Landslide disaster detection method based on improved remote sensing landslide image segmentation model

Abstract

The invention discloses a landslide disaster detection method based on an improved remote sensing landslide image segmentation model, and relates to the technical field of remote sensing image processing. Comprising the following steps: S1, constructing a multi-feature input data set based on gradient tensor data; s2, preprocessing the multi-feature input data set and generating an enhanced feature data set through a feature enhancement technology; s3, constructing a landslide detection model; s4, optimizing the landslide detection model by combining the training data set, the cross entropy loss and an Adam optimizer to generate an optimized landslide detection model; s5, correcting the optimized landslide detection model based on a segmentation evaluation result to generate a final landslide detection model; and S6, generating a current landslide disaster detection result based on the final landslide detection model and the acquired landslide image real-time data. Implicit feature information of remote sensing image data can be fully mined, semantic segmentation precision of a traditional model is effectively improved, and landslide extraction precision and accuracy are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to a landslide disaster detection method based on an improved remote sensing landslide image segmentation model. Background Art

[0002] Remote sensing images contain rich semantic information and play an important role in landslide hazard monitoring tasks. Traditional landslide identification is mainly based on remote sensing visual interpretation and human-computer interaction recognition, which has problems such as time-consuming and labor-intensive, highly subjective and low extraction accuracy. Semantic segmentation, as an important task in deep learning, has played an important role in the automated recognition of remote sensing images due to its end-to-end pixel-level classification capabilities. Existing semantic segmentation models for landslide disasters in remote sensing images usually cannot take into account multi-scale ground features. The recognition results have problems such as blurred boundaries, internal voids, and overall instance discontinuity. In addition, traditional dataset construction only uses the original remote sensing images, which makes it difficult to fully mine the implicit edge information of the image. In short, how to fully mine the implicit feature information of remote sensing image data and effectively improve the semantic segmentation accuracy of traditional models is the key to improving the precision and accuracy of landslide extraction.

[0003] Therefore, it is an urgent problem for those skilled in the art to propose a landslide disaster detection method based on an improved remote sensing landslide image segmentation model to solve the difficulties existing in the existing technology. Summary of the Invention

[0004] In view of this, the present invention provides a landslide hazard detection method based on an improved remote sensing landslide image segmentation model, which can fully mine the implicit feature information of remote sensing image data and effectively improve the semantic segmentation accuracy of traditional models, thereby improving the landslide extraction precision and accuracy.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions: A landslide hazard detection method based on an improved remote sensing landslide image segmentation model includes the following steps: S1. Calculate the corresponding gradient tensor data based on the acquired original landslide image, and construct a multi-feature dataset based on the gradient tensor data; S2. Preprocessing the multi-feature dataset to generate a preprocessed feature set, and processing the preprocessed feature set using feature enhancement technology to generate an enhanced feature dataset; S3. Build a landslide detection model by combining the U-Net network architecture and the CBAM attention mechanism; S4. Combining the acquired landslide detection model training dataset, cross entropy loss, and Adam optimizer to complete iterative training of the network and generate an optimized landslide detection model; The original landslide image data and gradient tensor data are used as the input data of the training dataset, and the landslide label data is used as the output data of the training dataset; S5. Use the optimized landslide detection model to segment the remote sensing image to generate landslide hazard detection results, and perform multi-dimensional evaluation on the segmentation results to generate segmentation evaluation results of the optimized landslide detection model.

[0006] Optionally, the corresponding gradient tensor data is calculated including: Calculate the horizontal gradient data of each pixel in the original remote sensing image based on the Sobel operator And the gradient data in the vertical direction ; And calculate the gradient magnitude of the pixel and the gradient direction of the pixel : ; Pixel-based gradient magnitude Construct a feature set of varying intensity based on the gradient direction of the pixel Construct a change direction feature set, and generate a multi-feature data set by integrating the change intensity feature set and the change direction feature set; Optionally, a landslide detection model can be constructed by combining the U-Net network architecture and the CBAM attention mechanism, including: The U-Net network architecture includes: input module, encoder, decoder, skip connection, and output module; Among them, the input module is used to: convert the enhanced feature dataset into a feature map; The encoder is used to extract features and compress spatial dimensions of the input feature map layer by layer to generate multi-scale high-dimensional semantic features, including multiple convolution blocks and maximum pooling layers; The decoder of the U-Net network architecture is optimized based on the CBAM attention mechanism to generate a CBAM attention decoder. The CBAM attention decoder is used to gradually restore the spatial resolution of the feature map and fuse multi-scale information to generate a segmentation mask of the same size as the input. It includes multiple upsampling blocks and convolution blocks. The CBAM attention mechanism performs attention screening on the enhanced feature dataset through the channel attention mechanism and the spatial attention mechanism; The CMUnet network architecture is built by combining the encoder and skip connection of the U-Net network architecture and the CBAM attention decoder, and the landslide detection model is generated based on the implementation data.

[0007] Optionally, attention filtering is performed on the enhanced feature dataset to enhance features related to landslide detection (such as edges and textures) and suppress irrelevant features. The process can be described as: ; ; Where, Represents the Kronecker product. That is, the feature map output by the previous convolutional layer After entering the CBAM module, it first passes through the channel attention mechanism Mapped to the first feature map , and then Enter the spatial attention mechanism for input , after processing by the spatial attention mechanism, the second feature map is obtained ; Features after CBAM attention optimization Then enter the subsequent convolutional layer to gradually restore the spatial resolution of the feature map and generate a segmentation mask.

[0008] Optional, channel attention mechanism Also includes: Channel Attention Mechanism Used to filter preset target features;

[0009] Spatial Attention Mechanism Used to filter the location of preset target features;

[0010] Where, represents the Sigmoid activation function, Represented as a convolutional layer operation, stands for Multilayer Perceptron; stands for average pooling; stands for max pooling.

[0011] Optionally, generate an optimized landslide detection model, including: Construct a training dataset for the landslide detection model. Combine the training dataset and cross-entropy loss to calculate the pixel-by-pixel error and generate the corresponding error gradient. Combine the error gradient and the Adam optimizer to modify the CMUnet network architecture weights to complete iterative training of the network and generate an optimized landslide detection model. Among them, the cross entropy loss function calculation formula is as follows: ; in, is the true label, Predict probabilities for the model.

[0012] It can be seen from the above technical solution that, compared with the prior art, the landslide disaster detection method based on the improved remote sensing landslide image segmentation model of the present invention has the following beneficial effects: Gradient tensor-derived features were constructed to reflect the edges, textures, and topographical variations of landslide areas, improving network modeling accuracy based on a multi-feature input dataset. The CMUnet network architecture was built using the CBAM attention mechanism, which combines channel and spatial attention mechanisms to filter the enhanced feature dataset. This helps further enhance the network's feature extraction capabilities, improve recognition accuracy, reduce internal voids, and enhance recognition continuity. While enhancing useful features, it suppresses the influence of irrelevant features, focusing more attention on the target area to improve classification accuracy. This results in an optimized landslide detection model for more accurate and complete landslide segmentation in remote sensing images. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] 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.

[0014] Figure 1 A flow chart of a landslide hazard detection method based on an improved remote sensing landslide image segmentation model provided by the present invention; Figure 2 A schematic diagram of the construction process of the improved remote sensing landslide image segmentation model provided by the present invention; Figure 3 A network structure diagram of the improved remote sensing landslide image segmentation model provided by the present invention; Figure 4 Schematic diagram of the principle of the CBAM attention mechanism provided by the present invention; Figure 5 The training loss graph provided for this embodiment; Figure 6 To improve the test results of the remote sensing landslide image segmentation model, 6a is an example of the original remote sensing image in the training set, 6b and 6c are the landslide label and model prediction results corresponding to 6a, respectively; 6d is an example of the original remote sensing image in the test set, 6e and 6f are the landslide label and model prediction results corresponding to 6d, respectively. DETAILED DESCRIPTION

[0015] 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.

[0016] In this application, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. The terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article or apparatus comprising the element.

[0017] The present invention can be used in a variety of general-purpose or special-purpose computing device environments or configurations, such as personal computers, server computers, handheld or portable devices, tablet devices, multi-processor devices, and distributed computing environments including any of the above.

[0018] Reference Figure 1 As shown, the present invention discloses a landslide disaster detection method based on an improved remote sensing landslide image segmentation model, comprising the following steps: S1. Calculate the corresponding gradient tensor data based on the acquired original landslide image, and construct a multi-feature input dataset based on the gradient tensor data; S2. Preprocessing the multi-feature input dataset to generate a preprocessed feature dataset, and processing the preprocessed feature dataset using a feature enhancement technique to generate an enhanced feature dataset; S3. Build a landslide detection model by combining the U-Net network architecture and the CBAM attention mechanism; S4. Construct a training dataset for the landslide detection model. This dataset is combined with the cross-entropy loss (which calculates the error pixel by pixel and guides model optimization) and the Adam optimizer (which adjusts network weights based on the error gradient) to iteratively train the network and generate an optimized landslide detection model. The original landslide image data and gradient tensor data are used as the input data of the training dataset, and the landslide label data is used as the output data of the training dataset; S5. Use the optimized landslide detection model to segment the remote sensing image to generate landslide disaster detection results, and perform multi-dimensional evaluation on the segmentation results to generate segmentation evaluation results of the optimized landslide detection model. Specifically, Figure 1This is a technical roadmap for constructing a landslide image segmentation model based on multi-feature input and an improved CMUnet model. It primarily includes three parts: dataset construction, model construction, and model training. The dataset construction part primarily involves calculating derivative features based on gradient tensors, data preprocessing, data augmentation, and training and test set partitioning. The model construction part primarily involves building a U-Net architecture and introducing the CBAM attention mechanism. The model training part primarily involves training the network to learn the mapping relationship between remote sensing images and landslide labels to establish a landslide recognition model.

[0019] This paper uses a landslide dataset from Bijie, Guizhou Province, as an example. This dataset contains satellite optical imagery and landslide boundary shape files. All images in this dataset—770 landslide images and 2,003 non-landslide images—were cropped from TripleSat satellite imagery captured between May and August 2018. The landslide sample data consists of landslide images (*.png) and landslide shape files (*.png). To fully exploit the implicit information in the raw remote sensing image data, this paper constructs derived features using gradient tensors to assist in network training.

[0020] Furthermore, the corresponding gradient tensor data is calculated including: Calculate the gradient data of each pixel in the original remote sensing image in the horizontal and vertical directions based on the Sobel operator; ; in, is the original image, For horizontal and are the gradient images in the horizontal and vertical directions respectively; Calculate the gradient magnitude of the pixel based on the horizontal and vertical gradients and the gradient direction of the pixel : ; Pixel-based gradient magnitude (reflects the intensity of the change, larger values ​​indicate stronger edges) and gradient direction (Indicates the direction of change and reflects the directionality of landslide texture) Constructs a multi-feature input dataset; Among them, based on the gradient amplitude Generate the first derived feature dataset based on the gradient direction A second derived feature dataset is generated.

[0021] Furthermore, the landslide detection model is constructed by combining the U-Net network architecture and the CBAM attention mechanism, including: The U-Net network architecture includes: input module, encoder, decoder, skip connection, and output module; Among them, the input module is used to: convert the enhanced feature dataset into a feature map; The encoder is used to extract features and compress spatial dimensions of the input feature map layer by layer to generate multi-scale high-dimensional semantic features, including multiple convolution blocks and maximum pooling layers; The decoder is symmetrical with the encoder and is used to gradually restore the spatial resolution of the feature map and fuse multi-scale information, ultimately generating a segmentation mask of the same size as the input. It includes multiple upsampling blocks and convolution blocks. Based on the CBAM attention mechanism, the decoder of the U-Net network architecture is optimized to generate a CBAM attention decoder; The CBAM attention mechanism performs attention screening on the enhanced feature dataset through the channel attention mechanism and the spatial attention mechanism; The CMUnet network architecture is built by combining the encoder and skip connection of the U-Net network architecture and the CBAM attention decoder, and the landslide detection model is generated based on the implementation data.

[0022] Furthermore, the enhanced feature dataset is screened for attention, which strengthens the features related to landslide detection (such as edges and textures) and suppresses irrelevant features. Figure 4 As shown, the process can be described as: ; ; Where, Represents the Kronecker product. That is, the feature map output by the previous convolutional layer After entering the CBAM module, it first passes through the channel attention mechanism Mapped to the first feature map , and then Enter the spatial attention mechanism for input , after processing by the spatial attention mechanism, the second feature map is obtained . Features after CBAM attention optimization Then enter the subsequent convolutional layer to gradually restore the spatial resolution of the feature map and generate a segmentation mask; Focus on which features are meaningful features, The focus is on which features are meaningful features. The specific calculation formula is as follows: ; ; Where, represents the Sigmoid activation function, Represented as a convolutional layer operation, stands for Multilayer Perceptron; stands for average pooling; stands for max pooling.

[0023] Specifically, the gradient magnitude reflects the intensity of changes in pixel intensity; larger values ​​indicate stronger edges. This method uses this as a first derived feature to assist network training. The gradient direction indicates the direction of change and can assist in analyzing the directionality of landslide textures. This method uses this as a second derived feature to assist network training. Furthermore, the sample data is preprocessed to a uniform 512*512 size and augmented using data augmentation techniques such as rotation and flipping. 70% of the sample data is selected as training data, and the remaining 30% as test data.

[0024] U-Net is a convolutional neural network with a classic encoder-decoder structure. Its core design concept is to fuse the low-level features in the encoder (downsampling path) with the high-level features in the decoder (upsampling path) through skip connections. This design eliminates the need for the network to separately calculate the loss value of high-level features during the decoding process. Instead, it directly utilizes low-level detail information and high-level semantic information through feature fusion to generate a feature map containing multi-scale features, thereby improving image segmentation accuracy. In order to further improve the feature extraction capability of the network, improve recognition accuracy, reduce internal voids, and improve recognition continuity, the present invention introduces the CBAM attention mechanism (Convolutional Block Attention Module) between the convolutional layers of the traditional U-Net decoder, and proposes the CMUnet network for remote sensing image segmentation, such as Figure 3 CBAM is an attention module that can be easily embedded in end-to-end convolutional neural networks. It combines the channel attention mechanism and the spatial attention mechanism to enhance useful features while suppressing the influence of irrelevant features, focusing more on the target area to improve classification accuracy.

[0025] Furthermore, an optimized landslide detection model is generated, including: Construct a training dataset for the landslide detection model. Combine the training dataset with the cross-entropy loss (which calculates the error pixel by pixel and guides model optimization) and the Adam optimizer (which adjusts network weights based on the error gradient) to iteratively train the network and generate an optimized landslide detection model. The original landslide image data and gradient tensor data serve as the input data for the training dataset, while the landslide label data serves as the output data. The cross entropy loss function is calculated as follows: ; in, is the true label, Predict probabilities for the model.

[0026] Specifically, the training process uses cross entropy loss and Adam optimizer (learning rate = 1×10 -4 ), through multiple iterative training, the model learns the pixel-level mapping relationship between remote sensing images and landslide labels, and finally achieves accurate segmentation of landslide areas. Figure 5 The training loss graph provided for this embodiment. The stable convergence of the loss function indicates that the network training is effective and can obtain a relatively reliable mapping relationship model between the input remote sensing image and the output landslide label.

[0027] The training set and test set are input into the above landslide segmentation model. After the model prediction, the corresponding landslide segmentation image is output, as shown in Figure 6 As shown in the figure, combined with the landslide labels, the model performs well on both the training set and the test set. The prediction results have clear boundaries, good internal integrity and continuity, and are consistent with the labels.

[0028] Through systematic statistical analysis of the prediction results, the model performance was quantitatively evaluated using three core evaluation indicators: accuracy, precision, and recall. Specific experimental data are shown in Table 1.

[0029] Table 1

[0030] Experimental results show that the multi-feature input CMUnet landslide segmentation model proposed in this paper shows significant advantages in both training and test sets. The three indicators of the training set all reach above 90% (accuracy 91.36%, precision 93.15%, recall rate 92.79%), and the three indicators of the test set all remain above 88% (accuracy 89.37%, precision 88.65%, recall rate 91.12%), indicating that the model has excellent generalization ability and strong adaptability to actual scenarios.

[0031] Each embodiment in this specification is described in a progressive manner. The same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple. For relevant parts, refer to the partial description of the method embodiment. The system and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0032] Those skilled in the art may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein may be implemented by electronic hardware, computer software, or a combination of both.

[0033] To clearly illustrate the interchangeability of hardware and software, the above descriptions have generally described the components and steps of each example by function. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Professionals may use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of the present invention.

[0034] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A landslide hazard detection method based on an improved remote sensing landslide image segmentation model, characterized in that: The following steps are involved: S1. Calculate the corresponding gradient tensor data based on the acquired original landslide image, and construct a multi-feature dataset based on the gradient tensor data; S2. Preprocessing the multi-feature dataset to generate a preprocessed feature set, and processing the preprocessed feature set using feature enhancement technology to generate an enhanced feature dataset; S3. Build a landslide detection model by combining the U-Net network architecture and the CBAM attention mechanism; S4. Combining the acquired landslide detection model training dataset, cross entropy loss, and Adam optimizer to complete iterative training of the network and generate an optimized landslide detection model; The original landslide image data and gradient tensor data are used as the input data of the training dataset, and the landslide label data is used as the output data of the training dataset; S5. Use the optimized landslide detection model to segment the remote sensing image to generate landslide hazard detection results, and perform multi-dimensional evaluation on the segmentation results to generate segmentation evaluation results of the optimized landslide detection model.

2. The landslide hazard detection method based on an improved remote sensing landslide image segmentation model according to claim 1, characterized in that: The corresponding gradient tensor data is calculated including: Calculate the horizontal gradient data of each pixel in the original remote sensing image based on the Sobel operator And the gradient data in the vertical direction ; And calculate the gradient magnitude of the pixel and the gradient direction of the pixel : ； Pixel-based gradient magnitude Construct a feature set of varying intensity based on the gradient direction of the pixel Construct a change direction feature set, and integrate the change intensity feature set and the change direction feature set to generate a multi-feature dataset.

3. The landslide hazard detection method based on an improved remote sensing landslide image segmentation model according to claim 1, characterized in that: Combining the U-Net network architecture and the CBAM attention mechanism to build a landslide detection model includes: The U-Net network architecture includes: input module, encoder, decoder, skip connection, and output module; Among them, the input module is used to: convert the enhanced feature dataset into a feature map; The encoder is used to extract features and compress spatial dimensions of the input feature map layer by layer to generate multi-scale high-dimensional semantic features, including multiple convolution blocks and maximum pooling layers; The decoder of the U-Net network architecture is optimized based on the CBAM attention mechanism to generate a CBAM attention decoder. The CBAM attention decoder is used to gradually restore the spatial resolution of the feature map and fuse multi-scale information to generate a segmentation mask of the same size as the input. It includes multiple upsampling blocks and convolution blocks. The CBAM attention mechanism performs attention screening on the enhanced feature dataset through the channel attention mechanism and the spatial attention mechanism; The CMUnet network architecture is built by combining the encoder and skip connection of the U-Net network architecture and the CBAM attention decoder, and the landslide detection model is generated based on the implementation data.

4. The landslide hazard detection method based on an improved remote sensing landslide image segmentation model according to claim 3, characterized in that: Perform attention screening on the enhanced feature dataset to enhance edge features and texture features related to landslide detection and suppress irrelevant features; ； ； Where, stands for Kronecker product; that is, the feature map output by the previous convolutional layer After entering the CBAM module, it first passes through the channel attention mechanism Mapped to the first feature map , and then Enter the spatial attention mechanism for input , after processing by the spatial attention mechanism, the second feature map is obtained ; Features after CBAM attention optimization Then enter the subsequent convolutional layer to gradually restore the spatial resolution of the feature map and generate a segmentation mask.

5. The landslide hazard detection method based on an improved remote sensing landslide image segmentation model according to claim 4, characterized in that: Channel Attention Mechanism Also includes: Channel Attention Mechanism Used to filter preset target features; ； Spatial Attention Mechanism Used to filter the location of preset target features; ； Where, represents the Sigmoid activation function, Represented as a convolutional layer operation, stands for Multilayer Perceptron; stands for average pooling; stands for max pooling.

6. The landslide hazard detection method based on an improved remote sensing landslide image segmentation model according to claim 5, characterized in that: Generate an optimized landslide detection model, including: Construct a training dataset for the landslide detection model. Combine the training dataset and cross-entropy loss to calculate the pixel-by-pixel error and generate the corresponding error gradient. Combine the error gradient and the Adam optimizer to modify the CMUnet network architecture weights to complete iterative training of the network and generate an optimized landslide detection model. Among them, the cross entropy loss function calculation formula is as follows: ； in, is the true label, Predict probabilities for the model.

Images