A landslide extraction method based on an improved U-Net network with asymmetric convolution.

Abstract

This invention discloses a landslide extraction method based on an improved U-Net network using asymmetric convolution. The method includes: multi-source data construction comprising acquiring six-channel data containing DEM data, slope information, and variation characteristics of the target area from high-resolution remote sensing images; and model construction employing a U-Net network structure overall, with the encoder using asymmetric convolution instead of the original standard convolutional layers, and the decoder using dense upsampling. This application first extracts variation characteristics and then inputs them into the model for training. In the prediction stage of the deep learning model training, DEM + slope + variation characteristics are used as constraints for landslides, enabling deep learning to extract deep data connections such as complex relationships between landslides and slope surfaces. It creatively utilizes asymmetric convolution to more accurately extract features in different directions without being affected by rotating targets, improving the network's robustness to rotational distortion.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing and intelligent extraction of remote sensing information, specifically relating to a landslide extraction method based on an improved U-Net network with asymmetric convolution. Background Technology

[0002] As a common and frequent geological disaster, landslides can cause severe damage to buildings, casualties, and property losses, with even more severe damage in hilly and mountainous areas. Statistics show that landslides, collapses, and debris flows account for a quarter of all natural disaster deaths in my country each year. In recent years, natural disasters such as heavy rainfall, volcanic activity, and earthquakes have become more frequent, triggering landslides and causing even more damage. Rapidly obtaining landslide information after a disaster is crucial for seizing the "golden 72 hours" for rescue, facilitating rapid organization of disaster relief efforts, and providing effective support for geological disaster prevention and control. Remote sensing Earth observation data is timely, has a wide observation range, and is easily accessible, making it an important data source for landslide detection. In recent years, thanks to advancements in computer vision and graphics processing units (GPUs), convolutional neural networks have achieved better results in almost all computer vision tasks.

[0003] Deep learning models have wide applications in satellite image classification, semantic segmentation, and change detection. Among them, U-Net, as a lightweight and easy-to-deploy model, is widely used in landslide segmentation tasks. However, the U-Net model itself has certain shortcomings, such as the semantic gap between the encoder and decoder. This limits the performance of the U-Net model in landslide detection tasks and restricts the accuracy of interpreting landslides in real-world scenarios, making it unable to meet practical needs.

[0004] The existing technology has the following shortcomings: (1) Landslide distribution has its own characteristics. The distribution of landslides is very irregular. The irregular boundary will be distorted when rotated. When the satellite takes pictures of landslides at different angles, the landslide distribution on the satellite image will have a large difference. This difference is caused by the rotational distortion caused by the different observation angles. In addition, in order to enhance the generalization ability of the neural network during the training process, it is necessary to perform multi-scale scaling, random rotation and random flipping operations on the input image. These operations also bring about the problem of rotational distortion. As a result, the semantic information of the encoder and decoder in the U-Net model is mismatched, which limits the model's ability to extract landslides in complex scenes, resulting in inaccurate landslide extraction and difficulty in boundary capture. (2) Landslides are a type of change, but not all changes are caused by landslides. Landslides occur on slopes, but not all slopes will produce landslides. The existing technology only divides landslide and non-landslide areas according to simple rules, without transforming the data into a high-dimensional space to obtain deeper data connections. Due to the various reasons mentioned above, the existing U-Net network often has low extraction accuracy when extracting landslides and cannot meet the actual needs. Summary of the Invention

[0005] The purpose of this invention is to provide a landslide extraction method based on an improved U-Net network using asymmetric convolution.

[0006] This improved landslide extraction method based on asymmetric convolution addresses the issues of semantic information mismatch and information loss during the upsampling stage in the U-Net model, thereby enhancing the network's ability to extract landslides in complex scenarios. It also solves the problems of inaccurate landslide extraction and difficulty in boundary capture in the U-Net network.

[0007] A landslide extraction method based on an improved U-Net network using asymmetric convolution, comprising:

[0008] The model is constructed using a U-Net network structure, with the encoder part employing asymmetric convolution and the decoder part employing dense upsampling.

[0009] The multi-source data construction consists of six channels of multi-source data, which are composed of RGB post-temporal remote sensing image data, RGB before and after temporal remote sensing image change characteristics, DEM data and slope information.

[0010] The model constructed by training with a six-channel multi-source dataset labeled with landslides is obtained as a trained model capable of extracting landslides from the target area.

[0011] This application makes full use of landslide data in the region, including accurate location, area, and distribution status. It first extracts the change features and then inputs them into the model for training. In the prediction stage of deep learning model training, DEM + slope + change features are used as constraints for landslides, which enables deep learning to extract deep data connections such as complex relationships between landslides and slopes from the data. It creatively uses asymmetric convolution to extract features in different directions more accurately without being affected by rotating targets, thereby improving the network's robustness to rotational distortion.

[0012] In the landslide extraction method based on the improved U-Net network of asymmetric convolution mentioned above, the six-channel multi-source data of the target area is used as input and fed into the trained model to extract landslides in the target area.

[0013] In the landslide extraction method based on the improved U-Net network of asymmetric convolution mentioned above, the change features are obtained based on the RGB pre-temporal remote sensing images and RGB post-temporal remote sensing images of the corresponding region. The RGB post-temporal remote sensing images and RGB pre-temporal remote sensing images are obtained by high-resolution remote sensing images of the corresponding time phases.

[0014] The extraction of the change features is as follows:

[0015] 1) Select temporal data from different periods for the region, define the neighborhood range, subtract the temporal data from each previous and subsequent periods channel by channel, calculate the difference within the neighborhood range of each pixel, and select the value with the largest difference as the variation feature of the region.

[0016] 2) The variation features calculated on each channel are averaged over the channel dimension to generate a single-channel variation feature map.

[0017] Landslide extraction from remote sensing images is essentially image segmentation. To improve segmentation performance, this application extracts landslide-related features from the preceding and following time-phase remote sensing data for model training and prediction. It also incorporates DEM data and slope information to address the issues of inaccurate landslide extraction and difficulty in boundary capture by the U-Net model. Furthermore, it uses a pixel-level change feature extraction method to extract the preceding and following time-phase remote sensing data, thereby further improving the accuracy of the U-Net model in interpreting landslides in real-world scenarios.

[0018] In the landslide extraction method based on the improved U-Net network using asymmetric convolution described above, the slope information is calculated as follows:

[0019]

[0020] p and q represent the rates of change (increments) of elevation in the horizontal and vertical directions, respectively.

[0021] In the aforementioned landslide extraction method based on the improved U-Net network using asymmetric convolution, the constructed model includes two convolutional modules. The outputs of the two convolutional modules are connected to an encoder, which consists of multiple asymmetric convolutional layers. The outputs of the encoder are connected to a decoder. The decoder uses a dense upsampling module composed of multiple convolutional layers, which are directly or indirectly connected to the outputs of the two convolutional modules and the outputs of each asymmetric convolutional layer. This allows for layer-by-layer decoding of the acquired high-resolution remote sensing image feature maps at different scales, reducing the feature map size to the original image size.

[0022] This application utilizes asymmetric convolution, which, compared to traditional convolution, can more accurately extract features from different directions without being affected by rotating targets. The U-Net model itself suffers from information loss during the decoder upsampling stage, which limits its accuracy in interpreting landslides in real-world scenarios, preventing it from meeting practical needs. Therefore, this application employs dense upsampling to address the information loss problem during the upsampling stage, improving the network's ability to extract landslides in complex scenes.

[0023] In the aforementioned landslide extraction method based on the improved U-Net network using asymmetric convolution, the two convolutional modules include a first convolutional layer and a second convolutional layer. Both the first and second convolutional layers are 3×3 2D convolutions, and each convolutional layer is followed by a batch normalization (BN) layer and a nonlinear activation layer.

[0024] The nonlinear activation layer preferably uses the ReLU activation function. After processing by two convolutional modules, the input 6 channels are expanded to 64 channels.

[0025] In the above-mentioned landslide extraction method based on the improved U-Net network of asymmetric convolution, the encoder includes four asymmetric convolutional layers, each of which includes three convolutional layers in parallel. The kernel sizes of the three convolutional layers are 1×3, 3×1 and 3×3, respectively, and each convolutional layer is followed by a nonlinear activation layer.

[0026] By feeding the feature map obtained after downsampling through two convolutional layers into this feature encoding network, feature maps at different scales can be obtained.

[0027] Firstly, 1×3 and 3×1 convolutions inherently enhance the attention to horizontal and vertical features, respectively, thus improving robustness against rotational distortion. Secondly, the asymmetric convolution still incorporates traditional 3×3 2D convolutions, ensuring the layer's ability to extract local and global features. This guarantees that the receptive field of the asymmetric convolution is the same as that of a single 3×3 2D convolution, compensating for the reduced receptive field of 1×3 and 3×1 convolutions. This application addresses the characteristics of landslide distribution and problems encountered during model training, enhancing the network's robustness to rotational distortion and resolving the semantic information mismatch between the encoder and decoder.

[0028] In the landslide extraction method based on the improved U-Net network of asymmetric convolution mentioned above, the dense upsampling module reduces the channel dimension of the low-resolution feature map by spatial rearrangement, while increasing the spatial resolution until it is the same as the size of the landslide label.

[0029] In the landslide extraction method based on the improved U-Net network with asymmetric convolution mentioned above, the dataset labeled with landslides is created in the following way:

[0030] Landslide areas are manually delineated on high-resolution remote sensing images, and binary label maps are created that correspond perfectly to the pixels of the selected high-resolution remote sensing images. Pixels corresponding to landslides are labeled as 1, and non-landslides are labeled as 0. Finally, the results are cropped into small blocks.

[0031] The dataset is constructed by grouping the cropped high-resolution remote sensing images, i.e., RGB-pre-temporal remote sensing image data, landslide labels and corresponding RGB-pre- and RGB-pre-temporal remote sensing image change characteristics, DEM data and slope information.

[0032] Creating the binary label map also involves enhancing the generated patches to produce the final samples. Enhancement processes include multi-scale scaling, random rotation, and vertical / horizontal flipping. The 6-channel data and its labeled raster are cropped into corresponding patches, with the patch size determined based on the specific requirements.

[0033] High-resolution remote sensing images from different time periods often suffer from registration problems. These problems arise from geometric and spectral deviations caused by differences in solar position and satellite attitude. Directly using remote sensing images with registration problems to extract landslides can interfere with the results. Manually correcting these registration errors is time-consuming, labor-intensive, and rarely yields satisfactory results. The method used in this invention can improve the robustness to registration errors during the extraction of change features.

[0034] In the above-mentioned improved U-Net network landslide extraction method based on asymmetric convolution, this method uses a cross loss function and employs backpropagation and gradient descent to update network parameters to train the model.

[0035] This method also includes a precision calculation method with flexible boundaries, used to detect the prediction accuracy of the trained model based on the test dataset:

[0036] Obtain the values ​​of TP (positive sample predicted as positive by the model), FP (negative sample predicted as positive by the model), FN (positive sample predicted as negative by the model), and TN (negative sample predicted as negative by the model), and then adjust TP, FP, TN, and FN to obtain TP. C FP C 、FN C and FP C A precision calculation method incorporating flexible boundaries is adopted to achieve automatic model accuracy detection while ensuring the accuracy of the precision calculation. Specifically, TP stands for True Positive, FP for False Positive, TN for True Negative, and FN for False Negative. A buffer zone with a distance of one pixel is created at the boundary of the landslide patch, and the number of pixels in the buffer zone is e. The corrected TP... C =TP+e,FP C =FP-e,FN C =FN-e,FP C =FP+e.

[0037] Accuracy metrics include Precision, which reflects the algorithm's ability to identify negative samples; Recall, which represents the proportion of all positive examples that are correctly predicted; and F1-score, which measures the accuracy of the binary classification model.

[0038] The function expression for Precision is as follows:

[0039]

[0040] The function expression for Recall is as follows:

[0041]

[0042] The function expression for the F1-score is as follows:

[0043]

[0044] The F1-score takes into account both the precision and recall of a classification model. It can be seen as a harmonic average of the two, with a value between 0 and 1.

[0045] The accuracy evaluation metrics also include the Interchange of Units (IoU), which represents the overlap between the set of true values ​​and model predictions of the labels. The functional expression for IoU is as follows:

[0046]

[0047] Because landslide boundaries are often indistinct, vegetation and gullies caused by precipitation increase the difficulty of manually interpreting landslide boundaries, affecting the accuracy of calculations. The aforementioned indicators include methods for calculating the accuracy of flexible boundaries, thus focusing the prediction results on the landslide itself rather than the conformity of the boundary.

[0048] Compared with the prior art, the beneficial effects of the present invention are reflected in:

[0049] (1) This application makes full use of the landslide data in the region, including accurate location, area, distribution status, etc., first extracts the change features, and then inputs them into the model training. In the prediction stage of deep learning model training, DEM + slope + change features are used as constraints for landslides, so that deep learning can extract deep data connections such as complex relationships between landslides and slopes from the data. It creatively uses asymmetric convolution to extract features in different directions more accurately without being disturbed by the rotating target, thus improving the robustness of the network to rotation distortion.

[0050] (2) Landslide extraction through remote sensing images is essentially image segmentation. In order to improve the segmentation effect, this application extracts landslide-related features from the remote sensing data before and after the event for model training and prediction. It also adds DEM data and slope information to solve the problems of inaccurate landslide extraction and difficulty in boundary capture by the U-Net model. The pixel-level change feature extraction method is used to extract the remote sensing data before and after the event, further improving the accuracy of the U-Net model in interpreting landslides in real-world scenarios.

[0051] (3) This application utilizes asymmetric convolution, which, compared to traditional 3x3 2D convolution, can extract features from different directions more accurately without being disturbed by the rotating target. This application makes improvements to address the characteristics of landslide distribution and the problems existing in the model training process, enhances the network's robustness to rotational distortion, and solves the problem of semantic information mismatch between the encoder and decoder.

[0052] (4) The U-Net model itself suffers from information loss during the decoder upsampling stage, which limits the accuracy of the U-Net model in interpreting landslides in real-world scenarios and makes it unable to meet actual needs. Therefore, this application uses dense upsampling to solve the problem of information loss during the upsampling stage and improves the network's ability to extract landslides in complex scenarios.

[0053] (5) High-resolution remote sensing images from different time phases often suffer from registration problems. These problems are caused by geometric and spectral deviations due to differences in solar position and satellite attitude. Directly using remote sensing images with registration problems to extract landslides will interfere with the results. Manually correcting such registration errors is time-consuming and laborious, and it is difficult to achieve good results. The method used in this invention can improve the robustness to registration errors during the extraction of change features.

[0054] (6) Since landslide boundaries are usually indistinct, vegetation and gullies caused by precipitation increase the difficulty of manually interpreting landslide boundaries, affecting the accuracy of calculation results. The above-mentioned indicators include the accuracy calculation method for flexible boundaries, so that the prediction results focus on the landslide itself rather than the fit of the boundary. Attached Figure Description

[0055] Figure 1 This is a flowchart of the improved U-Net network for landslide extraction based on asymmetric convolution, as described in this invention.

[0056] Figure 2 This is a schematic diagram of the landslide extraction model based on the improved U-Net network of asymmetric convolution according to the present invention.

[0057] Figure 3 This is a schematic diagram of the three basic modules in the landslide extraction model of this invention;

[0058] Figure 4 This is a schematic diagram of the dense upsampling module of the landslide extraction model of the present invention. Detailed Implementation

[0059] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0060] This embodiment provides a landslide extraction method based on an improved U-Net network with asymmetric convolution, referencing... Figure 1 The flowchart includes the following steps:

[0061] Step 1: Data preprocessing and data labeling.

[0062] Data acquisition involves obtaining high-resolution remote sensing images of the target area and performing data preprocessing. This includes mosaicking, color balancing, orthorectification, and image cropping to obtain a complete high-resolution remote sensing image of the target area in TIFF format. DEM data and previous temporal remote sensing images of the region are also acquired.

[0063] Step 2: Multi-source data construction

[0064] Slope information was extracted using DEM data of the region. A pixel-level change feature extraction method was used to extract the change features of the region based on the preceding and following time-phase data. The final data consisted of six channels, including RGB subsequent time-phase remote sensing image, DEM data, slope information, and change features.

[0065] The slope information is calculated as follows:

[0066]

[0067] p and q represent the rates of change (increments) of elevation in the horizontal and vertical directions, respectively.

[0068] Methods for extracting change features include:

[0069] 1) First, select the temporal data of the region before and after, set the neighborhood range, subtract the temporal data of the region before and after one channel, calculate the difference in the neighborhood range of each pixel, and select the value with the largest difference as the change feature of the region.

[0070] 2) The variation features calculated on each channel are averaged over the channel dimension to generate a single-channel variation feature map.

[0071] Step 3: Create a dataset labeled with landslides.

[0072] Landslide areas are manually delineated on remote sensing imagery, and binary label maps are created that perfectly correspond to the pixels of the selected high-resolution remote sensing imagery. Pixels corresponding to landslides are labeled 1, and non-landslide pixels are labeled 0. Finally, the results are cropped into small patches; more specifically:

[0073] 1) Select landslides in a portion of the high-resolution remote sensing image and manually delineate the boundaries. Finally, create a label map that corresponds exactly to the pixels of the selected high-resolution remote sensing image.

[0074] 2) Rasterize the outlined vector to generate a raster file corresponding to each pixel of the cropped area;

[0075] 3) The high-resolution remote sensing images and their annotated rasters are cropped into corresponding small blocks of 256*256 pixels each to generate the original dataset;

[0076] 4) Perform augmentation operations on the original dataset generated in 3) to produce the final dataset. Augmentation processes include multi-scale scaling, random rotation, and random flipping.

[0077] Step 4: Build the model.

[0078] The main body adopts the U-Net network structure. The encoder uses asymmetric convolutions instead of the original standard convolutional layers, and the decoder uses dense upsampling. The network structure is as follows: Figures 2-4As shown, the specific explanation is as follows:

[0079] 1) Network construction,

[0080] The network is constructed using 2D convolutions with a stride of 2 and a size of 3×3, with each layer followed by a BatchNormalization layer and a ReLU activation function. These two layers downsample the input image from 256×256 to 128×128, expanding the input from 6 channels (RGB remote sensing image, DEM data, slope information, and variation features) to 64 channels.

[0081] This structure reduces the resolution to reduce the size of the image that needs to be computed, while extracting information and expanding it to multiple channels.

[0082] 2) Construct a feature encoding module.

[0083] A feature encoder is constructed using four asymmetric convolutional layers. The feature encoding module is constructed using asymmetric convolutions. Convolutional kernels of sizes 1×3, 3×1, and 3×3 are used to replace the standard 3×3 convolution. Multiple non-linear activation layers are integrated to extract image features in different directions. The parameters of each layer are shown in Table 1.

[0084] Table 1. Parameter settings for each convolutional layer in the landslide extraction model, excluding the pixel shuffle layer.

[0085]

[0086] By feeding the feature map obtained after downsampling through two convolutional layers into this feature encoding network, feature maps at different scales can be obtained.

[0087] 3) Construct the decoder using dense upsampling.

[0088] The size of the densely upsampled input feature map is h×w×C, and the size of the feature map generated after the dense upsampling module is H×W×L.

[0089] Where h = H / d, w = W / d, and d is the downsampling factor;

[0090] In order for the output feature map to be able to be rearranged to match the label size, the size of the output feature map should be h×w×d. 2 ×L, this layer is called the Pixel shuffle layer;

[0091] The decoder decodes the feature maps of high-resolution remote sensing images at different scales layer by layer to reduce the size of the feature maps to the original image size. More specifically, the module reduces the channel dimension of the low-resolution feature maps by spatial rearrangement, thereby improving the spatial resolution until it is the same size as the landslide label.

[0092] Step 5: Model training.

[0093] The landslide extraction model designed in step four was trained using the samples created in step three. The cross-entropy loss function was used, and backpropagation and gradient descent methods were employed to update the network parameters in order to train the model.

[0094] The specific steps of model training include:

[0095] 1) Batch input of paired images (containing six-channel multi-source data) and tags

[0096] 2) After network calculation, the final output of the network is displayed.

[0097] The output is first normalized for each channel using the softmax method; then the error is calculated by calculating the loss function value for the normalized result and the label image respectively, and updating the weight of each parameter using backpropagation and gradient descent.

[0098] The loss function used is the cross-entropy loss function, and the formula is:

[0099]

[0100] Where N is the number of pixels in each image, px is the true class distribution of the corresponding pixel i, and qx is the predicted class distribution of the corresponding pixel i.

[0101] 3) Automatic detection of model accuracy;

[0102] By employing a precision calculation method that incorporates flexible boundaries, it is determined whether the prediction results focus on the landslide itself rather than the boundary fit. This method allows for a certain range of boundary errors; in the experiment, the allowed error range was 1 pixel.

[0103] Precision, Recall, F1-score, and IoU are calculated from TP, FP, TN, and FN;

[0104] A buffer zone with a distance of one pixel is established at the boundary of the landslide patch, and the number of pixels within the buffer zone is e. The corrected TP C =TP+e,FP C =FP-e,FN C =FN-e,FP C =FP+e;

[0105] Precision reflects the algorithm's ability to identify negative samples, and its function expression is as follows:

[0106]

[0107] Recall represents the proportion of all positive examples that were correctly predicted, and its function expression is as follows:

[0108]

[0109] The F1 score is a statistical metric used to measure the accuracy of a binary classification model. Its function expression is as follows:

[0110]

[0111] IoU represents the overlap ratio between the set of true values ​​and the set of model predictions for the labels. The function expression is as follows:

[0112]

[0113] Set appropriate values ​​for Precision, Recall, F1-score, and IoU to evaluate the accuracy of the results.

[0114] Step 6: Landslide extraction.

[0115] The trained landslide extraction model, along with high-resolution remote sensing imagery, DEM data, and change characteristics of the target area, is used to extract and identify landslides in the target area; including:

[0116] 1) Perform similar data preprocessing on the high-resolution image of the target area to be extracted as in step one, and then obtain the DEM data, previous time-phase remote sensing data, slope information and change characteristics of the target area. The method of obtaining slope information and change characteristics is the same as that used when constructing multi-source data, and will not be repeated here. The previous time-phase remote sensing data used to obtain change characteristics has been obtained and saved in advance. When calculating change characteristics, the corresponding previous time-phase remote sensing data can be extracted according to the target area to be extracted.

[0117] The processed high-resolution remote sensing image, DEM data, slope information and change characteristics of the target area to be extracted are overlaid together into 6-channel data, and then cropped and numbered.

[0118] 2) Input the cropped image into the network, obtain the recognition result, name it according to the name corresponding to the input number, and save it;

[0119] 3) Reassemble the regional prediction results according to the cropped numbers to obtain the landslide extraction results.

[0120] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A landslide extraction method based on an improved U-Net network with asymmetric convolution, characterized in that, include: The model is constructed using a U-Net network structure, with the encoder part employing asymmetric convolution and the decoder part employing dense upsampling. The constructed model includes two convolutional modules. The outputs of the two convolutional modules are connected to the encoder, which consists of multiple asymmetric convolutional layers. The outputs of the encoder are connected to the decoder. The decoder uses a dense upsampling module composed of multiple convolutional layers, which are directly or through connection layers connected to the outputs of the two convolutional modules and the outputs of each asymmetric convolutional layer. The model uses a layer-by-layer decoding method to decode the feature maps of high-resolution remote sensing images of different scales to the size of the original image. The dense upsampling module reduces the channel dimension of the low-resolution feature map by spatial rearrangement, while increasing the spatial resolution until it is the same size as the landslide label. The multi-source data construction consists of six channels of multi-source data, which are composed of RGB post-temporal remote sensing image data, RGB before and after temporal remote sensing image change characteristics, DEM data and slope information. Create a dataset labeled with landslides using the following method: Landslide areas are manually delineated on high-resolution remote sensing images, and binary label maps are created that correspond perfectly to the pixels of the selected high-resolution remote sensing images. Pixels corresponding to landslides are labeled as 1, and non-landslides are labeled as 0. Finally, the results are cropped into small blocks. The dataset is constructed by grouping the cropped high-resolution remote sensing images, namely RGB post-temporal remote sensing image data, landslide labels and corresponding RGB post-temporal remote sensing image change characteristics, DEM data and slope information. Using a six-channel multi-source dataset labeled with landslides, the model constructed is trained based on the cross loss function and backpropagation and gradient descent methods to obtain a trained model that can extract landslides in the target area. The six-channel multi-source data of the target area is used as input to the trained model, so that the trained model can be used to extract landslides in the target area.

2. The landslide extraction method based on the improved U-Net network with asymmetric convolution as described in claim 1, characterized in that, The change characteristics are obtained based on the RGB pre-temporal remote sensing images and RGB post-temporal remote sensing images of the corresponding region. The RGB post-temporal remote sensing images and RGB pre-temporal remote sensing images are obtained from the high-resolution remote sensing images of the corresponding time phases, respectively. The extraction of the change features is as follows: 1) Select temporal data from different periods for the region, define a neighborhood range, subtract the temporal data from each previous and subsequent period channel by channel, calculate the difference within the neighborhood range of each pixel, and select the value with the largest difference as the variation feature of the region. 2) The variation features calculated on each channel are averaged over the channel dimension to generate a single-channel variation feature map.

3. The landslide extraction method based on the improved U-Net network with asymmetric convolution as described in claim 2, characterized in that, The slope information is calculated as follows. p and q represent the rates of change of elevation in the horizontal and vertical directions, respectively.

4. The landslide extraction method based on the improved U-Net network with asymmetric convolution as described in claim 3, characterized in that, The two-layer convolutional module includes a first convolutional layer and a second convolutional layer. Both the first and second convolutional layers are 2D convolutions of size 3×3, and each convolutional layer is followed by a BN layer and a non-linear activation layer.

5. The landslide extraction method based on the improved U-Net network with asymmetric convolution as described in claim 3, characterized in that, The encoder includes four asymmetric convolutional layers, each of which includes three convolutional layers in parallel. The kernel sizes of the three convolutional layers are 1×3, 3×1, and 3×3, respectively, and each convolutional layer is followed by a nonlinear activation layer.

6. The landslide extraction method based on the improved U-Net network with asymmetric convolution according to claim 1, characterized in that, This method also includes a precision calculation method with flexible boundaries, used to detect the prediction accuracy of the trained model based on the test dataset: Obtain the values ​​of TP (positive sample predicted as positive by the model), FP (negative sample predicted as positive by the model), FN (positive sample predicted as negative by the model), and TN (negative sample predicted as negative by the model), and then adjust TP, FP, TN, and FN to obtain TP. C FP C 、FN C and FP C .

Images