High-resolution remote sensing image landslide identification method, device, system, and storage medium

Abstract

The present invention discloses a method, device, system, and storage medium for identifying landslides using high-resolution remote sensing images. The method comprises: Step S1, obtaining a high-resolution remote sensing landslide dataset; Step S2, constructing a high-resolution remote sensing landslide extraction network (HRLE-Net) based on the high-resolution remote sensing landslide dataset; Step S3, dividing the high-resolution remote sensing landslide dataset into a training set and a test set; Step S4, training the high-resolution remote sensing landslide extraction network (HRLE-Net) based on the training set; and Step S5, inputting the test set into the trained high-resolution remote sensing landslide extraction network (HRLE-Net) to extract landslide areas. The technical solution of the present invention enables accurate identification of landslide areas.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a method and device, a system, and a storage medium for identifying landslides using high-resolution remote sensing images. Background Art

[0002] Landslides, a serious geological disaster, often occur in mountainous areas with complex topography. Their suddenness and unpredictability pose significant challenges to post-disaster rescue efforts. Traditional landslide detection methods, such as manual field surveys and visual interpretation of remote sensing imagery, are no longer sufficient for rapid and efficient rescue efforts. Therefore, exploring new technologies to rapidly obtain information on geological disaster-prone areas and accurately identify landslide zones has become a crucial focus in disaster reduction, relief, and reconstruction efforts.

[0003] Remote sensing technology, due to its ability to provide rich surface information at large spatial scales, has become a crucial tool for monitoring and assessing landslide hazards. Remote sensing imagery can promptly capture signs of landslides before they occur, providing strong support for early warning systems and risk assessments. However, traditional landslide detection methods, such as manual interpretation of satellite imagery or aerial photographs, are time-consuming and susceptible to subjective factors, resulting in low accuracy and efficiency in landslide detection results. Furthermore, traditional methods lack the comprehensive landslide mapping and high resolution and wide coverage required for monitoring, limiting their application in landslide hazard monitoring. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method and device, system and storage medium for identifying landslides using high-resolution remote sensing images, so as to achieve accurate identification of landslide areas.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A landslide identification method based on high-resolution remote sensing images, comprising:

[0007] Step S1, obtaining a high-resolution remote sensing landslide dataset;

[0008] Step S2: constructing a high-resolution remote sensing landslide extraction network HRLE-Net based on the high-resolution remote sensing landslide dataset; wherein the high-resolution remote sensing landslide extraction network HRLE-Net includes: a remote sensing landslide shallow layer feature extraction module, a remote sensing landslide deep layer feature extraction module, and a remote sensing landslide multi-level feature integration module;

[0009] Step S3, dividing the high-resolution remote sensing landslide dataset into a training set and a test set;

[0010] Step S4, training a high-resolution remote sensing landslide extraction network HRLE-Net according to the training set;

[0011] Step S5: Input the test set into the trained high-resolution remote sensing landslide extraction network HRLE-Net to extract the landslide area.

[0012] Preferably, the remote sensing landslide shallow feature extraction module includes: three multi-class residual void convolution blocks, namely F-MRFCB, S-MRFCB and T-MRFCB. The three multi-class residual void convolution blocks use void convolution kernels with different void rates, and cooperate with batch normalization and activation function to jointly extract the shallow features of landslides in high-resolution remote sensing images.

[0013] Preferably, the remote sensing landslide shallow feature extraction module includes: a landslide two-dimensional feature extraction submodule LTFE and a landslide global feature extraction module LGFE; wherein, LTFE extracts deep features of landslides in high-resolution remote sensing images from horizontal and vertical dimensions and cooperates with void convolution; LGFE uses the residual idea and pooling operation to extract deep features of landslides in high-resolution remote sensing images; the output results of LTFE and LGFE are fused with the shallow features obtained in the remote sensing landslide shallow feature extraction module.

[0014] Preferably, the remote sensing landslide multi-level feature integration module includes: a landslide spatial attention integration sub-module LSAA, which uses the maximum pooling operation combined with the void convolution kernel, and multiplies the input element by element, and uses different convolution kernels to obtain multi-level feature maps of different scales to extract multi-level features of landslides in high-resolution remote sensing images.

[0015] The present invention also provides a high-resolution remote sensing image landslide identification device, comprising:

[0016] Acquisition module, used to obtain high-resolution remote sensing landslide datasets;

[0017] A construction module is used to construct a high-resolution remote sensing landslide extraction network HRLE-Net based on a high-resolution remote sensing landslide dataset;

[0018] The partitioning module is used to partition the high-resolution remote sensing landslide dataset into training and test sets;

[0019] A training module is used to train the high-resolution remote sensing landslide extraction network HRLE-Net based on the training set. The high-resolution remote sensing landslide extraction network HRLE-Net includes: a remote sensing landslide shallow feature extraction module, a remote sensing landslide deep feature extraction module, and a remote sensing landslide multi-level feature integration module;

[0020] The extraction module is used to input the test set into the trained high-resolution remote sensing landslide extraction network HRLE-Net to extract the landslide area.

[0021] An embodiment of the present invention further provides a high-resolution remote sensing image landslide identification system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a high-resolution remote sensing image landslide identification method when executed by the processor.

[0022] An embodiment of the present invention further provides a storage medium having a computer program stored thereon. When the computer program is run, the method for identifying landslides using high-resolution remote sensing images is executed.

[0023] The remote sensing landslide shallow feature extraction module in the HRLE-Net of the present invention is used to process the complex and changeable landslide features in high-resolution remote sensing images, and can adapt to changes in the shape, size and position of landslides caused by geographical conditions and time changes. The remote sensing landslide deep feature extraction module solves the problem of landslides being intertwined with background information such as surrounding terrain and vegetation in remote sensing images by extracting deep features, helping the model to accurately distinguish landslides from the background in complex background environments, thereby improving recognition accuracy and reliability. The remote sensing landslide multi-level feature integration module integrates shallow features and deep features and uses an attention mechanism to solve the problem of small and slender landslides being difficult to identify in remote sensing images. By integrating feature information at different levels, it can capture the subtle texture and morphological changes of landslides in images, thereby improving the model's recognition ability for such landslides. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0025] Figure 1 This is a flow chart of a method for identifying landslides using high-resolution remote sensing images according to an embodiment of the present invention;

[0026] Figure 2 Schematic diagram of the HRLE-Net structure;

[0027] Figure 3 This is the data processing flow chart of HRLE-Net;

[0028] Figure 4 Schematic diagram of the F-MRFCB structure;

[0029] Figure 5 Schematic diagram of the S-MRFCB structure;

[0030] Figure 6Schematic diagram of the T-MRFCB structure;

[0031] Figure 7 Schematic diagram of LTFE structure;

[0032] Figure 8 Schematic diagram of LGFE structure;

[0033] Figure 9 Schematic diagram of LSAA structure. DETAILED DESCRIPTION

[0034] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0035] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0036] Example 1:

[0037] like Figure 1 As shown, an embodiment of the present invention provides a method for identifying landslides using high-resolution remote sensing images, comprising:

[0038] Step S1, obtaining a high-resolution remote sensing landslide dataset;

[0039] Step S2: constructing a high-resolution remote sensing landslide extraction network HRLE-Net based on the high-resolution remote sensing landslide dataset;

[0040] Step S3, dividing the high-resolution remote sensing landslide dataset into a training set and a test set;

[0041] Step S4, training a high-resolution remote sensing landslide extraction network HRLE-Net according to the training set;

[0042] Step S5: Input the test set into the trained high-resolution remote sensing landslide extraction network HRLE-Net to extract the landslide area.

[0043] As an implementation method of an embodiment of the present invention, in step S1, high-resolution remote sensing landslide images (hereinafter referred to as landslide images) are obtained, and all of the above images are resized to the same size to construct a high-resolution remote sensing landslide classification dataset. Subsequently, for each landslide image in the high-resolution remote sensing landslide classification dataset, the landslide area is annotated to generate a high-resolution remote sensing landslide segmentation dataset. Next, a script program is used to semantically segment each landslide to obtain a high-resolution remote sensing landslide semantic segmentation dataset. The RGB color label of each image in the high-resolution remote sensing landslide semantic segmentation dataset is then converted into a corresponding high-resolution remote sensing landslide single-channel digital label. Finally, the high-resolution remote sensing landslide single-channel digital label is accurately matched with the high-resolution remote sensing landslide classification dataset to ensure that each landslide image in the high-resolution remote sensing landslide classification dataset corresponds to its unique, corresponding single-channel digital label, thereby obtaining a high-resolution remote sensing landslide dataset.

[0044] In the experiment of the present invention, a high-resolution remote sensing image of a specific area in Sichuan Province was selected as the data source. A Python script was written to uniformly resize the 2,000 remote sensing images of the data source to a size of 512×512 pixels, thereby constructing a high-resolution remote sensing landslide classification dataset. Each image in the high-resolution remote sensing landslide classification dataset was manually annotated using the SuperAnnotate software tool to obtain a high-resolution remote sensing landslide segmentation dataset. The high-resolution remote sensing landslide segmentation dataset was then converted into a high-resolution remote sensing landslide semantic segmentation dataset using a voc2mask script. Furthermore, the 2,000 RGB color labels in the high-resolution remote sensing landslide semantic segmentation dataset were converted into corresponding single-channel digital labels.

[0045] The final high-resolution remote sensing landslide dataset consists of 2,000 high-resolution remote sensing landslide images and their corresponding 2,000 high-resolution remote sensing landslide single-channel digital labels, ensuring that each high-resolution remote sensing landslide image in the high-resolution remote sensing landslide dataset corresponds to its unique and precisely matched high-resolution remote sensing landslide single-channel digital label.

[0046] As an implementation method of the embodiment of the present invention, in step S2, as Figure 2 、 3As shown in the figure, the High-resolution Remote Sensing Landslide Extraction Network (HRLE-Net) is used to extract landslides from high-resolution remote sensing images. The process of using HRLE-Net to extract landslides from high-resolution remote sensing images is divided into three steps: extracting shallow features from high-resolution remote sensing landslide images; extracting deep features from high-resolution remote sensing landslide images; and integrating multi-level features from high-resolution remote sensing landslide images. These three main steps correspond to the three modules of the HRLE-Net network model: shallow feature extraction module for remote sensing landslides; deep feature extraction module for remote sensing landslides; and multi-level feature integration module for remote sensing landslides.

[0047] Further, step S2 includes:

[0048] Step 2.1 Extract shallow features of high-resolution remote sensing landslide images

[0049] The high-resolution remote sensing landslide images from the high-resolution remote sensing landslide classification dataset constructed in Step 1 are input into HRLE-Net.

[0050] The task of extracting shallow features of landslide images is completed by the remote sensing landslide shallow feature extraction module of HRLE-Net.

[0051] The operation and construction process of the remote sensing landslide shallow feature extraction module is as follows:

[0052] Step 2.1.1: Input the landslide image into the remote sensing landslide shallow feature extraction module.

[0053] Step 2.1.2: Use a 3×3 convolution kernel to perform convolution on the landslide image, and then perform batch normalization (i.e. Figure 2 BatchNorm operation in) and activation function processing (i.e. Figure 2 The ReLU operation in is used to obtain the first landslide shallow layer feature map F1.

[0054] Step 2.1.3: Input F1 into the first multi-type residual dilated convolutional block F-MRFCB (First Multi-type Residual Dilated Convolutional Block). The structure of F-MRFCB is as follows Figure 4 As shown:

[0055] The operation and construction process of the F-MRFCB module is as follows:

[0056] First, a convolution operation is performed on F1 using a convolution kernel of size 3×3 and a dilation rate of 3, and then batch normalization is performed (i.e. Figure 4BatchNorm operation in) and activation function processing (i.e. Figure 4 The ReLU operation in the CNN is used to obtain the second landslide shallow feature map F1_1; a convolution operation is performed on F1_1 with a size of 5×5 and a void rate of 5, and then batch normalization and activation function processing are performed to obtain the third landslide shallow feature map F1_2; finally, F1 and F1_2 are spliced in the channel dimension (i.e. Figure 4 ), and perform convolution processing with a convolution kernel size of 3×3 to obtain the fourth landslide shallow layer feature map F2.

[0057] Among them, the dilation rate: The ordinary convolution operation calculates the pixel value of the corresponding position in the output feature map by multiplying the convolution kernel with each pixel point of the input feature map and summing them. The dilation convolution inserts additional intervals (or "holes") in the convolution kernel. These intervals can increase the effective size of the convolution kernel without increasing the number of parameters. The size of this interval is called the "dilation rate" and is usually expressed as an integer. It determines the size of the interval in the convolution kernel.

[0058] Concat channel dimension splicing: refers to splicing two or more features with the same spatial dimensions (height and width) on the channel dimension to generate a new feature map. This new feature map has the same spatial dimensions as the input feature map, but the number of channels is the sum of the number of channels of all input feature maps.

[0059] Step 2.1.4: Use a 3×3 convolution kernel to convolve F2, then perform normalization and activation function processing to obtain the shallow feature map F3 of the fifth landslide.

[0060] Step 2.1.5: Input F3 into the second multi-type residual dilated convolutional block S-MRFCB (Second Multi-type Residual Dilated Convolutional Block). The structure of S-MRFCB is as follows Figure 5 As shown:

[0061] The S-MRFCB operation and construction process is as follows:

[0062] First, a convolution operation is performed on F3 using a convolution kernel of size 5×5 and a dilation rate of 5, and then batch normalization is performed (i.e. Figure 5 BatchNorm operation in) and activation function processing (i.e. Figure 5The ReLU operation in the CNN is used to obtain the shallow feature map F3_1 of the sixth landslide. Then, a convolution operation is performed on F3_1 with a convolution kernel of size 7×7 and a void ratio of 7, and then batch normalization and activation function processing are performed to obtain the shallow feature map F3_2 of the seventh landslide. Finally, F3 and F3_2 are spliced in the channel dimension (i.e. Figure 5 ), and perform convolution processing with a convolution kernel size of 3×3 to obtain the eighth landslide shallow layer feature map F4.

[0063] Step 2.1.6: Use a 3×3 convolution kernel to convolve F4, then perform normalization and activation function processing to obtain the ninth landslide shallow layer feature map F5.

[0064] Step 2.1.7: Input F5 into the third multi-type residual dilated convolutional block T-MRFCB (Third Multi-type Residual Dilated Convolutional Block). The structure of T-MRFCB is as follows Figure 5 As shown:

[0065] The T-MRFCB operation and construction process is as follows:

[0066] First, a convolution operation is performed on F5 using a convolution kernel of size 7×7 and a dilation rate of 7, and then batch normalization is performed (i.e. Figure 6 BatchNorm operation in) and activation function processing (i.e. Figure 6 The ReLU operation in the network is used to obtain the shallow feature map F5_1 of the tenth landslide. Then, a convolution operation is performed on F5_1 with a convolution kernel of size 9×9 and a void ratio of 9. After that, batch normalization and activation function processing are performed to obtain the shallow feature map F5_2 of the eleventh landslide. Finally, F5 and F5_2 are spliced in the channel dimension (i.e. Figure 6 ), and perform convolution processing with a convolution kernel size of 3×3 to obtain the shallow feature map F6 of the twelfth landslide.

[0067] Step 2.2 Extract deep features of high-resolution remote sensing landslide images

[0068] The task of extracting deep features from high-resolution remote sensing landslide images is accomplished by the remote sensing landslide deep feature extraction module of HRLE-Net.

[0069] The operation and construction process of the remote sensing landslide deep feature extraction module is as follows:

[0070] Step 2.2.1: Input F2 into the Landslide Two-dimensional Feature Extraction block (LTFE). The structure of LTFE is as follows: Figure 7 As shown:

[0071] The LTFE running and building process is:

[0072] First, a convolution operation is performed on F2 using a convolution kernel of size 5×5 pixels and a void ratio of 5, and then a horizontal maximum pooling (i.e. Figure 7 The X_Max Pool operation in the pooling operation is performed to obtain the first landslide depth feature map F2_1; then, a convolution operation is performed on F2 using a convolution kernel of size 5×5 and a void rate of 5, and then a vertical maximum pooling operation (i.e. Figure 7 The Y_Max Pool operation in the pooling operation is performed to obtain the second landslide depth feature map F2_2; finally, F2_1 and F2_2 are spliced in the channel dimension (i.e. Figure 7 Concat operation in ), and apply the activation function (i.e. Figure 7 The third landslide depth feature map F7 is obtained by processing with the Sigmoid operation in

[15] .

[0073] The process of applying LTFE to F2 to obtain F7 is shown in the following formula (1):

[0074] F7=S(Concat(CMP(Conv5_5(F2)),YMP(Conv5_5(F2)))) (1)

[0075] Among them, Concat() represents the concatenation operation on the channel dimension; XMP() represents the X_Max Pool operation; YMP() represents the Y_Max Pool operation; Conv5_5() represents the convolution operation of the convolution kernel with a size of 5×5 and a void rate of 5; S represents the Sigmoid activation function;

[0076] F2 and F7 are concatenated in the channel dimension, and then a convolution kernel of size 1×1 is applied to adjust the number of channels to obtain the fourth landslide depth feature map F8.

[0077] Step 2.2.2: Apply bilinear interpolation to adjust the size of F6 to obtain the fifth landslide depth feature map F9.

[0078] Step 2.2.3: Input F9 into the Landslide Global Feature Extraction block LGFE. The structure of LGFE is as follows: Figure 8 As shown:

[0079] The LGFE operation and construction process is:

[0080] First, apply global max pooling (i.e. Figure 8 The Global Max Pool operation in ( ) performs a pooling operation on F9 to obtain a one-dimensional vector of length C, and then uses a one-dimensional convolution of size 1×1 on the above one-dimensional vector (i.e. Figure 8 After the Conv1d operation in

[15] and the activation function processing, the weight w of each channel is obtained, and finally the weight w is multiplied element by element with F9 to obtain the sixth landslide depth feature map F10.

[0081] The process of applying LGFE to F9 to obtain F10 is shown in the following formula (2):

[0082] F10=w⊙F9=S(Conv1d(GMP(F9)))⊙F9 (2)

[0083] Among them, S represents the Sigmoid activation function; GMP() represents the Global Max Pool operation; Conv1d() represents the one-dimensional convolution operation; ⊙ represents the element-by-element product.

[0084] Step 2.2.4: Concatenate F10 and F9 in the channel dimension, apply a 1×1 convolution kernel to adjust the number of channels, and then apply bilinear interpolation to resize to obtain the seventh landslide depth feature map F11.

[0085] Step 2.2.5: First, concatenate F11 and F8 in the channel dimension, then apply a 1×1 convolution kernel to adjust the number of channels, and finally apply bilinear interpolation to the feature size to obtain the eighth landslide depth feature map F12.

[0086] Step 2.3 Integrate multi-level features of high-resolution remote sensing landslide images

[0087] The task of integrating the multi-level features of high-resolution remote sensing landslide images is completed by the remote sensing landslide multi-level feature integration module of HRLE-Net.

[0088] The operation and construction process of the remote sensing landslide multi-level feature integration module is as follows:

[0089] Step 2.3.1: After convolving F6 with a 1×1 convolution kernel, convolution operations are performed using three convolution kernels of sizes 5×5, 9×9, and 11×11, respectively, to obtain the first landslide multi-level feature map F6_1, the second landslide multi-level feature map F6_2, and the third landslide multi-level feature map F6_3.

[0090] Step 2.3.2: Concatenate F6_1, F6_2, and F6_3 in the channel dimension, then apply a 1×1 convolution kernel and bilinear interpolation to adjust the number and size of channels to obtain the fourth landslide multi-level feature map F13.

[0091] Step 2.3.3: Input F13 into the landslide spatial attention aggregation block LSAA. The structure of LSAA is as follows: Figure 9 As shown:

[0092] The LSAA running and building process is:

[0093] First, apply max pooling (i.e. Figure 8 The Max Pool operation in the pooling operation is performed on F13, and then the weights of each channel are obtained after processing with a dilated convolution of size 5×5 and a dilation rate of 3 and an activation function. Finally, the weight Multiplying element-by-element with F13 yields the fifth landslide depth feature map F14.

[0094] Applying LSAA to process F13 to obtain F14 is shown in the following formula (3):

[0095]

[0096] Among them, S represents the Sigmoid activation function; MP() represents the Max Pool operation; Conv5_3() represents the dilated convolution operation with a size of 5×5 and a dilation rate of 3; ⊙ represents the element-by-element product.

[0097] Step 2.3.4: Concatenate F14 and F12 in the channel dimension, and then apply a 1×1 convolution kernel to adjust the number of channels to obtain the landslide semantic segmentation image.

[0098] In the experiment of the present invention, the high-resolution remote sensing landslide image is input into the remote sensing landslide shallow layer feature extraction module. The size of the high-resolution remote sensing landslide image is 512×512 pixels and 3 channels. The high-resolution remote sensing landslide image is convolved with a convolution kernel of size 3×3, and then batch normalization (i.e. Figure 2 BatchNorm operation in) and activation function processing (i.e. Figure 2The ReLU operation in the CNN is used to obtain the first landslide shallow feature map F1, which has a size of 256×256 pixels and 32 channels. F1 is input into the first multi-class residual void convolution block F-MRFCB to obtain the fourth landslide shallow feature map F2, which has a size of 256×256 pixels and 32 channels. A convolution operation is performed on F2 with a size of 3×3 convolution kernel, and then normalized and activated function processed to obtain the fifth landslide shallow feature map F3, which has a size of 128×128 pixels and 64 channels. F3 is input into the second In the multi-class residual void convolution block S-MRFCB, the eighth landslide shallow feature map F4 is obtained. The size of F4 is 128×128 pixels and 64 channels. The convolution operation is performed on F4 with a convolution kernel of size 3×3, and then normalized and activated. The ninth landslide shallow feature map F5 is obtained. The size of F5 is 64×64 pixels and 128 channels. F5 is input into the third multi-class residual void convolution block T-MRFCB to obtain the twelfth landslide shallow feature map F6. The size of F6 is 64×64 pixels and 128 channels.

[0099] In the remote sensing landslide deep feature extraction module, F2 is input into the landslide two-dimensional feature extraction submodule LTFE to obtain the third landslide depth feature map F7, which is 256×256 pixels and has 32 channels. F2 and F7 are concatenated in the channel dimension, and a 1×1 convolution kernel is applied to adjust the number of channels to obtain the fourth landslide depth feature map F8, which is 256×256 pixels and has 32 channels. Bilinear interpolation is applied to adjust the size of F6 to obtain the fifth landslide depth feature map F9, which is 128×128 pixels and has 64 channels. F9 is input into the landslide global feature extraction module LGFE to obtain the sixth landslide depth feature map F10, which is 128×128 pixels and 64 channels. F10 and F9 are concatenated in the channel dimension, and the number of channels is adjusted by a 1×1 convolution kernel. Bilinear interpolation is then applied to resize the image to obtain the seventh landslide depth feature map F11, which is 256×256 pixels and 32 channels. F11 is first concatenated in the channel dimension with F8, and the number of channels is adjusted by a 1×1 convolution kernel. Finally, bilinear interpolation is applied to the feature size to obtain the eighth landslide depth feature map F12, which is 512×512 pixels and 32 channels.

[0100] In the remote sensing landslide multi-level feature integration module, after applying a convolution kernel of size 1×1 to F6, three convolution kernels of sizes 5×5, 9×9, and 11×11 are used for convolution operations to obtain the first landslide multi-level feature map F6_1, the second landslide multi-level feature map F6_2, and the third landslide multi-level feature map F6_3. Among them, the size of F6_1 is 64×64 pixels and 128 channels, the size of F6_2 is 64×64 pixels and 128 channels, and the size of F6_3 is 64×64 pixels and 128 channels. F6_1, F6_2, and F6_3 are concatenated along the channel dimension, and a 1×1 convolution kernel and bilinear interpolation are applied to adjust the number and size of channels. This yields the fourth landslide multi-level feature map F13, which is 512×512 pixels and has 128 channels. This F13 is then fed into the landslide spatial attention integration submodule (LSAA) to yield the fifth landslide deep feature map F14, which is 512×512 pixels and has 128 channels. F14 is concatenated along the channel dimension with F12, and a 1×1 convolution kernel is applied to adjust the number of channels. This yields a landslide semantic segmentation image (removing background interference and retaining only the landslide; in the landslide semantic segmentation image, white represents the landslide and black represents the background). The landslide semantic segmentation image is 512×512 pixels and has one channel.

[0101] As an implementation method of an embodiment of the present invention, in step S3, the preprocessed high-resolution remote sensing landslide dataset is divided into a training set, a validation set, and a test set in a ratio of 70%:15%:15%, wherein the training set contains 1,400 images, accounting for 70% of the total, while the validation set and the test set each contain 300 images, accounting for 15% of the total dataset, and there is no overlap between the three subsets.

[0102] As an implementation method of an embodiment of the present invention, step S4 specifically includes: first initializing all neural network parameters and setting model-related hyperparameters, such as: training rounds, batch size, optimizer selection, learning rate, and total number of iterations.

[0103] Next, the High-Resolution Remote Sensing Landslide Extraction Network (HRLE-Net) is trained. After initializing parameters, the training and validation data sets are divided into batches. Each batch of training data is fed into the HRLE-Net for training, and the training loss for that batch is calculated. After a round of training has traversed all batches of the training set, the validation data is also fed into the HRLE-Net batch by batch, generating the corresponding batch loss values (batch_loss). These batch_loss values are used to monitor model overfitting and adjust training strategies accordingly, such as early termination or adjusting the learning rate. During training and validation, the HRLE-Net automatically learns and adjusts parameters based on the loss and batch_loss values. Training of the HRLE-Net concludes when the batch_loss values converge after one or more rounds of training.

[0104] As an implementation method of the embodiment of the present invention, in step S5, the test set is input into the trained HRLE-Net model, and the landslide area result, that is, the landslide semantic segmentation image, is output.

[0105] In the remote sensing landslide shallow feature extraction module of HRLE-Net, the embodiment of the present invention designs three multi-class residual void convolution blocks, namely F-MRFCB, S-MRFCB and T-MRFCB. The three multi-class residual void convolution blocks use void convolution kernels with different void rates, and cooperate with batch normalization and activation functions to jointly extract the shallow features of landslides in high-resolution remote sensing images.

[0106] In this embodiment of the present invention, HRLE-Net is used in the remote sensing landslide shallow feature extraction module, designing a two-dimensional landslide feature extraction submodule (LTFE) and a global landslide feature extraction module (LGFE). LTFE extracts deep features of landslides from high-resolution remote sensing images from horizontal and vertical dimensions using dilated convolution. LGFE extracts deep features of landslides from high-resolution remote sensing images using a residual algorithm combined with pooling. The outputs of LTFE and LGFE are fused with the shallow features obtained from the remote sensing landslide shallow feature extraction module. The high computational efficiency of shallow features combined with the strong representation capability of deep features enables the extraction of richer features.

[0107] In this embodiment of the present invention, a landslide spatial attention integration submodule (LSAA) is designed within the HRLE-Net remote sensing landslide multi-level feature integration module. LSAA utilizes a combination of a max pooling operation and a dilated convolution kernel, performing element-wise multiplication on the module's input. It also uses multi-level feature maps of varying scales generated by convolution with different convolution kernels to extract multi-level features of landslides from high-resolution remote sensing images.

[0108] The embodiments of the present invention can more accurately capture and identify complex and changeable landslide features when processing these complex features; the method can more accurately distinguish landslides from the background, solving the problem that landslides are often intertwined with the surrounding terrain, vegetation and other background information in remote sensing images, which makes the model easily confused during recognition; the method has stronger recognition of small and slender landslides, reduces the number of landslide classification errors, and solves the problem that the existing technology, due to excessive convolution and pooling processes, causes its texture information to gradually disappear on the feature map, resulting in the model being unable to accurately identify these landslides

[0109] Example 2:

[0110] The embodiment of the present invention further provides a device for identifying landslides using high-resolution remote sensing images, comprising:

[0111] Acquisition module, used to obtain high-resolution remote sensing landslide datasets;

[0112] A construction module is used to construct a high-resolution remote sensing landslide extraction network HRLE-Net based on a high-resolution remote sensing landslide dataset;

[0113] The partitioning module is used to partition the high-resolution remote sensing landslide dataset into training and test sets;

[0114] A training module is used to train the high-resolution remote sensing landslide extraction network HRLE-Net based on the training set. The high-resolution remote sensing landslide extraction network HRLE-Net includes: a remote sensing landslide shallow feature extraction module, a remote sensing landslide deep feature extraction module, and a remote sensing landslide multi-level feature integration module;

[0115] The extraction module is used to input the test set into the trained high-resolution remote sensing landslide extraction network HRLE-Net to extract the landslide area.

[0116] Example 3:

[0117] An embodiment of the present invention further provides a high-resolution remote sensing image landslide identification system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a high-resolution remote sensing image landslide identification method when executed by the processor.

[0118] Example 4:

[0119] An embodiment of the present invention further provides a storage medium having a computer program stored thereon. When the computer program is run, the method for identifying landslides using high-resolution remote sensing images is executed.

[0120] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A method for identifying landslides using high-resolution remote sensing images, characterized in that: include: Step S1, obtaining a high-resolution remote sensing landslide dataset; Step S2: constructing a high-resolution remote sensing landslide extraction network HRLE-Net based on the high-resolution remote sensing landslide dataset; wherein the high-resolution remote sensing landslide extraction network HRLE-Net includes: a remote sensing landslide shallow layer feature extraction module, a remote sensing landslide deep layer feature extraction module, and a remote sensing landslide multi-level feature integration module; Step S3, dividing the high-resolution remote sensing landslide dataset into a training set and a test set; Step S4, training a high-resolution remote sensing landslide extraction network HRLE-Net according to the training set; Step S5: input the test set into the trained high-resolution remote sensing landslide extraction network HRLE-Net to extract the landslide area; The remote sensing landslide shallow feature extraction module includes three multi-class residual dilated convolution blocks: F-MRFCB, S-MRFCB, and T-MRFCB. These three multi-class residual dilated convolution blocks use dilated convolution kernels with different dilation rates, and are combined with batch normalization and activation functions to extract shallow features of landslides from high-resolution remote sensing images. The remote sensing landslide shallow feature extraction module includes: the landslide two-dimensional feature extraction submodule LTFE and the landslide global feature extraction module LGFE. LTFE extracts deep features of landslides from high-resolution remote sensing images from horizontal and vertical dimensions using dilated convolution. LGFE uses the residual concept combined with pooling to extract deep features of landslides from high-resolution remote sensing images. The outputs of LTFE and LGFE are fused with the shallow features obtained from the remote sensing landslide shallow feature extraction module. The multi-level feature integration module of remote sensing landslides includes: the landslide spatial attention integration sub-module LSAA. LSAA uses the maximum pooling operation combined with the void convolution kernel, and multiplies the input element by element. It also uses the multi-level feature maps of different scales obtained by convolution with different convolution kernels to extract the multi-level features of landslides in high-resolution remote sensing images.

2. A high-resolution remote sensing image landslide identification device for implementing the high-resolution remote sensing image landslide identification method described in claim 1, characterized in that: include: Acquisition module, used to obtain high-resolution remote sensing landslide datasets; A construction module is used to construct a high-resolution remote sensing landslide extraction network HRLE-Net based on a high-resolution remote sensing landslide dataset; The partitioning module is used to partition the high-resolution remote sensing landslide dataset into training and test sets; A training module is used to train the high-resolution remote sensing landslide extraction network HRLE-Net based on the training set. The high-resolution remote sensing landslide extraction network HRLE-Net includes: a remote sensing landslide shallow feature extraction module, a remote sensing landslide deep feature extraction module, and a remote sensing landslide multi-level feature integration module; The extraction module is used to input the test set into the trained high-resolution remote sensing landslide extraction network HRLE-Net to extract the landslide area.

3. A high-resolution remote sensing image landslide identification system, characterized by: include: A memory and a processor, wherein the memory stores a computer program executed by the processor, and when the computer program is executed by the processor, the method for identifying landslides using high-resolution remote sensing images according to claim 1 is executed.

4. A storage medium, characterized in that The storage medium stores a computer program, which executes the high-resolution remote sensing image landslide identification method according to claim 1 when running.

Images