Landslide detection method, system and equipment based on target detection model and medium

Abstract

The invention discloses a landslide detection method, system and device based on a target detection model, and a medium. The method comprises the following steps: obtaining a remote sensing image; the remote sensing image is input into a landslide detection model for target detection, a landslide detection result is obtained, and the landslide detection model is constructed based on a traditional YOLOv9 model and a channel attention mechanism. According to the technical scheme of the invention, by introducing the attention mechanism, the interdependence between modeling channels can be defined, the channel feature response can be adaptively recalibrated, and the feature expression ability of the convolutional neural network is significantly enhanced; through the improvement of the deep learning algorithm, the precision and efficiency of remote sensing image analysis are improved, and the method is of great significance to application scenes such as geological disaster monitoring, environment evaluation and urban planning.

Description

Technical Field

[0001] The present invention belongs to the technical field of target detection, and in particular relates to a landslide detection method, system, equipment and medium based on a target detection model. Background Art

[0002] Remote sensing image target detection technology is one of the key technologies in the field of computer vision. It allows the automatic identification and location of ground targets in images acquired from high altitude or space. With the continuous advancement of remote sensing technology, the amount of image data acquired is becoming increasingly large, which puts higher requirements on the accuracy and processing speed of target detection models. Deep learning, especially CNN (Convolutional Neural Network), has shown great potential and advantages in this field.

[0003] The YOLO (You Only Look Once) series of models are popular for their fast and accurate target detection capabilities. As the latest version of the series, YOLOv9 further improves the detection speed and accuracy by optimizing the network structure and algorithm. However, in specific application scenarios, such as landslide detection in remote sensing images, there are more stringent requirements on the accuracy and feature expression capabilities of the model.

[0004] In remote sensing image landslide detection, the model needs to be able to capture subtle changes and features in the image to accurately locate the landslide area. Although the traditional YOLOv9 model performs well in the field of target detection, it still has the problem of insufficient feature expression when processing complex remote sensing image data. Summary of the invention

[0005] The purpose of the present invention is to provide a landslide detection method, system, equipment and medium based on a target detection model to solve the problems existing in the above-mentioned prior art.

[0006] To achieve the above object, the present invention provides a landslide detection method based on a target detection model, comprising:

[0007] Acquisition of remote sensing images;

[0008] The remote sensing image is input into a landslide detection model for target detection to obtain a landslide detection result, wherein the landslide detection model is constructed based on a traditional YOLOv9 model and a channel attention mechanism.

[0009] Optionally, the process of constructing the landslide detection model specifically includes:

[0010] Build a traditional YOLOv9 model, place the channel attention mechanism module after the 37th efficient layer aggregation network of the traditional YOLOv9 model as the 38th layer of the model, and change the inference layer to the introduced channel attention mechanism module.

[0011] Optionally, the training process of the landslide detection model specifically includes:

[0012] Acquiring training data, wherein the training data includes remote sensing training images and corresponding landslide results;

[0013] An initial landslide detection model is constructed, the training data is input into the initial landslide detection model for target detection, and training is performed with the goal of minimizing the loss between the initial training result after target detection and the landslide result corresponding to the remote sensing training image to obtain a trained landslide detection model.

[0014] Optionally, the processing of the landslide detection model specifically includes:

[0015] Inputting the remote sensing image into a landslide detection model to extract features and obtain a feature map;

[0016] Adjusting the feature map based on the channel attention mechanism module to obtain an adjusted feature map;

[0017] Landslide detection is performed based on the adjusted feature map to obtain the prediction result of the landslide area.

[0018] Optionally, the adjusting the feature map based on the channel attention mechanism module specifically includes:

[0019] Performing global average compression on the feature map to obtain the global features of each channel and calculate the weight of each channel;

[0020] Based on element-by-element multiplication, the obtained weight is multiplied by each channel of the original feature map to obtain the adjusted feature map.

[0021] A landslide detection system based on a target detection model, comprising:

[0022] Data acquisition module, used to obtain remote sensing images;

[0023] The landslide detection module is used to input the remote sensing image into a landslide detection model for target detection to obtain a landslide detection result, wherein the landslide detection model is constructed based on a traditional YOLOv9 model and a channel attention mechanism.

[0024] An electronic device comprises a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the landslide detection method based on a target detection model.

[0025] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the landslide detection method based on a target detection model is implemented.

[0026] The technical effects of the present invention are:

[0027] By introducing the attention mechanism, the present invention can clearly model the interdependence between channels and adaptively recalibrate the channel feature responses, significantly enhancing the feature expression capability of the convolutional neural network; through the improvement of the deep learning algorithm, the accuracy and efficiency of remote sensing image analysis are improved, which is of great significance for application scenarios such as geological disaster monitoring, environmental assessment and urban planning. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0029] Figure 1 1 is a partial network structure diagram of the traditional YOLOv9 model used in the embodiments of the present invention.

[0030] Figure 2 4 is a structural diagram of the SE module used in the embodiment of the present invention.

[0031] Figure 3 It is a network structure diagram of the YOLOv9 landslide detection optimization model based on the SE-ELAN structure constructed in an embodiment of the present invention.

[0032] Figure 4 This is an example of a multi-scale landslide dataset used in an embodiment of the present invention.

[0033] Figure 5 This is the prediction result obtained by applying the YOLOv9 hybrid landslide detection model based on the SE-ELAN structure on the multi-scale landslide dataset in an embodiment of the present invention.

[0034] Figure 6 It is the prediction result obtained by applying the original YOLOv9 model to the multi-scale landslide dataset in the embodiment of the present invention;

[0035] Figure 7 This is a flow chart of landslide detection in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] Various exemplary embodiments of the present invention will now be described in detail. This detailed description should not be considered as limiting the present invention, but should be understood as a more detailed description of certain aspects, features, and embodiments of the present invention.

[0037] It should be understood that the terms described in the present invention are only for describing special embodiments and are not intended to limit the present invention. In addition, for the numerical range in the present invention, it should be understood that each intermediate value between the upper and lower limits of the scope is also specifically disclosed. Each smaller range between the intermediate value in any stated value or stated range and any other stated value or intermediate value in the described range is also included in the present invention. The upper and lower limits of these smaller ranges can be independently included or excluded in the scope.

[0038] It will be apparent to those skilled in the art that various modifications and variations may be made to the specific embodiments of the present invention description without departing from the scope or spirit of the present invention. Other embodiments derived from the present invention description will be apparent to those skilled in the art. The present application description and examples are exemplary only.

[0039] The words “include,” “including,” “have,” “contain,” etc. used in this article are open-ended terms, meaning including but not limited to.

[0040] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0041] Embodiment 1

[0042] like Figure 1 - Figure 7 As shown, this embodiment provides a landslide detection method based on a target detection model, including: acquiring a remote sensing image; inputting the remote sensing image into a landslide detection model for target detection to obtain a landslide detection result, wherein the landslide detection model is constructed based on a traditional YOLOv9 model and a channel attention mechanism.

[0043] This embodiment proposes a YOLOv9 hybrid landslide detection model based on SE-ELAN structure, aiming to improve the accuracy and efficiency of landslide detection in remote sensing images. The model introduces a Squeeze-and-Excitation Networks (SE) module in a specific auxiliary reversible branch and places it after the Efficient Layer Aggregation Network (ELAN) module as the 38th layer of the model. In addition, this embodiment also adjusts the 37th layer ELAN module used as reasoning in the traditional YOLOv9 model to the 38th layer SE module, optimizing the structure of the traditional model. The SE module significantly enhances the feature expression ability of the convolutional neural network by explicitly modeling the interdependence between channels and adaptively recalibrating the channel feature response. The results of the remote sensing image landslide detection performance evaluation experiment show that the YOLOv9 hybrid landslide detection model based on SE-ELAN structure achieves a 1.1% improvement in detection accuracy compared with the traditional YOLOv9 model with almost no increase in computational cost. This embodiment not only improves the accuracy of landslide detection while maintaining the computational efficiency of the model, but also has important application value in the fields of geological disaster monitoring, environmental assessment and urban planning.

[0044] The specific implementation steps of this embodiment are as follows:

[0045] 1) An architecture based on a traditional YOLOv9 model, wherein the SE module is introduced in a specific auxiliary reversible branch, and the SE module is located after the ELAN module as the 38th layer of the YOLOv9 hybrid landslide detection model based on the SE-ELAN structure;

[0046] 2) The 37th layer ELAN module used for inference in the traditional model is adjusted to the 38th layer SE module. The SE module directly acts on the model output by optimizing the inference process, adjusting and optimizing the final prediction results;

[0047] 3) The SE module improves the interpretability of the model by learning the importance weights of the channels;

[0048] 4) The SE module dynamically adjusts the feature response according to the input data, thereby improving the adaptability of the model;

[0049] 5) The SE module strengthens important feature channels and suppresses unimportant channels, thereby enhancing the robustness of the model;

[0050] 6) The YOLOv9 hybrid landslide detection model based on the SE-ELAN structure improves the accuracy of landslide detection in remote sensing images by 1.1% compared with traditional models, and brings significant performance improvements to the existing most advanced Convolutional Neural Network (CNN) at a slightly increased computational cost.

[0051] The dataset used in the examples of this embodiment is derived from a public remote sensing image landslide detection dataset, which can be obtained through the website https: / / github.com / wenjieo / LS-YOLO. This dataset is publicly shared by researchers in related fields for the purpose of academic research and model training. This dataset has the advantages of a significant number of landslide samples, a high degree of intra-class variation, a wide range of landslide scales, and a complex background. This dataset is generated based on the Bijie landslide dataset and the multi-region landslide dataset using online enhancement and offline enhancement techniques. The offline enhancement technology includes geometric transformation (clockwise rotation consisting of 90°, 180°, 270°, flip operation, salt noise injection, and Gaussian noise injection) and color space transformation (image darkening, image brightening, and contrast adjustment). Online enhancement (Mosaic) is used during model training to better train the landslide detection model. Figure 4 As shown in the figure, it contains 7620 positive sample images with various landslide types and 2003 negative sample images. The Bijie landslide dataset was acquired by the TripleSat series satellites from May to August 2018, and contains 770 positive sample images with landslides and 2003 negative sample images without landslides. The ground resolution of these images is 0.8 meters, which can provide high-precision surface feature information. The multi-region landslide dataset has a wide distribution of study areas, covering five regions in China: Gansu, Sichuan, Guizhou, Yunnan and Tibet. The landslide dataset consists of 500 positive sample images, each of which contains one or more landslides.

[0052] Traditional YOLOV9 model and SE module such as Figure 1 and 2 As shown in the figure, in the traditional YOLOv9 model, in order to solve the problem of data loss during deep network transmission, especially the challenges of information bottleneck and reversible function, the concept of programmable gradient information (PGI) is proposed. PGI mainly consists of the following three components:

[0053] (1) Main Branch: responsible for the main reasoning process; (2) Auxiliary Reversible Branch: designed to deal with the problems caused by the deepening of the neural network. As the depth of the network increases, the information bottleneck may cause the loss function to be unable to generate reliable gradients; (3) Multi-level Auxiliary Information: designed to deal with the error accumulation problem caused by deep supervision, especially suitable for architectures with multiple prediction branches and lightweight models. In the reasoning process of PGI, reasoning only relies on the main branch, so no additional reasoning cost is incurred. At the same time, the auxiliary reversible branch effectively alleviates the problem of information loss in the deep network through its specific structure. Multi-level auxiliary information reduces the impact of error propagation by providing additional supervision signals, thereby improving the overall performance and stability of the model.

[0054] The network framework of the YOLOv9 hybrid landslide detection model based on SE-ELAN structure is shown in the figure Figure 3 shown.

[0055] This embodiment integrates the SE module into the YOLOv9-c model, and significantly improves the feature expression capability of the convolutional neural network by explicitly modeling the interdependence between channels and adaptively recalibrating the channel feature response. Specifically, the YOLOv9 hybrid landslide detection model based on the SE-ELAN structure introduces the Squeeze-and-Excitation Networks (SE) module in a specific auxiliary reversible branch and places it after the Efficient Layer Aggregation Network (ELAN) module as the 38th layer of the model. In addition, this embodiment also adjusts the 37th layer ELAN module used as reasoning in the traditional YOLOv9 model to the 38th layer SE module, optimizing the structure of the traditional YOLOv9 model.

[0056] The function of the SE module is to compress the input feature map to form a channel-level feature descriptor, thereby capturing the overall trend of the feature response of each channel and integrating global context information. Then, the module performs an excitation operation through an adaptive gating mechanism to generate a set of adjustment factors, which are then used to weight the original feature map to enhance the network's ability to recognize key features. Finally, the output of the SE module is directly passed to the subsequent layers of the network, which significantly improves the performance of the model in the task of landslide detection in remote sensing images. This process is mainly divided into two key steps:

[0057] 1) Squeeze. In the squeeze stage, the SE module compresses the spatial dimension of the input feature map through global average pooling to generate global information for each channel. Formally, the statistic It is generated by shrinking its spatial dimension H×W to U, and the calculation formula is:

[0058]

[0059] Among them, z c is the compression feature of channel c, u c (i,j) is the value at position (i,j) in the input feature map.

[0060] 2) Excitation. In order to make full use of the information summarized in the squeeze operation, a second operation is introduced after the squeeze operation to better capture the dependencies between channels. This operation is designed to meet the following two criteria: first, it must be flexible, especially able to learn non-linear interactions between channels; second, it must be able to learn non-exclusive relationships, thus ensuring that multiple channels can be emphasized rather than just activating a single channel. To achieve these goals, the SE module adopts a simple sigmoid function activation gating mechanism:

[0061] s=F ex (z,W)=σ(g(z,W))=σ(W2δ(W1z)) (2)

[0062] Among them, δ is the ReLU function,

[0063] At the same time, in order to limit the model complexity and improve generalization, the module forms a bottleneck of two fully connected (FC) layers around the nonlinearity to parameterize the gating mechanism, where the first fully connected layer reduces the number of channels from C to Where r is the reduction ratio, calculated as:

[0064] s=σ(W1·z) (3)

[0065] Among them, s is the output after processing by the first layer.

[0066] The second fully connected layer restores the number of output channels to C, calculated as:

[0067] u=σ(W2·s) (4)

[0068] Among them, u is the final channel weight.

[0069] After obtaining the channel weights u, the SE module applies these weights to each channel of the input feature map to generate a recalibrated feature map. The calculation formula is:

[0070]

[0071] in, is the recalibrated feature, u c is the weight of channel c, x c is the channel c of the input feature map.

[0072] In remote sensing image processing tasks, due to complex scenes and diverse target features, traditional target detection models often face the problem of insufficient feature expression. By integrating the SE module into the traditional YOLOv9 model, the SE module can effectively squeeze and stimulate the channel features in the feature map, improve the model's feature expression ability, and thus improve the overall performance.

[0073] It is feasible to adjust the 37th layer ELAN module used for inference in the traditional model to the 38th layer SE module. The specific process is as follows:

[0074] ① Input the remote sensing image I into the model to obtain the feature map F;

[0075] ② Input the feature map F into the SE module to obtain the adjusted feature map F';

[0076] ③ Use the adjusted feature map F' to detect landslides in remote sensing images and obtain the prediction results of the landslide area;

[0077] ④ Among them, the specific calculation process of the SE module in the YOLOv9 hybrid landslide detection model based on the SE-ELAN structure is:

[0078] Global averaging compresses the height and width of each feature map to obtain the global features of each channel:

[0079]

[0080] Among them, u c represents the feature map of the cth channel, H and W represent the height and width of the feature map respectively. In this model, H=W=20;

[0081] The weight ω of each channel is calculated through two fully connected layers and a ReLU activation function:

[0082] ω=F sq (z,W)=σ(W2δ(W1z)) (6)

[0083] Among them, δ represents the ReLU activation function, σ represents the Sigmoid activation function, W1 and W2 represent the weights of the fully connected layer. In this model, the dimension of W1 is [512, 32], and the dimension of W2 is [32, 512].

[0084] Perform element-wise multiplication of the obtained weight ω with the original feature map F to obtain the final feature map F':

[0085] F'=F·ω (7)

[0086] It is feasible that the SE module improves the interpretability of the model by learning the importance weights of the channels, which is achieved by the following formula:

[0087] The fully connected layer learns the nonlinear relationship between channels:

[0088] ω=σ(W2δ(W1z)) (8)

[0089] Among them, ω represents the weight of each channel, which reflects the contribution of each channel to the final prediction result.

[0090] It is feasible that the introduction of the SE module enables the model to more flexibly adjust the feature response in the inference stage to adapt to different remote sensing image input conditions, which is specifically manifested as: dynamically adjusting the weight ω according to the statistical characteristics of the input feature map F; the adjustment of the weight ω makes the model more sensitive to the detection of landslide features.

[0091] It can be implemented that the SE module strengthens important feature channels and suppresses unimportant channels to enhance the robustness of the model. This is achieved through the following formula:

[0092] Feature Recalibration:

[0093]

[0094] Among them, x c represents the feature map of the cth channel, s c is the corresponding channel weight, this mechanism makes the model more robust to noise and outliers.

[0095] The model achieves 1.1% higher accuracy than the traditional YOLOv9 model in landslide detection in remote sensing images, and brings significant performance improvements to the existing state-of-the-art CNN at a slightly increased computational cost:

[0096] like Figure 3 As shown, this embodiment provides a YOLOv9 hybrid landslide detection model based on SE-ELAN structure, and its specific application is as follows:

[0097] 1) Set up the training process: execute the train_dual.py script, which is a Python program for training dual models; use 8 worker processes to speed up the training process; specify GPU device number 0 for training;

[0098] Set each batch to contain 16 samples; specify the dataset configuration file as mydataset.yaml, which contains the dataset information required for training; set the input image size to 640 pixels; select the model configuration file as yolov9-SE.yaml, which defines the architecture of the YOLOv9 model combined with SE; use the hyperparameter settings in the hyp.scratch-high.yaml file; set --min-item 0; train the model for a total of 500 cycles;

[0099] Turn off the data augmentation technique "mosaic" (--close-mosaic 15) for the first 15 epochs. This is an image blending technique used to improve the generalization ability of the model.

[0100] 2) Input the remote sensing image I to the model and obtain the feature map F;

[0101] 3) Input the feature map F into the SE module to obtain the adjusted feature map F';

[0102] 4) Use the adjusted feature map F' to detect landslides in remote sensing images and obtain the prediction results of the landslide area as follows Figure 5 As shown, the red box represents the model's prediction results and confidence scores.

[0103] The model experiment results are shown in Table 1:

[0104] Table 1 Experimental results of YOLOv9 hybrid landslide detection model based on SE-ELAN structure

[0105]

[0106] The same training process is set in the original YOLOv9 model. The prediction results obtained in this embodiment are as follows Figure 6 shown.

[0107] The experimental results of the traditional YOLOv9 model are shown in Table 2:

[0108] Table 2 Experimental results of traditional YOLOv9 hybrid landslide detection model

[0109]

[0110] In the performance evaluation experiment in remote sensing image landslide detection, this embodiment implements a unified training and testing process for two models, the YOLOv9 hybrid landslide detection model based on the SE-ELAN structure and the traditional YOLOv9 model. The experimental results, as shown in Tables 1 and 2, reveal that the YOLOv9 hybrid landslide detection model based on the SE-ELAN structure achieves a significant improvement of 1.1% in detection accuracy compared to the traditional model. In addition, the two models are almost the same in terms of parameter quantity and computational complexity. This finding shows that the performance optimization of the improved model does not sacrifice the size of the model or increase significant computational cost. This result emphasizes that the YOLOv9 hybrid landslide detection model based on the SE-ELAN structure effectively improves the accuracy of landslide detection while maintaining computational efficiency, demonstrating its potential application value in remote sensing image analysis.

[0111] Comprehensive experimental results show that the YOLOv9 hybrid landslide detection model based on SE-ELAN structure has achieved a significant improvement in detection accuracy while maintaining the same computational efficiency as the traditional YOLOv9 model. This improvement is of great significance for improving the accuracy of landslide detection in remote sensing images and provides a more reliable technical means for geological disaster monitoring.

[0112] A landslide detection system based on a target detection model, comprising:

[0113] Data acquisition module, used to obtain remote sensing images;

[0114] The landslide detection module is used to input the remote sensing image into a landslide detection model for target detection to obtain a landslide detection result, wherein the landslide detection model is constructed based on a traditional YOLOv9 model and a channel attention mechanism.

[0115] An electronic device comprises a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the landslide detection method based on a target detection model.

[0116] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the landslide detection method based on a target detection model is implemented.

[0117] The above is only a preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A landslide detection method based on a target detection model, characterized in that: include: Acquisition of remote sensing images; The remote sensing image is input into a landslide detection model for target detection to obtain a landslide detection result, wherein the landslide detection model is constructed based on a traditional YOLOv9 model and a channel attention mechanism.

2. The landslide detection method based on target detection model according to claim 1 is characterized in that: The construction process of the landslide detection model specifically includes: Build a traditional YOLOv9 model, place the channel attention mechanism module after the 37th efficient layer aggregation network of the traditional YOLOv9 model as the 38th layer of the model, and change the inference layer to the introduced channel attention mechanism module.

3. A landslide detection method based on target detection model according to claim 2, characterized in that: The training process of the landslide detection model specifically includes: Acquiring training data, wherein the training data includes remote sensing training images and corresponding landslide results; An initial landslide detection model is constructed, the training data is input into the initial landslide detection model for target detection, and training is performed with the goal of minimizing the loss between the initial training result after target detection and the landslide result corresponding to the remote sensing training image to obtain a trained landslide detection model.

4. The landslide detection method based on target detection model according to claim 2 is characterized in that: The processing process of the landslide detection model specifically includes: Inputting the remote sensing image into a landslide detection model to extract features and obtain a feature map; Adjusting the feature map based on the channel attention mechanism module to obtain an adjusted feature map; Landslide detection is performed based on the adjusted feature map to obtain the prediction result of the landslide area.

5. A landslide detection method based on target detection model according to claim 4, characterized in that: The adjusting of the feature map based on the channel attention mechanism module specifically includes: Performing global average compression on the feature map to obtain the global features of each channel and calculate the weight of each channel; Based on element-by-element multiplication, the obtained weight is multiplied by each channel of the original feature map to obtain the adjusted feature map.

6. A landslide detection system based on a target detection model, characterized in that: include: Data acquisition module, used to obtain remote sensing images; The landslide detection module is used to input the remote sensing image into a landslide detection model for target detection to obtain a landslide detection result, wherein the landslide detection model is constructed based on a traditional YOLOv9 model and a channel attention mechanism.

7. An electronic device, characterized in that: The invention comprises a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute a landslide detection method based on a target detection model according to any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that: The device stores a computer program, which, when executed by a processor, implements a landslide detection method based on a target detection model as described in any one of claims 1 to 5.

Images