Multi-temporal unsupervised landslide detection method and device based on progressive label updating
Through the method of gradual label update, the problem of automatic identification problems and scarcity of high-resolution data in multi-time phase landslide detection is solved, the landslide detection accuracy is improved, and efficient multi-time phase landslide detection is achieved.
Patent Information
- Application Number
- CN202410959876.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-17
- Publication Date
- 2024-12-03
AI Technical Summary
The existing multi-time phase landslide detection methods are difficult to automatically identify landslide areas at different periods, and the data on high-resolution multi-time phase landslide cataloging and mapping are scarce, which limits the development of landslide recovery research.
Using a multi-time phase unsupervised landslide detection method based on progressive tag update, by acquiring multi-time phase high-resolution remote sensing images, dividing the source domain and the target domain, building a cross-time phase-style adaptive learning network and U-Net segmentation network, combining time Perceive pairing constraints, design progressive label update schemes, generate high-quality pseudo-labels, and guide landslide detection model training.
The recognition accuracy of the landslide detection model is improved, the dependence on a large amount of labeled data is reduced, and efficient high-resolution multi-time phase landslide detection is achieved.
Smart Images

Figure CN119068349A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of artificial intelligence and remote sensing information processing technology, and in particular to a multi-temporal unsupervised landslide detection method and device based on progressive label updating. Background Art
[0002] The purpose of multi-temporal landslide cataloging is to detect landslides in the same geographical location of multi-temporal data. This method can quickly estimate the trend of landslide changes, making it play an important role in post-disaster reconstruction, landslide prevention and control, and regional ecological restoration. The initial multi-temporal landslide method was mainly field investigation or visual interpretation based on aerial images. It consumed a lot of manpower and material resources and could not meet the requirements of rapid and large-scale landslide cataloging. Remote sensing images have shown great advantages in efficiency, objectivity, and timeliness. Therefore, landslide detection based on remote sensing images is becoming mainstream. At present, a large number of multi-temporal landslide cataloging methods based on remote sensing data have emerged, which can be roughly divided into two categories: traditional methods and deep learning methods.
[0003] Although the traditional method has achieved relatively satisfactory accuracy, it mainly relies on the professional knowledge of experts to select artificial feature variables, has poor automation capabilities, and is difficult to popularize. Nowadays, deep learning technology has shown potential advantages in automatic feature extraction. Compared with traditional methods, deep learning can automatically extract various features and reduce redundant manual operations. Considering that deep networks rely on a large amount of data-driven, it increases time and economic costs. Many scholars have studied the introduction of transfer learning to solve cross-scenario problems, ignoring the geographic location information shared by multi-temporal data, and cannot be fully applied to multi-temporal tasks. The goal of this invention is to distinguish the landslide situation in subsequent years based on the information of the previous year.
[0004] Due to differences in vegetation restoration, lighting conditions, image acquisition time, seasonal differences, background, etc., there are huge distribution differences between scenes, making it difficult for landslide detection models to mine robust landslide information from multi-temporal remote sensing data without label constraints. Therefore, it is a challenging task to automatically identify all landslide areas in different periods. In addition, high-resolution multi-period landslide data provides more accurate information and lays a solid foundation for scientific research. However, high-resolution multi-temporal landslide cataloging data is very rare, which limits the development of multi-temporal landslide restoration research. Therefore, it is urgent to explore automatic, efficient, and high-resolution multi-temporal landslide cataloging techniques. Summary of the invention
[0005] The purpose of the present invention is to provide a multi-temporal unsupervised landslide detection method and device based on progressive label updating in order to fill the gap in high-resolution multi-temporal landslide cataloguing and drawing technology.
[0006] The above-mentioned purpose of the present application is achieved through the following technical solutions: S1: Acquire multi-temporal and high-resolution remote sensing images of landslide hazard areas; S2: Divide the remote sensing images into source domain images and target domain images, and preprocess the remote sensing images to obtain landslide samples; the landslide samples include: source domain samples and target domain samples; S3: constructing a landslide detection model; the landslide detection model includes: a cross-temporal style adaptive learning network, a U-Net segmentation network, and a time-aware pairing constraint; S4: According to the cross-temporal style adaptive learning network and the landslide sample, the landslide sample after style conversion is obtained; S5: Input the landslide samples after style conversion and the original landslide samples into the U-Net segmentation network to obtain the features of the source domain samples and the features of the target domain samples; S6: Measuring the features of source domain samples and target domain samples through time-aware pairing constraints; S7: Design a progressive label updating scheme to generate high-quality landslide pseudo-labels to guide the training of landslide detection models; S8: Acquire the remote sensing data to be predicted; perform landslide detection on the remote sensing data to be predicted based on the trained landslide detection model.
[0007] Optionally, step S1 includes: the remote sensing images include: remote sensing images of the year when the landslide disaster occurred, and remote sensing images of each year after the year when the landslide disaster occurred.
[0008] Optionally, step S2 includes: S21: The remote sensing images of the year when the landslide disaster occurred are defined as the source domain images; the remote sensing images of each year after the year when the landslide disaster occurred are defined as the target domain images; S22: perform mean variance normalization processing on remote sensing images; S23: Crop the normalized remote sensing image into pixel blocks of a preset size to obtain individual landslide samples.
[0009] Optionally, step S4 includes: The cross-temporal style adaptive learning network adopts the StyTr2 network; The source domain images and target domain images in different years at the same geographical location in the landslide disaster area are defined as landslide data pairs; the landslide data pairs are input into the cross-temporal style adaptive learning network for style conversion to obtain landslide samples with different styles.
[0010] Optionally, the U-Net segmentation network adopts a binary cross entropy loss function, and the calculation formula is:
[0011] Where N is the number of samples; represents the true label of the source domain sample, 0 represents non-landslide and 1 represents landslide; Represents the probability that a pixel in the source domain image is predicted to be a landslide.
[0012] Optionally, step S6 includes: Through the time-aware pairing constraint, the distance between the source domain features and the target domain features is reduced. The calculation formula is as follows:
[0013] Where k represents the Gaussian radial basis function kernel for calculating feature distribution, m represents the number of paired samples across time, and a and b represent the features of source domain samples and target domain samples after style conversion at the same geographical location in the landslide disaster area, respectively.
[0014] Optionally, step S7 includes: S71: Since the source domain image and the target domain image have the same geographical location and a common landslide area, source domain image and target domain image pairs with the same geographical location are selected to be input into the landslide detection model. During the training phase, three weight files are empirically retained at a fixed iteration interval to obtain three different prediction results. S72: Calculate the intersection-and-union ratio between the third prediction result and the first two prediction results The intersection-combination ratio formula is as follows:
[0015] in is the true value of the source domain sample, is the prediction result of the target domain image in the third model; S73: Selecting the prediction results whose intersection-over-union ratio is greater than a preset threshold as high-quality landslide pseudo labels; S74: Use the prior knowledge of landslide pseudo-labels to simulate the landslide detection model and mine fine-grained landslide information.
[0016] Optionally, the total loss of the landslide detection model , which is expressed as follows:
[0017] In the formula, β represents the weight coefficient of the loss function; Represents the loss value of the prediction result of the source domain sample; Represents the loss value of the prediction result of the target domain sample; The distance loss value between the features of the domain sample and the features of the target domain sample;
[0018] in, is the number of samples, represents the pseudo labels of the target domain samples filtered using the progressive label update scheme, represents the probability that a pixel in the target domain is predicted to be a landslide.
[0019] An electronic device comprises a processor, a memory, a user interface and a network interface, wherein the memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory so that the electronic device executes a multi-temporal unsupervised landslide detection method based on progressive label updating.
[0020] A computer-readable storage medium stores instructions. When the instructions are executed, a multi-temporal unsupervised landslide detection method based on progressive label updating is performed.
[0021] The beneficial effects of the technical solution provided by this application are: The multi-temporal high-resolution remote sensing images of the landslide disaster area are obtained, and a landslide detection model is constructed; the landslide data when the disaster occurs is defined as the source domain, and the multi-temporal data that changes every year thereafter is defined as the target domain, and the collected remote sensing images are preprocessed to generate landslide samples that can be read by the model; at the inter-domain level, a cross-temporal style adaptive strategy is proposed to generate landslide samples with different styles to enrich sample diversity; the landslide data pairs are constructed from the style-converted samples and the corresponding cross-temporal samples, and are input into the landslide detection model; at the inter-domain level, a time-aware pairing constraint is constructed to further align the features of different domains; at the intra-domain level, a progressive label update scheme is designed to generate high-quality landslide pseudo-labels to guide the deep model to mine fine-grained landslide information; the model is optimized and continuously iterated and updated by combining segmentation loss and time-aware pairing constraint loss; landslide detection is performed on high-resolution multi-temporal remote sensing data based on the trained model and the detection result graph is output. The model prediction result can greatly improve the recognition accuracy of the model, so that the present invention does not rely on a large amount of labeled data driving and obtains ideal results. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] The present application will be further described below with reference to the accompanying drawings and embodiments, in which: Figure 1 is a step diagram of a multi-temporal unsupervised landslide detection method based on progressive label updating in an embodiment of the present application; Figure 2 is a structural diagram of a multi-temporal unsupervised landslide detection method based on progressive label updating in an embodiment of the present application; Figure 3is a result diagram of the multi-temporal unsupervised landslide detection method based on progressive label updating in an embodiment of the present application; Figure 4 It is a schematic diagram of the electronic device structure of the multi-temporal unsupervised landslide detection method based on progressive label updating in an embodiment of the present application. DETAILED DESCRIPTION
[0023] In order to have a clearer understanding of the technical features, purposes and effects of the present application, the specific implementation methods of the present application are now described in detail with reference to the accompanying drawings.
[0024] The embodiments of the present application provide a multi-temporal unsupervised landslide detection method based on progressive label updating.
[0025] Please refer to Figure 1 , Figure 1 : is a step diagram of a multi-temporal unsupervised landslide detection method based on progressive label updating in an embodiment of the present application, comprising: S1: Acquire multi-temporal and high-resolution remote sensing images of landslide hazard areas; S2: Divide the remote sensing images into source domain images and target domain images, and preprocess the remote sensing images to obtain landslide samples; the landslide samples include: source domain samples and target domain samples; Step S2 includes: S21: The remote sensing images of the year when the landslide disaster occurred are defined as the source domain images; the remote sensing images of each year after the year when the landslide disaster occurred are defined as the target domain images; Specifically, the study area used in the present invention is located in Hokkaido, Japan. The landslide disaster occurred in 2018, that is, the remote sensing image data in 2018 is the source domain, and the remote sensing image data in 2019, 2020, 2021 and 2022 are the target domain data.
[0026] S22: perform mean variance normalization processing on remote sensing images; Specifically, normalization is performed by subtracting the mean from the data point and dividing by the standard deviation. Different data sets are normalized to the same scale for better data analysis and model training. By calculating the mean and variance of the data set, and subtracting the mean from each data point and dividing by the variance, the mean of the data set is 0 and the standard deviation is 1.
[0027] S23: Crop the normalized remote sensing image into pixel blocks of a preset size to obtain individual landslide samples.
[0028] Specifically, the normalized remote sensing image is cropped to retain the edge information of the image as much as possible so as to input the model for training and testing.
[0029] S3: constructing a landslide detection model; the landslide detection model includes: a cross-temporal style adaptive learning network, a U-Net segmentation network, and a time-aware pairing constraint; S4: According to the cross-temporal style adaptive learning network and the landslide sample, the landslide sample after style conversion is obtained; Step S4 includes: The cross-temporal style adaptive learning network adopts the StyTr2 network; The source domain images and target domain images in different years at the same geographical location in the landslide disaster area are defined as landslide data pairs; the landslide data pairs are input into the cross-temporal style adaptive learning network for style conversion to obtain landslide samples with different styles.
[0030] Specifically, the present invention introduces the StyTr2 network as a style adaptive deep learning network. StyTr2 consists of a Transformer-based content encoder, a Transformer-based style encoder, and a Transformer-based decoder. The Transformer structure it uses can not only extract remote information in a specific field, but also alleviate the content loss problem of CNN-based methods. The source domain image is converted into an image with the style of the target domain image while retaining all the content information of the original image.
[0031] S5: Input the landslide samples after style conversion and the original landslide samples into the U-Net segmentation network to obtain the features of the source domain samples and the features of the target domain samples; The U-Net segmentation network uses a binary cross entropy loss function, calculated as:
[0032] Where N is the number of samples; represents the true label of the source domain sample, 0 represents non-landslide and 1 represents landslide; Represents the probability that a pixel in the source domain image is predicted to be a landslide.
[0033] Specifically, since the source domain image after style conversion and the corresponding source domain sample have consistent landslide labels, the present invention adopts Binary Cross Entropy Loss (BCE) to optimize the segmentation model.
[0034] Specifically, the encoder of the U-Net segmentation network consists of convolutional layers and pooling layers. The encoder gradually reduces the size of the feature map to capture the global information in the image. The decoder consists of deconvolutional layers and skip connections, which combines low-resolution feature maps with high-resolution feature maps, gradually restores the resolution of the feature maps, and generates more accurate segmentation results.
[0035] S6: Measuring the features of source domain samples and target domain samples through time-aware pairing constraints; Step S6 includes: Through the time-aware pairing constraint, the distance between the source domain features and the target domain features is reduced. The calculation formula is as follows:
[0036] Where k represents the Gaussian radial basis function kernel for calculating feature distribution, m represents the number of paired samples across time, and a and b represent the features of source domain samples and target domain samples after style conversion at the same geographical location in the landslide disaster area, respectively.
[0037] S7: Design a progressive label updating scheme to generate high-quality landslide pseudo-labels to guide the training of landslide detection models; Step S7 includes: S71: Since the source domain image and the target domain image have the same geographical location and a common landslide area, source domain image and target domain image pairs with the same geographical location are selected to be input into the landslide detection model. During the training phase, three weight files are empirically retained at a fixed iteration interval to obtain three different prediction results. Specifically, in addition, since the source domain has reliable label information and shares a geographical area with the target domain image, the landslide that is prone to occur in the source domain image is likely to still exist in the target domain image. Therefore, the image intersection ratio between the prediction results of the third model and the source domain label is further calculated.
[0038] S72: Calculate the intersection-and-union ratio between the third prediction result and the first two prediction results The intersection-combination ratio formula is as follows:
[0039] in is the true value of the source domain sample, is the prediction result of the target domain image in the third model; Specifically, since the source domain has reliable label information and shares a common geographical area with the target domain image, and the vegetation recovery speed in the landslide area is slow, the landslides that are prone to occur in the source domain image still exist in the target domain image. Therefore, the image intersection ratio between the output of the third model and the source domain label is further calculated.
[0040] S73: Selecting the prediction results whose intersection-over-union ratio is greater than a preset threshold as high-quality landslide pseudo labels; Specifically, the average value of the intersection ratio is calculated as the label confidence to evaluate the reliability of the third prediction result.
[0041] S74: Use the prior knowledge of landslide pseudo-labels to simulate the landslide detection model and mine fine-grained landslide information.
[0042] Specifically, due to the lack of effective supervision signals, it is difficult for the landslide detection model to learn a lot of fine-grained landslide information in the target domain image. At the intra-domain level, a progressive label update scheme is designed. Considering that the source domain and target domain images have the same geographical location and common landslide areas, the present invention selects source domain samples and target domain samples with the same geographical location as landslide sample pairs, and inputs them into the landslide detection model to obtain landslide prediction results. The quality of pseudo-labels is improved in the optimization and filtering process, thereby improving the optimization results of the landslide detection model.
[0043] S8: Acquire the remote sensing data to be predicted; perform landslide detection on the remote sensing data to be predicted based on the trained landslide detection model.
[0044] The total loss of the landslide detection model , which is expressed as follows:
[0045] In the formula, β represents the weight coefficient of the loss function; Represents the loss value of the prediction result of the source domain sample; Represents the loss value of the prediction result of the target domain sample; The distance loss value between the features of the domain sample and the features of the target domain sample;
[0046] in, is the number of samples, represents the pseudo labels of the target domain samples filtered using the progressive label update scheme, represents the probability that a pixel in the target domain is predicted to be a landslide.
[0047] The structural diagram of the technical solution of this application is as follows: Figure 2 shown.
[0048] The effects of the present invention are further described below in conjunction with simulation experiment examples.
[0049] (1) Simulation experiment conditions: The hardware test platform for this experiment is: Intel Core i7 processor, deep learning machine with 64 GB memory, and the software platform is: Linux operating system, Python 3.6.8. The calculation process is performed using two NVIDIA GeForce RTX 2080Ti GPUs (12 GB RAM) in the NVIDIA CUDA Toolkit 10.1 environment. The input image of the present invention is PlanetScope high-resolution remote sensing imagery. The study area is located in Hokkaido, Japan. The image acquisition time is 2018, 2019, 2020, 2021 and 2022. The image spatial resolution is 3 meters, and there are 4 bands covering the visible to near-infrared spectral range.
[0050] (2) Simulation content: In this embodiment, a multi-temporal unsupervised landslide detection method and device based on progressive label updating provided by the present invention are compared with three classic deep learning domain adaptation methods, which are DDC, AdaptSeg and PyCDA methods, as follows: The domain adaptation method proposed by E. Tzeng et al. in "Deep domain confusion: Maximizing for domain invariance, arXiv preprintarXiv:1412.3474, 2014." is referred to as the DDC method.
[0051] The domain adaptation method proposed by YH Tsai et al. in "Learning to adapt structured output space for semantic segmentation, 2018, pp. 7472–7481." is referred to as the AdaptSeg method.
[0052] The domain adaptation method proposed by Q. Lian et al. in "Constructing self-motivated pyramid curriculums for cross-domain semantic segmentation: A non-adversarial approach, 2019, pp.6758–6767." is referred to as the PyCDA method.
[0053] In order to evaluate the performance of the model, during the experiment, for the detection results obtained by different detection methods, the confusion matrix was constructed according to the true labels. We used four indicators, namely precision, recall, F1 score (F1) and mean intersection over union (mIoU), to quantitatively evaluate the performance of the method of the present invention. These indicators are calculated based on the true positive (TP), false positive (FP), false negative (FN) and true negative (TN) of each region. Among them, TP represents the pixels correctly detected as landslide areas, FP represents the pixels that identify non-landslide areas as landslides, FN represents the landslide pixels not detected by the deep model, and TN represents the pixels correctly detected as non-landslide areas. The calculation method of these indicators is as follows:
[0054]
[0055]
[0056]
[0057] in, m+ 1 represents the number of categories, P ij Indicates the truth i , but the prediction is j The number of P ii Indicates that both the real and predicted i The number of P ji Indicates the truth j , but the prediction is i The number of
[0058] Please refer to Figure 3 , Figure 3 is a result diagram of a multi-temporal high-resolution remote sensing image landslide detection method obtained by different methods in an embodiment of the present invention, wherein: Figure 3 (a) Landslide detection map obtained by DDC network method for 2019 image data; Figure 3 (b) Landslide detection map obtained by AdaptSeg network method using image data in 2019; Figure 3 (b) Landslide detection map obtained by PyCDA network method using image data in 2019; Figure 3 (d) is the landslide detection map obtained by the method of the present invention (PluTsa) based on the image data in 2019; Figure 3 (e) Landslide detection map obtained by DDC network method for 2020 image data; Figure 3(f) Landslide detection map obtained by AdaptSeg network method using image data in 2020; Figure 3 (g) Landslide detection map obtained by 2020 image data PyCDA network method; Figure 3 (h) is a landslide detection map obtained by the method (PluTsa) of the present invention using 2020 image data; Figure 3 (i) Landslide detection map obtained by DDC network method for 2021 data; Figure 3 (j) 2021Landslide detection map obtained by AdaptSeg network method; Figure 3 (k) 2021 landslide detection map obtained by PyCDA network method; Figure 3 (l) is a landslide detection map obtained by the method (PluTsa) of the present invention for 2021 image data; Figure 3 (m) Landslide detection map obtained by DDC network method for 2022 image data; Figure 3 (b) Landslide detection map obtained by AdaptSeg network method using 2022 image data; Figure 3 (n) 2022 landslide detection map obtained by PyCDA network method; Figure 3 (o) is the landslide detection map obtained by the method (PluTsa) of the present invention using 2022 image data.
[0059] (3) Analysis of experimental results See Table 1. Table 1 compares the Figure 3 The detection results of each method are evaluated.
[0060] Table 1 Evaluation results of detection accuracy of different methods for remote sensing images in 2019
[0061] Table 2 Evaluation results of detection accuracy of different methods for remote sensing images in 2020
[0062] Table 3 Evaluation results of detection accuracy of different methods for remote sensing images in 2021
[0063] Table 4 Evaluation results of detection accuracy of different methods for remote sensing images in 2022
[0064] Comprehensive Table 1~Table 4 and Figure 3 It can be seen that the DDC method Figure 3Although the accuracy rate of the area is high, the omission phenomenon is obvious, and a large number of landslides are not identified; the AdaptSeg method can identify more landslides, but it is difficult to completely eliminate the misclassification of homogeneous areas, and there is still a very obvious omission phenomenon; the PyCDA method can identify a large number of landslides, but there is a large-scale multi-detection phenomenon, that is, a large number of non-landslides are mistakenly identified as landslides. The method of the present invention is superior to the previous three detection methods in terms of visual effect and quantitative analysis. It can significantly improve the ability of landslide identification, identify a large number of landslide areas that are not detected by DDC and AdaptSeg, and correctly identify a large number of non-landslide areas that are mistakenly detected as landslides by PyCDA. It can be concluded that the method of the multi-phase unsupervised landslide detection method based on progressive label update of the present invention has the effect of improving the accuracy of landslide detection.
[0065] The present application also discloses an electronic device. Figure 4 , Figure 4 The electronic device 500 may include: at least one processor 501 , at least one network interface 504 , a user interface 503 , a memory 505 , and at least one communication bus 502 .
[0066] The communication bus 502 is used to realize the connection and communication between these components.
[0067] The user interface 503 may include a display screen, and the optional user interface 503 may also include a standard wired interface or a wireless interface.
[0068] The network interface 504 may optionally include a standard wired interface or a wireless interface (such as a WI-FI interface).
[0069] The present application also provides a computer-readable storage medium, such as a flash memory, a hard disk, a multimedia card, a card-type memory (for example, an SD or DX memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a disk, an optical disk, a server, an App application store, etc., on which a computer program is stored, and the program realizes corresponding functions when executed by a processor. The computer-readable storage medium of this embodiment is used to realize the high-resolution image landslide detection of the dual-branch network of the joint spectrum, digital elevation model and auxiliary features of the method embodiment when executed by the processor.
[0070] The above are only exemplary embodiments of the present disclosure and cannot be used to limit the scope of the present disclosure. That is, any equivalent changes and modifications made according to the teachings of the present disclosure are still within the scope of the present disclosure. After considering the disclosure of the specification and the truth of practice, those skilled in the art will easily think of other embodiments of the present disclosure.
[0071] This application is intended to cover any variation, use or adaptation of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary technical means in the art not described in the present disclosure. The description and examples are to be regarded as exemplary only, and the scope and spirit of the present disclosure are defined by the claims.
Claims
1. A multi-temporal unsupervised landslide detection method based on progressive label updating, characterized in that: The method comprises the following steps: S1: Acquire multi-temporal and high-resolution remote sensing images of landslide hazard areas; S2: Divide the remote sensing images into source domain images and target domain images, and preprocess the remote sensing images to obtain landslide samples; the landslide samples include: source domain samples and target domain samples; S3: constructing a landslide detection model; the landslide detection model includes: a cross-temporal style adaptive learning network, a U-Net segmentation network, and a time-aware pairing constraint; S4: According to the cross-temporal style adaptive learning network and the landslide sample, the landslide sample after style conversion is obtained; S5: Input the landslide samples after style conversion and the original landslide samples into the U-Net segmentation network to obtain the features of the source domain samples and the features of the target domain samples; S6: Measuring the features of source domain samples and target domain samples through time-aware pairing constraints; S7: Design a progressive label updating scheme to generate landslide pseudo labels to guide the training of landslide detection models; S8: Acquire the remote sensing data to be predicted; perform landslide detection on the remote sensing data to be predicted based on the trained landslide detection model.
2. The multi-temporal unsupervised landslide detection method based on progressive label updating according to claim 1, characterized in that: Step S1 includes: the remote sensing images include: remote sensing images of the year when the landslide disaster occurred, and remote sensing images of each year after the year when the landslide disaster occurred.
3. The multi-temporal unsupervised landslide detection method based on progressive label updating according to claim 2, characterized in that: Step S2 includes: S21: The remote sensing images of the year when the landslide disaster occurred are defined as the source domain images; the remote sensing images of each year after the year when the landslide disaster occurred are defined as the target domain images; S22: perform mean variance normalization processing on remote sensing images; S23: Crop the normalized remote sensing image into pixel blocks of a preset size to obtain individual landslide samples.
4. The multi-temporal unsupervised landslide detection method based on progressive label updating according to claim 1, characterized in that: Step S4 includes: The cross-temporal style adaptive learning network adopts the StyTr2 network; The source domain images and target domain images in different years at the same geographical location in the landslide disaster area are defined as landslide data pairs; the landslide data pairs are input into the cross-temporal style adaptive learning network for style conversion to obtain landslide samples with different styles.
5. The multi-temporal unsupervised landslide detection method based on progressive label updating according to claim 1, characterized in that: The U-Net segmentation network uses a binary cross entropy loss function, calculated as: Where N is the number of samples; represents the true label of the source domain sample, 0 represents non-landslide and 1 represents landslide; Represents the probability that a pixel in the source domain image is predicted to be a landslide.
6. The multi-temporal unsupervised landslide detection method based on progressive label updating according to claim 1, characterized in that: Step S6 includes: Through the time-aware pairing constraint, the distance between the source domain features and the target domain features is reduced. The calculation formula is as follows: Where k represents the Gaussian radial basis function kernel for calculating feature distribution, m represents the number of paired samples across time, and a and b represent the features of source domain samples and target domain samples after style conversion at the same geographical location in the landslide disaster area, respectively.
7. The multi-temporal unsupervised landslide detection method based on progressive label updating according to claim 1, characterized in that: Step S7 includes: S71: Since the source domain image and the target domain image have the same geographical location and a common landslide area, source domain image and target domain image pairs with the same geographical location are selected to be input into the landslide detection model. During the training phase, three weight files are empirically retained at a fixed iteration interval to obtain three different prediction results. S72: Calculate the intersection-and-union ratio between the third prediction result and the first two prediction results The intersection-combination ratio formula is as follows: in is the true value of the source domain sample, is the prediction result of the target domain image in the third model; S73: Selecting the prediction results whose intersection-over-union ratio is greater than a preset threshold as high-quality landslide pseudo labels; S74: Use the prior knowledge of landslide pseudo-labels to simulate the landslide detection model and mine fine-grained landslide information.
8. The multi-temporal unsupervised landslide detection method based on progressive label updating according to claim 1, characterized in that: The total loss of the landslide detection model , which is expressed as follows: In the formula, β represents the weight coefficient of the loss function; Represents the loss value of the prediction result of the source domain sample; Represents the loss value of the prediction result of the target domain sample; The distance loss value between the features of the domain sample and the features of the target domain sample; in, is the number of samples, represents the pseudo labels of the target domain samples filtered using the progressive label update scheme; Represents the probability that a pixel in the target domain is predicted to be a landslide.
9. An electronic device, characterized in that: The electronic device comprises a processor (501), a memory (505), a user interface (503) and a network interface (504), wherein the memory (505) is used to store instructions, the user interface (503) and the network interface (504) are used to communicate with other devices, and the processor (501) is used to execute the instructions stored in the memory (505) so that the electronic device executes the method according to any one of claims 1 to 8.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, and when the instructions are executed by a computer, the method according to any one of claims 1 to 8 is executed.
Citation Information
Patent Citations
Cross-scene-domain adaptive landslide detection method and device, equipment and storage medium
CN115760569A
Landslide identification method, system, equipment and medium
CN116738236A
