Cross-scene domain adaptive landslide detection method, device and equipment and storage medium
Patent Information
- Application Number
- CN202211379516.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-11-04
- Publication Date
- 2026-11-03
- Estimated Expiration
- 2042-11-04
AI Technical Summary
Existing technologies require significant marking time for cross-scenario landslide detection, impacting detection performance and efficiency.
By acquiring optical images and digital elevation models of the source and target domains, data merging and preprocessing are performed. Then, landslide location prediction is performed using a style transfer network, an improved U-Net3+ segmentation network, segmentation loss, and adversarial and domain distance minimization loss.
It improves the transferability accuracy and detection capability of landslide detection across different scenarios, and enhances the practicality of the model.
Smart Images
Figure CN115760569B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing technology, and in particular to a cross-scene domain adaptive landslide detection method, apparatus, device, and storage medium. Background Technology
[0002] Landslides are catastrophic natural disasters that cause enormous damage to personal safety, property, and the natural environment. Effective and accurate landslide detection is of great significance in supporting post-disaster relief and reconstruction. Remote sensing technology has significant advantages in terms of efficiency, objectivity, and timeliness, providing important support for geological disaster monitoring and investigation. However, as a special type of land cover, landslides are characterized by large intraclass variability, irregular shapes, and large scale differences, resulting in low accuracy in landslide logging and mapping. In addition, the significant differences between the source and target domain backgrounds cause the model accuracy to decrease when transferred to another scene.
[0003] Landslide detection is the primary step in landslide investigation, and its research can be broadly categorized into three types: First, manual visual interpretation, which relies on the differences in hue, texture, shape, shadow, and pattern between the landslide and its surrounding environment, using the expert experience of interpreters for identification. This method is time-consuming and labor-intensive, and difficult to meet the needs of large-scale landslide detection. Second, traditional machine learning, which uses remote sensing imagery and related topographic data to extract spectral, texture, and shape features of landslide hazards, combining statistical decision-making, probability estimation, and machine learning methods for landslide detection. The automation and accuracy of this method depend mainly on feature selection and parameter settings. Third, deep learning, due to its advantages in automatically extracting high-level semantic features from images, has demonstrated high performance in landslide detection. This method is gaining increasing attention and makes end-to-end landslide detection possible.
[0004] Current research has used deep learning techniques to detect landslides, achieving better results than traditional machine learning methods. However, this performance is driven by a large amount of labeled data and requires that the training and test data have the same distribution. Therefore, when the model is transferred to new landslide detection scenarios, it will experience a significant performance drop. However, labeling a large amount of data for each scenario is very time-consuming and impractical for real-world applications. Therefore, research on the transfer of landslide detection is an important research topic. Summary of the Invention
[0005] The main objective of this invention is to provide a cross-scene adaptive landslide detection method, apparatus, device, and storage medium, aiming to solve the technical problem that cross-scene landslide detection consumes a lot of labeling time, affecting the performance of landslide detection and resulting in low landslide detection efficiency in the prior art.
[0006] In a first aspect, the present invention provides a cross-scene domain adaptive landslide detection method, the cross-scene domain adaptive landslide detection method comprising the following steps:
[0007] Acquire optical images, digital elevation models, and source domain landslide label data from the source and target domains; merge the optical images, digital elevation models, and source domain landslide label data to obtain merged image data.
[0008] The merged image data is preprocessed, and the preprocessed image is cropped to obtain a cropped image. The cropped image is then input into a style transfer network for style transfer to obtain a target domain transformed image corresponding to the source domain style.
[0009] The target domain-transformed image is input into a domain adaptation network, which includes an improved U-Net3+ segmentation network, segmentation loss, and adversarial and domain distance minimization loss, to obtain landslide location prediction results.
[0010] Optionally, the step of acquiring optical images of the source and target domains, digital elevation models, and source domain landslide label data, and merging the optical images, digital elevation models, and source domain landslide label data to obtain merged image data, includes:
[0011] Acquire optical images of the source and target domains, digital elevation models, and landslide label data of the source domain; perform formula analysis on the optical images and the digital elevation models to extract feature variables;
[0012] The feature variables are merged with the source domain landslide label data to obtain merged image data.
[0013] Optionally, the acquisition of optical images of the source and target domains, digital elevation models, and source domain landslide label data, followed by formula analysis of the optical images and the digital elevation models to extract feature variables, includes:
[0014] Acquire optical images of the source and target domains, digital elevation models, and landslide label data of the source domain;
[0015] The vegetation index is calculated using the reflectance of the red and near-infrared bands of the optical image through the following formula:
[0016]
[0017] Wherein, NDVI is the vegetation index, NIR is the reflectance in the near-infrared band, and R is the reflectance in the red band.
[0018] Feature variables of terrain features in the optical image are extracted based on the vegetation index and the digital elevation model.
[0019] Optionally, the step of preprocessing the merged image data, cropping the preprocessed image to obtain a cropped image, and inputting the cropped image into a style transfer network for style transfer to obtain a target domain-transformed image corresponding to the source domain style includes:
[0020] The maximum and minimum values of each channel on which the merged image data is based are obtained, and the maximum and minimum values are normalized to obtain a normalized image.
[0021] Calculate the mean and variance of the normalized image, and perform deviation standardization on the mean and variance to obtain a standardized image;
[0022] The standardized image is cropped to obtain a cropped image. The cropped image is then input into a style transfer network for style transfer to obtain a target domain transformed image corresponding to the source domain style.
[0023] Optionally, cropping the standardized image to obtain a cropped image, and inputting the cropped image into a style transfer network for style transfer to obtain a target domain-transformed image corresponding to the source domain style, includes:
[0024] The standardized image is cropped to obtain a cropped image;
[0025] The style transfer network is input through the RGB channels of the optical image to obtain a trained style transfer model;
[0026] The cropped image style is converted to the source domain image style by the trained style transfer model, and a target domain converted image with the corresponding source domain style is obtained.
[0027] Optionally, the step of inputting the target domain-transformed image into a domain adaptation network, wherein the domain adaptation network includes an improved U-Net3+ segmentation network, segmentation loss, and adversarial and domain distance minimization loss, to obtain landslide location prediction results, includes:
[0028] An improved U-Net3+ segmentation network is obtained by replacing the max pooling layers of each coding layer in the original U-Net3+ network structure with Tokenized MLP.
[0029] The source domain image is input into the improved U-Net3+ segmentation network to obtain the source domain landslide features and source domain landslide location prediction results. The segmentation network is then optimized using the source domain real labels.
[0030] The style-transferred target domain image is input into the improved U-Net3+ segmentation network to obtain the target domain landslide features and target domain landslide location prediction results.
[0031] The countermeasure loss is calculated based on the predicted landslide locations in the source region and the predicted landslide locations in the target region.
[0032] Minimize the domain distance between the source domain landslide features and the target domain landslide features to obtain the domain distribution spatial distance loss between the source domain landslide features and the target domain landslide features;
[0033] The U-Net3+ segmentation network is optimized by combining the adversarial loss and the domain distance minimization loss to obtain the landslide location prediction result in the target domain.
[0034] Optionally, the step of constructing a counterbalancing loss based on the predicted landslide location in the source region and the predicted landslide location in the target region, and calculating the counterbalancing loss based on the predicted landslide location in the source region and the predicted landslide location in the target region, includes:
[0035] Based on the predicted landslide locations in the source region and the predicted landslide locations in the target region, an adversarial loss is constructed using the following formula:
[0036]
[0037] in, To combat the losses, To distinguish for the discriminator The probability of the source domain;
[0038] Accordingly, the step of minimizing the domain distance between the source domain landslide features and the target domain landslide features includes:
[0039] The maximum mean deviation constraint function is used to reduce the domain distance between the spatial distributions of landslide features in the source domain and the target domain by the following formula:
[0040]
[0041] Among them, L mmd The distance between the minimized domains is given by k, where k is the Gaussian radial basis function kernel function used to calculate the feature distribution, a is the landslide feature of the source domain, and b is the landslide feature of the target domain.
[0042] Secondly, to achieve the above objectives, the present invention also proposes a cross-scene domain adaptive landslide detection device, the cross-scene domain adaptive landslide detection device comprising:
[0043] The data acquisition module is used to acquire optical images, digital elevation models, and source domain landslide label data from the source and target domains, and to merge the optical images, digital elevation models, and source domain landslide label data to obtain merged image data.
[0044] The cropping and transformation module is used to preprocess the merged image data, crop the preprocessed image to obtain a cropped image, and input the cropped image into a style transfer network for style transfer to obtain a target domain transformed image corresponding to the source domain style.
[0045] The prediction module is used to input the target domain-transformed image into a domain adaptation network, which includes an improved U-Net3+ segmentation network, segmentation loss, and adversarial and domain distance minimization loss to obtain landslide location prediction results.
[0046] Thirdly, to achieve the above objectives, the present invention also proposes a cross-scene adaptive landslide detection device, which includes: a memory, a processor, and a cross-scene adaptive landslide detection program stored in the memory and executable on the processor. The cross-scene adaptive landslide detection program is configured to implement the steps of the cross-scene adaptive landslide detection method described above.
[0047] Fourthly, to achieve the above objectives, the present invention also proposes a storage medium storing a cross-scene domain adaptive landslide detection program, wherein the cross-scene domain adaptive landslide detection program, when executed by a processor, implements the steps of the cross-scene domain adaptive landslide detection method described above.
[0048] The proposed cross-scene domain adaptive landslide detection method acquires optical images, digital elevation models (DEMs), and source domain landslide label data from both the source and target domains. The method merges these data to obtain merged image data. The merged image data is then preprocessed and cropped to obtain a cropped image. This cropped image is input into a style transfer network for style transfer, resulting in a target domain-transformed image corresponding to the source domain style. Finally, the target domain-transformed image is input into a domain adaptation network, which includes an improved U-Net3+ segmentation network, segmentation loss, and adversarial and domain distance minimization losses, to obtain landslide location prediction results. This method fully utilizes the optical images and DEMs of high-resolution imagery to extract feature factors and mines domain-invariant features across different scenes, thereby improving the transferability accuracy of cross-scene landslide detection and enhancing the model's cross-scene detection capability and practicality. Attached Figure Description
[0049] Figure 1 This is a schematic diagram of the device structure of the hardware operating environment involved in the embodiments of the present invention;
[0050] Figure 2 This is a flowchart illustrating the first embodiment of the cross-scene domain adaptive landslide detection method of the present invention;
[0051] Figure 3 This is a flowchart illustrating the second embodiment of the cross-scene domain adaptive landslide detection method of the present invention;
[0052] Figure 4 This is a flowchart illustrating the third embodiment of the cross-scene domain adaptive landslide detection method of the present invention;
[0053] Figure 5 This is a flowchart illustrating the fourth embodiment of the cross-scene domain adaptive landslide detection method of the present invention;
[0054] Figure 6 This is a functional block diagram of the first embodiment of the cross-scene domain adaptive landslide detection device of the present invention.
[0055] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0056] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
[0057] The solution of this invention mainly involves: acquiring optical images, digital elevation models, and source domain landslide label data from the source and target domains; merging these data to obtain merged image data; preprocessing the merged image data and cropping the preprocessed image to obtain a cropped image; inputting the cropped image into a style transfer network for style transfer to obtain a target domain transformed image corresponding to the source domain style; and inputting the target domain transformed image into a domain adaptation network, which includes an improved U-Net3+ segmentation network, segmentation loss, and adversarial and domain distance minimization loss, to obtain landslide location prediction results. This solution fully utilizes the optical images and digital elevation models of high-resolution images to extract feature factors and mine domain-invariant features from different scenes, thereby improving the transferability accuracy of cross-scene landslide detection, enhancing the model's cross-scene detection capability and practicality, and solving the technical problem in existing technologies where cross-scene landslide detection consumes a large amount of labeling time, affecting landslide detection performance and resulting in low landslide detection efficiency.
[0058] Reference Figure 1 , Figure 1 This is a schematic diagram of the device structure of the hardware operating environment involved in the embodiments of the present invention.
[0059] like Figure 1As shown, the device may include: a processor 1001, such as a CPU; a communication bus 1002; a user interface 1003; a network interface 1004; and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be high-speed RAM or non-volatile memory, such as a disk drive. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.
[0060] Those skilled in the art will understand that Figure 1 The device structure shown does not constitute a limitation on the device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.
[0061] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating device, a network communication module, a user interface module, and a cross-scenario adaptive landslide detection program.
[0062] The device of the present invention calls the cross-scene domain adaptive landslide detection program stored in the memory 1005 through the processor 1001, and performs the following operations:
[0063] Acquire optical images, digital elevation models, and source domain landslide label data from the source and target domains; merge the optical images, digital elevation models, and source domain landslide label data to obtain merged image data.
[0064] The merged image data is preprocessed, and the preprocessed image is cropped to obtain a cropped image. The cropped image is then input into a style transfer network for style transfer to obtain a target domain transformed image corresponding to the source domain style.
[0065] The target domain-transformed image is input into a domain adaptation network, which includes an improved U-Net3+ segmentation network, segmentation loss, and adversarial and domain distance minimization loss, to obtain landslide location prediction results.
[0066] The device of the present invention, through processor 1001 calling the cross-scene domain adaptive landslide detection program stored in memory 1005, also performs the following operations:
[0067] Acquire optical images of the source and target domains, digital elevation models, and landslide label data of the source domain; perform formula analysis on the optical images and the digital elevation models to extract feature variables;
[0068] The feature variables are merged with the source domain landslide label data to obtain merged image data.
[0069] The device of the present invention, through processor 1001 calling the cross-scene domain adaptive landslide detection program stored in memory 1005, also performs the following operations:
[0070] Acquire optical images of the source and target domains, digital elevation models, and landslide label data of the source domain;
[0071] The vegetation index is calculated using the reflectance of the red and near-infrared bands of the optical image through the following formula:
[0072]
[0073] Wherein, NDVI is the vegetation index, NIR is the reflectance in the near-infrared band, and R is the reflectance in the red band.
[0074] Feature variables of terrain features in the optical image are extracted based on the vegetation index and the digital elevation model.
[0075] The device of the present invention, through processor 1001 calling the cross-scene domain adaptive landslide detection program stored in memory 1005, also performs the following operations:
[0076] The maximum and minimum values of each channel on which the merged image data is based are obtained, and the maximum and minimum values are normalized to obtain a normalized image.
[0077] Calculate the mean and variance of the normalized image, and perform deviation standardization on the mean and variance to obtain a standardized image;
[0078] The standardized image is cropped to obtain a cropped image. The cropped image is then input into a style transfer network for style transfer to obtain a target domain transformed image corresponding to the source domain style.
[0079] The device of the present invention, through processor 1001 calling the cross-scene domain adaptive landslide detection program stored in memory 1005, also performs the following operations:
[0080] The standardized image is cropped to obtain a cropped image;
[0081] The style transfer network is input through the RGB channels of the optical image to obtain a trained style transfer model;
[0082] The cropped image style is converted to the source domain image style by the trained style transfer model, and a target domain converted image with the corresponding source domain style is obtained.
[0083] The device of the present invention, through processor 1001 calling the cross-scene domain adaptive landslide detection program stored in memory 1005, also performs the following operations:
[0084] An improved U-Net3+ segmentation network is obtained by replacing the max pooling layers of each coding layer in the original U-Net3+ network structure with Tokenized MLP.
[0085] The source domain image is input into the improved U-Net3+ segmentation network to obtain the source domain landslide features and source domain landslide location prediction results. The segmentation network is then optimized using the source domain real labels.
[0086] The style-transferred target domain image is input into the improved U-Net3+ segmentation network to obtain the target domain landslide features and target domain landslide location prediction results.
[0087] The countermeasure loss is calculated based on the predicted landslide locations in the source region and the predicted landslide locations in the target region.
[0088] Minimize the domain distance between the source domain landslide features and the target domain landslide features to obtain the domain distribution spatial distance loss between the source domain landslide features and the target domain landslide features;
[0089] The U-Net3+ segmentation network is optimized by combining the adversarial loss and the domain distance minimization loss to obtain the landslide location prediction result in the target domain. The domain adaptation network includes the improved U-Net3+ segmentation network, segmentation loss, and adversarial and domain distance minimization losses to obtain the landslide location prediction result.
[0090] The device of the present invention, through processor 1001 calling the cross-scene domain adaptive landslide detection program stored in memory 1005, also performs the following operations:
[0091] Based on the predicted landslide locations in the source region and the predicted landslide locations in the target region, an adversarial loss is constructed using the following formula:
[0092]
[0093] in, To combat the losses, To distinguish for the discriminator The probability of the source domain;
[0094] Accordingly, the step of minimizing the domain distance between the source domain landslide features and the target domain landslide features includes:
[0095] The maximum mean deviation constraint function is used to reduce the domain distance between the spatial distributions of landslide features in the source domain and the target domain by the following formula:
[0096]
[0097] Among them, L mmd The distance between the minimized domains is given by k, where k is the Gaussian radial basis function kernel function used to calculate the feature distribution, a is the landslide feature of the source domain, and b is the landslide feature of the target domain.
[0098] This embodiment, through the above-described scheme, acquires optical images, digital elevation models, and source domain landslide label data from the source and target domains. It then merges these data to obtain merged image data. The merged image data is preprocessed, and the preprocessed image is cropped to obtain a cropped image. This cropped image is input into a style transfer network for style transfer, resulting in a target domain-converted image corresponding to the source domain style. Finally, the target domain-converted image is input into a domain adaptation network, which includes an improved U-Net3+ segmentation network, segmentation loss, and adversarial and domain distance minimization losses, to obtain landslide location prediction results. This approach fully utilizes the optical images and digital elevation models of high-resolution imagery to extract feature factors and mines domain-invariant features across different scenes, thereby improving the transferability accuracy of cross-scene landslide detection and enhancing the model's cross-scene detection capability and practicality.
[0099] Based on the above hardware structure, an embodiment of the cross-scene domain adaptive landslide detection method of the present invention is proposed.
[0100] Reference Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the cross-scene domain adaptive landslide detection method of the present invention.
[0101] In the first embodiment, the cross-scenario adaptive landslide detection method includes the following steps:
[0102] Step S10: Obtain optical images, digital elevation models, and source domain landslide label data of the source and target domains; merge the optical images, digital elevation models, and source domain landslide label data to obtain merged image data.
[0103] It should be noted that by acquiring optical images of the source and target domains, digital elevation models, and landslide label data of the source domain, the corresponding data can be merged to obtain merged image data.
[0104] In the specific implementation, the input consists of labeled and unlabeled high-resolution remote sensing images of the disaster acquired by the PlanetScope optical sensor and the corresponding digital elevation model. The labeled data is the source domain, which has landslide labels; the unlabeled data is the target domain, which is the image to be detected.
[0105] Step S20: Preprocess the merged image data and crop the preprocessed image to obtain a cropped image. Input the cropped image into a style transfer network for style transfer to obtain a target domain transformed image corresponding to the source domain style.
[0106] It is understood that after performing normalization and deviation standardization preprocessing on the merged image data, the preprocessed image can be cropped to obtain a corresponding cropped image. Then, the cropped image is input into a style transfer network for style transfer to obtain a target domain transformed image with the corresponding source domain style.
[0107] Step S30: Input the target domain transformed image into a domain adaptation network, which includes an improved U-Net3+ segmentation network, segmentation loss, and adversarial and domain distance minimization loss, to obtain the landslide location prediction result.
[0108] It should be understood that by inputting the target domain-transformed image into the domain adaptation network, which may include an improved U-Net3+ segmentation network, segmentation loss, and adversarial and domain distance minimization loss, the landslide location prediction result can be obtained.
[0109] This embodiment, through the above-described scheme, acquires optical images, digital elevation models, and source domain landslide label data from the source and target domains. It then merges these data to obtain merged image data. The merged image data is preprocessed, and the preprocessed image is cropped to obtain a cropped image. This cropped image is input into a style transfer network for style transfer, resulting in a target domain-converted image corresponding to the source domain style. Finally, the target domain-converted image is input into a domain adaptation network, which includes an improved U-Net3+ segmentation network, segmentation loss, and adversarial and domain distance minimization losses, to obtain landslide location prediction results. This approach fully utilizes the optical images and digital elevation models of high-resolution imagery to extract feature factors and mines domain-invariant features across different scenes, thereby improving the transferability accuracy of cross-scene landslide detection and enhancing the model's cross-scene detection capability and practicality.
[0110] Furthermore, Figure 3 This is a flowchart illustrating the second embodiment of the cross-scene domain adaptive landslide detection method of the present invention, as shown below. Figure 3As shown, based on the first embodiment, a second embodiment of the cross-scene domain adaptive landslide detection method of the present invention is proposed. In this embodiment, step S10 specifically includes the following steps:
[0111] Step S11: Obtain optical images of the source and target domains, digital elevation models, and landslide label data of the source domain; perform formula analysis on the optical images and the digital elevation models to extract feature variables.
[0112] It should be noted that after acquiring the optical images, digital elevation models, and landslide label data of the source and target domains, formula analysis can be performed based on the optical images and digital elevation models of the source and target domains to extract multiple feature variables.
[0113] Furthermore, step S11 specifically includes the following steps:
[0114] Acquire optical images of the source and target domains, digital elevation models, and landslide label data of the source domain;
[0115] The vegetation index is calculated using the reflectance of the red and near-infrared bands of the optical image through the following formula:
[0116]
[0117] Wherein, NDVI is the vegetation index, NIR is the reflectance in the near-infrared band, and R is the reflectance in the red band.
[0118] Feature variables of terrain features in the optical image are extracted based on the vegetation index and the digital elevation model.
[0119] It is understood that the vegetation index is calculated from the red and near-infrared bands of the source and target domain images; terrain feature correlation factors are extracted from the digital elevation model; the remote sensing image, digital elevation model, vegetation index, and terrain features are combined; the vegetation index NDVI is calculated from the red and near-infrared bands of the optical image; and multiple terrain features are extracted using the high-resolution remote digital elevation model of the source and target domains.
[0120] The extracted terrain features include: aspect, slope, mountain shadow, curvature, planar curvature, and profile curvature, described as follows:
[0121]
[0122]
[0123] Then, optical images, digital elevation models, vegetation indices, and terrain features are combined.
[0124] Step S12: Merge the feature variables with the source domain landslide label data to obtain merged image data.
[0125] It is understandable that merged image data can be obtained by merging the feature variables with the source domain landslide label data.
[0126] This embodiment, through the above-described scheme, acquires optical images, digital elevation models, and source domain landslide label data of the source and target domains. It then performs formula analysis on the optical images and digital elevation models to extract feature variables. Finally, it merges the feature variables with the source domain landslide label data to obtain merged image data. This approach can fully utilize the optical images and digital elevation models of high-resolution images to extract feature factors and explore the domain invariance features of different scenarios.
[0127] Furthermore, Figure 4 This is a flowchart illustrating the third embodiment of the cross-scene domain adaptive landslide detection method of the present invention, as shown below. Figure 4 As shown, based on the first embodiment, a third embodiment of the cross-scene domain adaptive landslide detection method of the present invention is proposed. In this embodiment, step S20 specifically includes the following steps:
[0128] Step S21: Obtain the maximum and minimum values of each channel on which the merged image data is based, and normalize the maximum and minimum values to obtain a normalized image.
[0129] It should be noted that, based on the merged data, i.e. the merged image data, the merged image data is normalized according to the maximum and minimum values of each channel, that is, the maximum and minimum values of each channel are normalized to obtain a normalized image.
[0130] In the specific implementation, channel merging is performed by merging the optical images, digital elevation models, landslide labels, and extracted feature factors of the source and target domains, respectively. The extracted feature factors include: vegetation index, slope, aspect, mountain shadow, curvature, planar curvature, and profile curvature. The maximum and minimum values of each channel in the merged image of the source and target domains are normalized to obtain a normalized merged image.
[0131] Step S22: Calculate the mean and variance of the normalized image, and perform deviation standardization on the mean and variance to obtain a standardized image.
[0132] It is understandable that after normalization, the mean and variance are calculated and then deviation standardization is performed. That is, after calculating the mean and variance of the normalized image, deviation standardization is performed to obtain a standardized image.
[0133] In practice, the mean and variance of each channel are calculated for the merged image of the normalized source and target domains, and deviation standardization is performed to adjust the data distribution.
[0134] Step S23: Crop the standardized image to obtain a cropped image, and input the cropped image into a style transfer network for style transfer to obtain a target domain transformed image corresponding to the source domain style.
[0135] It should be understood that cropping the standardized image can yield a cropped image, which can then be input into a style transfer network for style transfer to obtain a target domain-transformed image corresponding to the source domain style.
[0136] In the specific implementation, normalization is used to scale the range values of each channel of the merged image to [0, 1]. Of course, it can also be scaled to other values, and this embodiment does not limit this. The mean and variance of the normalized image are statistically analyzed, and deviation standardization is performed to adjust the data distribution. The deviation-standardized image data is cropped based on a certain degree of overlap to obtain the cropped image, so as to preserve the context information of the image edges.
[0137] Furthermore, step S23 specifically includes the following steps:
[0138] The standardized image is cropped to obtain a cropped image;
[0139] The style transfer network is input through the RGB channels of the optical image to obtain a trained style transfer model;
[0140] The cropped image style is converted to the source domain image style by the trained style transfer model, and a target domain converted image with the corresponding source domain style is obtained.
[0141] It should be understood that the style transfer model is trained by inputting the RGB channels of the source and target domain images into the style adaptation network; the RGB channels of the source and target domain images are then input into the trained style adaptation network to convert the image style of the target domain to the image style of the source domain; the style-transformed image is obtained, which combines the style of the source domain and the content of the target domain to obtain the style-transformed target domain image.
[0142] This embodiment, through the above-described scheme, obtains the maximum and minimum values of each channel on which the merged image data is based, normalizes the maximum and minimum values to obtain a normalized image, calculates the mean and variance of the normalized image, performs deviation standardization on the mean and variance to obtain a standardized image, crops the standardized image to obtain a cropped image, and inputs the cropped image into a style transfer network for style transfer to obtain a target domain transformed image corresponding to the source domain style. This approach can fully utilize the optical image and digital elevation model of high-resolution imagery to extract feature factors and obtain effective information such as tone, texture, and terrain.
[0143] Furthermore, Figure 5 This is a flowchart illustrating the fourth embodiment of the cross-scene domain adaptive landslide detection method of the present invention, as shown below. Figure 5 As shown, based on the first embodiment, a fourth embodiment of the cross-scene domain adaptive landslide detection method of the present invention is proposed. In this embodiment, step S30 specifically includes the following steps:
[0144] Step S31: Replace the max pooling layer of each coding layer in the original U-Net3+ network structure with a Tokenized MLP to obtain an improved U-Net3+ segmentation network.
[0145] It should be noted that the network is improved based on the original U-Net++++, and a landslide segmentation model is established based on it. Specifically, the Tokenized MLP can be reproduced. The Tokenized MLP can encode meaningful feature information using appropriate parameters and computation. The original U-Net++++ network structure is modified by replacing the max pooling layers of each encoding layer in the original U-Net3+ with the Tokenized MLP to obtain the improved U-Net3+ segmentation network.
[0146] Step S32: Input the source domain image into the improved U-Net3+ segmentation network to obtain the source domain landslide features and the source domain landslide location prediction results, and optimize the segmentation network using the source domain real labels.
[0147] It is understandable that by inputting the real and style-transformed source domain images into the improved U-Net3+ segmentation network, the source domain landslide features and source domain landslide location prediction results can be obtained, and the segmentation network can be optimized using the real source domain labels.
[0148] Step S33: Input the style-transferred target domain image into the improved U-Net3+ segmentation network to obtain the target domain landslide features and target domain landslide location prediction results.
[0149] It should be understood that by inputting the style-transferred target domain map into the improved U-Net3+ segmentation network, the target domain landslide features and target domain landslide location prediction results can be obtained.
[0150] Step S34: Calculate the countermeasure loss based on the predicted landslide location in the source region and the predicted landslide location in the target region.
[0151] It should be understood that, based on the landslide prediction results of the source domain and the target domain, an adversarial loss is constructed to constrain the landslide segmentation model, thereby learning the domain invariant characteristics of the landslide. Based on the landslide prediction results of the source domain and the true labels, a similarity coefficient is used to optimize the improved U-Net3+ segmentation network.
[0152] Furthermore, step S34 specifically includes the following steps:
[0153] Based on the predicted landslide locations in the source region and the predicted landslide locations in the target region, an adversarial loss is constructed using the following formula:
[0154]
[0155] in, To combat the losses, To distinguish for the discriminator The probability of the source domain.
[0156] In the specific implementation, considering the imbalance problem of landslide samples, similarity coefficient loss is used to further optimize and improve U-Net3+. The similarity coefficient loss function is defined as follows:
[0157]
[0158] Where p(x) represents the landslide surface data and g(x) represents the landslide prediction probability.
[0159] Since there are differences between decoding layers, the output results of different decoding layers will inevitably differ. To optimize the slippery output structure between the last decoding layer and the previous decoding layer, a two-layer similarity coefficient loss is designed, the expression of which is:
[0160] L seg =αL DSC1 +βL DSC2
[0161] L DSC1 L represents the landslide prediction probability from the last decoding layer. DSC2 This represents the landslide prediction probability from the previous decoding layer, and α and β represent weighting coefficients, which are 1 and 0.1, respectively.
[0162] Optimization against losses specifically includes:
[0163] Considering the consistency of the landslide output space between the source and target domains, an adversarial network is constructed to explore the domain invariance between the source and target domains. Specifically, the landslide prediction probabilities from the last decoding layer and the previous decoding layer are input into the adversarial network for optimization, and the last decoding layer is used as an example for description.
[0164] The discriminator network is optimized using the binary cross-entropy loss, taking the output space from the source domain and the output space from the target domain as inputs. The purpose is to determine whether the output space comes from the source domain or the target domain. The calculation formula is as follows:
[0165]
[0166] in, This represents the predicted landslide probability in the source domain. This represents the predicted landslide probability for the target domain. This represents the predicted probability distribution of the source domain. This represents the predicted probability distribution of the target domain. Indicates that the discriminator distinguishes Let be the probability of the source domain. Indicates the discriminator distinguishes The probability of the target domain.
[0167] To extract domain-invariant features and complete the adversarial process, the segmentation model is optimized by forcing the discriminator to identify the target domain's landslide output probability as the source domain, thus projecting different domains into similar feature spaces. The adversarial loss of the segmentation network is calculated as follows:
[0168]
[0169] in, To combat the losses, To distinguish for the discriminator The probability of the source domain;
[0170] The operations of the penultimate decoding layer are the same as those of the last decoding layer; therefore, the constraint functions for optimizing the discriminator and the segmentation network are defined as follows:
[0171]
[0172]
[0173] Therefore, the total loss of the adversarial network is:
[0174]
[0175] In the formula, λ1 and λ2 control the importance of the loss function, and are empirically defined as 0.01 and 0.001, respectively.
[0176] Preferably, the optimization of adversarial loss specifically includes:
[0177] The maximum mean deviation constraint function is used to reduce the distance between the feature space distributions of the source and target domains. Its formula is as follows:
[0178]
[0179] Among them, L mmd The distance between the minimized domains is given by k, where k represents the feature distribution calculated by the Gaussian radial basis function kernel function, a represents the source domain feature, and b represents the target domain feature.
[0180] Step S35: Minimize the domain distance between the source domain landslide features and the target domain landslide features to obtain the domain distribution spatial distance loss of the source domain landslide features and the target domain landslide features.
[0181] Understandably, during the testing phase, the target domain image is input into the optimized landslide detection model to obtain predicted landslide location data. Based on the landslide image features of the source and target domains, domain distance minimization is introduced to reduce the distance between the feature distribution spaces of the target and source domains.
[0182] Step S36: Optimize the U-Net3+ segmentation network by combining the adversarial loss and the domain distance minimization loss to obtain the landslide location prediction result of the target domain.
[0183] It should be understood that the transformed image is adjusted according to the spatial distance of the distribution to obtain an adjusted test image. By inputting the test image into the optimized U-Net3+ segmentation network, the U-Net3+ segmentation network is optimized by combining the adversarial loss and the domain distance minimization loss, and the landslide location prediction result of the target domain can be obtained.
[0184] This embodiment, through the above scheme, obtains an improved U-Net3+ segmentation network by replacing the max-pooling layers of each coding layer in the original U-Net3+ network structure with a Tokenized MLP; the source domain image is input into the improved U-Net3+ segmentation network to obtain source domain landslide features and source domain landslide location prediction results, and the segmentation network is optimized using the source domain ground truth labels; the style-transferred target domain image is input into the improved U-Net3+ segmentation network to obtain target domain landslide features and target domain landslide location prediction results; an adversarial loss is constructed based on the source domain landslide location prediction results and the target domain landslide location prediction results, and the adversarial loss is calculated based on the source domain landslide location prediction results and the target domain landslide location prediction results; for the... The source domain landslide features and the target domain landslide features are minimized to obtain the domain distribution spatial distance loss of the source domain landslide features and the target domain landslide features. The U-Net3+ segmentation network is jointly optimized based on the adversarial loss and the domain distance minimization loss to obtain the landslide location prediction result of the target domain. The domain adaptation network includes an improved U-Net3+ segmentation network, segmentation loss, and adversarial and domain distance minimization losses to obtain the landslide location prediction result. It can make full use of the optical image and digital elevation model of high-resolution imagery to extract feature factors and mine the domain invariant features of different scenes, thereby improving the transferability accuracy of cross-scene landslide detection and improving the cross-scene detection capability and practicality of the model.
[0185] Accordingly, the present invention further provides a cross-scenario adaptive landslide detection device.
[0186] Reference Figure 6 , Figure 6 This is a functional block diagram of the first embodiment of the cross-scene domain adaptive landslide detection device of the present invention.
[0187] In a first embodiment of the cross-scene domain adaptive landslide detection device of the present invention, the cross-scene domain adaptive landslide detection device includes:
[0188] The data acquisition module 10 is used to acquire optical images of the source domain and the target domain, digital elevation models, and source domain landslide label data, and to merge the optical images, digital elevation models, and source domain landslide label data to obtain merged image data.
[0189] The cropping and transformation module 20 is used to preprocess the merged image data, crop the preprocessed image to obtain a cropped image, and input the cropped image into a style transfer network for style transfer to obtain a target domain transformed image corresponding to the source domain style.
[0190] The prediction module 30 is used to input the target domain transformed image into a domain adaptation network, which includes an improved U-Net3+ segmentation network, segmentation loss, and adversarial and domain distance minimization loss to obtain landslide location prediction results.
[0191] The data acquisition module 10 is also used to acquire optical images of the source domain and the target domain, digital elevation models, and source domain landslide label data; perform formula analysis on the optical images and the digital elevation models to extract feature variables; and merge the feature variables with the source domain landslide label data to obtain merged image data.
[0192] The data acquisition module 10 is also used to acquire optical images of the source domain and the target domain, digital elevation models, and source domain landslide label data.
[0193] The vegetation index is calculated using the reflectance of the red and near-infrared bands of the optical image through the following formula:
[0194]
[0195] Wherein, NDVI is the vegetation index, NIR is the reflectance in the near-infrared band, and R is the reflectance in the red band.
[0196] Feature variables of terrain features in the optical image are extracted based on the vegetation index and the digital elevation model.
[0197] The cropping and conversion module 20 is further configured to obtain the maximum and minimum values of each channel on which the merged image data is based, normalize the maximum and minimum values to obtain a normalized image; calculate the mean and variance of the normalized image, perform deviation standardization on the mean and variance to obtain a standardized image; crop the standardized image to obtain a cropped image, and input the cropped image into a style transfer network for style transfer to obtain a target domain converted image corresponding to the source domain style.
[0198] The cropping and conversion module 20 is further configured to crop the standardized image to obtain a cropped image; input the RGB channel of the optical image into a style conversion network to obtain a trained style conversion model; and convert the image style of the cropped image into the source domain image style through the trained style conversion model to obtain a target domain converted image with the corresponding source domain style.
[0199] The prediction module 30 is further configured to replace the max-pooling layers of each coding layer in the original U-Net3+ network structure with a Tokenized MLP to obtain an improved U-Net3+ segmentation network; input the source domain image into the improved U-Net3+ segmentation network to obtain source domain landslide features and source domain landslide location prediction results, and optimize the segmentation network using the source domain ground truth labels; input the style-transferred target domain image into the improved U-Net3+ segmentation network to obtain target domain landslide features and target domain landslide location prediction results; and construct an anti-damage network based on the source domain landslide location prediction results and the target domain landslide location prediction results. The process involves: calculating an adversarial loss based on the predicted landslide location in the source domain and the predicted landslide location in the target domain; minimizing the domain distance between the landslide features in the source domain and the landslide features in the target domain to obtain the domain distribution spatial distance loss between the landslide features in the source domain and the landslide features in the target domain; and optimizing the U-Net3+ segmentation network by combining the adversarial loss and the domain distance minimization loss to obtain the predicted landslide location in the target domain. The domain adaptation network includes an improved U-Net3+ segmentation network, segmentation loss, and adversarial and domain distance minimization losses to obtain the predicted landslide location.
[0200] The prediction module 30 is further configured to construct an adversarial loss based on the predicted landslide location in the source domain and the predicted landslide location in the target domain using the following formula:
[0201]
[0202] in, To combat the losses, To distinguish for the discriminator The probability of the source domain;
[0203] Accordingly, the step of minimizing the domain distance between the source domain landslide features and the target domain landslide features includes:
[0204] The maximum mean deviation constraint function is used to reduce the domain distance between the spatial distributions of landslide features in the source domain and the target domain by the following formula:
[0205]
[0206] Among them, L mmd The distance between the minimized domains is given by k, where k is the Gaussian radial basis function kernel function used to calculate the feature distribution, a is the landslide feature of the source domain, and b is the landslide feature of the target domain.
[0207] The steps for implementing each functional module of the cross-scene domain adaptive landslide detection device can be referred to in the various embodiments of the cross-scene domain adaptive landslide detection method of the present invention, and will not be repeated here.
[0208] Furthermore, this embodiment of the invention also proposes a storage medium storing a cross-scene domain adaptive landslide detection program, which, when executed by a processor, performs the following operations:
[0209] Acquire optical images, digital elevation models, and source domain landslide label data from the source and target domains; merge the optical images, digital elevation models, and source domain landslide label data to obtain merged image data.
[0210] The merged image data is preprocessed, and the preprocessed image is cropped to obtain a cropped image. The cropped image is then input into a style transfer network for style transfer to obtain a target domain transformed image corresponding to the source domain style.
[0211] The target domain-transformed image is input into a domain adaptation network, which includes an improved U-Net3+ segmentation network, segmentation loss, and adversarial and domain distance minimization loss, to obtain landslide location prediction results.
[0212] Furthermore, when the cross-scenario adaptive landslide detection program is executed by the processor, it also performs the following operations:
[0213] Acquire optical images of the source and target domains, digital elevation models, and landslide label data of the source domain; perform formula analysis on the optical images and the digital elevation models to extract feature variables;
[0214] The feature variables are merged with the source domain landslide label data to obtain merged image data.
[0215] Furthermore, when the cross-scenario adaptive landslide detection program is executed by the processor, it also performs the following operations:
[0216] Acquire optical images of the source and target domains, digital elevation models, and landslide label data of the source domain;
[0217] The vegetation index is calculated using the reflectance of the red and near-infrared bands of the optical image through the following formula:
[0218]
[0219] Wherein, NDVI is the vegetation index, NIR is the reflectance in the near-infrared band, and R is the reflectance in the red band.
[0220] Feature variables of terrain features in the optical image are extracted based on the vegetation index and the digital elevation model.
[0221] Furthermore, when the cross-scenario adaptive landslide detection program is executed by the processor, it also performs the following operations:
[0222] The maximum and minimum values of each channel on which the merged image data is based are obtained, and the maximum and minimum values are normalized to obtain a normalized image.
[0223] Calculate the mean and variance of the normalized image, and perform deviation standardization on the mean and variance to obtain a standardized image;
[0224] The standardized image is cropped to obtain a cropped image. The cropped image is then input into a style transfer network for style transfer to obtain a target domain transformed image corresponding to the source domain style.
[0225] Furthermore, when the cross-scenario adaptive landslide detection program is executed by the processor, it also performs the following operations:
[0226] The standardized image is cropped to obtain a cropped image;
[0227] The style transfer network is input through the RGB channels of the optical image to obtain a trained style transfer model;
[0228] The cropped image style is converted to the source domain image style by the trained style transfer model, and a target domain converted image with the corresponding source domain style is obtained.
[0229] Furthermore, when the cross-scenario adaptive landslide detection program is executed by the processor, it also performs the following operations:
[0230] An improved U-Net3+ segmentation network is obtained by replacing the max pooling layers of each coding layer in the original U-Net3+ network structure with Tokenized MLP.
[0231] The source domain image is input into the improved U-Net3+ segmentation network to obtain the source domain landslide features and source domain landslide location prediction results. The segmentation network is then optimized using the source domain real labels.
[0232] The style-transferred target domain image is input into the improved U-Net3+ segmentation network to obtain the target domain landslide features and target domain landslide location prediction results.
[0233] The countermeasure loss is calculated based on the predicted landslide locations in the source region and the predicted landslide locations in the target region.
[0234] Minimize the domain distance between the source domain landslide features and the target domain landslide features to obtain the domain distribution spatial distance loss between the source domain landslide features and the target domain landslide features;
[0235] The U-Net3+ segmentation network is optimized by combining the adversarial loss and the domain distance minimization loss to obtain the landslide location prediction result in the target domain. The domain adaptation network includes the improved U-Net3+ segmentation network, segmentation loss, and adversarial and domain distance minimization losses to obtain the landslide location prediction result.
[0236] Furthermore, when the cross-scenario adaptive landslide detection program is executed by the processor, it also performs the following operations:
[0237] Based on the predicted landslide locations in the source region and the predicted landslide locations in the target region, an adversarial loss is constructed using the following formula:
[0238]
[0239] in, To combat the losses, To distinguish for the discriminator The probability of the source domain;
[0240] Accordingly, the step of minimizing the domain distance between the source domain landslide features and the target domain landslide features includes:
[0241] The maximum mean deviation constraint function is used to reduce the domain distance between the spatial distributions of landslide features in the source domain and the target domain by the following formula:
[0242]
[0243] Among them, L mmd The distance between the minimized domains is given by k, where k is the Gaussian radial basis function kernel function used to calculate the feature distribution, a is the landslide feature of the source domain, and b is the landslide feature of the target domain.
[0244] This embodiment, through the above-described scheme, acquires optical images, digital elevation models, and source domain landslide label data from the source and target domains. It then merges these data to obtain merged image data. The merged image data is preprocessed, and the preprocessed image is cropped to obtain a cropped image. This cropped image is input into a style transfer network for style transfer, resulting in a target domain-converted image corresponding to the source domain style. Finally, the target domain-converted image is input into a domain adaptation network, which includes an improved U-Net3+ segmentation network, segmentation loss, and adversarial and domain distance minimization losses, to obtain landslide location prediction results. This approach fully utilizes the optical images and digital elevation models of high-resolution imagery to extract feature factors and mines domain-invariant features across different scenes, thereby improving the transferability accuracy of cross-scene landslide detection and enhancing the model's cross-scene detection capability and practicality.
[0245] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.
[0246] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.
[0247] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.
Claims
1. A cross-scenario adaptive landslide detection method, characterized in that, The cross-scenario adaptive landslide detection method includes: Acquire optical images, digital elevation models, and source domain landslide label data from the source and target domains; merge the optical images, digital elevation models, and source domain landslide label data to obtain merged image data. The merged image data is preprocessed, and the preprocessed image is cropped to obtain a cropped image. The cropped image is then input into a style transfer network for style transfer to obtain a target domain transformed image corresponding to the source domain style. The target domain-transformed image is input into a domain adaptation network, which includes an improved U-Net3+ segmentation network, segmentation loss, and adversarial and domain distance minimization loss, to obtain landslide location prediction results. The step of inputting the target domain-transformed image into a domain adaptation network, wherein the domain adaptation network includes an improved U-Net3+ segmentation network, segmentation loss, and adversarial and domain distance minimization loss, to obtain landslide location prediction results includes: An improved U-Net3+ segmentation network is obtained by replacing the max pooling layers of each coding layer in the original U-Net3+ network structure with Tokenized MLP. The source domain image is input into the improved U-Net3+ segmentation network to obtain the source domain landslide features and source domain landslide location prediction results. The segmentation network is then optimized using the source domain real labels. The style-transferred target domain image is input into the improved U-Net3+ segmentation network to obtain the target domain landslide features and target domain landslide location prediction results. The countermeasure loss is calculated based on the predicted landslide locations in the source region and the predicted landslide locations in the target region. Minimize the domain distance between the source domain landslide features and the target domain landslide features to obtain the domain distribution spatial distance loss between the source domain landslide features and the target domain landslide features; The U-Net3+ segmentation network is optimized by combining the adversarial loss and the domain distance minimization loss to obtain the landslide location prediction result in the target domain.
2. The cross-scenario adaptive landslide detection method as described in claim 1, characterized in that, The process involves acquiring optical images of the source and target domains, a digital elevation model (DEM), and source domain landslide label data. The process then merges these data to obtain merged image data, including: Acquire optical images of the source and target domains, digital elevation models, and landslide label data of the source domain; perform formula analysis on the optical images and the digital elevation models to extract feature variables; The feature variables are merged with the source domain landslide label data to obtain merged image data.
3. The cross-scenario adaptive landslide detection method as described in claim 2, characterized in that, The process involves acquiring optical images of the source and target domains, digital elevation models, and landslide label data in the source domain. Formula analysis is then performed on the optical images and the digital elevation model to extract feature variables, including: Acquire optical images of the source and target domains, digital elevation models, and landslide label data of the source domain; The vegetation index is calculated using the reflectance of the red and near-infrared bands of the optical image through the following formula: in, The vegetation index, The reflectivity is in the near-infrared band. The reflectivity is in the red band. Feature variables of terrain features in the optical image are extracted based on the vegetation index and the digital elevation model.
4. The cross-scenario adaptive landslide detection method as described in claim 1, characterized in that, The process of preprocessing the merged image data, cropping the preprocessed image to obtain a cropped image, and inputting the cropped image into a style transfer network for style transfer to obtain a target domain-transformed image corresponding to the source domain style includes: The maximum and minimum values of each channel on which the merged image data is based are obtained, and the maximum and minimum values are normalized to obtain a normalized image. Calculate the mean and variance of the normalized image, and perform deviation standardization on the mean and variance to obtain a standardized image; The standardized image is cropped to obtain a cropped image. The cropped image is then input into a style transfer network for style transfer to obtain a target domain transformed image corresponding to the source domain style.
5. The cross-scenario adaptive landslide detection method as described in claim 4, characterized in that, The process of cropping the standardized image to obtain a cropped image, and then inputting the cropped image into a style transfer network for style transfer to obtain a target domain-transformed image corresponding to the source domain style, includes: The standardized image is cropped to obtain a cropped image; The style transfer network is input through the RGB channels of the optical image to obtain a trained style transfer model; The cropped image style is converted to the source domain image style by the trained style transfer model, thus obtaining a target domain converted image with the corresponding source domain style.
6. The cross-scenario adaptive landslide detection method as described in claim 1, characterized in that, The step of constructing an adversarial loss based on the predicted landslide location in the source region and the predicted landslide location in the target region, and calculating the adversarial loss based on the predicted landslide location in the source region and the predicted landslide location in the target region, includes: Based on the predicted landslide locations in the source region and the predicted landslide locations in the target region, an adversarial loss is constructed using the following formula: in, To combat the losses, To distinguish for the discriminator The probability of the source domain; Accordingly, the step of minimizing the domain distance between the source domain landslide features and the target domain landslide features includes: The maximum mean deviation constraint function is used to reduce the domain distance between the spatial distributions of landslide features in the source domain and the target domain by the following formula: in, The distance between the minimized domains is given by k, where k is the Gaussian radial basis function kernel function used to calculate the feature distribution, a is the landslide feature of the source domain, and b is the landslide feature of the target domain.
7. A cross-scenario adaptive landslide detection device, characterized in that, The cross-scenario adaptive landslide detection device includes: The data acquisition module is used to acquire optical images, digital elevation models, and source domain landslide label data from the source and target domains, and to merge the optical images, digital elevation models, and source domain landslide label data to obtain merged image data. The cropping and transformation module is used to preprocess the merged image data, crop the preprocessed image to obtain a cropped image, and input the cropped image into a style transfer network for style transfer to obtain a target domain transformed image corresponding to the source domain style. The prediction module is used to input the target domain-transformed image into the domain adaptation network, which includes an improved U-Net3+ segmentation network, segmentation loss, and adversarial and domain distance minimization loss to obtain the landslide location prediction result. The prediction module is further configured to replace the max-pooling layers of each coding layer in the original U-Net3+ network structure with a Tokenized MLP to obtain an improved U-Net3+ segmentation network; input the source domain image into the improved U-Net3+ segmentation network to obtain source domain landslide features and source domain landslide location prediction results, and optimize the segmentation network using the source domain ground truth labels; input the style-transferred target domain image into the improved U-Net3+ segmentation network to obtain target domain landslide features and target domain landslide location prediction results; calculate adversarial loss based on the source domain landslide location prediction results and the target domain landslide location prediction results; minimize the domain distance between the source domain landslide features and the target domain landslide features to obtain the domain distribution spatial distance loss between the source domain landslide features and the target domain landslide features; and optimize the U-Net3+ segmentation network by combining the adversarial loss and the domain distance minimization loss to obtain the target domain landslide location prediction results.
8. A cross-scenario adaptive landslide detection device, characterized in that, The cross-domain adaptive landslide detection device includes: a memory, a processor, and a cross-domain adaptive landslide detection program stored in the memory and executable on the processor, wherein the cross-domain adaptive landslide detection program is configured to implement the steps of the cross-domain adaptive landslide detection method as described in any one of claims 1 to 6.
9. A storage medium, characterized in that, The storage medium stores a cross-scene domain adaptive landslide detection program, which, when executed by a processor, implements the steps of the cross-scene domain adaptive landslide detection method as described in any one of claims 1 to 6.