Landslide detection method, device, program product and electronic device

Abstract

The present disclosure provides a landslide detection method, device, program product and electronic equipment, and relates to the technical field of computers. The method comprises the following steps: determining a to-be-detected remote sensing image, then performing feature extraction on the to-be-detected remote sensing image through a backbone network of a landslide detection model to obtain a plurality of initial feature maps with different sizes; performing fusion processing on the plurality of initial feature maps with different sizes through a neck network of the landslide detection model to obtain a plurality of fused feature maps with different sizes; performing feature processing on the plurality of fused feature maps with different sizes through a receptive field attention submodule of the landslide detection model to obtain target receptive field features; and performing detection processing on the target receptive field features through a detection network of the landslide detection model to obtain a landslide detection frame and a landslide category label, thereby improving the accuracy of detecting landslides in remote sensing images to a certain extent.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of computer, and particularly relates to a landslide detection method, device, program product and electronic device. BACKGROUND

[0002] Landslide is a natural geological disaster caused by earthquakes, rainfall, human activities and other factors, which seriously threatens human life and property safety. Therefore, how to quickly and accurately detect landslides to provide great convenience and help for disaster relief and governance has become a problem to be solved.

[0003] In optical images, landslides usually exhibit the following characteristics due to the destruction of the original terrain integrity: (1) partially covered bare soil by vegetation; (2) broken ground surface; (3) rough and uneven texture; (4) discontinuous landscape pattern. Moreover, over time, the distinctive surface features of the aforementioned landslides gradually weaken due to factors such as vegetation regeneration and surface weathering. For such landslides, the existing detection methods mainly extract landslide features from optical images and rely on the professional knowledge and experience of landslide experts to manually analyze remote sensing images to delineate the location and boundary of landslides to determine the landslide detection results. SUMMARY

[0004] However, the aforementioned method of manually analyzing to determine the location and boundary of landslides is not only time-consuming and laborious, but the landslide detection results determined based on this method are affected by the subjective judgment of individual experts, resulting in low accuracy of landslide judgment.

[0005] The present disclosure provides a landslide detection method, device, computer program product and electronic device to improve the accuracy of landslide detection.

[0006] According to a first aspect of the present disclosure, a landslide detection method is provided, the method comprising:

[0007] determining a remote sensing image to be detected;

[0008] extracting features of the remote sensing image to be detected through a backbone network of a landslide detection model to obtain a plurality of initial feature maps with different sizes; and performing fusion processing on the plurality of initial feature maps with different sizes through a neck network of the landslide detection model to obtain a plurality of fusion feature maps with different sizes;

[0009] performing feature processing on the plurality of fusion feature maps with different sizes through a receptive field attention sub-module of the landslide detection model to obtain target receptive field features;

[0010] performing detection processing on the target receptive field features through a detection network of the landslide detection model to obtain a landslide detection box and a landslide class label.

[0011] In a possible implementation, the receptive field attention submodule of the landslide detection model is used to perform feature processing on the plurality of fused feature maps of different sizes to obtain target receptive field features, including:

[0012] The plurality of fused feature maps of different sizes are input into a first sub-branch of the receptive field attention submodule to obtain a first feature matrix;

[0013] The plurality of fused feature maps of different sizes are input into a second sub-branch of the receptive field attention submodule to obtain a second feature matrix;

[0014] The second feature matrix is input into a third sub-branch of the receptive field attention submodule to obtain a third feature matrix;

[0015] The target receptive field features are determined according to the first feature matrix and the third feature matrix.

[0016] In a possible implementation, the plurality of fused feature maps of different sizes are input into a first sub-branch of the receptive field attention submodule to obtain a first feature matrix, including:

[0017] The plurality of fused feature maps of different sizes are input into a nonlinear activation function of the receptive field attention submodule to obtain a channel attention vector of original features corresponding to the fused feature maps, and the channel attention vector is taken as the first feature matrix.

[0018] In a possible implementation, the plurality of fused feature maps of different sizes are input into a second sub-branch of the receptive field attention submodule to obtain a second feature matrix, including:

[0019] Global average pooling and one-dimensional convolution processing are performed on the plurality of fused feature maps of different sizes to obtain a first sub-feature matrix;

[0020] Different sizes of convolution kernels and dilated convolution are used to process the first sub-feature matrix to obtain a second feature matrix.

[0021] In a possible implementation, the target receptive field features are determined according to the first feature matrix and the third feature matrix, including:

[0022]

[0023] wherein, X out are used to represent the target receptive field features; wherein, , X F are used to represent the second feature matrix, XUsed to characterize the feature matrix corresponding to the fused feature map; X a Used to characterize the third characteristic matrix; Used to characterize the first characteristic matrix; Re LU ( X a ) is used to represent the application of the ReLU activation function to each element in the third feature matrix.

[0024] In a possible implementation, the landslide detection model is trained based on the following method:

[0025] Collecting a remote sensing image data set and dividing the remote sensing image data set into a training set, a validation set, and a test set;

[0026] Training a preset landslide detection model according to the training set to obtain multiple prediction information;

[0027] Determining comparison information between the plurality of prediction information and the labeled information in the training set through a loss function, and adjusting the preset landslide detection model according to the comparison information to obtain an adjusted preset landslide detection model;

[0028] Adjusting the hyperparameters in the adjusted preset landslide detection model according to the validation set, and obtaining an updated preset landslide detection model when it is determined that the validation set loss reaches a first preset condition;

[0029] The updated preset landslide detection model is evaluated according to the test set, and when it is determined that the evaluation result meets the second preset condition, the landslide detection model is obtained.

[0030] In one possible implementation, the loss function is determined based on the following formula:

[0031]

[0032] in, N 1 Used to represent prediction information, N 2 Used to represent annotation information, C 2 A constant used to characterize the relationship between the training set and the training set. Used to characterize the second-order Wasserstein distance between the predicted information and the labeled information; NWD ( N 1 , N 2 ) is used to characterize the similarity between the predicted information and the labeled information.

[0033] According to a second aspect of the present disclosure, a landslide detection device is provided, the device comprising:

[0034] A determination unit, used for determining a remote sensing image to be detected;

[0035] The detection unit is used to perform feature extraction on the remote sensing image to be detected through the backbone network of the landslide detection model to obtain multiple initial feature maps of different sizes; and to fuse the multiple initial feature maps of different sizes through the neck network of the landslide detection model to obtain multiple fused feature maps of different sizes; to perform feature processing on the multiple fused feature maps of different sizes through the receptive field attention submodule of the landslide detection model to obtain target receptive field features; and to detect and process the target receptive field features through the detection network of the landslide detection model to obtain a landslide detection frame and a landslide category label.

[0036] In a possible implementation, the detection unit is specifically configured to:

[0037] Inputting the multiple fused feature maps of different sizes into the first sub-branch of the receptive field attention sub-module to obtain a first feature matrix;

[0038] Inputting the multiple fused feature maps of different sizes into the second sub-branch of the receptive field attention sub-module to obtain a second feature matrix;

[0039] Input the second feature matrix into the third sub-branch of the receptive field attention sub-module to obtain a third feature matrix;

[0040] Determine target receptive field features based on the first feature matrix and the third feature matrix.

[0041] In a possible implementation, the detection unit is specifically configured to:

[0042] The multiple fused feature maps of different sizes are input into the nonlinear activation function of the receptive field attention sub-module to obtain the channel attention vector of the original feature corresponding to the fused feature map, and the channel attention vector is used as the first feature matrix.

[0043] In a possible implementation, the detection unit is specifically configured to:

[0044] Performing global average pooling and one-dimensional convolution processing on the multiple fused feature maps of different sizes to obtain a first sub-feature matrix;

[0045] The first sub-feature matrix is ​​processed using convolution kernels of different sizes and dilated convolution to obtain a second feature matrix.

[0046] In a possible implementation, the detection unit is specifically configured to:

[0047]

[0048] in, X out Used to characterize the target receptive field characteristics; among them, , X F Used to characterize the second characteristic matrix, X Used to characterize the feature matrix corresponding to the fused feature map; X a Used to characterize the third characteristic matrix; Used to characterize the first characteristic matrix; Re LU ( X a ) is used to represent the application of the ReLU activation function to each element in the third feature matrix.

[0049] In a possible implementation, the device further includes a training unit configured to:

[0050] Collecting a remote sensing image data set and dividing the remote sensing image data set into a training set, a validation set, and a test set;

[0051] Training a preset landslide detection model according to the training set to obtain multiple prediction information;

[0052] Determining comparison information between the plurality of prediction information and the labeled information in the training set through a loss function, and adjusting the preset landslide detection model according to the comparison information to obtain an adjusted preset landslide detection model;

[0053] Adjusting the hyperparameters in the adjusted preset landslide detection model according to the validation set, and obtaining an updated preset landslide detection model when it is determined that the validation set loss reaches a first preset condition;

[0054] The updated preset landslide detection model is evaluated according to the test set, and when it is determined that the evaluation result meets the second preset condition, the landslide detection model is obtained.

[0055] In one possible implementation, the loss function is determined based on the following formula:

[0056]

[0057] in, N 1 Used to represent prediction information, N 2 Used to represent annotation information, C 2A constant used to characterize the relationship between the training set and the training set. Used to characterize the second-order Wasserstein distance between the predicted information and the labeled information; NWD ( N 1 , N 2 ) is used to characterize the similarity between the predicted information and the labeled information.

[0058] According to a third aspect of the present disclosure, a computer program product is provided, comprising a computer program, wherein when the computer program is executed by a processor, the method of the first aspect and possible implementations thereof are implemented.

[0059] According to a fourth aspect of the present disclosure, an electronic device is provided, comprising: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to execute the method of the above-mentioned first aspect and its possible implementation methods by executing the executable instructions.

[0060] The technical solution disclosed in this disclosure has the following beneficial effects:

[0061] In an embodiment of the present disclosure, a remote sensing image to be detected can be determined, and then feature extraction can be performed on the remote sensing image to be detected through the backbone network of the landslide detection model to obtain multiple initial feature maps of different sizes; and, the multiple initial feature maps of different sizes can be fused through the neck network of the landslide detection model to obtain multiple fused feature maps of different sizes; the receptive field attention submodule of the landslide detection model can be used to perform feature processing on the multiple fused feature maps of different sizes to obtain target receptive field features; and the target receptive field features can be detected and processed through the detection network of the landslide detection model to obtain a landslide detection frame and a landslide category label. It can be seen that the present disclosure adopts a landslide detection model, which can not only automatically detect the remote sensing image to be detected, avoiding errors caused by manual analysis and improving the accuracy of detection, but also combines the receptive field attention submodule to perform landslide detection on the remote sensing image to be detected based on the features determined by deep mining of the remote sensing image to be detected, further improving the accuracy of detection of the remote sensing image to be detected.

[0062] Other features and advantages of the present disclosure will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present disclosure. The purposes and other advantages of the present disclosure can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following briefly introduces the drawings required for use in the embodiments of the present disclosure. Obviously, the drawings introduced below are only some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0064] Figure 1 A schematic diagram of an application scenario in this exemplary embodiment is shown;

[0065] Figure 2 A schematic diagram showing a process of obtaining a landslide detection model in this exemplary embodiment is shown;

[0066] Figure 3 A schematic diagram showing a remote sensing image in an initial data set according to this exemplary embodiment;

[0067] Figure 4 A schematic diagram showing a remote sensing image in a processed data set according to this exemplary embodiment;

[0068] Figure 5 A schematic flow chart showing a landslide detection method in this exemplary embodiment is shown;

[0069] Figure 6 A schematic diagram showing the structure of an efficient channel attention model in this exemplary embodiment is shown;

[0070] Figure 7 A schematic diagram showing a landslide detection model in this exemplary embodiment;

[0071] Figure 8 A schematic structural diagram of a landslide detection device according to this exemplary embodiment is shown;

[0072] Figure 9 A schematic structural diagram of an electronic device in this exemplary embodiment is shown. DETAILED DESCRIPTION

[0073] In order to make the purpose, technical solutions and advantages of the present disclosure clearer, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, rather than all of the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present disclosure. In the absence of conflict, the embodiments in the present disclosure and the features in the embodiments can be arbitrarily combined with each other. In addition, although the logical order is shown in the flow chart, in some cases, the steps shown or described can be performed in an order different from that here.

[0074] The terms "comprises" and "comprising" and any variations thereof in the specification and claims of this disclosure are intended to cover non-exclusive protection. For example, a process, method, system, product, or apparatus comprising a series of steps or elements is not limited to the listed steps or elements, but may optionally include steps or elements not listed, or may optionally include other steps or elements inherent to the process, method, product, or apparatus.

[0075] In the embodiments of the present disclosure, one or more, "many" refers to two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent: the existence of A alone, the existence of A and B at the same time, and the existence of B alone, where A and B can be singular or plural. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, c can be single or plural.

[0076] It should be noted that the terms "first", "second", "third", etc. in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order, sequence, size and priority. For example, the first sub-branch, the second sub-branch and the third sub-branch in the embodiment of the present disclosure are only used to distinguish different branches. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0077] The following describes exemplary embodiments of the present disclosure in conjunction with the accompanying drawings. The accompanying drawings are schematic diagrams of the present disclosure and are not necessarily drawn to scale. Some of the block diagrams shown in the accompanying drawings may be functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, or in hardware modules or integrated circuits, or in networks, processors or microcontrollers. The embodiments can be implemented in various forms and should not be construed as being limited to the examples set forth herein. The features, structures or characteristics described in the present disclosure can be combined in one or more embodiments in any suitable manner. In the description below, many specific details are provided to provide a full description of the embodiments of the present disclosure. However, those skilled in the art will appreciate that one or more specific details may be omitted when implementing the technical solutions of the present disclosure, or that other methods, components, devices, steps, etc. may be used to replace one or more specific details.

[0078] It should be noted that in the embodiments of this disclosure, certain software, components, models, and other existing industry solutions may be mentioned. These should be considered exemplary and are intended solely to illustrate the feasibility of implementing the technical solutions of this disclosure. This does not mean that the applicant has or will necessarily use such solutions. The collection, dissemination, and use of data in this technical solution comply with relevant national laws and regulations.

[0079] In recent years, deep learning methods have been increasingly applied to remote sensing, including the automated detection of landslides. This process typically involves two steps: first, landslide areas are labeled based on prior knowledge, and then the labeled data is used to train a deep learning model to detect landslides in other areas.

[0080] Specifically, to improve detection performance, a context enhancement module composed of a deconvolution is introduced in the regression branch of the decoupled head, and the improved decoupled head is used to replace the coupled head in YOLOv5s to improve the effectiveness of the model. In addition, by using a multi-scale attention model based on Swin Transformer (a visual model based on Transformer architecture, which performs well in computer vision tasks such as image classification, object detection, etc., and its core feature is to improve the computational efficiency and the ability to capture local features through the shifted window mechanism), it can comprehensively capture and learn multi-scale landslide features. At the same time, the Multi-Scale Feature Fusion Lightweight Neural Network (MFFLnet) adopts a Deep Transfer Learning (DTL) strategy, which enables MFFLnet to utilize prior landslide knowledge from the source domain to alleviate overfitting.

[0081] However, these models mentioned above show inconsistent performance when dealing with small targets, objects with extreme shapes (such as thin, narrow or high), and datasets containing different target types in the same class.

[0082] Therefore, the current research also proposes corresponding methods for specific landslide challenges. First, morphological edge recognition is combined with the Swin Transformer deep learning model to enhance the delineation of boundaries, which can effectively solve the problem of irregular boundaries and feature discretization in landslide detection.

[0083] However, due to the existence of interference factors such as vegetation coverage, the model often has difficulty in fully understanding the optical characteristics of the landslide. To solve the above problems, the texture features of high-resolution images can be used, combined with auxiliary features such as Normalized Difference Vegetation Index (NDVI) and Gray-Level Co-Occurrence Matrix (GLCM), to improve the performance of the model.

[0084] However, these methods are only targeted at specific areas, and the performance of large-scale landslide detection has not been fully verified, which has great limitations.

[0085] To explore more effective landslide detection solutions, a network that simultaneously focuses on local and global features of landslides is developed by combining the advantages of convolutional neural networks (CNNs) and transformer models. Despite their advantages, transformer-based models and their improved variants typically require large datasets and long training times to achieve optimal performance.

[0086] In view of this, an exemplary embodiment of the present disclosure provides a landslide detection method, by which a remote sensing image to be detected can be determined, and then feature extraction is performed on the remote sensing image to be detected through the backbone network of the landslide detection model to obtain multiple initial feature maps of different sizes; and the multiple initial feature maps of different sizes are fused and processed through the neck network of the landslide detection model to obtain multiple fused feature maps of different sizes; the multiple fused feature maps of different sizes are feature processed through the receptive field attention submodule of the landslide detection model to obtain target receptive field features; and the target receptive field features are detected and processed through the detection network of the landslide detection model to obtain a landslide detection frame and a landslide category label. It can be seen that the present disclosure adopts a landslide detection model, which can not only automatically detect the remote sensing image to be detected, avoiding errors caused by manual analysis and improving detection accuracy, but also combines the receptive field attention submodule to perform landslide detection on the remote sensing image to be detected based on the features determined by deep mining of the remote sensing image to be detected, further improving the accuracy of detection of the remote sensing image to be detected.

[0087] To better understand the technical solutions provided by the embodiments of the present disclosure, the following briefly introduces the application scenarios to which the technical solutions provided by the embodiments of the present disclosure are applicable. It should be noted that the application scenarios introduced below are only used to illustrate the embodiments of the present disclosure and are not limiting. In specific implementations, the technical solutions provided by the embodiments of the present disclosure can be flexibly applied according to actual needs.

[0088] See also Figure 1 As shown, Figure 1 This is an application scenario to which the technical solutions of the embodiments of the present disclosure can be applied. In this scenario diagram, a terminal device 110 and an electronic device 120 are included. The terminal device 110 and the electronic device 120 can be directly or indirectly connected to each other via one or more networks 130. Optionally, the application scenario may also include other electronic devices, which can, for example, perform linked processing based on the output results of the electronic device 120, but this is not limited in the embodiments of the present disclosure.

[0089] In the embodiment of the present disclosure, the user can send a to-be-detected remote sensing image based on the terminal device 110, and then the electronic device 120 can determine the to-be-detected remote sensing image; the to-be-detected remote sensing image is subjected to feature extraction through a backbone network of the landslide detection model to obtain a plurality of initial feature maps of different sizes; and the plurality of initial feature maps of different sizes are subjected to fusion processing through a neck network of the landslide detection model to obtain a plurality of fusion feature maps of different sizes; the plurality of fusion feature maps of different sizes are subjected to feature processing through a receptive field attention sub-module of the landslide detection model to obtain a target receptive field feature; and the target receptive field feature is subjected to detection processing through a detection network of the landslide detection model to obtain a landslide detection frame and a landslide category label.

[0090] In the embodiment of the present disclosure, Figure 1 The terminal device 110 can be a mobile phone, a tablet computer (PAD), a personal computer (PC), a smart television, a smart watch, a smart speaker, a smart vehicle device, a wearable device, or the like, but is not limited thereto.

[0091] In the embodiment of the present disclosure, the electronic device 120 can be a server, which can be a standalone physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server or a cloud server cluster providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and basic cloud computing services such as big data and artificial intelligence platforms, but is not limited thereto.

[0092] Of course, the method provided by the embodiment of the present disclosure is not limited to the application scenarios shown in Figure 1 The embodiment of the present disclosure does not limit the application scenarios, for example, only the electronic device 120 implements the landslide detection method.

[0093] To further illustrate the technical solutions provided by the embodiments of the present disclosure, the following will be described in detail in conjunction with the drawings and specific embodiments. Although the embodiments of the present disclosure provide the following method operation steps as shown in the embodiments or drawings, more or fewer operation steps can be included in the method based on conventional or non-creative labor. The execution order of the steps is not limited to the execution order provided by the embodiments of the present disclosure in the logical sense. The method can be executed in sequence or in parallel during actual processing or device execution.

[0094] The following will be described in detail in conjunction with Figure 5The method flow chart shown in FIG. 1 illustrates the landslide detection method in the embodiment of the present disclosure. Figure 5 The steps shown can be performed as follows Figure 1 The electronic device 120 is shown performing.

[0095] In the embodiment of the present disclosure, it is possible to determine the collection of a remote sensing image dataset, divide the remote sensing image dataset into a training set, a validation set, and a test set, and then use the training set, the validation set, and the test set to train a preset landslide detection model to obtain a landslide detection model. The landslide detection model is then used to detect new remote sensing images to be detected. Before introducing the landslide detection method, the following first introduces the process of obtaining the landslide detection model in the embodiment of the present disclosure. Please refer to Figure 2 , Figure 2 Schematic diagram of the process of obtaining a landslide detection model in an embodiment of the present disclosure.

[0096] Step 201: Collect a remote sensing image dataset and divide the remote sensing image dataset into a training set, a validation set, and a test set.

[0097] In the disclosed embodiments, remote sensing image datasets can be collected from different regions. For example, remote sensing image datasets from three regions can be collected: the Bijie landslide dataset, the Luding landslide dataset, and the Southwest landslide dataset. The first three datasets are derived from different satellite imagery and cover landslide events in different regions. In other words, the remote sensing image datasets in the disclosed embodiments have a relatively rich foundation of data, providing a good foundation for subsequent identification of landslide events in different regions.

[0098] See also Figure 3 As shown, Figure 3 A schematic diagram of a remote sensing image of a Bijie landslide dataset provided by an embodiment of the present disclosure. Specifically, Figure 3 The Bijie landslide dataset shown here is from the Bijie Landslide Dataset published by Wuhan University. The dataset covers the entire city of Bijie, Guizhou Province, China, with a total area of ​​26,853 square kilometers. The area has an average elevation of 1,600 meters and is surrounded by mountains and rivers. Due to its steep slopes, fragile ecosystem, and unstable geology, the region is one of the most landslide-prone areas in China. The dataset contains 770 optical remote sensing images of landslide samples, acquired by the TripleSat satellite between May and August 2018. The images are composed of three primary RGB color bands and have a resolution of 0.8 meters.

[0099] In the embodiments of the present disclosure, when the remote sensing data image set is obtained, considering that the sizes of the images in the remote sensing data image set are different, the data in the remote sensing data image set can be standardized to ensure that the mean of the input data is 0 and the standard deviation is 1. Optionally, all samples can be uniformly adjusted to 320x320 pixels using a letterbox filling strategy. The letterbox filling strategy is an image processing technology for adjusting images of different sizes to a fixed size while maintaining the original aspect ratio of the image to avoid image geometric distortion. For example, see Figure 4 Figure 4 The remote sensing image in the remote sensing data image set adjusted according to the letterbox strategy is shown in

[0100] Optionally, the remote sensing data image set can be randomly divided into three groups according to a preset ratio for training, verification and testing, that is, a training set, a verification set and a test set are obtained. The preset ratio is, for example, 8:1:1, and of course, other ratios are also possible, which are not limited in the embodiments of the present disclosure.

[0101] Step 202: training the preset landslide detection model according to the training set to obtain a plurality of prediction information.

[0102] In the embodiments of the present disclosure, the preset landslide detection model can be based on a YOLOv11 model, which introduces a receptive field attention, a normalized Wasserstein distance, a CrossStage Partial with Pyramid Squeeze Attention (C2PSA) module, and an improved YOLOv11. That is, the preset landslide detection model provided in the embodiments of the present disclosure retains the advantages of the YOLOv11 model while reconstructing the network structure to enhance the feature extraction and small target recognition capability.

[0103] Step 203: determining the comparison information of the plurality of prediction information and the labeled information in the training set through a loss function, and adjusting the preset landslide detection model according to the comparison information to obtain an adjusted preset landslide detection model.

[0104] In one possible implementation, the loss function is determined based on the following formula one:

[0105] Formula one

[0106] wherein, N 1 is used to represent the prediction information, N 2 is used to represent the labeled information, C ​2 A constant used to characterize the relationship between the training set and Used to characterize the second-order Wasserstein distance between the predicted information and the labeled information; NWD ( N 1 , N 2 ) is used to characterize the similarity between the predicted information and the labeled information. Optionally, Usually set to the mean absolute value of the training set.

[0107] In the disclosed embodiment, NWD first models the bounding box as a two-dimensional Gaussian distribution, and then uses the normalized Wasserstein distance to measure the similarity of the derived Gaussian distributions.

[0108] Specifically, given two two-dimensional Gaussian distributions and , the second-order Wasserstein distance between them is defined as:

[0109] Formula 2.

[0110] To convert this distance into a similarity metric between 0 and 1, NWD uses an exponentially normalized loss function. In the disclosed embodiments, this loss function design makes NWD insensitive to targets of varying scales, making it more suitable for measuring the similarity between the true and predicted values ​​of small targets in remote sensing imagery within the model. In other words, the preset landslide detection model in this disclosure utilizes a normalized Wasserstein distance loss instead of the traditional intersection-over-union loss, improving the preset landslide detection model's ability to distinguish between positive and negative samples and promoting more efficient convergence of the preset landslide detection model.

[0111] Step 204: adjusting the hyperparameters in the adjusted preset landslide detection model according to the validation set, and obtaining an updated preset landslide detection model when it is determined that the validation set loss reaches a first preset condition.

[0112] Step 205: Evaluate the updated preset landslide detection model according to the test set, and when it is determined that the evaluation result meets the second preset condition, obtain the landslide detection model.

[0113] In the embodiment of the present disclosure, the first preset condition and the second preset condition can be set accordingly based on actual implementation, and this is not limited in the embodiment of the present disclosure.

[0114] Furthermore, in the embodiment of the present disclosure, after obtaining the landslide detection model, the remote sensing image to be detected can be detected according to the landslide detection model. For details, please refer to Figure 5 The flowchart shown.

[0115] Step 501: determining a remote sensing image to be detected.

[0116] In the embodiments of the present disclosure, the remote sensing image to be detected can be sent by a user to the electronic device through a terminal device, so that the electronic device determines the remote sensing image to be detected. Alternatively, the electronic device can determine the remote sensing image to be detected from a set of collected remote sensing images according to a preset rule. The embodiments of the present disclosure do not limit this.

[0117] Step 502: performing feature extraction on the remote sensing image to be detected through a backbone network of the landslide detection model to obtain a plurality of initial feature maps of different sizes, and performing fusion processing on the plurality of initial feature maps of different sizes through a neck network of the landslide detection model to obtain a plurality of fusion feature maps of different sizes.

[0118] Step 503: performing feature processing on the plurality of fusion feature maps of different sizes through a receptive field attention sub-module of the landslide detection model to obtain a target receptive field feature.

[0119] In the embodiments of the present disclosure, the plurality of fusion feature maps of different sizes can be input into a first sub-branch of a receptive field attention (RFA) module to obtain a first feature matrix; the first feature matrix can be input into a second sub-branch of the receptive field attention sub-module to obtain a second feature matrix; the second feature matrix can be input into a third sub-branch of the receptive field attention sub-module to obtain a third feature matrix; and the target receptive field feature can be determined according to the first feature matrix and the third feature matrix.

[0120] Optionally, inputting the plurality of fusion feature maps of different sizes into the first sub-branch of the receptive field attention sub-module to obtain the first feature matrix includes: inputting the fusion feature maps into a nonlinear activation function of the receptive field attention sub-module to obtain a channel attention vector of an original feature corresponding to the plurality of fusion feature maps of different sizes, and taking the channel attention vector as the first feature matrix. For example, the first sub-branch processes the fusion feature maps through a Sigmoid function to obtain the channel attention vector of the original feature corresponding to the fusion feature maps, and generates a corresponding weight value to obtain the first feature matrix.

[0121] Optionally, inputting the plurality of fusion feature maps of different sizes into the second sub-branch of the receptive field attention sub-module to obtain the second feature matrix includes: performing global average pooling and one-dimensional convolution processing on the plurality of fusion feature maps of different sizes to obtain a first sub-feature matrix; and performing processing on the first sub-feature matrix using convolution kernels of different sizes and dilated convolution to obtain the second feature matrix.

[0122] Specifically, the second sub-branch consists of Efficient Channel Attention (ECA) and Receptive Field Block (RFB) to fully extract feature information. ECA consists of global average pooling and one-dimensional convolution modules. Specifically, for input X, ECA is calculated as shown in Formula 3 and Formula 4 below:

[0123] Formula 3

[0124] Formula 4

[0125] in, F ECA ( X ) is used to represent efficient channel attention processing on the input X, σ is used to represent the Sigmoid function, GAP(X) Often indicates input X Perform a global average pooling operation, Used to represent the channel attention vector, X Used to represent multiple fusion feature maps of different sizes, Y Used to characterize the first sub-feature matrix, Conv1D Used to represent a one-dimensional convolution, the size of the convolution kernel is k, k By channel dimension C Adaptive determination, where ，k The calculation formula is shown in Formula 5 below:

[0126] Formula 5

[0127] Among them, |∙|odd means the nearest odd number to |∙|, gamma and b is a hyperparameter. Optionally, in the embodiment of the present disclosure, gamma and b The values ​​of are set to 2 and 1 respectively.

[0128] For example, see Figure 6 As shown, k When it is equal to 5, ECA can perform Figure 6 The processing shown in FIG5 is to collect, for the input X with a height of H, a width of W, and a channel dimension of C, first perform a global average pooling operation, then perform a one-dimensional convolution operation with a convolution kernel of 5, and then process the result of the convolution operation through the Sigmoid function to obtain the first sub-feature matrix.

[0129] In the embodiments of the present disclosure, the RFB adopts a multi-branch pool, wherein different branches use different size convolution kernels to respond to different size receptive fields. That is, the RFB designs different padding and expansion to process small, medium and large range feature information respectively. Finally, the three branches of the RFB are spliced to form a complete receptive field feature as the second feature matrix. Moreover, the relationship between the receptive field size and the eccentric distance is also considered in the RFB, which is beneficial to highlight the importance of the central area, improve the insensitivity to small spatial displacement, and help the network focus on identifying high-level abstract features without being affected by specific spatial positions.

[0130] That is, in the RFA, the feature map is first processed by the ECA, and then the RFB fully reshapes the feature map to highlight the importance of the area close to the center, thereby allowing the landslide detection model to focus on identifying high-level abstract features.

[0131] In the embodiments of the present disclosure, when the first feature matrix, the second feature matrix and the third feature matrix are obtained, the target receptive field feature can be determined based on the following Formula Six:

[0132] Formula Six

[0133] wherein, X out is used to represent the target receptive field feature; wherein, , X F is used to represent the second feature matrix, X is used to represent the feature matrix corresponding to the fusion feature map; X a is used to represent the third feature matrix; is used to represent the first feature matrix; Re LU ( X a is used to represent the application of the Rectified Linear Unit (ReLU) activation function to each element in the third feature matrix.

[0134] It can be seen that the receptive field attention (RFA) is introduced at three positions of the neck of the landslide detection model, which is connected to the detection head to detect small targets, medium targets and large targets. In this way, the RFA can better capture multi-scale landslide features by combining multi-scale convolution operation and channel attention to simulate human visual receptive field, so as to improve the detection accuracy of the remote sensing image to be detected.

[0135] Step 604: detecting and processing the target receptive field feature through the detection network of the landslide detection model to obtain a landslide detection frame and a landslide category label.

[0136] In the embodiments of the present disclosure, the detection network of the landslide detection model is responsible for predicting the position, class and confidence of each bounding box. Specifically, the detection network simultaneously predicts the class of each bounding box, maps the feature map to the class probability distribution using a soft-max function, removes redundant bounding boxes through non-maximum suppression (NMS), and retains the bounding box with the highest confidence. Wherein, when training the detection network, the normalized wasserstein distance (NWD) loss function can be used to replace the traditional intersection over union (IoU) loss function, so as to improve the detection performance of small targets.

[0137] In the embodiments of the present disclosure, by simulating the human visual receptive field, combining multi-scale convolution operation and channel attention mechanism, complex and multi-scale landslide features can be effectively extracted, so that the landslide detection model performs well in processing landslide targets of different sizes and shapes, especially the detection ability of small targets is significantly improved. Moreover, the present disclosure introduces NWD in the loss function to replace the traditional IoU calculation method, therefore, the landslide detection model provided in the embodiments of the present disclosure performs well in processing small targets, can more accurately calculate the similarity between the predicted box and the real box, reduce the error caused by small targets, and improve the detection accuracy of small targets.

[0138] In order to better introduce the scheme provided in the embodiments of the present disclosure, referring to FIG. 1, the landslide detection scheme is introduced below with a specific example. Figure 7

[0139] Referring to FIG. 1, the landslide detection scheme is introduced below with a specific example. Figure 7 ​It can be known that the landslide detection model provided in the embodiment of the disclosure: using CBS (an abbreviation of a basic module, which is composed of three core components: C: Convolution (convolution layer); B: Batch Normalization (batch normalization); S: SiLU / Swish (activation function, usually SiLU, which is (Sigmoid Linear Unit)) as the basic feature extraction, C3k2 / C3k (wherein, C3 (also known as CSP Bottleneck) is a variant of Cross Stage Partial Network (CSPNet), mainly composed of the following three parts: main branch: containing multiple residual blocks (Bottleneck); shortcut branch: directly connected to the input, preserving the original features; final fusion: merging the features of the two branches through the Concat operation. C3k or C3k2 is a variant of C3) to construct a multi-branch residual block to enrich the feature expression, and to realize cross-scale feature interaction through "up-sampling + feature fusion"; using a fast spatial pyramid pooling module (Spatial Pyramid Pooling - Fast, SPPF) to compress features, RFA to enhance attention, and finally using NWD loss to optimize the detection task. In addition, the C2PSA block is deleted from YOLOv11, and the removal of the C2PSA block simplifies the model structure and improves the overall efficiency. These improvements enable the landslide detection model provided in the embodiment of the disclosure to achieve excellent landslide detection performance in various scenarios. Among them, the aforementioned various scenarios are, for example, multi-scale targets, complex optical features, limited sample conditions, and the like, which are not limited in the embodiment of the disclosure.

[0140] In the embodiment of the disclosure, the remote sensing image to be detected can be input into the backbone network in the landslide detection model, and then the backbone network performs down-sampling (i.e., size reduction) and feature extraction processing on the remote sensing image to be detected, to generate feature maps of different levels, such as 320x..., 160x... and the like shown in FIG. 11, i.e., to obtain initial feature maps. Then, based on the neck network in the landslide detection model, the multi-scale feature maps output by the backbone network are cross-level fused to obtain three fusion feature maps of different sizes. Figure 7

[0141] In the embodiment of the disclosure, the three fusion feature maps of different sizes can be input into the first sub-branch of the receptive field attention sub-module of the landslide detection model to obtain a first feature matrix, i.e., the feature matrix corresponding to the leftmost branch in RFA in FIG. 11. In addition, the three fusion feature maps of different sizes can also be input into the second sub-branch of the receptive field attention sub-module, i.e., the second sub-branch of the receptive field attention sub-module in FIG. 11. Figure 7 Figure 7 ​​The ECA in the middle branch of the RFA in FIG. 1 is obtained to obtain a first sub-feature matrix, and the first sub-feature matrix corresponding to the three feature maps with different sizes is input into different branch pools, and different branches use different size convolution kernels to respond to different size receptive fields.

[0142] For example, Figure 7 The convolution configuration of the left branch in the RFB shown in FIG. 1 is 3x3 convolution (dilation rate = 1, padding = 1), and the receptive field of the left branch is the smallest, which is used for detecting small targets and details (such as edges and textures). The convolution configuration of the middle branch in the RFB is 3x3 convolution (dilation rate = 2, padding = 2) -> 3x3 convolution (dilation rate = 1, padding = 1), and the receptive field of the middle branch is medium, which is used for detecting medium targets and semi-global semantics. In addition, the convolution configuration of the right branch in the RFB is three layers of 3x3 convolution, and the dilation rates are 1 -> 2 -> 3 in sequence, and the padding is increased synchronously. The receptive field of the left branch is the largest, which is used for detecting large targets and global structures (such as object contours). Then, the channel dimension reduction and feature fusion are performed on the left branch, the middle branch and the right branch in the RFB to obtain the complete receptive field feature, that is, the second feature matrix.

[0143] In the embodiments of the present disclosure, the feature maps of different receptive fields can be directly added in the "spatial dimension", so that the features of the second feature matrix and the first feature matrix are complementary to determine the third feature matrix, and the target receptive field feature is determined according to the third feature matrix and the first feature matrix. Further, the target receptive field feature can be input into a detection network to obtain a landslide detection frame and a landslide category label.

[0144] In the specific implementation process, the inventors used high-resolution remote sensing image data to verify the detection effect of the landslide detection method based on the landslide detection model provided in the embodiments of the present disclosure and other landslide detection methods in experiments. The verification results are shown in Table 1. Obviously, the comprehensive performance of the improved YOLOv11, that is, the landslide detection model provided in the embodiments of the present disclosure, is the best.

[0145] Table 1: Verification table of landslide detection model detection effect

[0146]

[0147] Specifically, referring to Table 1, in terms of precision, the improved YOLO is 6.1% higher than the original YOLOv11. Although the precision of the improved YOLOv11 is lower than the 96.17% of the LA-YOLO-LLL model, the precision and recall rate of the F1 and mAP50 (mean Average Precision at IoU=0.5, i.e., the AP (Average Precision) is calculated only when the IoU of the predicted box and the true box is ≥0.5, which is considered as "correct detection") parameters in the embodiment of the present disclosure are 3.17% and 3.7% higher than those of the LA-YOLO-LLL model, respectively. In terms of recall rate, the improved YOLOv11, i.e., the landslide detection model provided in the embodiment of the present disclosure, is 3.4% higher than the second-best YOLOv8, reaching 99.5%, which is significantly higher than all other models. In addition, the mAP50 of the improved YOLOv11, i.e., the landslide detection model provided in the embodiment of the present disclosure, is 99.1%, 3.4% and 3% higher than that of YOLOv8 and YOLOv11, respectively, and the mAP@50-95 (mean Average Precision at IoU=0.5:0.95, i.e., the average value of mAP when the IoU threshold is from 0.5 to 0.95 (step 0.05)) is 3.2% and 3.8% higher than that of YOLOv8 and YOLOv11, respectively.

[0148] It can be seen that the improved YOLOv11, i.e., the landslide detection model, has good performance stability for various challenging landslide images. For example, the target smaller than 32*32 pixels is defined as a small target, and due to the limitation of features, it is difficult for general models to learn these targets, but the improved YOLOv11 provided in the present disclosure can still accurately and stably detect these targets. In addition, the shape and size of the object are different, and even some very small, large or extreme object shapes (such as long, narrow and high) may exist, which brings great difficulty to the accurate identification and accurate positioning of the target, but the improved YOLOv11 provided in the present disclosure still performs well in this type of detection. In addition, due to the influence of factors such as vegetation sheltering and surface differentiation, the boundary between part of the target and the background is not clear, and the overall feature expression of the target is weak, which is easy to cause the model to miss detection and false detection, but the improved YOLOv11 provided in the present disclosure still performs well in this case. Therefore, when the improved YOLOv11 provided in the present disclosure is used for landslide detection on the to-be-detected remote sensing image, the landslide type of the to-be-detected remote sensing image can be detected efficiently and accurately.

[0149] The exemplary embodiments of the present disclosure also provide a landslide detection device. Referring to FIG. 8, Figure 8 As shown in FIG. 8, the landslide detection device 800 includes the following program units:

[0150] The determination unit 801 is configured to determine a remote sensing image to be detected.

[0151] The detection unit 802 is configured to perform feature extraction on the remote sensing image to be detected through a backbone network of a landslide detection model to obtain a plurality of initial feature maps of different sizes; perform fusion processing on the plurality of initial feature maps of different sizes through a neck network of the landslide detection model to obtain a plurality of fusion feature maps of different sizes; perform feature processing on the plurality of fusion feature maps of different sizes through a receptive field attention submodule of the landslide detection model to obtain target receptive field features; and perform detection processing on the target receptive field features through a detection network of the landslide detection model to obtain a landslide detection frame and a landslide category label.

[0152] In a possible implementation, the detection unit 802 is specifically configured to:

[0153] input the plurality of fusion feature maps of different sizes into a first sub-branch of the receptive field attention submodule to obtain a first feature matrix;

[0154] input the plurality of fusion feature maps of different sizes into a second sub-branch of the receptive field attention submodule to obtain a second feature matrix;

[0155] input the second feature matrix into a third sub-branch of the receptive field attention submodule to obtain a third feature matrix;

[0156] determine the target receptive field features according to the first feature matrix and the third feature matrix.

[0157] In a possible implementation, the detection unit 802 is specifically configured to:

[0158] input the plurality of fusion feature maps of different sizes into a nonlinear activation function of the receptive field attention submodule to obtain a channel attention vector of an original feature corresponding to the fusion feature map, and take the channel attention vector as the first feature matrix.

[0159] In a possible implementation, the detection unit 802 is specifically configured to:

[0160] perform global average pooling and one-dimensional convolution processing on the plurality of fusion feature maps of different sizes to obtain a first sub-feature matrix;

[0161] perform processing on the first sub-feature matrix using convolution kernels of different sizes and dilated convolution to obtain a second feature matrix.

[0162] In a possible implementation, the detection unit 802 is specifically configured to:

[0163]

[0164] wherein, X out for representing a target receptive field feature; wherein, , X F for representing a second feature matrix, X for representing a feature matrix corresponding to the fusion feature map; X a for representing a third feature matrix; for representing a first feature matrix; Re LU ( X a ) for representing applying a ReLU activation function element by element in the third feature matrix.

[0165] In a possible implementation, the apparatus further includes a training unit configured to:

[0166] collect a remote sensing image dataset, and divide the remote sensing image dataset to obtain a training set, a validation set, and a test set;

[0167] train a preset landslide detection model according to the training set to obtain a plurality of prediction information;

[0168] determine comparison information of the plurality of prediction information and labeled information in the training set through a loss function, and adjust the preset landslide detection model according to the comparison information to obtain an adjusted preset landslide detection model;

[0169] adjust hyperparameters in the adjusted preset landslide detection model according to the validation set, and when it is determined that a validation set loss reaches a first preset condition, obtain an updated preset landslide detection model;

[0170] evaluate the updated preset landslide detection model according to the test set, and when it is determined that an evaluation result reaches a second preset condition, obtain the landslide detection model.

[0171] In a possible implementation, the loss function is determined based on the following formula:

[0172]

[0173] wherein, N 1 for representing prediction information, N 2 for representing labeled information, C 2 for representing a constant related to the training set, for characterizing a solution of a second order Wasserstein distance between prediction information and annotation information. NWD N 1 N 2 for characterizing a degree of similarity between prediction information and annotation information.

[0174] The specific details of the parts of the above apparatus have been described in detail in the method part embodiments, and the undisclosed details can be referred to the content of the method part embodiments, thus not described again.

[0175] It should be noted that, although several modules or units of the device for action execution are mentioned in the above detailed description, such division is not mandatory. In fact, according to the example embodiments of the present disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided into embodied by multiple modules or units.

[0176] The example embodiments of the present disclosure also provide a computer program product. The computer program product includes a computer program, which, when executed by a processor, implements the above landslide detection method.

[0177] In an embodiment, the computer program product can be a tangible product containing the computer program, such as a computer readable storage medium storing the computer program. The readable storage medium can be a storage medium based on electrical, magnetic, optical, electromagnetic, infrared, etc. signals, including but not limited to: random access memory (RAM), read-only memory (ROM), magnetic tape, floppy disk, flash memory (Flash), mechanical hard disk (HDD), solid state disk (SSD), etc. For example, the computer program product can be implemented as a non-volatile storage medium storing the computer program, such as a read-only memory, a NAND flash memory, etc.

[0178] In an embodiment, the computer program product can be an intangible product containing the computer program. For example, the computer program product can be implemented as a virtual digital product, such as an executable file, an installation package, etc. digital file storing the computer program.

[0179] ​​Code of computer program can be written in one or more programming languages. Programming languages such as C language, Java, C++, etc. Program code can be executed entirely on a user computing device, or partially on a user computing device, or as a separate software package, or partially on a user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In the case involving a remote computing device, the remote computing device can be connected to the user computing device through any kind of network, such as a local area network (LAN), a wide area network (WAN), etc., or can be connected to an external computing device (for example, through an Internet connection provided by an operator).

[0180] Computer program can be carried or transmitted by electronic, magnetic, optical, electromagnetic, infrared, etc. signals. Electronic device can convert signals carrying computer program into digital signals, and then run computer program. When computer program runs on electronic device, its code is used to make electronic device execute (more specifically, can make processor of electronic device execute) method steps of various exemplary embodiments of the present disclosure, such as can execute the above landslide detection method, which includes the following steps: step 501: determining a to-be-detected remote sensing image; step 502: performing feature extraction on the to-be-detected remote sensing image through the backbone network of the landslide detection model to obtain a plurality of initial feature maps with different sizes; and performing fusion processing on the plurality of initial feature maps with different sizes through the neck network of the landslide detection model to obtain a plurality of fusion feature maps with different sizes; step 504: performing feature processing on the plurality of fusion feature maps with different sizes through the receptive field attention submodule of the landslide detection model to obtain target receptive field features; step 505: performing detection processing on the target receptive field features through the detection network of the landslide detection model to obtain a landslide detection frame and a landslide category label.

[0181] By implementing the above method steps through a computer program, the remote sensing image to be detected can be determined, and then the remote sensing image to be detected is subjected to feature extraction through a backbone network of the landslide detection model to obtain a plurality of initial feature maps of different sizes; and the plurality of initial feature maps of different sizes are subjected to fusion processing through a neck network of the landslide detection model to obtain a plurality of fusion feature maps of different sizes; the plurality of fusion feature maps of different sizes are subjected to feature processing through a receptive field attention sub-module of the landslide detection model to obtain a target receptive field feature; and the target receptive field feature is subjected to detection processing through a detection network of the landslide detection model to obtain a landslide detection frame and a landslide category label. It can be seen that, by adopting the landslide detection model, the remote sensing image to be detected can be automatically detected, errors caused by manual analysis are avoided, the detection accuracy is improved, and the landslide detection on the remote sensing image to be detected is performed on the features determined by deep mining of the remote sensing image to be detected in combination with the receptive field attention sub-module, so that the detection accuracy of the remote sensing image to be detected is further improved.

[0182] The exemplary embodiments of the present disclosure also provide an electronic device, which can include a processor and a memory. The memory stores executable instructions of the processor, such as a computer program. The processor executes the method steps of various exemplary embodiments of the present disclosure by executing the executable instructions.

[0183] The following description refers to the accompanying drawings, which illustrate examples of the present disclosure. Figure 9 The electronic device is exemplarily illustrated in the form of a general computing device. It should be understood that, Figure 9 The electronic device 120 shown is merely an example and should not limit the functions and use ranges of the embodiments of the present disclosure.

[0184] As Figure 9 shown, the electronic device 120 can include a processor 910, a memory 920, a bus 930, an I / O (input / output) interface 940, and a network adapter 950.

[0185] The memory 920 can include a volatile memory, such as a RAM 921, a cache unit 922, and a non-volatile memory, such as a ROM 923. The memory 920 can also include one or more program modules 924, which include but are not limited to an operating system, one or more application programs, other program modules, and program data, each of which or some combination thereof can include implementation of a network environment. For example, the program modules 924 can include the units in the above-described apparatus.

[0186] The processor 910 can include one or more processing units, for example: the processor 910 can include an AP (Application Processor, application processor), a modem processor, a GPU (Graphics Processing Unit, graphics processing unit), an ISP (Image Signal Processor, image signal processor), a controller, an encoder, a decoder, a DSP (Digital Signal Processor, digital signal processor), a baseband processor and / or an NPU (Neural-Network Processing Unit, neural network processor), etc.

[0187] The processor 910 can be used to execute executable instructions stored in the memory 920, such as the above landslide detection method, which includes the following steps: step 501: determining a to-be-detected remote sensing image; step 502: performing feature extraction on the to-be-detected remote sensing image through the backbone network of the landslide detection model to obtain a plurality of initial feature maps of different sizes; and performing fusion processing on the plurality of initial feature maps of different sizes through the neck network of the landslide detection model to obtain a plurality of fusion feature maps of different sizes; step 504: performing feature processing on the plurality of fusion feature maps of different sizes through the receptive field attention submodule of the landslide detection model to obtain target receptive field features; step 505: performing detection processing on the target receptive field features through the detection network of the landslide detection model to obtain a landslide detection frame and a landslide category label.

[0188] By executing the above method steps through the processor 910, the to-be-detected remote sensing image can be determined, and then the to-be-detected remote sensing image is subjected to feature extraction through the backbone network of the landslide detection model to obtain a plurality of initial feature maps of different sizes; and the plurality of initial feature maps of different sizes are subjected to fusion processing through the neck network of the landslide detection model to obtain a plurality of fusion feature maps of different sizes; the plurality of fusion feature maps of different sizes are subjected to feature processing through the receptive field attention submodule of the landslide detection model to obtain target receptive field features; the target receptive field features are subjected to detection processing through the detection network of the landslide detection model to obtain a landslide detection frame and a landslide category label. It can be seen that the landslide detection model adopted by the present disclosure can not only automatically detect the to-be-detected remote sensing image, avoiding errors caused by manual analysis, and improving the accuracy of detection, but also combines the receptive field attention submodule to perform landslide detection on the to-be-detected remote sensing image based on the features determined by deep mining of the to-be-detected remote sensing image, further improving the accuracy of detection of the to-be-detected remote sensing image.

[0189] The bus 930 is used to realize the connection between different components of the electronic device 120, and can include a data bus, an address bus and a control bus.

[0190] The electronic device 120 can communicate with one or more external devices 1000 (e.g., a keyboard, a mouse, an external controller, etc.) through the I / O interface 940.

[0191] The electronic device 120 can communicate with one or more networks through the network adapter 950, for example, the network adapter 950 can provide a mobile communication solution such as 3G / 4G / 5G, or provide a wireless communication solution such as a wireless local area network, Bluetooth, near field communication, etc. The network adapter 950 can communicate with other modules of the electronic device 120 through the bus 930.

[0192] Although Figure 9 Other hardware and / or software modules can also be provided in the electronic device 120, which are not shown in the figure, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, etc.

[0193] As can be seen from the above, the technical solutions of the present disclosure can be implemented as methods, apparatuses, systems, computer program products, storage media, electronic devices, and the like. Those skilled in the art can understand that various aspects of the present disclosure can be implemented in the following forms: a complete hardware implementation, a complete software implementation (including firmware, microcode, etc.), or a combination of hardware and software aspects, such as can be referred to as "circuitry", "modules" or "systems".

[0194] It should be understood that the present disclosure is not limited to the specific method steps or structures described above and shown in the drawings, and various modifications and changes can be made without departing from the scope thereof. Those skilled in the art, based on the specific embodiments provided by the present disclosure, will easily think of other embodiments. Therefore, the specific embodiments provided by the present disclosure are only exemplary, the scope and spirit of the present disclosure are indicated by the claims, and should cover any variations, uses or adaptations of the present disclosure that follow the general principles of the present disclosure, and include common knowledge or conventional technical means in the technical field that are not disclosed by the present disclosure.

Claims

1. A landslide detection method, characterized in that: The method comprises: Determine the remote sensing image to be detected; Performing feature extraction on the remote sensing image to be detected through a backbone network of the landslide detection model to obtain a plurality of initial feature maps of different sizes; and performing fusion processing on the plurality of initial feature maps of different sizes through a neck network of the landslide detection model to obtain a plurality of fused feature maps of different sizes; Performing feature processing on the multiple fused feature maps of different sizes through the receptive field attention submodule of the landslide detection model to obtain target receptive field features; Detecting and processing the target receptive field features through the detection network of the landslide detection model to obtain a landslide detection frame and a landslide category label; The step of performing feature processing on the plurality of fused feature maps of different sizes by the receptive field attention submodule of the landslide detection model to obtain target receptive field features includes: Inputting the multiple fused feature maps of different sizes into the first sub-branch of the receptive field attention sub-module to obtain a first feature matrix; inputting the multiple fused feature maps of different sizes into the second sub-branch of the receptive field attention sub-module to obtain a second feature matrix; inputting the second feature matrix into the third sub-branch of the receptive field attention sub-module to obtain a third feature matrix; determining target receptive field features based on the first feature matrix and the third feature matrix; The landslide detection model is trained based on the following method: A remote sensing image data set is collected and divided into a training set, a validation set and a test set; a preset landslide detection model is trained according to the training set to obtain a plurality of prediction information; comparison information between the plurality of prediction information and the annotation information in the training set is determined by a loss function, and the preset landslide detection model is adjusted according to the comparison information to obtain an adjusted preset landslide detection model; hyperparameters in the adjusted preset landslide detection model are adjusted according to the validation set, and when it is determined that the validation set loss reaches a first preset condition, an updated preset landslide detection model is obtained; the updated preset landslide detection model is evaluated according to the test set, and when it is determined that the evaluation result reaches a second preset condition, the landslide detection model is obtained.

2. The method according to claim 1, characterized in that Inputting the multiple fused feature maps of different sizes into the first sub-branch of the receptive field attention sub-module to obtain a first feature matrix, including: The multiple fused feature maps of different sizes are input into the nonlinear activation function of the receptive field attention sub-module to obtain the channel attention vector of the original feature corresponding to the fused feature map, and the channel attention vector is used as the first feature matrix.

3. The method according to claim 1, characterized in that Inputting the multiple fused feature maps of different sizes into the second sub-branch of the receptive field attention sub-module to obtain a second feature matrix, including: Performing global average pooling and one-dimensional convolution processing on the multiple fused feature maps of different sizes to obtain a first sub-feature matrix; The first sub-feature matrix is ​​processed using convolution kernels of different sizes and dilated convolution to obtain a second feature matrix.

4. The method according to claim 1, wherein Determining target receptive field features according to the first feature matrix and the third feature matrix includes: in, X out Used to characterize the target receptive field characteristics; among them, , X F Used to characterize the second characteristic matrix, X Used to characterize the feature matrix corresponding to the fused feature map; X a Used to characterize the third characteristic matrix; Used to characterize the first characteristic matrix; Re LU ( X a ) is used to represent the application of the ReLU activation function to each element in the third feature matrix.

5. The method according to claim 1, wherein The loss function is determined based on the following formula: in, N 1 Used to represent prediction information, N 2 Used to represent annotation information, C 2 A constant used to characterize the relationship between the training set and the training set. Used to characterize the second-order Wasserstein distance between the predicted information and the labeled information; NWD ( N 1 , N 2 ) is used to characterize the similarity between the predicted information and the labeled information.

6. A landslide detection device, characterized in that: The device comprises: A determination unit, used for determining a remote sensing image to be detected; A detection unit is configured to perform feature extraction on the remote sensing image to be detected through a backbone network of a landslide detection model to obtain a plurality of initial feature maps of different sizes; and to perform fusion processing on the plurality of initial feature maps of different sizes through a neck network of the landslide detection model to obtain a plurality of fused feature maps of different sizes; perform feature processing on the plurality of fused feature maps of different sizes through a receptive field attention submodule of the landslide detection model to obtain a target receptive field feature; and perform detection processing on the target receptive field feature through a detection network of the landslide detection model to obtain a landslide detection frame and a landslide category label; Wherein, the detection unit is specifically used for: Inputting the multiple fused feature maps of different sizes into the first sub-branch of the receptive field attention sub-module to obtain a first feature matrix; inputting the multiple fused feature maps of different sizes into the second sub-branch of the receptive field attention sub-module to obtain a second feature matrix; inputting the second feature matrix into the third sub-branch of the receptive field attention sub-module to obtain a third feature matrix; determining target receptive field features based on the first feature matrix and the third feature matrix; The landslide detection model is trained based on the following method: A remote sensing image data set is collected and divided into a training set, a validation set and a test set; a preset landslide detection model is trained according to the training set to obtain a plurality of prediction information; comparison information between the plurality of prediction information and the annotation information in the training set is determined by a loss function, and the preset landslide detection model is adjusted according to the comparison information to obtain an adjusted preset landslide detection model; hyperparameters in the adjusted preset landslide detection model are adjusted according to the validation set, and when it is determined that the validation set loss reaches a first preset condition, an updated preset landslide detection model is obtained; the updated preset landslide detection model is evaluated according to the test set, and when it is determined that the evaluation result reaches a second preset condition, the landslide detection model is obtained.

7. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

8. An electronic device, characterized in that: include: processor; a memory for storing executable instructions of the processor; The processor is configured to perform the method according to any one of claims 1 to 5 by executing the executable instructions.

Images