Landslide detection method and system based on glaf and sa-iou optimization

Abstract

The application discloses a landslide detection method and system based on GLAF and SA-IoU optimization, relates to the fields of machine vision and artificial intelligence, and comprises the following steps: acquiring a high-resolution remote sensing image; performing feature extraction on the input image; introducing a global-local adaptive fusion module into a backbone network to cooperatively model global context information and local detail information; combining a lightweight convolution structure to reduce the calculation complexity; after a detection head outputs a prediction result, adopting a sample perception intersection-over-union loss function to jointly optimize classification and positioning, so as to improve the detection precision of small-scale, irregular and weak-boundary landslide targets in a complex terrain background. The landslide detection method and system based on GLAF and SA-IoU optimization can enhance the distinguishing ability of landslide targets and similar backgrounds in a remote sensing scene, improve the difficult sample learning ability and boundary positioning precision, and improve the accuracy of landslide detection while ensuring the detection efficiency.

Description

Technical Field

[0001] This invention relates to the fields of machine vision and artificial intelligence, and in particular to a landslide detection method and system based on GLAF and SA-IoU optimization. Background Technology

[0002] Landslides are a common and highly destructive geological hazard in mountainous areas, easily causing damage to roads, houses, infrastructure, and the ecological environment. With the development of high-resolution remote sensing imaging technology, automatic landslide detection using remote sensing images has become an important technical approach in geological hazard monitoring. Remote sensing images have advantages such as wide coverage, fast acquisition speed, and repeatable observation, providing a data foundation for large-scale landslide identification.

[0003] Among existing landslide detection methods, deep learning-based target detection methods have been widely used. One-stage target detection models, due to their end-to-end nature and fast inference speed, are well-suited for landslide monitoring tasks in emergency scenarios. One-stage detectors, represented by the YOLO series, have achieved good results in general target detection, but their direct application to remote sensing landslide detection still has significant limitations.

[0004] The reasons are as follows: First, landslide targets in remote sensing images typically exhibit large scale variations, irregular shapes, blurred boundaries, and weak textures, and possess strong visual similarity to background areas such as bare soil, river valleys, and roadside slopes, making the model prone to false positives and false negatives. Second, traditional convolutional feature extraction usually relies on a fixed receptive field, making it difficult to simultaneously consider local detail representation and global context modeling, resulting in insufficient feature discrimination ability against complex terrain backgrounds. Third, most existing loss functions employ a relatively uniform optimization strategy for samples of different difficulty levels, failing to fully consider the differences between small targets, weak boundary targets, and difficult samples during training. Furthermore, in cases of extreme imbalance between positive and negative samples, training is easily dominated by a large number of simple background samples, thereby affecting the accuracy of bounding box localization and classification learning performance.

[0005] Furthermore, although some improved methods have enhanced detection performance by introducing attention mechanisms, lightweight structures, or improved IoU loss, they are still mostly focused on local enhancement or single optimization steps. They have not yet effectively solved the problem of insufficient collaborative modeling of global context information and local detail information, nor have they been able to balance localization accuracy, learning ability for difficult samples, and computational efficiency.

[0006] Therefore, there is an urgent need to provide a new landslide detection method to improve the detection accuracy, robustness, and computational efficiency of small-scale, irregular, and weakly boundaryed landslide targets in complex backgrounds in high-resolution remote sensing images. Summary of the Invention

[0007] The purpose of this invention is to provide a landslide detection method and system based on GLAF and SA-IoU optimization, thereby solving the problems mentioned in the background art.

[0008] To achieve the above objectives, this invention provides a landslide detection method based on GLAF and SA-IoU optimization, comprising the following steps: S1. Data Acquisition and Preprocessing: Acquire high-resolution remote sensing images and preprocess them to obtain standardized image data; S2 and GLA-Det model feature extraction: The YOLO backbone network is used, and DS-C3K2 and DSConv structures are embedded in the YOLO backbone network to extract features from the standardized image data and obtain the basic feature map; S3. Feature Fusion: The Global-Local Adaptive Fusion (GLAF) module is used to divide the basic feature map into a local detail extraction branch and a global context aggregation branch. Through multi-scale dilated convolution, double pooling guided weighting, and residual fusion, global semantic information and local details are combined to complete feature fusion. S4. Model Training Optimization: The sample-aware intersection-union ratio (SA-IoU) loss function is adopted, combined with the dynamic box regression mechanism and the adaptive hard sample weighting mechanism, to calculate the total sample-aware intersection-union ratio loss and optimize the model's localization and classification capabilities.

[0009] Preferably, the preprocessing in S1 includes image size normalization, pixel value standardization, channel arrangement unification, and data format conversion.

[0010] Preferably, the basic feature map in S2 is specifically represented as follows: ; in, X A three-dimensional tensor composed of real numbers. R For the real number field, C , H , W These represent the number of channels, height, and width of the feature, respectively.

[0011] Preferably, the specific steps of S3 are as follows: S31, Through Convolution projects the base feature map to obtain projected features, and then divides the projected features into detail-preserving branch features and adaptive enhancement branch input features along the channel dimension. S32. Input the adaptive enhancement branch input features into the stacked adaptive enhancement unit, which includes multiple adaptive bottleneck modules, to obtain progressively enhanced features. S33. Within each adaptive bottleneck module, channel compression is performed on the input features of the adaptive enhancement branch, and the features are mapped to the hidden feature space to obtain compressed features. S34. Input the compressed features into the adaptive module, use multiple parallel dilated convolution branches with different dilation rates to extract multi-scale features, and concatenate them to obtain the fused multi-receptive field value features. S35. Use double pooling to extract global contextual information of value features, and obtain content features and auxiliary guidance features through convolution projection. Based on the two types of features, construct spatial response and channel response, and then construct a spatial-channel joint regulation mechanism with global semantic guidance. S36. Based on the constructed spatial and channel responses, a weighted mask is generated. The value features are adaptively recalibrated using the weighted mask to enhance the effective response of the landslide target and suppress background interference response, thereby obtaining enhanced features. S37. By restoring the convolution, the enhanced features are mapped to the target output space to obtain the restored features. The restored features are then fused with the input features of the current adaptive bottleneck module using residual fusion to obtain the output features of the single-level adaptive bottleneck module. S38. The detailed branch features, original split features and output features of each level of adaptive bottleneck module are concatenated and then fused by convolution to obtain the final features that have both local details and global context information.

[0012] Preferably, the specific formula for the final feature in S38 is as follows: ; in, F As the final feature, Indicates fused convolution, To preserve branching features in detail, To adaptively enhance the input features of the branch, The output characteristics of each level of adaptive bottleneck module.

[0013] Preferably, the specific steps of S4 are as follows: S41. Input the enhanced features into the detection head to obtain the predicted bounding box, classification logits and corresponding confidence of the candidate target, and at the same time determine the true bounding box and the true label; S42. Introduce a scaling factor that changes with the training phase. Under the condition that the center of the box remains unchanged, scale the width and height of the predicted bounding box to obtain the dynamically scaled predicted bounding box. S43. Calculate the intersection-union ratio (IUU) between the dynamically scaled predicted bounding box and the true bounding box, and construct the regression loss term by combining the scaling weight function and the quality-aware modulation function. S44. Apply a Sigmoid transformation to the classification logits to obtain the classification probability, and calculate the unweighted basic binary cross-entropy loss. S45. Calculate the average intersection-union ratio of all positive samples in the current batch to evaluate the overall matching quality of the samples; S46. Based on the relative relationship between the crossover ratio (CUP) of a single sample and the average CUP, adaptive classification weights are constructed using a piecewise function. The specific formula for the piecewise function is as follows: ; in, As a classification weighting factor, The crossover ratio for a single sample. For the average crossover ratio, This is a boundary bandwidth hyperparameter used to define the transition region near the average intersection-union ratio. e It is a natural constant; S47. Apply the adaptive classification weights to the basic classification loss to obtain the weighted classification loss term; S48. According to the preset balance coefficient, the regression loss term and the weighted classification loss term are weighted and summed to obtain the total loss of the sample perception intersection ratio.

[0014] Preferably, the specific formula for the total loss of sample-perceived crossover ratio in S48 is as follows: ; in, The total loss is calculated by the crossover and union ratio of the samples. For the regression loss term, For weighted classification loss items, and These represent the weight coefficients of the regression loss term and the weighted classification loss term, respectively.

[0015] A landslide detection method system based on GLAF and SA-IoU optimization includes: Data preprocessing unit: used to preprocess the input image to ensure image consistency; Feature extraction unit: Based on the YOLO backbone network, embedding DS-C3K2 and DSConv modules to extract features from the image; Global-Local Adaptive Fusion Unit (GLAF): Includes the global-to-local design module (GL) and the adaptive controllable receptive field module (ACRM). By adaptively adjusting the receptive field range, it fuses global semantic information with local details to improve the detection accuracy of landslide targets in complex backgrounds. Sample-aware Cross-Union Ratio (SA-IoU) Loss Unit: This unit employs a sample-aware SA-IoU loss function, combined with a quality-aware dynamic box regression mechanism and an adaptive hard sample weighting mechanism based on SA-IoU, to improve localization accuracy and the learning effect of hard samples.

[0016] Therefore, the landslide detection method and system based on GLAF and SA-IoU optimization described above have the following beneficial effects: (1) To address the problem of large redundancy and high computational cost in existing methods, this invention introduces the DS-C3K2 and DSConv lightweight structures to reduce the number of parameters and computational cost while ensuring feature extraction capability, thereby improving the model running efficiency.

[0017] (2) In view of the problems of insufficient global semantic information extraction and insufficient local detail expression in the existing YOLO model, this invention introduces the global-local adaptive feature fusion module GLAF. Through multi-scale dilated convolution, dual pooling guided weighting and residual fusion, local detail extraction is combined with global context aggregation, thereby enhancing the feature expression of landslide targets, suppressing interference from complex backgrounds, and improving target discrimination ability.

[0018] (3) In view of the problem that existing loss functions do not pay enough attention to difficult samples and have low positioning accuracy, this invention designs sample-aware intersection-union ratio loss (SA-IoU), improves the bounding box regression accuracy through dynamic scaling mechanism, and enhances the learning ability of difficult samples through classification reweighting mechanism, thereby improving the detection effect of small-scale, irregular and weak boundary landslide targets.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating an embodiment of the landslide detection method and system optimized by GLAF and SA-IoU according to the present invention. Figure 2 This is a schematic diagram of the global-local adaptive feature fusion module structure of an embodiment of the landslide detection method and system based on GLAF and SA-IoU optimization of the present invention; Figure 3 This is a schematic diagram of the dynamic scaling of the prediction box in the SA-IoU loss of the landslide detection method and system based on GLAF and SA-IoU optimization according to the present invention. Detailed Implementation

[0021] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0023] Example Please see Figures 1-3 This invention provides a landslide detection method and system based on GLAF and SA-IoU optimization, aiming to effectively improve the detection capability of the model in complex scenarios without increasing model complexity. Figure 1 As shown, the system includes a data preprocessing unit and a detection network designed based on the YOLO model. The data preprocessing unit preprocesses the input image, including normalizing the image size, standardizing pixel values, unifying channel arrangement, and converting data format to ensure image consistency.

[0024] The detection network designed based on the YOLO model includes: using YOLOv12 as the base network, introducing a global-local adaptive feature fusion module into the backbone network, incorporating DS-C3K2 and DSConv structures into the convolutional units, and using sample-aware cross-union loss for optimization during the training phase; the input image passes through the backbone network to obtain intermediate feature maps. Perform feature map processing After convolutional projection, the algorithm is divided into local and global branches. The local branch extracts fine-grained boundaries and textures through parallel depthwise convolution; the global branch extracts global contextual information through channel compression, multi-scale dilated convolution, double pooling guided weighting, channel recovery, and residual connections; finally, the two types of features are fused into enhanced output features. During the training phase, bounding box regression is performed on the predicted bounding boxes using a unified intersection-union ratio (IU) loss with dynamic scaling and quality awareness; classification optimization is performed on the classification branch using a dynamic reweighted loss based on IU statistics; the weighted sum of the two is used as the total loss to update the network parameters.

[0025] The landslide detection method based on GLAF and SA-IoU optimization includes the following steps: S1. Acquire the high-resolution remote sensing image to be detected and perform preprocessing operations on the remote sensing image. Preprocessing operations include image size normalization, pixel value standardization, channel arrangement unification, and data format conversion to obtain standardized image data for input to the detection network. The specific steps are as follows: S11. Image size normalization: Scaling remote sensing images of different sizes to the size required by the model, such as 640×640.

[0026] S12, Pixel value standardization: Adjust the pixel values ​​to a uniform range, such as converting them from 0-255 to 0-1, to facilitate model calculation.

[0027] S13. Unified channel arrangement: Unify the image channel order, such as RGB format, to avoid BGR / RGB confusion.

[0028] S14. Data format conversion: Convert the image into a data format that the model can read, such as converting the image format into a Tensor.

[0029] It should be noted that S11-S14 are optional serial preprocessing steps, processed as needed based on the original state of the remote sensing image: 1) Verify step by step whether the remote sensing image matches the model input specification, and only perform the corresponding operation on items that do not meet the conditions; 2) Processing follows a fixed sequence from S11 to S14; 3) If the parameters of the remote sensing image itself already meet the standards of the corresponding step, then skip this step.

[0030] Example: If the input remote sensing image resolution is 320×320, which does not conform to the standard 640×640 size, then perform image size normalization in step S11; if the pixel value, channel order, and data format all meet the standards, then only step S11 is performed, and the other steps are skipped directly, and the rest of the process is the same.

[0031] S2. Input the preprocessed remote sensing image into a YOLO-based backbone network to extract intermediate feature maps. : ; in, C Indicates the number of channels. H Indicates the feature map height. W Indicates the width of the feature map. It is a three-dimensional tensor composed of real numbers, containing the boundary, texture, shape and background semantic information of the landslide area, providing input for subsequent feature enhancement.

[0032] S3, Transfer the intermediate feature map The input global-local adaptive feature fusion module (GLAF) performs detail preservation, context enhancement, and multi-level fusion on the input features to obtain the enhanced feature map. .

[0033] The Global-Local Adaptive Feature Fusion (GLAF) module includes an input projection layer, feature splitting units, and stacked adaptive enhancement units. This module's role is to simultaneously preserve the fine-grained local structural information of the landslide target while progressively introducing broader contextual semantic information to improve the separability of the landslide target from similar backgrounds in complex remote sensing scenes. Figure 2 As shown, the specific steps are as follows: S301, Given the input feature map First through The convolution is linearly projected onto it to obtain the projected features. : ; in, Indicates the input projection convolution.

[0034] S302, Projecting features Divide into two sub-features along the channel dimension, denoted as detail-preserving branch features respectively. and adaptive augmentation of branch input features : ; in, It mainly preserves fine-grained structural information such as boundaries, texture discontinuities, and local morphological undulations; It serves as input to subsequent adaptive enhancement units, used to extract a wide range of contextual information step by step.

[0035] S303, Adaptively enhance branch input features Input by In a stacked adaptive enhancement unit composed of several adaptive bottleneck modules, progressively enhanced features are obtained: ; in, , Indicates the first The AdaptiveBottleNeck is designed to allow feature representations to evolve from local responses to a more comprehensive understanding of the context.

[0036] Each AdaptiveBottleNeck consists of four stages: channel compression, adaptive multi-scale enhancement, channel restoration, and residual fusion.

[0037] S304, Regarding the first Each AdaptiveBottleNeck has the following input features: First through Convolution maps input features to a hidden feature space, resulting in compressed features. : ; in, This is a channel compression function.

[0038] Let the number of input and output channels be... The number of hidden channels is ,but: ; in, This is the channel reduction factor.

[0039] S305, Compress features The input adaptive block uses multiple parallel dilated convolutional branches with different dilation rates to extract multi-scale responses. Let the set of dilation rates be: ; For each expansion rate ,have: ; in, Indicates the expansion rate Dilated convolution. The outputs of each branch are concatenated to obtain a multi-scale feature representation: ; in, The value features are used to preserve the original multi-scale responses under different receptive fields and serve as the objects for subsequent adaptive recalibration.

[0040] S306, Multi-scale features Performing global average pooling and global max pooling respectively yields two global descriptors: ; ; And a compact global description is obtained by averaging element by element: ; Global context summary information is extracted from multi-scale features to provide global guidance for subsequent spatial-channel joint recalibration.

[0041] Through two Convolutional projection generates content features and auxiliary guidance features : ; ; in, Used to characterize the local content response at the current location. Used to provide auxiliary guidance information generated from the global description.

[0042] S307, Based on content features and auxiliary guidance features Construct spatial responses respectively and channel response , can be represented as: ; ; in, This represents element-wise multiplication. and These represent the spatial response map and the channel response map, respectively. The spatial response is used to characterize important spatial locations under global guidance, while the channel response is used to characterize the interaction strength between different channels.

[0043] S308, Spatial Response and channel response The input is concatenated, gated convolution is passed to Sigmoid activation, and a weighted mask is generated. : ; in, Represents gated convolution mapping, This represents the Sigmoid activation function.

[0044] Then, using a weighted mask logarithmic features Adaptive recalibration is performed to obtain enhanced features. : ; The purpose of this step is to enhance the effective response related to landslides and suppress background disturbance responses such as bare soil, river valleys, and road slopes.

[0045] S309, Enhanced Features pass Convolution maps back to the target output space to obtain the recovered features. : ; in, This indicates channel-recovery convolution. This step is used to restore the hidden space features to the output channel dimension so that they can be passed to subsequent layers.

[0046] S310: Recovery characteristics Input features of the current AdaptiveBottleNeck Perform residual summation to obtain the first... Output of each AdaptiveBottleNeck: ; This is to preserve low-level semantic information and enhance the stability of gradient propagation.

[0047] S311: Preserve details of branching features Primitive splitting characteristics and the output features of each level of AdaptiveBottleNeck Perform channel splicing, and through Convolutional processes are used to fuse the features, resulting in the final output features of the GLAF module. : ; in, This represents a fused convolution. The final output feature contains both local detail information and global contextual information, which is used for subsequent landslide target detection.

[0048] S4. Enhance the feature map Input the neck network and the detection head to obtain the predicted bounding box, classification output and target confidence of the candidate landslide target.

[0049] S5. During the training phase, a sample-aware intersection-union loss (SA-IoU) is constructed based on the output of the detection head, and the network parameters are updated by backpropagation.

[0050] SA-IoU includes a quality-aware dynamic bounding box regression mechanism and an adaptive hard sample weighting mechanism based on the intersection-union ratio. This module aims to simultaneously improve bounding box localization quality and hard sample learning ability under conditions of severe positive-negative sample imbalance. The specific steps are as follows: S501, use the candidate target prediction results output by the detection head as the input of the sample perceptual intersection-union ratio (SA-IoU) loss. Let the first... The predicted bounding boxes for each sample are: ; The true bounding box is: ; in, These represent the coordinates of the center point of the predicted bounding box. These represent the width and height of the predicted bounding box, respectively. These represent the coordinates of the center point of the actual bounding box. These represent the width and height of the actual bounding box, respectively.

[0051] Meanwhile, let the classification logits corresponding to this sample be... The real label is The prediction confidence level is The predicted bounding box is used for subsequent localization regression optimization, while the classification logits, ground truth labels, and prediction confidence scores are used for subsequent classification loss calculation and quality modulation.

[0052] S502. To improve the localization accuracy of small-scale, irregular, and weakly bounded landslide targets, during the training phase, the predicted bounding box is... Introducing dynamic scaling factor The scaled predicted bounding box is obtained. : ; in, This indicates that a scaling operation is performed on the predicted bounding box. Indicates the first The dynamic scaling factor corresponding to each training phase.

[0053] S503, Utilizing the scaled predicted bounding box With the true bounding box Calculate the scaling crossover / union ratio: ; in, Indicates the first The intersection-over-union ratio (IoU) between the predicted bounding boxes and the ground truth bounding boxes for each sample under dynamic scaling conditions. This IoU serves as a measure of the quality of predicted bounding box localization in the regression branch and is used to subsequently construct the quality-aware dynamic bounding box regression loss term.

[0054] S504, Based on Scaling Intersection over Union Ratio Dynamic scaling factor and prediction confidence Construct the regression loss term The expression for the regression loss term is: ; in, This indicates the number of positive samples in the current batch; Indicates dynamic scaling factor The determined scaling weight function is used to adjust the strength of regression constraints at different training stages; Indicates the prediction confidence level A quality-aware modulation function is constructed to adjust the contribution of the sample to the regression loss based on the current confidence level of the prediction box.

[0055] S505. Apply a Sigmoid transformation to the classification logits to obtain the... The classification probability of each sample: ; Further calculation of the basic binary cross-entropy loss: ; in, This represents the Sigmoid activation function. Indicates the first The basic classification error of a sample under unweighted conditions.

[0056] S506. In the current training batch, calculate the average crossover ratio (CR) of all positive samples. Its expression is: ; in, Indicates the first Each positive sample corresponds to the intersection-over-union ratio between the predicted bounding box and the ground truth bounding box. Used to characterize the overall matching quality level of the current batch of positive samples.

[0057] S507, Based on the crossover ratio of a single sample Crossover ratio with average Differences, constructing classification weighting factors Classification weighting factor The following piecewise function is used to represent it: ; in, This is the boundary bandwidth hyperparameter, used to define the transition interval near the average intersection-union ratio.

[0058] Classification weighting factors Application to basic classification loss The weighted classification loss item is obtained. : ; in, This indicates the total number of samples in the current batch.

[0059] S508, Regression Loss Item With classification loss item According to the preset balance coefficient and Weighted summation yields the total loss of the sample perceived intersection-union ratio. : ; in, This represents the weighting coefficient of the regression loss term. The weight coefficients of the classification loss term are used to coordinate the contribution relationship between bounding box localization optimization and difficult sample classification learning.

[0060] This embodiment also provides a system for landslide detection based on GLAF and SA-IoU optimization, including: Data preprocessing unit: used to preprocess the input image to ensure image consistency; Feature extraction unit: Based on the YOLO backbone network, embedding DS-C3K2 and DSConv modules to extract features from the image; Global-Local Adaptive Fusion Unit (GLAF): Includes the global-to-local design module (GL) and the adaptive controllable receptive field module (ACRM). By adaptively adjusting the receptive field range, it fuses global semantic information with local details to improve the detection accuracy of landslide targets in complex backgrounds. Sample-aware Cross-Union Ratio (SA-IoU) Loss Unit: This unit employs a sample-aware SA-IoU loss function, combined with a quality-aware dynamic box regression mechanism and an adaptive hard sample weighting mechanism based on SA-IoU, to improve localization accuracy and the learning effect of hard samples.

[0061] Therefore, the landslide detection method and system based on GLAF and SA-IoU optimization described above can enhance the ability to distinguish landslide targets from similar backgrounds such as bare soil, river valley slopes, and road slopes in remote sensing scenes. At the same time, it can improve the learning ability of difficult samples and the accuracy of boundary positioning, thereby improving the accuracy of landslide detection while ensuring detection efficiency.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A landslide detection method based on GLAF and SA-IoU optimization, characterized in that, Includes the following steps: S1. Data Acquisition and Preprocessing: Acquire high-resolution remote sensing images and preprocess them to obtain standardized image data; S2 and GLA-Det model feature extraction: The YOLO backbone network is used, and DS-C3K2 and DSConv structures are embedded in the YOLO backbone network to extract features from the standardized image data and obtain the basic feature map; S3. Feature Fusion: The Global-Local Adaptive Fusion (GLAF) module is used to divide the basic feature map into a local detail extraction branch and a global context aggregation branch. Through multi-scale dilated convolution, double pooling guided weighting, and residual fusion, global semantic information and local details are combined to complete feature fusion. The specific steps of S3 are as follows: S31, Through Convolution projects the base feature map to obtain projected features, and then divides the projected features into detail-preserving branch features and adaptive enhancement branch input features along the channel dimension. S32. Input the adaptive enhancement branch input features into the stacked adaptive enhancement unit, which includes multiple adaptive bottleneck modules, to obtain progressively enhanced features. S33. Within each adaptive bottleneck module, channel compression is performed on the input features of the adaptive enhancement branch, and the features are mapped to the hidden feature space to obtain compressed features. S34. Input the compressed features into the adaptive module, use multiple parallel dilated convolution branches with different dilation rates to extract multi-scale features, and concatenate them to obtain the fused multi-receptive field value features. S35. Use double pooling to extract global contextual information of value features, and obtain content features and auxiliary guidance features through convolution projection. Based on the two types of features, construct spatial response and channel response, and then construct a spatial-channel joint regulation mechanism for global semantic guidance. S36. Based on the constructed spatial and channel responses, a weighted mask is generated. The value features are adaptively recalibrated using the weighted mask to enhance the effective response of the landslide target and suppress background interference response, thereby obtaining enhanced features. S37. By restoring the convolution, the enhanced features are mapped to the target output space to obtain the restored features. The restored features are then fused with the input features of the current adaptive bottleneck module using residual fusion to obtain the output features of the single-level adaptive bottleneck module. S38. The detail-preserving branch features, original split features and output features of each level of adaptive bottleneck module are concatenated and then fused by convolution to obtain the final features that have both local details and global context information. S4. Model Training Optimization: The sample-aware intersection-union ratio (SA-IoU) loss function is adopted, combined with the dynamic box regression mechanism and the adaptive hard sample weighting mechanism, to calculate the total sample-aware intersection-union ratio loss and optimize the model's localization and classification capabilities.

2. The landslide detection method based on GLAF and SA-IoU optimization according to claim 1, characterized in that: The preprocessing in S1 includes image size normalization, pixel value standardization, channel arrangement unification, and data format conversion.

3. The landslide detection method based on GLAF and SA-IoU optimization according to claim 1, characterized in that, The specific representation of the basic feature map in S2 is as follows: ； in, X A three-dimensional tensor composed of real numbers. R For the real number field, C , H , W These represent the number of channels, height, and width of the feature, respectively.

4. The landslide detection method based on GLAF and SA-IoU optimization according to claim 1, characterized in that, The specific formula for the final feature in S38 is as follows: ； in, F As the final feature, Indicates fused convolution, To preserve branching features in detail, To adaptively enhance the input features of the branch, The output characteristics of each level of adaptive bottleneck module.

5. The landslide detection method based on GLAF and SA-IoU optimization according to claim 1, characterized in that, The specific steps of S4 are as follows: S41. Input the enhanced features into the detection head to obtain the predicted bounding box, classification logits and corresponding confidence of the candidate target, and at the same time determine the true bounding box and the true label; S42. Introduce a scaling factor that changes with the training phase. Under the condition that the center of the box remains unchanged, scale the width and height of the predicted bounding box to obtain the dynamically scaled predicted bounding box. S43. Calculate the intersection-union ratio (IUU) between the dynamically scaled predicted bounding box and the true bounding box, and construct the regression loss term by combining the scaling weight function and the quality-aware modulation function. S44. Apply a Sigmoid transformation to the classification logits to obtain the classification probability, and calculate the unweighted basic binary cross-entropy loss. S45. Calculate the average intersection-union ratio of all positive samples in the current batch to evaluate the overall matching quality of the samples; S46. Based on the relative relationship between the crossover ratio (CR) of a single sample and the average CR, adaptive classification weights are constructed using a piecewise function. The specific formula for the piecewise function is as follows: ； in, As a classification weighting factor, The crossover ratio for a single sample. For the average crossover ratio, This is a boundary bandwidth hyperparameter used to define the transition region near the average intersection-union ratio. e It is a natural constant; S47. Apply the adaptive classification weights to the basic classification loss to obtain the weighted classification loss term; S48. According to the preset balance coefficient, the regression loss term and the weighted classification loss term are weighted and summed to obtain the total loss of sample perception intersection ratio.

6. The landslide detection method based on GLAF and SA-IoU optimization according to claim 5, characterized in that, The specific formula for the total loss of sample-perceived cross-union ratio in S48 is as follows: ； in, The total loss is calculated by the crossover and union ratio of the samples. For the regression loss term, For weighted classification loss items, and These represent the weight coefficients of the regression loss term and the weighted classification loss term, respectively.

7. A system applying the landslide detection method based on GLAF and SA-IoU optimization as described in any one of claims 1-6, characterized in that, include: Data preprocessing unit: used to preprocess the input image to ensure image consistency; Feature extraction unit: Based on the YOLO backbone network, embedding DS-C3K2 and DSConv modules to extract features from the image; Global-Local Adaptive Fusion Unit (GLAF): Includes the global-to-local design module (GL) and the adaptive controllable receptive field module (ACRM). By adaptively adjusting the receptive field range, it fuses global semantic information with local details to improve the detection accuracy of landslide targets in complex backgrounds. Sample-aware Cross-Union Ratio (SA-IoU) Loss Unit: This unit employs a sample-aware SA-IoU loss function, combined with a quality-aware dynamic box regression mechanism and an adaptive hard sample weighting mechanism based on SA-IoU, to improve localization accuracy and the learning effect of hard samples.

Images