A plateau mountainous area landslide detection method based on an improved YOLOv11 deep learning model

Abstract

The present application relates to the technical field of remote sensing image, and particularly relates to a highland mountainous area landslide detection method based on an improved YOLOv11 deep learning model, comprising: obtaining remote sensing images of a target area and performing pretreatment; constructing a YOLOv11 target detection model, and introducing a spatial adaptive feature modulation module and a convolution channel mixer into the model, the former is used for adaptively recalibrating the spatial dimension of a feature map to enhance the landslide area feature representation, and the latter is used for extracting local context information; using an improved point distance-based intersection over union loss function as a positioning loss, and training the model in combination with a multi-scale training strategy; inputting the image to be detected into the trained model, and outputting a landslide detection result. The cascade design of long-range feature modeling and local feature extraction enhances the feature extraction capability of the model for landslide targets, the loss function improves the boundary box regression accuracy, and effectively improves the accuracy and robustness of landslide detection in the complex background of highland mountainous areas.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image technology, specifically to a landslide detection method for plateau mountainous areas based on an improved YOLOv11 deep learning model. Background Technology

[0002] Landslides often cause severe geological disasters in mountainous areas, especially plateau regions, and have a significant impact on residents' lives and transportation. Traditional landslide monitoring methods usually rely on field surveys and ground sampling; however, these methods have drawbacks such as high cost, poor timeliness, and limited monitoring range. With the development of remote sensing technology, landslide detection methods based on remote sensing imagery have gradually become an important means of monitoring landslide disasters.

[0003] Currently, most landslide detection methods based on remote sensing imagery employ traditional image processing techniques or machine learning models. However, in complex mountainous terrain, due to the large scale differences of landslide targets, unclear boundaries, and strong background interference, existing detection models generally suffer from insufficient detection accuracy and a high number of missed and false detections. YOLO (You Only LookOnce), as an efficient target detection algorithm, performs well in various target detection tasks, but improving the detection accuracy of small-scale landslides and irregularly shaped landslides in complex mountainous backgrounds remains a challenge.

[0004] Therefore, in view of the shortcomings of the prior art, the present invention provides a landslide detection method based on the YOLOv11-SAFM model, which can effectively improve the detection accuracy of landslide targets in complex backgrounds in mountainous areas, especially the ability to detect small-scale landslides and identify them under complex background conditions. Summary of the Invention

[0005] The purpose of this invention is to provide a landslide detection method for plateau mountainous areas based on an improved YOLOv11 deep learning model. By introducing a spatial adaptive feature modulation module and a convolutional channel mixer into the YOLOv11 model, and combining an improved localization loss function and a multi-scale training strategy, the detection accuracy, boundary localization accuracy, and scale adaptability of landslide targets in plateau mountainous areas under complex backgrounds are improved.

[0006] To achieve the above-mentioned technical objectives and effects, the present invention is implemented through the following technical solution: A landslide detection method for high-altitude mountainous areas based on an improved YOLOv11 deep learning model includes the following steps: Acquire remote sensing images of the target area and preprocess the remote sensing images to obtain standardized image data; An initial target detection model based on the YOLOv11 model is constructed, and a feature enhancement module is introduced into the initial target detection model. At the same time, an improved localization loss function and a multi-scale training strategy are combined to enhance the model's ability to express the features of landslide targets under complex terrain conditions and improve detection accuracy. The feature enhancement module is used to perform adaptive spatial dimension recalibration on the remote sensing image feature map extracted by the initial target detection model to enhance the feature representation of the landslide area. The target detection model with the feature enhancement module was trained using the labeled landslide remote sensing image dataset to obtain a trained landslide detection model. The remote sensing image of the plateau and mountainous area to be detected is input into the trained landslide detection model, and the landslide detection results are output.

[0007] Furthermore, the feature enhancement module is a spatial adaptive feature modulation module, which adopts a residual connection structure and captures the global context information of the landslide target through long-range feature modeling. The processing is represented as follows: Where X is the input feature map, LN represents layer normalization, SAFM represents spatial adaptive feature modulation function, and Y is the modulated output feature.

[0008] Furthermore, the target detection model also incorporates a convolutional channel mixer cascaded with the spatial adaptive feature modulation module to extract local contextual information of the landslide target. The processing procedure is as follows: Where Y is the output of the spatial adaptive feature modulation module, CCM represents the convolutional channel mixer function, and Z is the fused output feature.

[0009] Furthermore, the training further includes: optimizing the target detection model using an improved localization loss function, wherein the improved localization loss function combines the geometric similarity information between the predicted box and the ground truth box during the bounding box regression process to improve localization accuracy.

[0010] Furthermore, the improved localization loss function is a cross-union ratio (CURRR) loss function based on point distance, and its calculation method is as follows: Where IoU is the intersection-union ratio between the predicted bounding box and the ground truth bounding box; and , respectively, are the squares of the Euclidean distances between the top-left corner and the bottom-right corner of the predicted bounding box and the ground truth bounding box; w and h are the width and height of the minimum bounding rectangle between the predicted bounding box and the ground truth bounding box, respectively.

[0011] Furthermore, the training further includes employing a multi-scale training strategy, that is, randomly selecting remote sensing image samples of different sizes to input into the network for training during the training process, in order to enhance the model's ability to detect landslide targets at different scales.

[0012] Furthermore, the preprocessing includes one or more combinations of operations such as geometric correction, radiometric correction, orthorectification, image cropping, and cloud removal on the remote sensing image.

[0013] Furthermore, the training and validation further include: dynamically adjusting the learning rate using an adaptive learning rate adjustment strategy, and evaluating the model's generalization ability using a cross-validation method.

[0014] Furthermore, the landslide detection results include information on the location and boundary range of the landslide; the method further includes: classifying the risk level based on the landslide area characteristics in the landslide detection results, and tracking and updating the landslide detection results over a long period of time to provide data support for landslide disaster prevention and control.

[0015] On the other hand, the present invention proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned deep learning-based landslide detection method for high-altitude mountainous areas.

[0016] The beneficial effects of this invention are: This invention addresses the challenge of feature extraction for landslide targets in complex geological contexts by introducing a spatially adaptive feature modulation module. To address the interference of background noise such as mountain shadows and vegetation cover in remote sensing images of plateau and mountainous areas on landslide identification, as well as the characteristics of blurred landslide target boundaries and large scale differences, this invention integrates a spatially adaptive feature modulation module into the YOLOv11 model. Based on a long-range feature modeling mechanism, it adaptively recalibrates the spatial dimension of the feature maps output by the convolutional layers, enabling the model to overcome the limitations of local receptive fields and capture the global contextual dependencies between the landslide target and its surrounding environment. This enhances the feature response of the landslide area and suppresses background noise at the feature level. Simultaneously, a convolutional channel mixer cascaded with this module further extracts local contextual information, strengthening the perception of detailed features such as landslide edges and textures. Through this cascaded structure combining global modulation and local enhancement, this invention effectively improves the model's sensitivity to landslide targets, especially small-scale landslides, and reduces the false negative and false positive rates in complex backgrounds.

[0017] This invention improves the regression accuracy of landslide target bounding boxes by employing a point-distance-based intersection-union (CUI) loss function. Addressing the issue that traditional CUI loss functions are insensitive to geometrical deviations between predicted and ground truth bounding boxes, this invention introduces an improved loss function that adds Euclidean distances between the top-left and bottom-right corners of the predicted and ground truth bounding boxes as penalty terms. By imposing constraints on the diagonal points representing the spatial location and size of the bounding boxes, the model not only optimizes the overlap area of ​​the boxes during regression but also forces the width, height, and overall proportions of the predicted boxes to closely approximate the geometry of the actual landslide. For irregular, elongated landslides commonly found in high-altitude mountainous areas, this point-distance-based constraint mechanism can more sensitively reflect bounding box misalignment and scale deviations, thereby driving the model to output detection results with more accurate boundaries and a geometry that better reflects the actual terrain features.

[0018] This invention enhances the model's adaptability to landslide targets at different scales and multi-source remote sensing imagery by incorporating a multi-scale training strategy. Considering the significant differences in landslide target scale due to the undulating terrain of plateau and mountainous areas, and the frequent need to process multi-source remote sensing data from different sensors in practical applications, this invention employs a multi-scale training strategy during the training phase. This involves randomly selecting sample images of different sizes for input into the network. This allows the model to learn scale-invariant feature representations of landslide targets at different spatial scales, synergizing with the inherent multi-scale pyramid structure and global receptive field capability of the spatial adaptive feature modulation module in the YOLOv11 model. By exposing the model to rich scale samples, the internal parameters are optimized in a broader feature space, enabling it to maintain stable recognition capabilities during the inference phase, whether facing small rock landslides in high-resolution imagery or large sedimentary landslides in medium-resolution imagery. This effectively overcomes performance degradation caused by differences in image sources or changes in target scale.

[0019] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the overall process of the landslide detection method in high-altitude mountainous areas using the YOLOv11-SAFM model of this invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1 This embodiment describes a high-precision landslide detection method based on the YOLOv11-SAFM model, such as... Figure 1 As shown, the specific implementation steps are as follows: Step 1: Data Acquisition and Preprocessing Acquire high-resolution remote sensing image data, including GF-2, Sentinel-2, Landsat-8, and other remote sensing image data; Geometric correction of remote sensing images is performed using Geographic Information System (GIS) technology to ensure the spatial accuracy of the images; Using existing landslide datasets, we generate labeled data for landslide areas to ensure the accuracy and representativeness of the labels. Preprocessing of remote sensing images, including image cropping, cloud removal, radiometric correction, and orthorectification, is performed to remove image noise and improve the accuracy of subsequent analysis.

[0024] Step 2: Model Building and Optimization Based on the YOLOv11 model, it is trained using open-source deep learning frameworks (such as PyTorch, TensorFlow, etc.) as the base network for landslide target detection tasks. Introducing a Spatial Adaptive Feature Modulation (SAFM) module to optimize feature extraction: This invention introduces a Spatial Adaptive Feature Modulation (SAFM) module into the YOLOv11 model to optimize the feature extraction process. The SAFM module improves the representation of key features in landslide areas through long-range feature modeling. The module's output features are added to the input features of the convolutional layers, and their distribution is adjusted through layer normalization. This process can be represented by the following formula: in, As input features, The features are those processed by SAFM. Furthermore, to further optimize feature fusion, this invention introduces a Convolutional Channel Mixer (CCM) to extract local contextual information, described by the following formula: (2) Where Y represents the feature after CCM processing. To theoretically illustrate the advantages of SAFM and CCM modules in enhancing feature representation and preserving details, this embodiment constructs an exemplary feature processing framework based on ideas from the field of image restoration. Within this framework, (3) This represents the initial features extracted from the input image I, after passing through the feature mixing module. After processing (consisting of SAFM and CCM), the enhanced features are fused with the initial features via residual connections to obtain the enhanced features. This process can be abstractly represented as: (4) Accordingly, an optimization objective combining spatial and frequency domain losses can be defined to ensure that the details and texture information of the features are effectively preserved. In the landslide detection task of this invention, the above feature processing logic is specifically applied to the YOLOv11 feature extraction network, and the feature maps are refined through SAFM and CCM modules, thereby improving the final detection accuracy.

[0025] in, It is the input image. This represents the initial features extracted by the convolutional layer. It is a feature hybrid module composed of SAFM and CCM. This is an upsampling operation. To optimize image reconstruction quality, the loss function combines spatial and frequency domain loss terms, defined as follows: (5) in, For high-resolution ground truth images, Fast Fourier Transform (FFT) is used to capture the frequency domain characteristics of an image. is the weighting factor. This loss function ensures high-quality image reconstruction while preserving detail and texture information; A multi-scale training strategy is adopted, which involves randomly selecting images of different sizes for training to ensure that the model can handle landslide targets of different scales, especially improving the detection accuracy of small-scale landslide targets. The loss function of YOLOv11 is optimized by adopting the MPDIoU (Mean Point-based Distance Intersection over Union) loss function to improve the regression accuracy of the bounding boxes. The IoU formula is as follows: (6) in, It is a prediction box. These are the ground truth bounding boxes. MPDIoU optimizes the bounding box regression process by incorporating Euclidean distance information between the two endpoints of the diagonal. The formula for MPDIoU is: (7) in, and These are the squares of the Euclidean distances between the top left and bottom right corners of the predicted bounding box and the ground truth bounding box, respectively. and These are the width and height of the smallest bounding rectangles of the predicted and ground truth bounding boxes, respectively, used to normalize the distance term to maintain the same dimensions as IoU. By simultaneously measuring the spatial deviation of the two diagonal points, MPDIoU can more sensitively reflect the misalignment and scale bias of the bounding boxes, thereby enhancing the stability of the regression.

[0026] Step 3: Model Training and Validation The training dataset includes remote sensing imagery data from different regions (such as Zhenxiong in Yunnan and plateau mountainous areas in Guizhou). The dataset covers landslides of different types and scales to ensure the diversity and representativeness of the training set. An adaptive learning rate adjustment strategy is adopted, which dynamically adjusts the learning rate as the training process progresses, ensuring that the model converges quickly in the early stage and is finely optimized in the later stage, thus avoiding inefficiency or overfitting. Cross-validation is used to randomly divide the dataset into training and validation sets to ensure that the data is not overfitted during training and to evaluate the model's generalization ability. The model is evaluated using metrics such as precision, recall, and F1-score, and its performance is tested on the validation set to ensure that the trained model has good practical application capabilities.

[0027] Step 4: Evaluation and application of landslide detection results: The trained YOLOv11-SAFM model is applied to the real-time monitoring and early warning system for landslide disasters, providing efficient and accurate landslide monitoring results and supporting the construction of disaster early warning systems. Based on the characteristics of landslide areas, the regions are divided into areas, and the landslide risk levels of different areas are analyzed to provide decision support for emergency management. By tracking and updating landslide monitoring results over a long period, we can provide continuous technical support and data assurance for the prevention and control of landslide disasters.

[0028] In one specific embodiment of the present invention, the SAFM module and the CCM module are cascaded to form a feature refinement unit, which is embedded between the backbone and neck of the YOLOv11 model. Specifically, the multi-scale feature maps extracted by the YOLOv11 backbone are first input to the feature refinement unit: after layer normalization, the feature maps are fed into the SAFM module for long-range context modeling, and the output is added to the input, then normalized again before being fed into the CCM module to extract local details, finally outputting an enhanced feature map. This enhanced feature map is then fed into the YOLOv11 neck network for multi-scale feature fusion, and finally the location and category of the landslide target are output by the detection head. Through this embedding method, the SAFM and CCM modules can effectively enhance the feature representation of landslide targets in complex terrain without disrupting the original architecture of YOLOv11.

[0029] Those skilled in the art will understand that the above-mentioned embedding positions are merely examples, and the SAFM and CCM modules can also replace some convolutional modules in the backbone network or be added to various layers of the neck network. As long as spatial adaptive modulation and local context enhancement of the feature map can be achieved, they all fall within the protection scope of this invention.

[0030] In summary, this invention proposes a landslide detection method for high-altitude mountainous areas based on an improved YOLOv11 deep learning model. The method includes: acquiring and preprocessing remote sensing images of the target area; constructing a YOLOv11 target detection model, incorporating a spatial adaptive feature modulation module and a convolutional channel mixer. The former is used to adaptively recalibrate the spatial dimensions of the feature map to enhance the feature representation of the landslide area, while the latter is used to extract local contextual information; an improved intersection-union loss function based on point distance is used as the localization loss, and the model is trained using a multi-scale training strategy; the image to be detected is input into the trained model, and the landslide detection result is output. This invention enhances the model's feature extraction capability for landslide targets through a cascaded design of long-range feature modeling and local feature extraction, and improves the bounding box regression accuracy by combining the improved localization loss function. This effectively improves the accuracy and robustness of landslide detection in complex backgrounds of high-altitude mountainous areas.

[0031] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A landslide detection method for plateau mountainous areas based on an improved YOLOv11 deep learning model, characterized in that, Includes the following steps: Acquire remote sensing images of the target area and preprocess the remote sensing images to obtain standardized image data; An initial target detection model based on the YOLOv11 model is constructed, and a feature enhancement module is introduced into the initial target detection model. At the same time, an improved localization loss function and a multi-scale training strategy are combined to enhance the model's ability to express the features of landslide targets under complex terrain conditions and improve detection accuracy. The feature enhancement module is used to perform adaptive spatial dimension recalibration on the remote sensing image feature map extracted by the initial target detection model to enhance the feature representation of the landslide area. The target detection model with the feature enhancement module was trained using the labeled landslide remote sensing image dataset to obtain a trained landslide detection model. The remote sensing image of the plateau and mountainous area to be detected is input into the trained landslide detection model, and the landslide detection results are output.

2. The method as described in claim 1, characterized in that, The feature enhancement module is a spatial adaptive feature modulation module, which adopts a residual connection structure and captures the global context information of the landslide target through long-range feature modeling. The processing is represented as follows: Where X is the input feature map, LN represents layer normalization, SAFM represents spatial adaptive feature modulation function, and Y is the modulated output feature.

3. The method as described in claim 2, characterized in that, The target detection model also incorporates a convolutional channel mixer cascaded with the spatial adaptive feature modulation module to extract local contextual information of the landslide target. The processing procedure is as follows: Where Y is the output of the spatial adaptive feature modulation module, CCM represents the convolutional channel mixer function, and Z is the fused output feature.

4. The method as described in claim 1, characterized in that, The training further includes: optimizing the target detection model using an improved localization loss function, wherein the improved localization loss function combines the geometric similarity information between the predicted box and the ground truth box during the bounding box regression process to improve localization accuracy.

5. The method as described in claim 4, characterized in that, The improved localization loss function is a cross-union ratio (CURRR) loss function based on point distance, and its calculation method is as follows: Where IoU is the intersection-union ratio between the predicted bounding box and the ground truth bounding box; and , respectively, are the squares of the Euclidean distances between the top-left and bottom-right corners of the predicted and ground truth boxes; w and h are the width and height of the minimum bounding rectangle between the predicted and ground truth boxes, respectively, used to normalize the distance terms.

6. The method as described in claim 1, characterized in that, The training further includes employing a multi-scale training strategy, which involves randomly selecting remote sensing image samples of different sizes to input into the network for training during the training process, in order to enhance the model's ability to detect landslide targets at different scales.

7. The method as described in claim 1, characterized in that, The preprocessing includes one or more operations or combinations thereof on the remote sensing image, such as geometric correction, radiometric correction, orthorectification, image cropping, and cloud removal.

8. The method as described in claim 1, characterized in that, The training and validation further include: dynamically adjusting the learning rate using an adaptive learning rate adjustment strategy, and evaluating the model's generalization ability using a cross-validation method.

9. The method as described in claim 1, characterized in that, The landslide detection results include information on the location and boundary range of the landslide; the method further includes: classifying the risk level based on the landslide area characteristics in the landslide detection results, and tracking and updating the landslide detection results over a long period of time to provide data support for landslide disaster prevention and control.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the landslide detection method for plateau and mountainous areas based on the improved YOLOv11 deep learning model as described in any one of claims 1-9.

Images