Landslide detection method of joint spectrum, digital elevation model double branch network

Abstract

The application provides a kind of high-resolution image landslide detection method of combined spectrum, digital elevation model and auxiliary feature dual-branch network, including obtaining the high-resolution remote sensing image to be detected, digital elevation model and landslide true label;Extract auxiliary features, mine multiple characteristic factors from them and merge all images in channel;The image after merging is preprocessed, and landslide sample is generated;Different data enhancement means are used to increase the number and diversity of landslide data set;Residual and channel attention are added to the network model based on U-Net+++ to improve the network model;The image sample pair is respectively input into the improved U-Net+++ branch network, and the prediction result is obtained;The consistency of sample pair is mined and the model is optimized using the loss function;Based on the test sample, the model accuracy is evaluated and the landslide detection result map is output.The application provides a novel and effective landslide detection method, which can fully mine the invariance features of high-resolution landslide images and effectively improve the recognition accuracy of landslide.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a high-resolution image landslide detection method using a dual-branch network that combines spectral analysis, digital elevation model, and auxiliary features. Background Technology

[0002] High-resolution satellite remote sensing technology is characterized by its "macroscopic, rapid, and accurate" nature. With the increasing availability of high spatial and temporal resolution sensors, it records high-precision geometric structures and topographic information within a region, providing crucial support for geological disaster monitoring and investigation. However, unlike surface environments, landslide data is irregular, multi-scale, and obscured by vegetation, making direct observation and interpretation difficult, resulting in low accuracy in landslide logging and mapping. Furthermore, the complex background of landslide-affected areas reduces the generalization performance of models.

[0003] Disaster identification and extraction have become routine methods for landslide emergency investigation and disaster assessment. Related research can be summarized into two categories. The first category is based on manual visual interpretation, which relies on the differences in image tone, texture, shape, shadow, and pattern between the landslide and its background environment. Landslide disaster identification and judgment are achieved through the expert experience of interpreters, including direct interpretation, comparative analysis, information compositing, and comprehensive inference methods. This type of method is time-consuming and labor-intensive, and difficult to meet the needs of large-scale disasters. The second category is computer interpretation, which uses the spectral, texture, and shape characteristics of landslide disasters in remote sensing images, employing pattern recognition methods such as statistical decision-making, probability estimation, machine learning, and deep learning for semi-automatic or automatic identification. The degree of automation and detection accuracy of this type of method mainly depends on the selection of segmentation rules and the setting of model parameters. Therefore, there is an urgent need to explore automatic and efficient landslide identification technologies. Summary of the Invention

[0004] This invention provides an improved high-resolution image landslide detection method that utilizes a dual-branch network combining spectral analysis, digital elevation model (DEM), and auxiliary features to avoid information redundancy. The method includes the following steps:

[0005] S1: Obtain high-resolution remote sensing images, digital elevation models, and corresponding rasterized landslide real labels for the area to be detected;

[0006] S2: Extract auxiliary features, including calculating vegetation indices from high-resolution remote sensing images, extracting topographic features from digital elevation models, and merging all auxiliary features with high-resolution remote sensing images.

[0007] S3: Normalize the merged image and the landslide real label separately to obtain the normalized merged image and the normalized landslide real label; perform deviation standardization on the normalized merged image, and then crop it based on a certain degree of overlap to preserve the contextual information of the image boundary.

[0008] S4: Two data augmentation strategies are used to process the cropped images to increase the data and diversity of the landslide dataset, and finally, multiple training image sample pairs of two different qualities are obtained.

[0009] S5: The residual network and channel attention module are embedded in the original U-Net+++ to further improve the performance of the segmentation network. A robust bi-branch landslide detection model is built based on the modified U-Net+++.

[0010] S6: Based on the obtained training image sample pairs, define the landslide area of ​​the training image sample pair as a positive region pair, input the training image sample pairs into the bi-branch landslide detection model, and obtain the landslide prediction results of each training image sample pair;

[0011] S7: Based on the positive region pairs and landslide prediction results, a contrast loss is designed to constrain the consistency of landslide region pairs, and the cross-entropy and similarity coefficient loss are combined to optimize the bi-branch landslide detection model. After training, a well-trained robust model is obtained.

[0012] S8: Test the trained robust model using test samples, evaluate the detection accuracy, and output the landslide detection result map.

[0013] Following the above technical solution, step S2, the step of calculating the vegetation index from the high-resolution remote sensing image, includes:

[0014] The vegetation index NDVI is calculated using spectral bands. The formula is as follows:

[0015]

[0016] In the formula: NIR and R represent the reflectance of the near-infrared band and red band of the high-resolution remote sensing image to be detected, respectively.

[0017] Following the above technical solution, step S2, the step of extracting terrain features from the digital elevation model, includes:

[0018] Multiple terrain features are extracted using a digital elevation model of the area to be detected;

[0019] The extracted topographic features include: aspect, slope, hill shadow, curvature, planar curvature, profile curvature, depression-filling digital elevation model, flow direction, and accumulation, as detailed in Table 1 below:

[0020] Table 1 Description of Topographic Features

[0021]

[0022]

[0023] Following the above technical solution, step S4 specifically includes:

[0024] S4.1: Based on the cropped data, data augmentation is performed using methods such as horizontal flipping, vertical flipping, and diagonal mirroring without changing the pixel values;

[0025] S4.2: For the augmented image, use random noise addition, image dithering, and random erasure methods to change the image pixel values ​​and reduce image quality;

[0026] S4.3 defines images generated by different enhancement strategies as training image sample pairs, which have different quality image data and the same landslide true label.

[0027] Following the above technical solution, step S5, constructing a robust landslide detection model, specifically includes the following steps:

[0028] S5.1 Modify the original U-Net+++ network structure and replace the original convolution extraction module with a residual module;

[0029] S5.2 adds a channel attention module after the residual module, enabling the model to automatically focus on effective features;

[0030] S5.3 utilizes a modified U-Net+++ model to construct a dual-branch landslide detection model, which integrates multi-scale feature coupling, residual modules, and channel attention modules.

[0031] Following the above technical solution, in step S6, the landslide region of the obtained training image sample pair is defined as a positive region pair. The training image sample pairs are then input into their respective corresponding bi-branch landslide detection models to obtain the landslide prediction results of the training image samples.

[0032] Following the above technical solution, step S7, the optimization steps of the dual-branch landslide detection model include:

[0033] S7.1 uses the binary cross-entropy loss function to optimize the network model of each branch, and its calculation formula is as follows:

[0034]

[0035] In the formula, N is the number of samples, Yi represents the true label, and Pi represents the probability that the predicted sample belongs to the landslide.

[0036] S7.2 addresses the data imbalance problem in landslide detection by employing similarity coefficient loss for further optimization. The similarity coefficient loss function is defined as follows:

[0037]

[0038] In the formula, p(x) and g(x) represent the true label and predicted probability of the landslide, respectively;

[0039] S7.3 To maintain consistency in the predicted landslide area pairs across different branches, a comparative similarity coefficient loss is designed, the expression of which is:

[0040]

[0041] In the formula, b(x) and g(x) represent the prediction results from the two branches, respectively;

[0042] S7.4 combines the above three loss methods to optimize the model, and the total loss is shown below:

[0043] L Total =α·L BCE +β·L DSC +γ·L CDSC

[0044] In the formula, α, β and γ represent the weighting coefficients of the three types of losses, respectively.

[0045] This invention also provides a high-resolution image landslide detection system that combines spectral analysis, digital elevation model, and auxiliary features using a dual-branch network, characterized by comprising:

[0046] The data acquisition module is used to acquire high-resolution remote sensing images, digital elevation models, and corresponding rasterized landslide real labels for the area to be detected.

[0047] The auxiliary feature extraction module is used to extract auxiliary features, including calculating vegetation indices from high-resolution remote sensing images, extracting terrain features from digital elevation models, and merging auxiliary features.

[0048] The preprocessing module is used to normalize the merged image and the landslide real labels respectively to obtain the normalized merged image and the normalized landslide real labels; the normalized merged image is subjected to deviation standardization, and then cropped based on a certain degree of overlap to preserve the contextual information of the image boundaries.

[0049] The data augmentation module is used to process the cropped images using two data augmentation strategies to increase the data and diversity of the landslide dataset, ultimately obtaining multiple training image sample pairs of two different qualities.

[0050] A dual-branch landslide detection model building module is used to embed residual networks and channel attention modules into the original U-Net+++ to further improve the performance of the segmentation network, and to build a robust dual-branch landslide detection model based on the modified U-Net+++.

[0051] The training module is used to define the landslide area of ​​the training image sample pair as a positive region pair based on the obtained training image sample pairs, input the training image sample pairs into the bi-branch landslide detection model, and obtain the landslide prediction results of each training image sample pair.

[0052] The optimization module is used to design a contrast loss constraint on the consistency of landslide region pairs based on positive region pairs and landslide prediction results, and to optimize the bi-branch landslide detection model by combining cross-entropy and similarity coefficient loss. After training, a trained robust model is obtained.

[0053] The testing module is used to test the trained robust model with test samples, evaluate the detection accuracy, and output landslide detection result maps.

[0054] Following the above technical solution, the terrain features extracted from the digital elevation model include: aspect, slope, mountain shadow, curvature, plane curvature, profile curvature, depression-filling digital elevation model, flow direction, and accumulation.

[0055] The present invention also provides a storage medium storing a computer program executable by a processor, the computer program executing the high-resolution image landslide detection method of the dual-branch network of combined spectral, digital elevation model and auxiliary features described in the above technical solution.

[0056] The technical solution provided by this invention has the following beneficial effects: This invention can improve the landslide detection accuracy of semantic segmentation models, fully explore the rich spectral and auxiliary feature information in high-resolution images and digital elevation models, and provide a reliable information source for high-resolution remote sensing technology. First, data is supplemented by extracting vegetation index NDVI and terrain-related factors. Then, different data augmentation strategies are used to enhance the quantity and diversity of landslide data. Next, the original U-Net++ data is improved by adding residual and attention modules to enhance the model's feature extraction capability, and a dual-branch network model is constructed based on this. Finally, the model is optimized using cross-entropy, similarity coefficient loss, and contrastive similarity coefficient loss. These methods and strategies enable the model to deeply explore the robust features of landslides, thus giving the model higher generalization ability. The model's prediction results can maintain regional homogeneity to a greater extent, improving the model's recognition accuracy, giving this invention the advantages of finely extracting landslide boundaries and improving recognition accuracy. Attached Figure Description

[0057] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0058] Figure 1 This is a flowchart of a high-resolution image landslide detection method based on a dual-branch network combining spectral data, digital elevation model, and auxiliary features, according to an embodiment of the present invention.

[0059] Figure 2 This is a network structure diagram of the high-resolution image landslide detection method with a dual-branch network in an embodiment of the present invention;

[0060] Figure 3 These are the results of landslide detection methods using high-resolution remote sensing images obtained through different methods in embodiments of the present invention. Figure 3 (a) Landslide detection map obtained by the U-Net network method; Figure 3 (b) is the landslide detection map obtained by the depth U-Net+++ method; Figure 3 (c) is a landslide detection map obtained by the method of the present invention; Figure 3 (d) is Figure 3 (a)-3(c) Explanation of directions and markings;

[0061] Figure 4 This is a schematic diagram of the high-resolution image landslide detection system based on a dual-branch network that combines spectral analysis, digital elevation model, and auxiliary features, according to an embodiment of the present invention. Detailed Implementation

[0062] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0063] Please refer to Figure 1 The present invention provides a high-resolution image landslide detection method using a dual-branch network that combines spectral analysis, digital elevation model, and auxiliary features, comprising the following steps:

[0064] S1: Obtain the high-resolution remote sensing image to be detected and the corresponding rasterized landslide real label;

[0065] In this embodiment, a high-resolution post-disaster remote sensing image acquired by the RapidEye optical sensor to be segmented and a rasterized set of artificially marked landslide samples corresponding to the image are input.

[0066] S2: Extract the spectral features of the remote sensing image to be detected and calculate the vegetation index; extract the terrain feature related factors from the digital elevation model; combine the vegetation index and terrain features into auxiliary features, and merge the auxiliary features with the high-resolution remote sensing image;

[0067] Step S2, the steps for calculating the vegetation index, include:

[0068] The vegetation index NDVI is calculated using spectral bands. The formula is as follows:

[0069]

[0070] In the formula: NIR and R represent the reflectance of the near-infrared band and red band of the high-resolution remote sensing image to be detected, respectively.

[0071] Step S2, the step of extracting terrain features includes:

[0072] Multiple terrain features are extracted using the high-resolution remote digital elevation model to be detected, and the optical features, digital elevation model, vegetation index and terrain features are merged;

[0073] The extracted terrain features include: aspect, slope, hill shadow, curvature, planar curvature, profile curvature, depression digital elevation model, flow direction, and accumulation, described as follows:

[0074] Topographic features describe Slope Slope direction derived from DEM slope Steepness of surface features derived from DEM Mountain shadow Shadow relief under the influence of light source angle and shadow curvature Curvature of the grid surface Plane curvature Curvature of the grid surface perpendicular to the slope direction Section curvature Curvature of the grid surface along the slope direction Cut and fill digital elevation model Fill the recessed areas of the original digital elevation model with a grid Flow direction Flow direction grid from each grid to its adjacent point on the steepest downslope Cumulative amount Cumulative flow of each grid

[0075] S3: Preprocess the merged image and the rasterized landslide real labels;

[0076] Specifically, different preprocessing methods are used, including;

[0077] S3.1 Normalization is used to scale the range values ​​of each channel of the merged image to [0,1]. Normalization is used to process the landslide real label, setting the landslide area label value to 1 and the non-landslide area label value to 0.

[0078] S3.2 Perform deviation standardization on the normalized image and adjust the data;

[0079] S3.3 The image data with the deviation standardization and the normalized landslide real label are cropped based on a certain degree of overlap to obtain the cropped image and landslide real label data;

[0080] S4: The cropped image data is enhanced using two different image enhancement strategies to expand the amount of landslide data and increase the diversity of the data. The two images obtained by the two image enhancement strategies are called image sample pairs.

[0081] Specifically, different data augmentation strategies include:

[0082] S4.1: Use horizontal flipping, vertical flipping, and diagonal mirroring to expand the amount of landslide data. This type of method does not change the image pixel values, but only changes the pixel positions.

[0083] S4.2: Data augmentation is performed on the cropped image by randomly adding noise, image jittering, and random erasing. This type of method not only changes the image pixel values ​​but also reduces the image quality.

[0084] S4.3: Define the images generated by the two strategies as image sample pairs.

[0085] S5: Landslide Detection Model Construction;

[0086] Specifically, the network is improved based on the original U-Net+++, and a two-branch model is established accordingly, including:

[0087] S5.1 improves the original U-Net+++ network model by replacing the original feature extraction module with a residual module and adding a channel attention module after the residual module to automatically focus on effective features.

[0088] S5.2 uses the improved U-Net+++ to construct a robust bi-branch landslide detection model. The entire bi-branch landslide detection model framework is as follows: Figure 2 As shown;

[0089] S6: Define the landslide region of the image sample pair as a positive region pair, and input the image sample pair into the branch network of the dual-branch landslide detection model respectively to obtain the landslide prediction result of the image pair;

[0090] S7: Based on the positive region pair and the prediction results of the bi-branch landslide, design a contrast loss to constrain the consistency of the landslide region pair, and combine cross-entropy and similarity coefficient loss to optimize the model. After training, a trained robust model is obtained.

[0091] In step S7, the optimization steps for the dual-branch landslide detection model include:

[0092] S7.1 uses the binary cross-entropy loss function to optimize the network model of each branch, and its calculation formula is as follows:

[0093]

[0094] In the formula, N is the number of samples, Y represents the true label, and P represents the probability that the predicted sample belongs to a landslide.

[0095] S7.2 addresses the data imbalance problem in landslide detection by employing similarity coefficient loss for further optimization. The similarity coefficient loss function is defined as follows:

[0096]

[0097] In the formula, p(x) and g(x) represent the true label and predicted probability of the landslide, respectively;

[0098] S7.3 To maintain consistency in the predicted landslide area pairs across different branches, a comparative similarity coefficient loss is designed, the expression of which is:

[0099]

[0100] In the formula, b(x) and d(x) represent the prediction results from the two branches, respectively;

[0101] S7.4 combines the above three loss methods to optimize the model, and the total loss is:

[0102] L Total =α·L BCE +β·L DSC +γ·L CDSC

[0103] In the formula, α, β and γ represent the weighting coefficients of the three types of loss, which can be set to 0.5, 1 and 1 respectively based on experience.

[0104] S8: Use the test samples to test the trained robust model, evaluate the detection accuracy, and output the landslide detection result map.

[0105] The effects of the present invention will be further explained below with reference to simulation experiment examples.

[0106] (1) Simulation experimental conditions:

[0107] The hardware testing platform for this experiment was an Intel Core i7 processor deep learning machine with 64GB of memory, and the software platform was a Linux operating system and Python 3.6.8. The computation was performed using two NVIDIA GeForce RTX 2080Ti GPUs (12GB RAM) in the NVIDIA CUDA Toolkit 10.1 environment. The input image for this invention was a RapidEye high-resolution remote sensing image, acquired on August 13, 2011. The study area is located in the mountainous region of Rio de Janeiro, Brazil. The image has a spatial resolution of 5 meters, covering five bands from the visible to near-infrared spectrum, and a size of 5000×5000 pixels. The digital elevation model was derived from the ALOS phased array l-band synthetic aperture radar (PALSAR) sensor, with a spatial resolution of 12.5. For ease of analysis, it was resampled to 5 meters using bilinear interpolation to match the spatial resolution of the optical image.

[0108] (2) Simulation content:

[0109] In this embodiment, the high-resolution image landslide detection method provided by the present invention, which combines spectral density, digital elevation model, and auxiliary features using a dual-branch network, is compared with two traditional detection methods: the U-Net method and the deep U-Net+++ method, as follows:

[0110] The image semantic segmentation method proposed by O. Ronneberger et al. in "U-net: Convolutional networks for biomedical imagesegmentation, 2015, pp. 234-241" is referred to as the U-Net method.

[0111] H. Huang et al. proposed a multi-scale feature map aggregation framework in “Unet++: A nested u-net architecture for medical imagesegmentation. Unet 3+: A full-scale connected unet for medical imagesegmentation, 2018, pp. 1055-1059”, which has higher segmentation accuracy and boundary refinement capability, and is referred to as the U-Net+++ method.

[0112] During the experiment, a confusion matrix was constructed based on real ground reference data for the detection results obtained by different detection methods. The performance of the method of the present invention was quantitatively evaluated by calculating precision, recall, F1 score and mean intersection-union ratio (mIoU).

[0113] Precision is the ratio of the number of samples predicted as positive to the number of samples that are actually positive. Its expression is:

[0114]

[0115] The range of P is [0, 1], and the larger the P value, the better the model's performance. Besides precision, recall is also an important metric for evaluating model performance. Recall represents the proportion of samples predicted as positive out of the total number of actual landslides; the higher the value, the better the model.

[0116]

[0117] The F1 score (F1) is the harmonic mean of precision and recall, and the formula is as follows.

[0118] The larger the F1 value, the better the segmentation effect.

[0119]

[0120] The mean intersection-union ratio (mIoU) represents the average ratio of the intersection to the union of all categories. Its calculation formula is as follows:

[0121]

[0122] Among them, T P It is the number of cases that are correctly classified as positive examples; T N F is the number of instances that are correctly classified as negative. P F is the number of instances incorrectly classified as positive. N It represents the number of instances incorrectly classified as negative; m+1 represents the number of categories, P ij P represents the number of instances where the actual value is i but the predicted value is j. ii P represents the number of both actual and predicted values ​​of i. ji This represents the number of instances where the actual value is j but the prediction is i.

[0123] Please refer to Figure 3 , Figure 3 These are the results of landslide detection methods using high-resolution remote sensing images obtained through different methods in embodiments of the present invention. Figure 3 (a) is a landslide detection map obtained by U-Net; Figure 3 (b) is the landslide detection map obtained by U-Net++; Figure 3 (c) is a landslide detection map obtained by the method of the present invention.

[0124] (3) Analysis of experimental results

[0125] See Table 1. Table 1 evaluates the objective indicators... Figure 3 The detection results of each method were evaluated.

[0126] Table 1 Evaluation results of detection accuracy of different methods

[0127]

[0128] Combined Table 1 and Figure 3 It can be seen that the U-Net method in Figure 3 While U-Net+++ achieves high accuracy in identifying landslides, it suffers from significant omissions, failing to detect a large number of landslides. Although it can identify more landslides, it struggles to completely eliminate misclassifications of homogeneous regions, still exhibiting noticeable omissions. Figure 3(a)-(b). The method of this invention is superior to the previous two detection methods in terms of visual effects and quantitative analysis. It can significantly improve the ability to identify landslides, extract a large number of landslide areas that were not detected by U-Net and U-Net+++, and its landslide segmentation boundary positioning is clear and accurate. Therefore, the method of this invention, which combines spectral analysis, digital elevation model, and auxiliary features in a dual-branch network, has the effect of improving the accuracy of landslide detection.

[0129] To implement the detection method described in the above embodiments, this invention provides a high-resolution image landslide detection system that combines spectral analysis, digital elevation models, and auxiliary features using a dual-branch network. Figure 4 As shown, it includes:

[0130] The data acquisition module is used to acquire high-resolution remote sensing images, digital elevation models, and corresponding rasterized landslide real labels for the area to be detected.

[0131] The auxiliary feature extraction module is used to extract auxiliary features, including calculating vegetation index from high-resolution remote sensing images, extracting terrain features from digital elevation models, and merging all auxiliary features with high-resolution remote sensing images to obtain a merged image.

[0132] The preprocessing module is used to normalize the merged image and the landslide real label respectively, perform deviation standardization on the normalized merged image, and then crop it based on a certain degree of overlap to preserve the contextual information of the image boundary.

[0133] The data augmentation module is used to process the cropped images using two data augmentation strategies to increase the data and diversity of the landslide dataset, ultimately obtaining multiple training image sample pairs of two different qualities.

[0134] A dual-branch landslide detection model building module is used to embed residual networks and channel attention modules into the original U-Net+++ to further improve the performance of the segmentation network, and to build a robust dual-branch landslide detection model based on the modified U-Net+++.

[0135] The training module is used to define the landslide area of ​​the training image sample pair as a positive region pair based on the obtained training image sample pairs, input the training image sample pairs into the bi-branch landslide detection model, and obtain the landslide prediction results of each training image sample pair.

[0136] The optimization module is used to design a contrast loss constraint on the consistency of landslide region pairs based on positive region pairs and landslide prediction results, and to optimize the bi-branch landslide detection model by combining cross-entropy and similarity coefficient loss. After training, a trained robust model is obtained.

[0137] The testing module is used to test the trained robust model with test samples, evaluate the detection accuracy, and output landslide detection result maps.

[0138] The terrain features extracted from the digital elevation model include: aspect, slope, mountain shadow, curvature, plane curvature, profile curvature, depression filling digital elevation model, flow direction, and accumulation.

[0139] The various modules in the system are mainly used to implement the specific methods of the above embodiments, which will not be elaborated here.

[0140] This application also provides a computer-readable storage medium, such as flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, server, App store, etc., which stores a computer program. When the program is executed by a processor, it implements the corresponding function. The computer-readable storage medium of this embodiment is used to implement high-resolution image landslide detection of the dual-branch network of joint spectrum, digital elevation model, and auxiliary features of the method embodiment when executed by a processor.

[0141] In summary, the method of this invention can make full use of the spectral features, digital elevation models and auxiliary features of high-resolution images to mine the semantic robustness features of deep landslides, thereby effectively improving the accuracy and ability of landslide identification. It can solve the problems of insufficient utilization of high-resolution image feature information, low landslide detection accuracy and poor generalization ability in existing methods, and is a very practical and effective high-resolution image landslide detection method.

[0142] 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 system 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 system. 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 system that includes that element.

[0143] 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. In the unit claims listing several devices, several of these devices may be embodied by the same hardware item. The use of the terms first, second, and third, etc., does not indicate any order and can be interpreted as identifiers.

[0144] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A high-resolution image landslide detection method using a dual-branch network combining spectral analysis, digital elevation model, and auxiliary features, characterized in that... Includes the following steps: S1: Obtain high-resolution remote sensing images, digital elevation models, and corresponding rasterized landslide real labels for the area to be detected; S2: Extract auxiliary features, including calculating vegetation indices from high-resolution remote sensing images, extracting topographic features from digital elevation models, and merging all auxiliary features with high-resolution remote sensing images. S3: Normalize the merged image and the landslide real label respectively; The normalized merged image is standardized by deviation, and then cropped based on a certain degree of overlap to preserve the contextual information of the image boundary; S4: Two data augmentation strategies are used to process the cropped images to increase the data and diversity of the landslide dataset, ultimately resulting in multiple training image sample pairs of two different qualities; S5: The residual network and channel attention module are embedded in the original U-Net+++ to further improve the performance of the segmentation network. A robust bi-branch landslide detection model is built based on the modified U-Net+++. S6: Based on the obtained training image sample pairs, define the landslide area of ​​the training image sample pair as a positive region pair, input the training image sample pairs into the dual-branch landslide detection model, and obtain the landslide prediction results of each training image sample pair; S7: Based on the positive region pairs and landslide prediction results, a contrast loss is designed to constrain the consistency of landslide region pairs, and a bi-branch landslide detection model is optimized by combining cross-entropy and similarity coefficient loss. After training, a well-trained robust model is obtained. S8: Test the trained robust model using test samples, evaluate the detection accuracy, and output the landslide detection result map.

2. The high-resolution image landslide detection method using a dual-branch network combining spectral analysis, digital elevation model, and auxiliary features as described in claim 1, is characterized in that... Step S2, the step of calculating vegetation indices from high-resolution remote sensing imagery, includes: Calculating vegetation index using spectral bands NDVI The calculation formula is: In the formula: NIR and R These represent the reflectance of the near-infrared band and the red band of the high-resolution remote sensing image to be detected, respectively.

3. The high-resolution image landslide detection method using a dual-branch network combining spectral analysis, digital elevation model, and auxiliary features as described in claim 1, is characterized in that... Step S2, the steps for extracting terrain features from the digital elevation model, include: Multiple terrain features are extracted using a digital elevation model of the area to be detected; The extracted terrain features include: aspect, slope, hill shadow, curvature, planar curvature, profile curvature, depression-filled digital elevation model, flow direction, and accumulation. Among them, aspect is the slope direction derived from the DEM, slope is the steepness of the surface feature derived from the DEM, hill shadow is the shadow relief under the influence of light source angle and shadow, curvature is the curvature of the raster surface, planar curvature is the curvature of the raster surface perpendicular to the slope direction, profile curvature is the curvature of the raster surface along the slope direction, depression-filled digital elevation model is the raster after filling the depression area of ​​the original digital elevation model, flow direction is the flow direction raster from each raster to its adjacent point on the steepest slope, and accumulation is the cumulative flow of each raster.

4. The high-resolution image landslide detection method using a dual-branch network combining spectral analysis, digital elevation model, and auxiliary features as described in claim 1, is characterized in that... Step S4 specifically includes: S4.1: Based on the cropped data, data augmentation is performed using methods such as horizontal flipping, vertical flipping, and diagonal mirroring without changing the pixel values; S4.2: For images that have been augmented with data, random noise addition, image dithering, or random erasure methods are used to change the image pixel values ​​and reduce image quality; S4.3 defines images generated by different enhancement strategies as training image sample pairs, which have different quality image data and the same landslide true label.

5. The high-resolution image landslide detection method using a dual-branch network combining spectral analysis, digital elevation model, and auxiliary features as described in claim 1, is characterized in that... Step S5, constructing a robust dual-branch landslide detection model, specifically includes the following steps: S5.1 Modify the original U-Net+++ network structure and replace the original convolution extraction module with a residual module; S5.2 adds a channel attention module after the residual module, enabling the model to automatically focus on effective features; S5.3 utilizes a modified U-Net+++ model to construct a dual-branch landslide detection model, which integrates multi-scale feature coupling, residual modules, and channel attention modules.

6. The high-resolution image landslide detection method using a dual-branch network combining spectral analysis, digital elevation model, and auxiliary features as described in claim 1, is characterized in that... In step S7, the optimization steps for the dual-branch landslide detection model include: S7.1 uses the binary cross-entropy loss function to optimize the network model of each branch, and its calculation formula is as follows: In the formula, N is the number of samples. Indicates the true label, This indicates the probability that the predicted sample belongs to a landslide. S7.2 addresses the data imbalance problem in landslide detection by employing similarity coefficient loss for further optimization. The similarity coefficient loss function is defined as follows: In the formula, p(x) and g(x) These represent the true label and predicted probability of the landslide, respectively. S7.3 To maintain consistency in the predicted landslide area pairs across different branches, a comparative similarity coefficient loss is designed, the expression of which is: In the formula, b(x) and d(x) These represent the prediction results from the two branches respectively; S7.4 combines the above three loss methods to optimize the model, and the total loss is: In the formula, α, β and γ represent the weighting coefficients of the three types of losses, respectively.

7. A high-resolution image landslide detection system using a dual-branch network combining spectral analysis, digital elevation model, and auxiliary features, characterized in that... include: The data acquisition module is used to acquire high-resolution remote sensing images, digital elevation models, and corresponding rasterized landslide real labels for the area to be detected. The auxiliary feature extraction module is used to extract auxiliary features, including calculating vegetation indices from high-resolution remote sensing images, extracting terrain features from digital elevation models, and merging all auxiliary features with high-resolution remote sensing images. The preprocessing module is used to normalize the merged image and the landslide real label respectively, perform deviation standardization on the normalized merged image, and then crop it based on a certain degree of overlap to preserve the contextual information of the image boundary. The data augmentation module is used to process the cropped images using two data augmentation strategies to increase the data and diversity of the landslide dataset, ultimately obtaining multiple training image sample pairs of two different qualities. A dual-branch landslide detection model building module is used to embed residual networks and channel attention modules into the original U-Net+++ to further improve the performance of the segmentation network, and to build a robust dual-branch landslide detection model based on the modified U-Net+++. The training module is used to define the landslide area of ​​the training image sample pair as a positive region pair based on the obtained training image sample pairs, input the training image sample pairs into the bi-branch landslide detection model, and obtain the landslide prediction results of each training image sample pair. The optimization module is used to design a contrast loss constraint on the consistency of landslide region pairs based on positive region pairs and landslide prediction results, and to optimize the bi-branch landslide detection model by combining cross-entropy and similarity coefficient loss. After training, a trained robust model is obtained. The testing module is used to test the trained robust model with test samples, evaluate the detection accuracy, and output landslide detection result maps.

8. The high-resolution image landslide detection system with a dual-branch network combining spectral analysis, digital elevation model, and auxiliary features as described in claim 7, is characterized in that... Topographic features extracted from digital elevation models include: aspect, slope, mountain shadow, curvature, planar curvature, profile curvature, depression digital elevation model, flow direction, and accumulation.

9. A storage medium, characterized in that, It contains a computer program that can be executed by a processor, which performs the high-resolution image landslide detection method based on a dual-branch network of combined spectral, digital elevation model and auxiliary features as described in any one of claims 1-6.

Images