Semi-supervised remote sensing image historical landslide detection method based on three-dimensional joint disturbance

Abstract

The application provides a kind of semi-supervised remote sensing image historical landslide detection method based on three-dimensional joint disturbance, comprising: a small amount of landslide sample marking and sliding segmentation are carried out on the remote sensing image of the landslide area to be detected, and a landslide data set is established;The training set and the validation set containing unmarked and marked data are jointly input into the model, then input disturbance, feature disturbance and model disturbance with strong and weak data enhancement are introduced in the training process, a joint weighted loss function of supervised learning and consistency regularization is established;The test set is input into the trained teacher model, and the final regional historical landslide detection result is output. The application fully explores and learns the landslide features in unmarked data, makes the feature space distribution learned by the model more extensive, improves the generalization ability of the model, effectively reduces the number of artificial marking required for historical landslide detection, and realizes accurate historical landslide detection under the condition of a small amount of marked data and a large amount of unmarked data.

Description

Technical Field

[0001] This invention relates to the field of remote sensing detection technology, and in particular to a semi-supervised remote sensing image historical landslide detection method based on three-dimensional joint perturbation. Background Technology

[0002] Landslides, as a widely distributed geological hazard, not only cause significant economic and property losses after they collapse and slide, but also pose a great threat to personal safety. Therefore, rapid, accurate, and automated detection and identification of landslides are of great practical significance for geological disaster prevention and control, emergency rescue and disaster relief, and the protection of people's lives and property.

[0003] Landslide detection can be divided into two categories: detection of historical landslides and detection of early landslides. Detection of historical landslides can provide important basic data support for establishing a regional landslide inventory, reflecting the distribution of landslides, assessing landslide susceptibility, and exploring landslide mechanisms. Early detection of historical landslides is mainly carried out through visual interpretation of remote sensing images or on-site surveys. Although these methods have high accuracy, they are time-consuming, labor-intensive, and greatly affected by human subjectivity, making it difficult to meet the practical needs of the existing disaster prevention and mitigation system.

[0004] In recent years, with the development and progress of remote sensing and computer vision technologies, computer-aided methods for automatically detecting historical landslides from remote sensing images have gradually become mainstream. Currently, computer-based methods for detecting historical landslides using remote sensing mainly include four categories: change detection, feature thresholding, machine learning, and deep learning. Change detection methods have high requirements for image resolution and conditions such as imaging time, illumination, and sensor quality between images from different periods. Feature thresholding methods have poor universality because the selection of feature thresholds and segmentation scales is usually done manually. Machine learning methods require experts with specialized knowledge to design and extract features, thus limiting their efficiency and generalization ability. Deep learning methods utilize deep convolutional neural networks (DCNNs) and a large amount of manually labeled data to automatically extract high- and low-dimensional features from images and perform classification, significantly improving the efficiency and accuracy of historical landslide detection. Although DCNN-based historical landslide detection has been widely used and achieved good results, it largely depends on the availability of a large amount of labeled data, leading to a surge in model training costs. To reduce reliance on high-quality expert-labeled data, transfer learning and semi-supervised learning (SSL) have been applied to historical landslide detection tasks.

[0005] Transfer learning reduces the need for labeled data in the target domain by transferring knowledge learned from a source domain with abundant labeled data. However, currently, there are few open-source landslide datasets that can be used as source domains, and the geological, topographical, and imaging environments of different landslide areas vary significantly. Therefore, landslide detection using transfer learning still faces considerable challenges. SLL (Simulated Landslide Learning) is a method that learns decision boundaries from a small amount of labeled data and a large amount of unlabeled data. Currently, SLL methods mainly include three categories: consistent regularization, pseudo-label self-training, and generative adversarial methods. However, erroneous features learned from generative landslide imagery may reduce the accuracy of landslide detection. Furthermore, pseudo-labels during self-training are not entirely accurate, which may lead to cognitive biases in the model. Pseudo-labels after threshold filtering can cause class imbalance problems, ultimately affecting the accuracy of landslide detection. Summary of the Invention

[0006] This invention provides a semi-supervised remote sensing image historical landslide detection method based on three-dimensional joint perturbation, in order to overcome the deficiencies in the existing technology.

[0007] In a first aspect, the present invention provides a semi-supervised remote sensing image historical landslide detection method based on three-dimensional joint perturbation, comprising:

[0008] Remote sensing images of the landslide area to be detected are collected, and the remote sensing images of the landslide area to be detected are divided into a training set, a validation set and a test set. The training set includes labeled image data and unlabeled image data, the validation set includes labeled image data, and the test set includes unlabeled image data.

[0009] Based on the training set, preset weak data augmentation perturbation and preset strong data augmentation perturbation are performed, and the perturbation is input into the student model for training. After feature perturbation, the perturbation is input into the main decoder and auxiliary decoder for training, and the perturbation is input into the teacher model for training to realize model perturbation. The joint weighted loss function is obtained, and the model training is completed by combining the validation set to obtain the historical landslide detection model.

[0010] The target set is input into the historical landslide detection model, and the final historical landslide detection results for the landslide area to be detected are output.

[0011] According to the present invention, a semi-supervised remote sensing image historical landslide detection method based on three-dimensional joint perturbation is provided, which involves acquiring remote sensing images of the landslide area to be detected, and dividing the remote sensing images of the landslide area to be detected into a training set, a validation set, and a test set, including:

[0012] The labeling results are obtained by manually labeling some landslide samples at the pixel level in the remote sensing image of the landslide area to be detected.

[0013] The remote sensing image of the landslide area to be detected and the labeled results are segmented using a fixed-size sliding window to obtain labeled image data. and unlabeled image data Where M is the number of labeled image data and N is the training set D. T The total number of image data in the middle, and M << NM, Representing a single landslide image patch, H = W represent the length and width of the single landslide image patch, respectively, C represents the number of bands possessed by the image patch, y i ∈{0,1} H×W×2 Represents pixel-level manual markings, with 0 indicating the background and 1 indicating the landslide area.

[0014] According to the present invention, a semi-supervised remote sensing image historical landslide detection method based on three-dimensional joint perturbation is provided. Based on the training set, preset weak data augmentation perturbation and preset strong data augmentation perturbation are applied, and input into a student model for training. After feature perturbation, the perturbation is applied to a main decoder and an auxiliary decoder for training. The perturbation is then applied to a teacher model for training to achieve model perturbation, obtaining a joint weighted loss function. The model is then trained using the validation set to obtain a historical landslide detection model. The method includes:

[0015] For the labeled image data X in the training set L The preset weak data enhancement perturbation is performed to obtain Will Input student model Landslide detection results obtained from labeled image data use and the corresponding pixel-level manual label Y L Constructing a supervised loss function

[0016] For the unlabeled image data X in the training set u The weak data augmentation perturbation is performed to obtain Will Input student model Weakly enhanced landslide detection results obtained from unlabeled image data Will Two randomized, preset strong data augmentation perturbations were performed to obtain... and Will and Input student model Obtain the results of markerless imagery strong data enhancement for disturbed landslide detection. and by Construct an input perturbation loss function for the target and

[0017] Will Input student model The encoder output is obtained. To each Apply Drop and Cut perturbations to obtain and Will Input student model The decoder gets Will Input student model The auxiliary decoder is obtained by Construct a feature perturbation loss function for the target and

[0018] exist After adding Gaussian noise perturbation N, input it into the teacher model. Obtain the model perturbation results by Construct a model perturbation loss function for the target

[0019] For the supervision loss function The input perturbation loss function and The characteristic perturbation loss function and and the model perturbation loss function The joint weighted loss function is constructed by performing a weighted summation, which constrains the update of the model parameters to obtain the historical landslide detection model.

[0020] According to the present invention, a semi-supervised remote sensing image historical landslide detection method based on three-dimensional joint perturbation is provided, which performs the detection on labeled image data X in the training set. L The preset weak data enhancement perturbation is performed to obtain Will Input student model Landslide detection results obtained from labeled image data use and the corresponding pixel-level manual label Y L Constructing a supervised loss function include:

[0021] For X L Perform random vertical flip, horizontal mirror flip, n×90° rotation, random cropping, and transpose operations to obtain... Will enter get

[0022] Constructing a supervised loss function by combining binary cross-entropy loss and Dice loss

[0023]

[0024] Wherein, CE is the binary cross-entropy loss function, and DICE is the Dice loss function, which is used to balance the imbalance between the number of landslide and background pixels in the image patch.

[0025] According to the present invention, a semi-supervised remote sensing image historical landslide detection method based on three-dimensional joint perturbation is provided, which performs the detection on the unlabeled image data X in the training set. u The weak data augmentation perturbation is performed to obtain Will Input student model Weakly enhanced landslide detection results obtained from unlabeled image data Will Two randomized, preset strong data augmentation perturbations were performed to obtain... and Will and Input student model Obtain the results of markerless imagery strong data enhancement for disturbed landslide detection. and by Construct an input perturbation loss function for the target and include:

[0026] For X u Perform random vertical flip, horizontal mirror flip, n×90° rotation, random cropping, and transpose operations to obtain... Will enter get

[0027] right Perform two random Gaussian blur, color dithering, and CutMix operations. The color dithering includes random adjustments to brightness, contrast, saturation, and hue. The color dithering is executed randomly, while the Gaussian blur and CutMix are executed sequentially.

[0028] The input perturbation loss function is constructed using Dice loss. and

[0029]

[0030] Where τ is the confidence threshold for filtering noise in the landslide detection results caused by strong data augmentation disturbances, and DICE is the Dice loss function.

[0031] According to the present invention, a semi-supervised remote sensing image-based historical landslide detection method based on three-dimensional joint perturbation is provided, which will... Input student model The encoder output is obtained. right Apply Drop perturbation to obtain Will Input student model The decoder gets by Construct a feature perturbation loss function for the target include:

[0032] Determine the sampling threshold γ;

[0033] right New features are obtained by summing and normalizing the channel dimensions.

[0034] generate Drop mask

[0035] Generate encoder output after Drop perturbation

[0036]

[0037] Will Input student model The decoder D obtains the final output of the Drop perturbation.

[0038] by As the objective, a feature perturbation loss function is established.

[0039]

[0040] According to the present invention, a semi-supervised remote sensing image-based historical landslide detection method based on three-dimensional joint perturbation is provided, which will... Input student model The encoder output is obtained. right Applying the Cut perturbation yields Will Input student model The auxiliary decoder is obtained by Construct a feature perturbation loss function for the target include:

[0041] right Perform the Cutout operation to obtain

[0042] Will Input student model The auxiliary decoder G obtains the Cutout perturbation and finally outputs the output. G represents PixelShuffle upsampling, and the number of upsampling operations is consistent with the number of downsampling operations in the decoder.

[0043] by As the objective, a feature perturbation loss function is established.

[0044]

[0045] Where MSE is the mean squared error.

[0046] According to the present invention, a semi-supervised remote sensing image-based historical landslide detection method based on three-dimensional joint perturbation is provided. After adding Gaussian noise perturbation N, input it into the teacher model. Obtain the model perturbation results by Construct a model perturbation loss function for the target include:

[0047] Determine the mean and standard deviation of the Gaussian noise perturbation;

[0048] Determine the teacher model With the student model All are U-Net models, the teacher model The student model is initialized using a layer-by-layer traversal method. The teacher model uses the Kaiming normal distribution for parameter initialization. The parameters are obtained by analyzing the student model. The parameters are implemented using an exponential moving average.

[0049] by Construct a model perturbation loss function for the target

[0050]

[0051] According to the present invention, a semi-supervised remote sensing image historical landslide detection method based on three-dimensional joint perturbation is provided, wherein the supervised loss function is... The input perturbation loss function and The characteristic perturbation loss function and and the model perturbation loss function Perform weighted summation to construct the joint weighted loss function, including:

[0052]

[0053] in, Let α, β, μ, η, and λ represent the joint weighted loss function, respectively. and The corresponding weights.

[0054] Secondly, the present invention also provides a semi-supervised remote sensing image historical landslide detection system based on three-dimensional joint perturbation, comprising:

[0055] The data processing module is used to acquire remote sensing images of the landslide area to be detected, and divide the remote sensing images of the landslide area to be detected into a training set, a validation set and a test set. The training set includes labeled image data and unlabeled image data, the validation set includes labeled image data, and the test set includes unlabeled image data.

[0056] The training module is used to perform preset weak data augmentation perturbation and preset strong data augmentation perturbation based on the training set, input the perturbation into the student model for training, input the perturbation into the main decoder and auxiliary decoder for training, input the perturbation into the teacher model for training to realize model perturbation, obtain the joint weighted loss function, and complete the model training by combining the validation set to obtain the historical landslide detection model.

[0057] The detection module is used to input the set to be detected into the historical landslide detection model and output the final historical landslide detection results for the landslide area to be detected.

[0058] The present invention provides a semi-supervised remote sensing image historical landslide detection method based on three-dimensional joint perturbation. By fully exploring and learning the landslide features in unlabeled data, the feature space learned by the model is more extensive, improving the model's generalization ability and effectively reducing the number of manual labels required for historical landslide detection. This enables accurate historical landslide detection under conditions of a small amount of labeled data and a large amount of unlabeled data. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0060] Figure 1This is one of the flowcharts of the semi-supervised remote sensing image historical landslide detection method based on three-dimensional joint perturbation provided by the present invention;

[0061] Figure 2 This is the second flowchart of the semi-supervised remote sensing image historical landslide detection method based on three-dimensional joint perturbation provided by the present invention;

[0062] Figure 3 This is a schematic diagram of the model training process provided by the present invention;

[0063] Figure 4 This invention provides a dataset of landslides in a certain region, including partial images, corresponding DEMs, and manually labeled maps.

[0064] Figure 5 These are schematic diagrams illustrating different input perturbations provided by the present invention;

[0065] Figure 6 This is a bar chart of quantitative evaluation results from a landslide dataset for a certain region provided by this invention.

[0066] Figure 7 This is a partial visualization of landslide detection results from a landslide dataset provided by this invention.

[0067] Figure 8 This is a schematic diagram of the structure of the semi-supervised remote sensing image historical landslide detection system based on three-dimensional joint perturbation provided by the present invention;

[0068] Figure 9 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0070] To address the limitations of existing technologies in the detection of historical landslides in remote sensing images, this invention proposes a semi-supervised automatic detection method for historical landslides in remote sensing images based on three-dimensional joint perturbation, starting from the theory of consistency regularization and through the joint use of diverse and independent perturbation strategies. This method can be applied to the accurate detection of historical landslides from a small number of images with pixel-level landslide markers and a large number of unmarked images.

[0071] Figure 1 This is one of the flowcharts illustrating the semi-supervised remote sensing image historical landslide detection method based on three-dimensional joint perturbation provided in this embodiment of the invention, such as... Figure 1As shown, it includes:

[0072] Step 100: Collect remote sensing images of the landslide area to be detected, and divide the remote sensing images of the landslide area to be detected into a training set, a validation set and a test set. The training set includes labeled image data and unlabeled image data, the validation set includes labeled image data, and the test set includes unlabeled image data.

[0073] Step 200: Based on the training set, perform preset weak data augmentation perturbation and preset strong data augmentation perturbation, input them into the student model for training, perform feature perturbation and input them into the main decoder and auxiliary decoder for training, input them into the teacher model for training to realize model perturbation, obtain the joint weighted loss function, and combine it with the validation set to complete the model training and obtain the historical landslide detection model;

[0074] Step 300: Input the test set into the historical landslide detection model and output the final historical landslide detection result of the landslide area to be detected.

[0075] Specifically, the overall logical process of the embodiments of the present invention is as follows: Figure 2 As shown, it includes:

[0076] The first stage involves data preparation, which involves labeling a small number of landslide samples and performing sliding segmentation on remote sensing images of the landslide-to-be-inspected area to establish a landslide dataset.

[0077] In the second stage, model training, the training set and validation set containing both unlabeled and labeled data are input into the model. Then, during the training process, input perturbations with strong and weak data augmentation, feature perturbations, and model perturbations are introduced to establish a joint weighted loss function of supervised learning and consistency regularization, so as to learn landslide features from a small amount of labeled data and a large amount of unlabeled data.

[0078] The third stage, landslide detection, involves inputting the target set into the trained teacher model and outputting the final historical landslide detection results for the region.

[0079] This invention fully explores and learns landslide features in unlabeled data, enabling the model to learn a wider feature space distribution, improving the model's generalization ability, effectively reducing the number of manual labels required for historical landslide detection, and achieving accurate historical landslide detection under conditions of a small amount of labeled data and a large amount of unlabeled data.

[0080] Based on the above embodiments, remote sensing images of the landslide area to be detected are acquired, and the remote sensing images of the landslide area to be detected are divided into a training set, a validation set, and a test set, including:

[0081] The labeling results are obtained by manually labeling some landslide samples at the pixel level in the remote sensing image of the landslide area to be detected.

[0082] The remote sensing image of the landslide area to be detected and the labeled results are segmented using a fixed-size sliding window to obtain labeled image data. and unlabeled image data Where M is the number of labeled image data and N is the training set D. T The total number of image data in the middle, and M << NM, Representing a single landslide image patch, H = W represent the length and width of the single landslide image patch, respectively, C represents the number of bands possessed by the image patch, y i ∈{0,1} H×W×2 Represents pixel-level manual markings, with 0 indicating the background and 1 indicating the landslide area.

[0083] Based on the above embodiments, such as Figure 3 The model training process shown includes:

[0084] For the labeled image data X in the training set L The preset weak data enhancement perturbation is performed to obtain Will Input student model Landslide detection results obtained from labeled image data use and the corresponding pixel-level manual label Y L Constructing a supervised loss function

[0085] For the unlabeled image data X in the training set u The weak data augmentation perturbation is performed to obtain Will Input student model Weakly enhanced landslide detection results obtained from unlabeled image data Will The preset strong data enhancement perturbation was performed twice randomly to obtain... and Will and Input student model Obtain the results of markerless imagery strong data enhancement for disturbed landslide detection. and by Construct an input perturbation loss function for the target and

[0086] Will Input student model The encoder output is obtained. To each Apply Drop and Cut perturbations to obtain and Will Input student model The decoder gets Will Input student model The auxiliary decoder is obtained by Construct a feature perturbation loss function for the target and

[0087] exist After adding Gaussian noise perturbation N, input it into the teacher model. Obtain the model perturbation results by Construct a model perturbation loss function for the target

[0088] For the supervision loss function The input perturbation loss function and The characteristic perturbation loss function and and the model perturbation loss function The joint weighted loss function is constructed by performing a weighted summation, which constrains the update of the model parameters to obtain the historical landslide detection model.

[0089] Among them, the labeled image data X in the training set L The preset weak data enhancement perturbation is performed to obtain Will Input student model Landslide detection results obtained from labeled image data use and the corresponding pixel-level manual label Y L Constructing a supervised loss function include:

[0090] For X L Perform random vertical flip, horizontal mirror flip, n×90° rotation, random cropping, and transpose operations to obtain... Will enter get

[0091] Constructing a supervised loss function by combining binary cross-entropy loss and Dice loss

[0092]

[0093] Wherein, CE is the binary cross-entropy loss function, and DICE is the Dice loss function, which is used to balance the imbalance between the number of landslide and background pixels in the image patch.

[0094] For the unlabeled image data X in the training set u The weak data augmentation perturbation is performed to obtain Will Input student model Weakly enhanced landslide detection results obtained from unlabeled image data Will Two randomized, preset strong data augmentation perturbations were performed to obtain... and Will and Input student model Obtain the results of markerless imagery strong data enhancement for disturbed landslide detection. and by Construct an input perturbation loss function for the target and include:

[0095] For X u Perform random vertical flip, horizontal mirror flip, n×90° rotation, random cropping, and transpose operations to obtain... Will enter get

[0096] right Perform two random Gaussian blur, color dithering, and CutMix operations. The color dithering includes random adjustments to brightness, contrast, saturation, and hue. The color dithering is executed randomly, while the Gaussian blur and CutMix are executed sequentially.

[0097] The input perturbation loss function is constructed using Dice loss. and

[0098]

[0099] Where τ is the confidence threshold for filtering noise in the landslide detection results caused by strong data augmentation disturbances, and DICE is the Dice loss function.

[0100] X u Input student model The encoder output is obtained. right Apply Drop perturbation to obtain Will Input student model The decoder gets by Construct a feature perturbation loss function for the target include:

[0101] Determine the sampling threshold γ;

[0102] right New features are obtained by summing and normalizing the channel dimensions.

[0103] generate Drop mask

[0104] Generate encoder output after Drop perturbation

[0105]

[0106] Will Input student model The decoder D obtains the final output of the Drop perturbation.

[0107] by As the objective, a feature perturbation loss function is established.

[0108]

[0109] X U Input student model The encoder output is obtained. right Applying the Cut perturbation yields Will Input student model The auxiliary decoder is obtained by Construct a feature perturbation loss function for the target include:

[0110] right Perform the Cutout operation to obtain

[0111] Will Input student model The auxiliary decoder G obtains the Cutout perturbation and finally outputs the output. G represents PixelShuffle upsampling, and the number of upsampling operations is consistent with the number of downsampling operations in the decoder.

[0112] by As the objective, a feature perturbation loss function is established.

[0113]

[0114] Where MSE is the mean squared error.

[0115] In X U After adding Gaussian noise perturbation N, input it into the teacher model. Obtain the model perturbation results by Construct a model perturbation loss function for the target include:

[0116] Determine the mean and standard deviation of the Gaussian noise perturbation;

[0117] Determine the teacher model With the student model All are U-Net models, the teacher model The student model is initialized using a layer-by-layer traversal method. The teacher model uses the Kaiming normal distribution for parameter initialization. The parameters are obtained by analyzing the student model. The parameters are implemented using an exponential moving average.

[0118] by Construct a model perturbation loss function for the target

[0119]

[0120] Finally, the supervision loss function The input perturbation loss function and The characteristic perturbation loss function and and the model perturbation loss function Perform weighted summation to construct the joint weighted loss function, including:

[0121]

[0122] in, Let α, β, μ, η, and λ represent the joint weighted loss function, respectively. and The corresponding weights.

[0123] This invention employs a semi-supervised remote sensing image-based automatic detection method for historical landslides, based on three-dimensional joint perturbation. This method enables accurate automatic detection of historical landslides from a small number of images with pixel-level landslide labels and a large number of unlabeled images, reducing reliance on high-quality expert-labeled data and lowering the cost of geological hazard investigations. Furthermore, by exploring a broader perturbation space through diverse and independent image-level inputs, features, and model perturbation strategies, the invention forces the model to learn more robust features from unlabeled data, enabling the model to grasp a wider feature space and improving its generalization ability for detecting landslides in different regions and types. The introduction of teacher and student models allows the teacher model to better represent mid-to-high-level semantic information, avoiding oscillations in prediction results and reducing the risk of overfitting.

[0124] The following uses a specific city as an example to illustrate the method of this invention. It employs TripleSat satellite imagery of the city provided by a landslide dataset, along with precise landslide markers and Digital Elevation Model (EMA) data for landslide detection. Specific implementation details and steps are as follows:

[0125] All experiments were conducted in a deep learning environment using Python 3.8, PyTorch 1.12.1, and CUDA 11.3, and an NVIDIA GeForce RTX 3060 graphics card with 12GB of VRAM was used to accelerate model training. No pre-trained weights were used in any model training. The specific experimental parameters are shown in Table 1.

[0126] Table 1

[0127]

[0128] Step S1: Use the images and corresponding labels provided in the city dataset as images and pixel-level manual labels for the landslide detection area, such as... Figure 4 As shown, all image patches are then resampled to a fixed size sliding window, resulting in a small number of landslide images with corresponding labels and a large number of unlabeled landslide images. The results are then divided into a training set D. T Validation set D V and the set to be inspected D W The training set D T It contains a small amount of labeled image data X L and a large amount of unlabeled image data X U , verification set D V For labeled image data, the retrieval set D W This is unlabeled image data;

[0129] In step S1, there is labeled image data. Unlabeled image data Where M is the number of labeled image data, and M is taken as 10, 30, 104, 154 respectively, and N = 514 is the training set D. T The total number of image data in China, in addition Representing a single landslide image patch, H=W=256 pixels represent the length and width of the image patch respectively, C=4 indicates that the image patch has four bands: red, green, blue, and DEM. i ∈{0,1} H×W×2 The pixel-level manual markers represent the landslide image blocks, with 0 for the background and 1 for the landslide area;

[0130] Step S2: For X L Perform weak data augmentation perturbation Then input the student model. Landslide detection results obtained from labeled image data Then use and the corresponding pixel-level manual label Y L Constructing a supervised loss function

[0131] In step S2, the weak data augmentation perturbation includes vertical and horizontal mirror flipping, n×90° rotation, random pruning within the range (0.8, 1), and transposition. During each data loading process, the probability of executing all weak data augmentation perturbation terms is 0.5. The supervised loss function... The combined binary cross-entropy loss and Dice loss are calculated using the following formula:

[0132]

[0133] Where CE is the binary cross-entropy loss function and DICE is the Dice loss function, used to balance the imbalance between the number of landslide and background pixels in the image patch.

[0134] Step S3: For X U Perform weak data augmentation perturbation Then input the student model. Weakly enhanced landslide detection results obtained from unlabeled image data Then Two random strong data augmentation perturbations were performed to obtain and Next and Input student model Obtain the results of markerless imagery strong data enhancement for disturbed landslide detection. and Finally As the objective, construct the input perturbation loss function. and

[0135] In step S3, the weak data augmentation perturbation remains consistent with step S2. The strong data augmentation perturbation includes a Gaussian blur with an intensity of (0.1, 2), color jitter, and a CutMix with a size of (0.02, 0.4). The color jitter includes brightness, contrast, and saturation adjustments with an intensity of (0.5, 1.5), and a hue adjustment with an intensity of 0.25. During each data loading process, the probability of executing the color jitter perturbation is 0.8, while the probability of executing the Gaussian blur and CutMix perturbations is 1. The input perturbation loss function is... The calculation formula is:

[0136]

[0137] Where: τ=0.3 is the confidence threshold for filtering noise in the landslide detection results caused by strong data augmentation disturbances;

[0138] Step S4: Input student model Obtain encoder output Then, two different types of characteristic perturbations are applied to it to obtain... Then As a student model The input to the decoder is obtained Will Input auxiliary decoder to obtain Next As the objective, construct a feature perturbation loss function. and

[0139] In step S4, the feature perturbation Drop process is as follows:

[0140] S4.1: Randomly determine the sampling threshold γ from the range (0.01, 0.1);

[0141] S4.2: Output of the student model encoder Summing the channel dimensions and normalizing them yields new features.

[0142] S4.3: Generate Drop Mask

[0143] S4.5: Generate encoder output after FeatureDrop perturbation

[0144] S4.6:

[0145] S4.7: with As the objective, the feature perturbation loss function The calculation formula is:

[0146] S4.8:

[0147] In step S4, the feature perturbation Cutout process is as follows:

[0148] S4.9: Output of the student model encoder Perform a Cutout operation to generate the encoder output after the Cutout perturbation.

[0149] S4.10: Will Input student model The auxiliary decoder G obtains the Cutout perturbation and finally outputs the output. Where G represents PixelShuffle upsampling, and the number of upsampling operations is consistent with the number of downsampling operations in the decoder.

[0150] S4.11: with As the objective, the feature perturbation loss function The calculation formula is:

[0151] S4.12:

[0152] S4.13: Where MSE is the mean squared error.

[0153] Step S5: The mean of the Gaussian noise perturbation N is 0, and the standard deviation is 0.1. Teacher model With student model The structure remains consistent, using the U-Net model. Manually initialize parameters using a hierarchical traversal method. Parameter initialization is performed using the Kaiming normal distribution. The parameters are obtained by adjusting the parameters. The parameters are implemented using an exponential moving average.

[0154] Model perturbation loss function The calculation formula is:

[0155]

[0156] The above-mentioned disturbances, such as Figure 5 As shown, (b)–(d) represent weak data augmentation perturbations, and (f)–(h) represent strong data augmentation perturbations.

[0157] Step S6: Monitor a loss and five consistency losses A joint weighted loss function is constructed by applying specific weights to each component. The model parameters are updated under common constraints, and the model training is completed.

[0158] In step S6, the joint weighted loss function The five consistency loss terms contribute equally to feature learning on unlabeled data, α=β=μ=η=λ=0.2. The parameters are obtained by adjusting the parameters. Perform backpropagation and update using the gradient descent rule;

[0159] Joint weighted loss function The calculation formula is:

[0160]

[0161] Input the images to be examined into the trained teacher model. The final historical landslide detection results for the region were obtained;

[0162] Step S7: Quantitatively evaluate the obtained historical landslide detection results using Overall Accuracy (OA), Precision (P), Recall (R), F1 score, and Intersection over Union (IoU). Figure 6 As shown.

[0163] Table 2 presents the quantitative evaluation results of this invention compared to supervised learning and several other semi-supervised learning methods. As shown in the table, all SSL methods achieve more accurate historical landslide detection results than supervised learning. In four sets of experiments with different amounts of labeled data and the same parameters, the OA, Recall, IoU, and F1-Score of this invention are superior to the compared SSL methods. This indicates that this invention can better utilize unlabeled data and broaden the feature space distribution by using three-dimensional joint perturbation, assisting researchers in detecting more landslides. With the increase in the amount of labeled data, the performance of all methods improves, and the gap between supervised learning and SSL methods gradually narrows. This is because the increased labeled data provides the model with more learnable features, reducing the gap between the detection set and the mastered feature space. However, it still lacks the ability to learn features from unlabeled data, so the overall accuracy is still lower than that of this invention. Furthermore, by stabilizing the training process through exponential moving averages on unlabeled data, this invention obtains more stable landslide detection results, as shown in the visualization results below. Figure 7 As shown.

[0164] Table 2

[0165]

[0166] The following describes the semi-supervised remote sensing image historical landslide detection system based on three-dimensional joint perturbation provided by the present invention. The semi-supervised remote sensing image historical landslide detection system based on three-dimensional joint perturbation described below can be referred to in correspondence with the semi-supervised remote sensing image historical landslide detection method based on three-dimensional joint perturbation described above.

[0167] Figure 8 This is a schematic diagram of the structure of a semi-supervised remote sensing image historical landslide detection system based on three-dimensional joint perturbation provided in an embodiment of the present invention, as shown below. Figure 8 As shown, it includes: a data processing module 81, a training module 82, and a detection module 83, wherein:

[0168] The data processing module 81 is used to acquire remote sensing images of the landslide area to be detected, and divide the remote sensing images of the landslide area to be detected into a training set, a validation set, and a test set. The training set includes labeled image data and unlabeled image data, the validation set includes labeled image data, and the test set includes unlabeled image data. The training module 82 is used to perform preset weak data augmentation perturbation and preset strong data augmentation perturbation based on the training set, input them into the student model for training, input them into the main decoder and auxiliary decoder for training after feature perturbation, input them into the teacher model for training to realize model perturbation, obtain the joint weighted loss function, and complete the model training by combining it with the validation set to obtain the historical landslide detection model. The detection module 83 is used to input the test set into the historical landslide detection model and output the final historical landslide detection result of the landslide area to be detected.

[0169] Figure 9 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 9As shown, the electronic device may include: a processor 910, a communication interface 920, a memory 930, and a communication bus 940, wherein the processor 910, the communication interface 920, and the memory 930 communicate with each other through the communication bus 940. The processor 910 can call logic instructions in the memory 930 to execute a semi-supervised remote sensing image historical landslide detection method based on three-dimensional joint perturbation. This method includes: acquiring remote sensing images of the landslide area to be detected; dividing the remote sensing images of the landslide area to be detected into a training set, a validation set, and a test set; the training set includes labeled image data and unlabeled image data; the validation set includes labeled image data; and the test set includes unlabeled image data; performing preset weak data augmentation perturbation and preset strong data augmentation perturbation based on the training set, inputting them into a student model for training; performing feature perturbation and inputting them into a main decoder and an auxiliary decoder for training; inputting them into a teacher model for training to achieve model perturbation, obtaining a joint weighted loss function; combining the validation set to complete model training, obtaining a historical landslide detection model; inputting the test set into the historical landslide detection model, and outputting the final historical landslide detection result for the landslide area to be detected.

[0170] Furthermore, the logical instructions in the aforementioned memory 930 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0171] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0172] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0173] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A semi-supervised remote sensing image historical landslide detection method based on three-dimensional joint perturbation, characterized in that, include: Remote sensing images of the landslide area to be detected are collected, and the remote sensing images of the landslide area to be detected are divided into a training set, a validation set and a test set. The training set includes labeled image data and unlabeled image data, the validation set includes labeled image data, and the test set includes unlabeled image data. Based on the training set, preset weak data augmentation perturbations and preset strong data augmentation perturbations are applied, and then input into the student model for training. After feature perturbation, the data is input into the main decoder and auxiliary decoder for training, and then input into the teacher model for training to achieve model perturbation. A joint weighted loss function is obtained, and the model training is completed using the validation set to obtain the historical landslide detection model, including: For the labeled image data in the training set The preset weak data enhancement perturbation is performed to obtain ,Will Input student model Landslide detection results were obtained from labeled image data. ,use and corresponding pixel-level manual markings Constructing a supervised loss function ; For the unlabeled image data in the training set The weak data augmentation perturbation is performed to obtain ,Will Input student model We obtained weakly enhanced landslide detection results from unlabeled image data. ,Will Two randomized, preset strong data augmentation perturbations were performed to obtain... and ,Will and Input student model Obtain the results of markerless imagery strong data enhancement for disturbed landslide detection. and ,by Construct an input perturbation loss function for the target and ; Will Input student model The encoder output is obtained. , respectively Apply Drop and Cut perturbations to obtain and ,Will Input student model The decoder gets ,Will Input student model The auxiliary decoder is obtained ,by Construct a feature perturbation loss function for the target and ; Will Input student model The encoder output is obtained. ,right Apply Drop perturbation to obtain ,Will Input student model The decoder gets ,by Construct a feature perturbation loss function for the target ,include: Determine the sampling threshold ; right New features are obtained by summing and normalizing the channel dimensions. ; generate Drop mask ; Generate encoder output after Drop perturbation : Will Input student model decoder The final output of the Drop perturbation is obtained. ; by As the objective, a feature perturbation loss function is established. : ； Will Input student model The encoder output is obtained. ,right Applying the Cut perturbation yields ,Will Input student model The auxiliary decoder is obtained ,by Construct a feature perturbation loss function for the target ,include: right Perform the Cutout operation to obtain ; Will Input student model Auxiliary decoder The final output of the Cutout perturbation is obtained. , Upsample PixelShuffle, with the number of upsampling operations matching the number of downsampling operations performed by the decoder; by As the objective, a feature perturbation loss function is established. : in, Mean square error; exist Add Gaussian noise perturbation Then input the teacher model Obtain the model perturbation results ,by Construct a model perturbation loss function for the target ; For the supervision loss function The input perturbation loss function and The characteristic perturbation loss function and and the model perturbation loss function The joint weighted loss function is constructed by performing weighted summation, and the model parameter updates are constrained to obtain the historical landslide detection model. The target set is input into the historical landslide detection model, and the final historical landslide detection results for the landslide area to be detected are output.

2. The method for detecting historical landslides based on three-dimensional joint perturbation in semi-supervised remote sensing images according to claim 1, characterized in that, Remote sensing images of the landslide area to be detected are acquired, and these images are divided into a training set, a validation set, and a test set, including: The labeling results are obtained by manually labeling some landslide samples at the pixel level in the remote sensing image of the landslide area to be detected. The remote sensing image of the landslide area to be detected and the labeled results are segmented using a fixed-size sliding window to obtain labeled image data. and unlabeled image data ,in, The number of labeled image data. For training set The total number of image data in China, and , Represents a single landslide image block. These represent the length and width of a single landslide image block, respectively. This represents the number of bands that an image block possesses. Represents pixel-level manual markings, with 0 indicating the background and 1 indicating the landslide area.

3. The method for detecting historical landslides based on three-dimensional joint perturbation in semi-supervised remote sensing images according to claim 1, characterized in that, For the labeled image data in the training set The preset weak data enhancement perturbation is performed to obtain ,Will Input student model Landslide detection results were obtained from labeled image data. ,use and corresponding pixel-level manual markings Constructing a supervised loss function ,include: right Perform random vertical flipping and horizontal mirror flipping. Rotation, random cropping, and transpose operations yield the following results: ,Will enter get ; Constructing a supervised loss function by combining binary cross-entropy loss and Dice loss : in, The binary cross-entropy loss function is... This is the Dice loss function, used to balance the imbalance between the number of landslide and background pixels in an image patch.

4. The method for detecting historical landslides based on three-dimensional joint perturbation in semi-supervised remote sensing images according to claim 1, characterized in that, For the unlabeled image data in the training set The weak data augmentation perturbation is performed to obtain ,Will Input student model We obtained weakly enhanced landslide detection results from unlabeled image data. ,Will Two randomized, preset strong data augmentation perturbations were performed to obtain... and ,Will and Input student model Obtain the results of markerless imagery strong data enhancement for disturbed landslide detection. and ,by Construct an input perturbation loss function for the target and ,include: right Perform random vertical flipping and horizontal mirror flipping. Rotation, random cropping, and transpose operations yield the following results: ,Will enter get ; right Perform two random Gaussian blur, color dithering, and CutMix operations. The color dithering includes random adjustments to brightness, contrast, saturation, and hue. The color dithering is executed randomly, while the Gaussian blur and CutMix are executed sequentially. The input perturbation loss function is constructed using Dice loss. and : in, To filter out strong data and enhance the confidence threshold for noise in landslide detection results, This is the Dice loss function.

5. The method for detecting historical landslides based on three-dimensional joint perturbation in semi-supervised remote sensing images according to claim 1, characterized in that, exist Add Gaussian noise perturbation Then input the teacher model Obtain the model perturbation results ,by Construct a model perturbation loss function for the target ,include: Determine the mean and standard deviation of the Gaussian noise perturbation; Determine the teacher model With the student model All are U-Net models, the teacher model The student model is initialized using a layer-by-layer traversal method. The teacher model uses the Kaiming normal distribution for parameter initialization. The parameters are obtained by analyzing the student model. The parameters are implemented using an exponential moving average. by Construct a model perturbation loss function for the target : 。 6. The method for detecting historical landslides based on three-dimensional joint perturbation in semi-supervised remote sensing images according to claim 1, characterized in that, For the supervision loss function The input perturbation loss function and The characteristic perturbation loss function and and the model perturbation loss function Perform weighted summation to construct the joint weighted loss function, including: in, Denotes the joint weighted loss function. , , , and They are respectively , , , and The corresponding weights.

7. A semi-supervised remote sensing image historical landslide detection system based on three-dimensional joint perturbation, based on the semi-supervised remote sensing image historical landslide detection method based on three-dimensional joint perturbation as described in any one of claims 1 to 6, characterized in that, include: The data processing module is used to acquire remote sensing images of the landslide area to be detected, and divide the remote sensing images of the landslide area to be detected into a training set, a validation set and a test set. The training set includes labeled image data and unlabeled image data, the validation set includes labeled image data, and the test set includes unlabeled image data. The training module is used to perform preset weak data augmentation perturbation and preset strong data augmentation perturbation based on the training set, input the perturbation into the student model for training, input the perturbation into the main decoder and auxiliary decoder for training, input the perturbation into the teacher model for training to realize model perturbation, obtain the joint weighted loss function, and complete the model training by combining the validation set to obtain the historical landslide detection model. The detection module is used to input the set to be detected into the historical landslide detection model and output the final historical landslide detection results for the landslide area to be detected.

Images