the system or electronic device of the present disclosure, or may exist independently.

[0097] Computer-readable storage media may be any tangible media that contains or stores a program, which may be an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device. More specific examples include, but are not limited to, an electrical connection with one or more wires, a portable computer disk, a hard disk, optical fiber, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0098] The computer-readable storage medium may also include a data signal propagated in baseband or as part of a carrier wave, in which the computer-readable program code is carried. Specific examples include but are not limited to electromagnetic signals, optical signals, or any suitable combination thereof.

[0099] It is to be understood that the above embodiments are merely exemplary embodiments used to illustrate the principles of the present disclosure, but the present disclosure is not limited thereto. For those of ordinary skill in the art, various modifications and improvements can be made without departing from the spirit and substance of the present disclosure, and these modifications and improvements are also considered to be within the scope of protection of the present disclosure.

Claims

1. A method for ground feature classification and change detection, characterized in that: The method comprises: Acquire a remote sensing image; wherein the remote sensing image includes a first phase remote sensing image and a second phase remote sensing image; The remote sensing image is input into a pre-established ground feature classification and change detection model, and the ground feature classification information and change information are output; wherein, The land feature classification and change detection model is obtained based on the fusion of semi-supervised learning and active learning, sample enhancement based on scene understanding, the combination of self-supervised learning and few-sample learning, and an improved interpretation model.

2. The method according to claim 1, characterized in that The ground feature classification and change detection model is pre-established through the following steps: Acquire a remote sensing image sample set; wherein the sample set includes labeled samples and unlabeled samples; Using the labeled samples to train a self-supervised model to obtain an initial model, and then using the initial model to screen out high-quality samples from the unlabeled samples; Active learning is used to select high-information samples from the unlabeled samples, and the high-information samples are labeled to update the remote sensing image sample set; The high-quality samples are enhanced by using a SimCLR network, feature vectors of the enhanced high-quality samples are extracted by using a convolutional neural network, and the initial model is subjected to few-sample learning using the feature vectors and the labeled samples; The interpretation algorithm of the initial model is improved by combining a multi-scale convolutional neural network, a two-stream network and a contrastive learning algorithm to obtain a ground feature classification and change detection model.

3. The method according to claim 2, characterized in that The using the initial model to select high-quality samples from the unlabeled samples includes: Performing sample enhancement on the unlabeled sample to obtain an enhanced unlabeled sample; Inputting the enhanced unlabeled column sample into the initial model to obtain the confidence of each pixel; A consistency regularization method is used to compare the confidences of the unlabeled samples and the enhanced unlabeled samples, and samples with confidences greater than a preset threshold are screened out as high-quality samples.

4. The method according to claim 2, characterized in that: The adopting active learning to select high-information samples from the unlabeled samples and labeling the high-information samples includes: Use the current model to predict the category probability distribution of the unlabeled samples and calculate the entropy value; A batch of unlabeled samples with the highest entropy value are determined as high-information samples, and the high-information samples are labeled.

5. The method according to claim 4, characterized in that The entropy value is calculated by the following formula: Where P i is the predicted probability of the i-th class.

6. The method according to claim 2, characterized in that Before using the SimCLR network to enhance the high-quality sample, the method further includes: The U-Net network, generative adversarial network and data consistency judgment are used to perform sample enhancement based on scene understanding on the labeled samples.

7. The method according to any one of claims 1 to 6, characterized in that: After obtaining the ground feature classification information and change information, the method further includes: The classification information and change information of the ground feature elements are subjected to marginalization post-processing.

8. A ground feature classification and change detection system, characterized in that: The system comprises: An image acquisition module, used to acquire remote sensing images; wherein the remote sensing images include remote sensing images of a first time phase and remote sensing images of a second time phase; The classification detection module is used to input the remote sensing image into a pre-established ground feature classification and change detection model, and output ground feature classification information and change information; wherein, The land feature classification and change detection model is obtained based on the fusion of semi-supervised learning and active learning, sample enhancement based on scene understanding, the combination of self-supervised learning and few-sample learning, and an improved interpretation model.

9. An electronic device, characterized in that: include: at least one processor; as well as, A memory communicatively connected to the at least one processor is used to store one or more programs, and when the one or more programs are executed by the at least one processor, the at least one processor can implement the method for land feature classification and change detection described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method for classifying and detecting changes of land features described in any one of claims 1 to 7 is implemented.

Images