Landslide detection method and system based on multi-scale feature fusion

Abstract

The present invention discloses a landslide detection method and system based on multi-scale feature fusion, relating to the field of remote sensing. The method comprises: acquiring a remote sensing image of a target area; inputting the remote sensing image into a trained landslide detection model; extracting bottom-level, mid-level, and high-level semantic features in the encoder portion of the model based on a local geometry module, a residual learning module, and a global correlation module; and outputting a landslide segmentation result after fusing the different-scale features extracted in the encoder portion based on the local geometry module and the global correlation module. The local geometry module and the residual learning module are constructed based on a convolutional neural network, and the global correlation module is constructed based on a Transformer network. The landslide detection model is obtained by training remote sensing images labeled with landslides. This method not only fully extracts multi-level features but also effectively fuses multiple features, avoiding computational redundancy and gradient vanishing problems, thereby improving the accuracy of the landslide detection model.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of remote sensing, and particularly relates to a landslide detection method and system based on multi-scale feature fusion. BACKGROUND

[0002] Landslides not only bring serious losses to the national economy, but also pose a serious threat to people's lives and safety. Effective landslide detection is of great significance for post-disaster rescue, post-disaster reconstruction and vegetation restoration planning. Traditional landslide detection relies on field investigation, which is time-consuming and labor-intensive, and is difficult to meet the requirements of large-scale landslide mapping. With the development of science and technology, massive remote sensing data provides effective data support for large-scale landslide detection. Compared with traditional computer landslide interpretation methods, deep learning-based methods do not require human feature extraction and parameter evaluation, making end-to-end automated landslide detection possible. In recent years, the application of deep learning methods to landslide detection has greatly improved the landslide recognition rate.

[0003] The method of using deep learning for landslide detection includes post-disaster single-time landslide detection, which locates and detects areas with landslide features, reducing the workload to some extent and attracting much attention in practical applications. Although there are many methods for landslide detection using post-disaster single-time data, most of them currently focus on using convolutional neural networks for landslide detection and using deep networks and attention mechanisms to learn global features.

[0004] Among them, the convolutional neural network aggregates each pixel and its neighborhood pixels through convolution operation to mine robust local features. Due to the limited receptive field of convolution operation, it is difficult to mine rich landslide information from a global perspective. When using deep networks and attention mechanisms to learn global features, deep networks are prone to cause computational redundancy and gradient vanishing, while attention mechanisms tend to focus on large continuous regions, making it difficult to explore diverse landslide semantic information. Therefore, the current methods for landslide detection using post-disaster single-time data often have low accuracy. SUMMARY

[0005] The main purpose of the present application is to solve the problem of inaccurate landslide detection results caused by insufficient mining of local and global features of landslide remote sensing images.

[0006] The application provides a landslide detection method based on multi-scale feature fusion, comprising: acquiring a remote sensing image of a target area; inputting the remote sensing image into a trained landslide detection model, extracting low-level, middle-level and high-level semantic features based on a local geometry module, a residual learning module and an overall correlation module in an encoder part of the model, and outputting a landslide segmentation result after fusing different scale features extracted in the encoder based on the local geometry module and the overall correlation module in a decoder part; wherein the local geometry module and the residual learning module are constructed based on a convolutional neural network, the overall correlation module is constructed based on a Transformer network, and the landslide detection model is obtained after being trained according to a remote sensing image labeled with a landslide label.

[0007] In the landslide detection method based on multi-scale feature fusion of the application, the remote sensing image of the target area is acquired, comprising: acquiring multi-channel images of the target area about blue light bands, green light bands, red edge bands, normalized vegetation indexes, digital elevations, slope directions, slope gradients and mountain shadows; after pre-processing and channel merging of the multi-channel images, the remote sensing image is obtained; wherein the pre-processing includes size cropping and normalization.

[0008] In the landslide detection method based on multi-scale feature fusion of the application, the output result of the local geometry module after processing the input feature map comprises:

[0009] Y VGM = RL (BN (Conv (RL (BN (Conv (X))))))

[0010] Wherein, X is the input feature map, Conv is the convolution operation, BN is the batch normalization, RL is the linear rectifier unit activation function, Y VGM is the output result of the local geometry module.

[0011] In the landslide detection method based on multi-scale feature fusion of the application, the output result of the residual learning module after processing the input feature map comprises:

[0012] Y RLM = RL (BN (Conv (BN (Conv (BN (Conv (X)))))) + X)

[0013] Wherein, X is the input feature map, Conv is the convolution operation, BN is the batch normalization, RL is the linear rectifier unit activation function, Y RLM is the output result of the residual learning module.

[0014] In the landslide detection method based on multi-scale feature fusion of the application, the output result of the overall correlation module after processing the input feature map comprises:

[0015]

[0016] wherein,

[0017] wherein, X is an input feature map, MSA is a multi-head self-attention mechanism, LN is normalization, Conv is convolution operation, BN is batch normalization, RL is linear correction unit activation function, F TRM is the output result of the overall correlation module.

[0018] In the landslide detection method based on multi-scale feature fusion, the local geometry module, the residual learning module and the overall correlation module are used to extract the low-level, middle-level and high-level semantic features in the encoder part, including: inputting the remote sensing image into the local geometry module, the residual learning module and the overall correlation module respectively to obtain the first feature map, the second feature map and the third feature map; inputting the second feature map into the residual learning module again and inputting the third feature map into the overall correlation module again to obtain the fourth feature map and the fifth feature map.

[0019] Correspondingly, the local geometry module and the overall correlation module are used to fuse the features of different scales extracted in the encoder in the decoder part, including: fusing the first feature map, the second feature map, the third feature map, the fourth feature map and the fifth feature map in the decoder part based on the local geometry module and the overall correlation module.

[0020] In the landslide detection method based on multi-scale feature fusion, before the remote sensing image of the target area is obtained, the method further includes: obtaining a sample remote sensing image with a landslide label; inputting any sample remote sensing image into a landslide detection model constructed based on the local geometry module, the residual learning module and the overall correlation module, calculating the loss value between the output result of the sample remote sensing image and the landslide label, and updating the model parameters of the constructed landslide detection model based on the calculated loss value; repeating the process of inputting any sample remote sensing image into the constructed landslide detection model until the model parameters of the constructed landslide detection model are updated until the constructed landslide detection model meets the preset standard condition to obtain the trained landslide detection model; wherein the loss value is obtained by weighting the cross-entropy loss and the similarity coefficient loss.

[0021] The application provides a landslide detection system based on multi-scale feature fusion, comprising: an input module for acquiring a remote sensing image of a target area; a processing module for inputting the remote sensing image into a trained landslide detection model, extracting low-level, middle-level and high-level semantic features based on a local geometry module, a residual learning module and an overall correlation module in the encoder part of the model, and outputting a landslide segmentation result after fusing different scale features extracted in the encoder based on the local geometry module and the overall correlation module in the decoder part; wherein the local geometry module and the residual learning module are constructed based on a convolutional neural network, the overall correlation module is constructed based on a Transformer network, and the landslide detection model is obtained after being trained according to a remote sensing image labeled with a landslide label.

[0022] In the landslide detection system based on multi-scale feature fusion of the application, the input module is specifically used for: acquiring multi-channel images of the target area about blue light bands, green light bands, red edge bands, normalized vegetation indexes, digital elevations, slope directions, slope gradients and mountain shadows; obtaining the remote sensing image after preprocessing and channel merging of the multi-channel images; wherein the preprocessing includes size cropping and normalization.

[0023] The application has the beneficial effects that: through the convolutional neural network and the Transformer technology, the local geometry module, the residual learning module and the overall correlation module are set to construct the multi-scale feature fusion landslide detection model, the low-level and middle-level features of the landslide remote sensing image and the high-level global features are extracted from the model, the problem of limited receptive field of the traditional convolution operation is avoided, at the same time, the extracted local features and overall features are fused based on the overall correlation and the local geometry module, not only the multi-level features are fully extracted, but also the purpose of effectively fusing multiple features is achieved, the calculation redundancy and gradient disappearance problem caused by the deep network are avoided, and finally the overall precision of the landslide detection model is improved. BRIEF DESCRIPTION OF DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0025] Figure 1 is a flowchart of the landslide detection method based on multi-scale feature fusion of the embodiment of the application;

[0026] Figure 2 is a network structure diagram of the landslide detection model of the embodiment of the application;

[0027] Figure 3 are schematic diagrams of landslide detection results of high-resolution remote sensing images obtained by different methods of embodiments of the present application. DETAILED DESCRIPTION

[0028] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0029] The present application will be described below with reference to the accompanying drawings and embodiments. Figures 1 to 3 The present application provides a landslide detection method and system based on multi-scale feature fusion, Figure 1 is a flowchart of a landslide detection method based on multi-scale feature fusion of an embodiment of the present application, as Figure 1 indicated, the present application provides a landslide detection method based on multi-scale feature fusion, comprising:

[0030] S1, obtaining a remote sensing image of a target area;

[0031] S2, inputting the remote sensing image into a trained landslide detection model, extracting low-level, middle-level and high-level semantic features based on a local geometry module, a residual learning module and an overall correlation module in an encoder part of the model, and outputting a landslide segmentation result after fusing different scale features extracted in the encoder based on the local geometry module and the overall correlation module in a decoder part.

[0032] Among them, the local geometry module and the residual learning module are constructed based on a convolutional neural network, the overall correlation module is constructed based on a Transformer network, and the landslide detection model is obtained after being trained according to a remote sensing image labeled with a landslide label.

[0033] First, before S1, the present application pre-sets a local geometry module, a residual learning module and an overall correlation module using a convolutional neural network and Transformer technology, and then constructs a multi-scale feature fusion landslide detection model based on this.

[0034] In S1, after obtaining the remote sensing image of the target area in the data set, the corresponding pretreatment can be performed first, and then in S2, the pretreated remote sensing image is input into the constructed multi-scale feature fusion landslide detection model (referred to as landslide detection model) to obtain the detection result of the landslide.

[0035] For the processing of the model, in the encoder, the local geometry module and the residual learning module are constructed to learn the low-level and middle-level semantic information, and the overall correlation module is constructed to learn the high-level semantic information. In the decoder, the local geometry module and the overall correlation module are constructed to perform deep fusion on the features of different scales extracted in the encoder, learn diversified landslide features, and generate a landslide prediction result.

[0036] The local geometry module constructed based on convolution operation learns shallow features such as color, texture, edge, shape and position information. The residual network can effectively capture semantic features and avoid the gradient vanishing problem in the training process. The residual learning module constructed based on convolution operation is further proposed to mine middle-level semantic information. However, it is difficult for convolution operation to learn long-distance dependencies in a global range, while the Transformer can mine global features. Therefore, the overall correlation module is constructed based on the Transformer technology to mine global features.

[0037] The landslide detection method based on multi-scale feature fusion of the present application avoids the problem of limited receptive field of traditional convolution operation, and fuses the extracted local features and overall features based on the overall correlation and local geometry modules, which not only fully extracts multi-level features, but also effectively fuses multiple features, avoids the calculation redundancy and gradient vanishing problem caused by deep network, and finally improves the overall accuracy of the landslide detection model.

[0038] In an embodiment, the remote sensing image of the target area is obtained, including: obtaining multi-channel images of the target area about blue light band, green light band, red edge band, normalized vegetation index, digital elevation, slope direction, slope and mountain shadow; after pre-processing and channel merging of the multi-channel images, the remote sensing image is obtained; wherein the pre-processing includes size cropping and normalization.

[0039] Firstly, the normalized vegetation index is calculated by using the red light band and the near-infrared band in the optical image, and then the slope, the slope direction and the mountain shadow are extracted by using the digital elevation model, and combined with the images of the blue light band, the green light band and the red edge band, the above multi-channel images are obtained, and the specific description of each feature factor is shown in Table 1.

[0040] Table 1 Feature factor description

[0041]

[0042]

[0043] The extracted normalized vegetation index, slope, slope direction, mountain shadow, optical image and digital elevation model are merged in channels to become a single file.

[0044] The merged file is subjected to data normalization operation, in which five channels in the optical image are subjected to standard deviation normalization according to a unified maximum value and minimum value to retain spectral information as much as possible, and other characteristic factors are subjected to standard deviation normalization according to the maximum value and minimum value of the channel itself, data is regularized, and is scaled down to a preset range to the interval of 0 to 1. The normalized image is cropped into an image with a size of, for example, 128x128 pixels, and the cropping is performed with a 20% overlap to retain the background information of the image edge as much as possible, and finally a remote sensing image for inputting into a landslide detection model is obtained.

[0045] The landslide detection method based on multi-scale feature fusion in the embodiment of the application fuses optical features and spatial features (digital elevation, aspect, and slope features), and is beneficial to obtaining an accurate landslide detection result.

[0046] In one embodiment, the output result of the local geometry module (VGM) after processing the input feature map includes:

[0047] Y VGM = RL(BN(Conv(RL(BN(Conv(X)))))

[0048] wherein X is the input feature map, Conv is convolution operation, BN is batch normalization, RL is linear rectifier unit activation function, and Y VGM is the output result of the local geometry module.

[0049] Specifically, the preprocessed remote sensing image is input into the local geometry module for bottom layer local feature extraction. The local geometry module can be composed of convolution operation with a convolution kernel size of 3, batch normalization, and linear rectifier unit activation function. For a given input image X∈R H×W×C wherein H, W, and C represent height, width, and channel number respectively, the preset image size is 128x128x8, wherein 128 is the image height, 128 is the image width, and 8 is the image channel number, X is converted into a fine-grained feature map by using the local geometry module, and the specific operation process of the constructed geometry module is as shown in the above formula.

[0050] In one embodiment, the output result of the residual learning module (RLM) after processing the input feature map includes:

[0051] Y RLM = RL(BN(Conv(BN(Conv(BN(Conv(X))))))+X)

[0052] wherein X is the input feature map, Conv is convolution operation, BN is batch normalization, RL is linear rectifier unit activation function, and YRLM is the output result of the residual learning module.

[0053] The local geometry module focuses on extracting low-level features using fewer convolution operations, lacking semantic information. The residual learning module can efficiently capture semantic information, and the preprocessed remote sensing image is input into the residual learning module for middle-level local feature extraction. The residual learning module includes two convolution operations, batch normalization and linear rectification unit activation function. Its specific operation process is shown in the above formula.

[0054] The landslide detection method based on multi-scale feature fusion in the embodiment of the application can effectively capture fine-grained features, and the extracted features have high-precision spatial position information and rich texture information, having strong detection capability for small landslides.

[0055] In one embodiment, the output result of the whole correlation module (TRM) after processing the input feature map includes:

[0056]

[0057] wherein,

[0058] wherein, X is the input feature map, MSA is the multi-head self-attention mechanism, LN is the normalization, Conv is the convolution operation, BN is the batch normalization, RL is the linear rectification unit activation function, F TRM is the output result of the whole correlation module.

[0059] The local geometry module and the residual learning module based on convolution operation can learn long-distance dependencies in the global range, but it is difficult to learn long-distance dependencies in the global range. The Transformer has great advantages in global modeling. Therefore, the whole correlation module is constructed based on the Transformer technology to mine the global features of the landslide image. The preprocessed image is input into the whole correlation module for global feature extraction. For a given input image X H×W×C , the input image is adjusted to a series of sub-blocks by using the block embedding operation, and the operation is shown in the above formula . Then, normalization, multi-layer perception and skip connection operations are performed for feature processing, and finally the result of F TRM is obtained.

[0060] In one embodiment, the local geometry module, the residual learning module and the whole correlation module are used to extract low-level, middle-level and high-level semantic features in the encoder part, respectively, including:

[0061] The remote sensing image is input into a local geometry module, a residual learning module and a global correlation module respectively to obtain a first feature map, a second feature map and a third feature map respectively; the second feature map is input into the residual learning module again, and the third feature map is input into the global correlation module again to obtain a fourth feature map and a fifth feature map respectively; correspondingly, the fusion of different scale features extracted in the encoder based on the local geometry module and the global correlation module in the decoder part includes: the fusion of the first feature map, the second feature map, the third feature map, the fourth feature map and the fifth feature map based on the local geometry module and the global correlation module in the decoder part.

[0062] Figure 2 is a network structure diagram of a landslide detection model of an embodiment of the present application, as shown in Figure 2 In the embodiment of the present application, two residual learning modules are used for feature extraction, the output result of the first residual learning module is input into the second residual learning module, and two layers of feature outputs, i.e., the second feature map and the fourth feature map, are obtained. At the same time, two global correlation modules are used for feature extraction, the output result of the first global correlation module is input into the second global correlation module, and two layers of feature outputs, i.e., the third feature map and the fifth feature map, are obtained. Then, the five feature maps are fused in the decoder part through the local geometry module and the global correlation module. Based on this, the landslide detection method based on multi-scale feature fusion of the embodiment of the present application fully mines the features of each level of the remote sensing image, and the mining of local features and global features is significantly improved, while avoiding the problems of calculation redundancy and gradient disappearance caused by deep network.

[0063] In one embodiment, before the remote sensing image of the target area is obtained, the method further includes: obtaining a sample remote sensing image with a landslide label; inputting any sample remote sensing image into a landslide detection model constructed based on the local geometry module, the residual learning module and the global correlation module, calculating a loss value between the output result of the sample remote sensing image and the landslide label, and updating the model parameters of the constructed landslide detection model based on the calculated loss value; repeating the process of inputting any sample remote sensing image into the constructed landslide detection model until the model parameters of the constructed landslide detection model are updated until the constructed landslide detection model meets a preset standard condition to obtain the trained landslide detection model; wherein the loss value is obtained by weighting the cross-entropy loss and the similarity coefficient loss.

[0064] Before the landslide detection model in S2 is used, the landslide detection model needs to be trained. The cross-entropy loss function can not only measure the effect of the model, but also conveniently calculate the loss value, and is widely used in segmentation tasks. Based on the landslide label and the sample remote sensing image, the cross-entropy loss is used to calculate the prediction loss value, and the formula is as follows:

[0065]

[0066] Wherein, N is the number of samples, Y i,j represents the true label, P i,j represents the probability that the predicted sample belongs to landslide.

[0067] Because the cross-entropy loss is when the number of current scene pixels is far less than the number of background pixels, that is, the number of non-landslide pixels is much larger than the number of landslide pixels, the non-landslide pixels dominate in the loss function, so that the model is seriously biased towards non-landslide, resulting in poor model training effect. Therefore, the application also uses a similarity coefficient loss to solve the sample imbalance problem, and its formula is as follows:

[0068]

[0069] Wherein, p(x) represents the true label of landslide, and the value is 1, which represents landslide, and the value is 0, which represents non-landslide; g(x) represents the predicted probability of landslide.

[0070] The application combines the two kinds of losses calculated, and optimizes the constructed multi-scale feature fusion landslide detection model together, and the calculation formula of the loss value is as follows

[0071] L Total =L BCE +β·L DSC

[0072] Wherein, beta represents the weight coefficient of the similarity coefficient loss, and its value can be set to 0.5.

[0073] Then, the model parameters are updated by using back propagation, and the multi-scale feature fusion landslide detection model is optimized by reducing the loss value, and finally when the landslide detection model meets the preset standard reaching condition, it can participate in the detection of S1 and S2. The standard reaching condition can be set according to requirements, including but not limited to the accuracy, precision and recall rate of detection.

[0074] Next, the effect of the application will be further described in combination with a simulation experiment example.

[0075] (1) Simulation experiment conditions: The hardware test platform of the experiment is: Intel Core i7 processor, and the memory is 64GB deep learning machine. The software platform is: Linux operating system, Python 3.6.8. The calculation process is executed in the NVIDIACUDA Toolkit 10.1 environment using two NVIDIA GeForce RTX 2080Ti GPUs (12GB RAM). The present application uses RapidEye high-resolution remote sensing satellite images, the remote sensing image has a spatial resolution of 5 meters, has a red light band, a green light band, a blue light band, a near-infrared band and a red edge band, and the image size is 5000*5000 pixels; the digital elevation model is derived from the ALOS phased array L-band synthetic aperture radar (PALSAR) sensor, and the spatial resolution is 12.5m. In order to facilitate analysis, it is matched with the spatial resolution of the optical image, and resampled to 5 meters.

[0076] (2) Simulation content:

[0077] In this embodiment, the landslide detection method provided by the present application is compared with two traditional detection methods, which are UNet method and Deeplabv3 plus method, as follows:

[0078] During the experiment, the landslide inventory data prediction map obtained based on different methods is combined with the real landslide label to construct a confusion matrix, so as to calculate the overall accuracy, precision, recall, F1 score, Kappa coefficient and average intersection over union to quantitatively evaluate the performance of the present application method.

[0079] Figure 3 is a schematic diagram of the landslide detection result of the high-resolution remote sensing image obtained by different methods of the present application, as shown in Figure 3 , wherein (a) is a landslide remote sensing image, (b) is a real landslide label, (c) is a landslide detection map obtained by UNet, (d) is a landslide detection map obtained by Deeplabv3 plus, and (e) is a landslide detection map obtained by the method of the present application.

[0080] (3) Analysis of experimental results

[0081] Table 2 is the detection accuracy evaluation result of different methods, as shown in Table 2, the detection results of each method in Figure 3 can be evaluated from the objective evaluation index by Table 2.

[0082] Table 2

[0083]

[0084] Comprehensive Table 2 and Figure 3It can be seen that, since landslide detection is a sample distribution imbalance task, the overall precision of each detection method is relatively small. In terms of precision, the landslide prediction catalog obtained by the Deeplabv3 plus method has high precision, but the recall rate is only 45.68%, which means that a large number of landslides are not identified. In terms of recall rate, the method proposed in the present application can identify more landslides than other methods, which shows that the model of the present application greatly improves the identification rate of landslide detection. In the field of landslide detection, more attention is often paid to the recall rate of landslide detection than to the accuracy. The F1 score is the harmonic value of precision and recall, and is the most important parameter for evaluating the accuracy of the model in various competitions. In addition, the values of Kappa coefficient and average intersection over union are also significantly higher than those of other methods. Therefore, the landslide detection method based on multi-scale feature fusion proposed in the present application can significantly improve the identification accuracy of landslide detection.

[0085] The present application also provides a landslide detection system based on multi-scale feature fusion, comprising: an input module for obtaining a remote sensing image of a target area; a processing module for inputting the remote sensing image into a trained landslide detection model, extracting low-level, middle-level and high-level semantic features based on a local geometry module, a residual learning module and an overall correlation module in the encoder part of the model, and outputting a landslide segmentation result after fusing different scale features extracted in the encoder based on the local geometry module and the overall correlation module in the decoder part; wherein the local geometry module and the residual learning module are constructed based on a convolutional neural network, the overall correlation module is constructed based on a Transformer network, and the landslide detection model is obtained after being trained according to a remote sensing image labeled with a landslide label.

[0086] In one system embodiment, the input module is specifically configured to: obtain multi-channel images of the target area related to blue light bands, green light bands, red edge bands, normalized vegetation indexes, digital elevations, slope directions, slope gradients and mountain shadows; and obtain the remote sensing image after preprocessing and channel merging of the multi-channel images; wherein the preprocessing includes size cropping and normalization.

[0087] The landslide detection system based on multi-scale feature fusion provided in the embodiments of the present application has the same implementation principle, technical effects and the foregoing landslide detection method based on multi-scale feature fusion. For brevity, the parts of the landslide detection system based on multi-scale feature fusion that are not mentioned in the foregoing landslide detection method based on multi-scale feature fusion can be referred to the corresponding contents in the foregoing landslide detection method based on multi-scale feature fusion.

[0088] The application also provides a computer readable storage medium, such as a flash memory, a hard disk, a multimedia card, a card memory (for example, an SD or DX memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a magnetic disk, an optical disk, a server, an App application store, and the like, which stores a computer program, and the program is executed by a processor to realize corresponding functions. The computer readable storage medium of the embodiment is executed by the processor to realize the method embodiment based on the landslide detection method of multi-scale feature fusion.

[0089] In summary, the method of the application can fully combine the advantages of convolutional neural network and Transformer technology, design local geometric module, residual learning module and overall correlation module, construct multi-scale feature fusion landslide detection network, thereby effectively improve the accuracy and ability of landslide recognition, solve the problems of lack of global information and low landslide detection precision in the prior art method, and is a very practical and effective high-resolution image landslide detection method.

[0090] It should be noted that, according to the needs of implementation, each step / component described in the application can be split into more steps / components, or two or more steps / components or part of the operation of the steps / components can be combined into a new step / component, to realize the purpose of the application.

[0091] The size of the serial number of each step in the above embodiment does not mean the order of execution, the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiment of the application.

[0092] It should be understood that those skilled in the art can improve or change according to the above description, and all these improvements and changes should belong to the protection scope of the appended claims of the application.

Claims

1. A landslide detection method based on multi-scale feature fusion, characterized in that: include: Acquire remote sensing images of the target area; Inputting the remote sensing image into the trained landslide detection model, extracting bottom-level, middle-level, and high-level semantic features based on a local geometry module, a residual learning module, and an overall correlation module in the encoder part of the model, and fusing the different scale features extracted from the encoder based on the local geometry module and the overall correlation module in the decoder part to output a landslide segmentation result; Among them, the local geometry module and the residual learning module are constructed based on convolutional neural networks, the overall correlation module is constructed based on the Transformer network, and the landslide detection model is obtained after training based on remote sensing images marked with landslide labels; the local geometry module includes a convolution operation with a convolution kernel size of 3, batch normalization and a linear correction unit activation function; the residual learning module includes two convolution operations, batch normalization and a linear correction unit activation function; the overall correlation module is used to use block embedding operations to adjust the input image into a series of sub-blocks, and then implement normalization, multi-layer perception and jump connection operations for feature processing.

2. The landslide detection method based on multi-scale feature fusion according to claim 1 is characterized in that: The step of acquiring a remote sensing image of a target area includes: Obtain multi-channel images of the target area including blue band, green band, red edge band, normalized difference vegetation index, digital elevation, aspect, slope and hill shadow; After preprocessing and channel merging the multi-channel image, the remote sensing image is obtained; The preprocessing includes size cropping and normalization.

3. The landslide detection method based on multi-scale feature fusion according to claim 1 is characterized in that: The output result after the local geometry module processes the input feature map includes: Among them, X is the input feature map, Conv is the convolution operation, BN is batch normalization, RL is the linear rectification unit activation function, It is the output result of the local geometry module.

4. The landslide detection method based on multi-scale feature fusion according to claim 1 is characterized in that: The output result after the residual learning module processes the input feature map includes: Among them, X is the input feature map, Conv is the convolution operation, BN is batch normalization, RL is the linear rectification unit activation function, is the output result of the residual learning module.

5. The landslide detection method based on multi-scale feature fusion according to claim 1 is characterized in that: The output result after the overall correlation module processes the input feature map includes: in, , X is the input feature map, MSA is the multi-head self-attention mechanism, LN is normalization, Conv is the convolution operation, BN is batch normalization, RL is the linear rectification unit activation function, Output results for the overall relevant module.

6. The landslide detection method based on multi-scale feature fusion according to claim 1 is characterized in that: The encoder part extracts low-level, mid-level, and high-level semantic features based on the local geometry module, residual learning module, and overall correlation module, including: Inputting the remote sensing image into a local geometry module, a residual learning module, and an overall correlation module respectively to obtain a first feature map, a second feature map, and a third feature map; Input the second feature map into the residual learning module again, and input the third feature map into the overall correlation module again, to obtain a fourth feature map and a fifth feature map respectively; Accordingly, the decoder part fuses the different scale features extracted from the encoder based on the local geometry module and the global correlation module, including: In the decoder part, the first feature map, the second feature map, the third feature map, the fourth feature map and the fifth feature map are fused based on the local geometry module and the global correlation module.

7. The landslide detection method based on multi-scale feature fusion according to claim 1 is characterized in that: Before acquiring the remote sensing image of the target area, the method further includes: Obtain sample remote sensing images with landslide labels; Inputting any sample remote sensing image into the landslide detection model constructed based on the local geometry module, the residual learning module and the overall correlation module, calculating the loss value between the sample remote sensing image output result and the landslide label, and updating the model parameters of the constructed landslide detection model based on the calculated loss value; Repeating the process of inputting any sample remote sensing image into the constructed landslide detection model to update the model parameters of the constructed landslide detection model until the constructed landslide detection model meets the preset compliance conditions, thereby obtaining the trained landslide detection model; The loss value is obtained by weighting the cross entropy loss and the similarity coefficient loss.

8. A landslide detection system based on multi-scale feature fusion, characterized in that: include: An input module is used to obtain remote sensing images of the target area; a processing module for inputting the remote sensing image into a trained landslide detection model; extracting bottom-level, middle-level, and high-level semantic features based on a local geometry module, a residual learning module, and an overall correlation module in an encoder portion of the model; and fusing the different-scale features extracted from the encoder based on the local geometry module and the overall correlation module in a decoder portion to output a landslide segmentation result; Among them, the local geometry module and the residual learning module are constructed based on convolutional neural networks, the overall correlation module is constructed based on the Transformer network, and the landslide detection model is obtained after training based on remote sensing images marked with landslide labels; the local geometry module includes a convolution operation with a convolution kernel size of 3, batch normalization and a linear correction unit activation function; the residual learning module includes two convolution operations, batch normalization and a linear correction unit activation function; the overall correlation module is used to use block embedding operations to adjust the input image into a series of sub-blocks, and then implement normalization, multi-layer perception and jump connection operations for feature processing.

9. The landslide detection system based on multi-scale feature fusion according to claim 8, characterized in that: The input module is specifically used for: Obtain multi-channel images of the target area including blue band, green band, red edge band, normalized difference vegetation index, digital elevation, aspect, slope and hill shadow; After preprocessing and channel merging the multi-channel image, the remote sensing image is obtained; The preprocessing includes size cropping and normalization.

10. A computer storage medium, characterized in that A computer program executable by a processor is stored therein, and the computer program executes the landslide detection method based on multi-scale feature fusion according to any one of claims 1 to 7.

Images