Geometry, context and frequency domain multi-source feature fused landslide detection method

Abstract

The invention discloses a landslide detection deep learning method fusing geometry, context and frequency domain multi-source features, and belongs to the technical field of geological disaster recognition and landslide remote sensing information processing. The problem that a traditional landslide detection method in the prior art is low in recognition precision is solved. The method comprises the following steps: constructing a landslide detection model based on a GCF-Net architecture, and inputting a to-be-detected image into the landslide detection model; capturing feature information of different scales in the to-be-detected image through a ghost pyramid feature fusion module; an output characteristic pattern is input to a decoder through a GCFmodule; and an EdgeECI module in the decoder is used to process the output feature map of the GCFI module, and a final output feature map after fusion is obtained. The landslide detection accuracy is effectively improved, and the method can be applied to geological disaster monitoring and early warning, mountain traffic and infrastructure safety, water conservancy and energy engineering operation monitoring, remote sensing big data processing platforms and other scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to a landslide detection deep learning method, in particular to a landslide detection method fusing geometric, contextual and frequency domain multi-source features, and belongs to the technical field of geological disaster identification and landslide remote sensing information processing. BACKGROUND

[0002] Landslide is one of the most common and destructive geological disasters, especially in the mountainous areas and hilly regions of Southwest China. It is often caused by the combined effects of geological structure, heavy rainfall, earthquakes and human engineering activities, and can cause road blockage, house collapse, water and energy facilities damage in a short time, posing a serious threat to people's life and property safety and regional economic development. Therefore, how to use efficient and accurate technical means to carry out landslide disaster identification and monitoring has always been an important problem in the field of geological disaster prevention and remote sensing application research. With the rapid development of remote sensing technology, high-resolution satellite images, unmanned aerial images and multi-source heterogeneous remote sensing data have gradually become an important data source for landslide identification, providing a wide range of non-contact and fast updating observation means for disaster detection and emergency response.

[0003] However, the performance of landslide in remote sensing image is extremely complex. On the one hand, the shape of landslide is significantly different, including large-scale landslides and small and scattered local landslides. On the other hand, the boundary of landslide area is often fuzzy, the internal texture is weakened, and the spectrum is similar to the surrounding bare soil or rock. Under the condition of vegetation coverage or shadow interference, the above targets are more likely to be confused with the background, leading to difficult identification. Traditional classification methods based on spectral difference, such as threshold method and object-oriented image analysis, are prone to misjudgment of bare land, farmland or building ruins as landslide, and have poor generalization performance. Machine learning methods such as support vector machine, random forest and k-nearest neighbor, although introducing texture, shape and terrain factors, rely on manual features and are difficult to express complex spatial context relationships, so traditional landslide identification methods still have the problem of insufficient accuracy when facing complex and variable landslide shapes in high-resolution remote sensing images.

[0004] In recent years, deep learning techniques have shown strong feature extraction and prediction capabilities in remote sensing image interpretation and geological disaster identification. In particular, the encoder-decoder type convolutional neural network (such as U-Net and its improved architecture) is widely used to process high-resolution remote sensing data. Landslide, as a typical geological disaster, has complex shape, fuzzy boundary and large scale span. Traditional methods that rely solely on spectral or texture differences have been difficult to meet the needs of high-precision automatic identification. In this context, the semantic segmentation framework based on deep learning has gradually become the mainstream technology path. Convolutional neural networks represented by U-Net can achieve multi-scale feature fusion through skip connection, which can balance semantic representation and spatial resolution to some extent, and has been widely used in landslide extraction tasks. UNet++ introduces dense skip connection to enhance feature interaction and multi-level expression. Attention U-Net improves the focusing ability on target areas through attention mechanism. SwinUNet and TransUNet, etc. Transformer model uses global self-attention structure to realize long-range dependency modeling and enhance the adaptability to complex scenes. However, the methods introduced deep learning in landslide identification still have common limitations: first, the use of geometric structure features is insufficient. Most methods only use terrain or spatial location information as fixed additional input, and cannot dynamically adjust the model's attention to the boundary or local structure during training. Second, the calculation of traditional convolution or attention weight mainly depends on the similarity measurement in feature space, which lacks direct modeling of geometric boundary, high-frequency texture and other information in landslide area, leading to missed detection and misjudgment in vegetation coverage, shadow interference or small-scale landslide scenes. Third, due to repeated downsampling operations, the model's receptive field is limited, making it difficult to balance the long-range context relationship and fine-grained boundary features at the same time. The problems of boundary ambiguity and target structure loss are still prominent.

[0005] Currently, there are studies that attempt to introduce spatial pyramid pooling, multi-scale context awareness and frequency domain attention mechanism into convolutional neural models to enhance the model's ability to adapt to different scales and complex textures. For example, the structure of dilated convolution and pyramid pooling improves the capture of multi-scale features to some extent, and the frequency domain enhancement method can highlight some edges and details. However, the improved models generally have two problems: first, the fusion of geometric and frequency domain features is still mainly static splicing, lacking a dynamically learnable regulation mechanism; second, the combination of edge information and semantic features is not close enough, and it performs poorly in small-scale target segmentation and fuzzy boundary segmentation. Therefore, in actual complex terrain scenarios, the improved model often has certain advantages in large-scale landslide identification, but it still lacks in small-scale landslide target detection and boundary recovery. Moreover, the improved model still faces several technical bottlenecks in the landslide segmentation task: first, due to the limitations of network downsampling operations and limited receptive fields, the model lacks in long-range dependency modeling, making it difficult to balance large-scale context and fine-grained details; second, the boundaries of landslide areas are often fuzzy and fragmented, and traditional methods lack the ability to recover boundaries and maintain details, leading to inaccurate boundary segmentation; third, for small-scale landslides with small areas and irregular edges, existing networks generally have low recognition accuracy and are prone to missed detection and false detection; in addition, they lack effective utilization of geometric structures and frequency domain information, often relying only on spatial domain convolution, which cannot fully express the complex geometric semantics and high-frequency texture features of landslide targets. In summary, although existing deep learning methods have promoted the development of landslide remote sensing segmentation technology, they still have shortcomings in global semantic modeling, boundary detail recovery, small target recognition, and multi-source interference suppression.

[0006] In summary, existing landslide segmentation methods based on U-Net, Attention mechanism or Transformer have made progress in semantic modeling and global dependency expression, but still struggle to solve core problems such as fuzzy boundaries, weak textures, small target easy to miss detection, and insufficient geometric structure expression in landslide remote sensing images. Therefore, there is an urgent need for a landslide detection method that integrates geometric, contextual and frequency domain multi-source features to improve the accuracy and robustness of landslide identification in complex terrain conditions and provide more reliable technical support for geological disaster monitoring and emergency management. SUMMARY

[0007] A brief summary of the present application is given below to provide a basic understanding of some aspects of the present application. It should be understood that this summary is not an exhaustive overview of the present application. It is not intended to identify key or important parts of the present application, nor is it intended to limit the scope of the present application. Its purpose is simply to give some concepts in a simplified form as a prelude to the more detailed description discussed later.

[0008] In view of this, in order to solve the problem of low recognition accuracy of traditional landslide detection methods in the prior art, the present invention provides a landslide detection method that integrates geometric, context and frequency domain multi-source features.

[0009] The technical solution is as follows: a landslide detection method integrating geometric, contextual, and frequency domain multi-source features, including the following steps:

[0010] S1. Construct a landslide detection model based on the GCF-Net architecture, which includes an encoder and a decoder, and input the image to be detected to the landslide detection model;

[0011] S2. By using the ghost pyramid feature fusion module in the encoder, multi-scale pyramid pooling and ghost convolution mechanisms are employed to capture feature information at different scales in the image to be detected.

[0012] S3. Through the GCF² module in the encoder, feature information of different scales is further fused in the connecting bridge, and the resulting GCF² module output feature map is input into the decoder;

[0013] S4. The EdgeECI module in the decoder performs depthwise separable convolution and point convolution operations on the output feature map of the GCF² module step by step. The results of the operation are processed by an efficient channel spatial attention mechanism and an edge-guided mechanism to obtain the fused final output feature map.

[0014] Furthermore, in S2, a ghost convolution mechanism is used. For the input image to be detected The process is performed to obtain the processed feature map. Multi-scale pyramid pooling processes the feature maps at each scale. Perform average pooling operation Post-fusion features yield pyramid pooling output. Perform convolution operation on it The feature map after the convolution operation is obtained. ;

[0015] Processed feature map Represented as:

[0016]

[0017] in, , , , These represent the image height, image width, and number of channels, respectively. This indicates the number of intermediate channels in the Ghost Pyramid feature fusion module;

[0018] Pyramid pooling output is represented as:

[0019]

[0020] wherein, represents the window size of the average pooling operation, represents the number of channels of the convolution operation output in the ghost pyramid feature fusion module;

[0021] the feature map after the convolution operation is represented as:

[0022] .

[0023] Further, in the S3, the GCF2 module includes a branch A and a branch B, the branch A processes the feature map after the convolution operation through a frequency domain enhancement and an autonomous learning position encoding to obtain the branch A output ;

[0024] The frequency domain enhancement process is represented as:

[0025]

[0026] wherein, represents the feature map output by the frequency domain enhancement, i.e., the frequency domain feature, is a frequency domain high-frequency mask, is a 1x1 convolution operation;

[0027] The autonomous learning position encoding process is represented as:

[0028]

[0029] wherein, represents the position encoding, is the encoding result of the frequency domain feature;

[0030] The branch A output

[0031]

[0032] wherein, represents group normalization;

[0033] The branch B extracts the context information of the feature map after the convolution operation through a multi-scale convolution operation, to obtain feature maps , , under different convolution kernels, which are fused through a fusion convolution layer and a concatenation operation to obtain a fusion feature map ;

[0034] Feature maps under different convolution kernels , , Represented as:

[0035]

[0036] in, This represents a 3×3 convolution operation. This represents a 5×5 convolution operation;

[0037] Fusion Feature Map Represented as:

[0038]

[0039] Based on the feature map after convolution operation Foreground encoding The feature map after convolution operation and fusion feature map Enhance contrast The enhanced feature map is obtained. ;

[0040]

[0041] Enhanced feature map and branch A output The feature maps are fused by stitching and then weighted by a gating mechanism to obtain the gated fused feature maps. ;

[0042] Feature map after gating fusion

[0043]

[0044] in, This represents the gating weight of branch A. This represents the gating weight of branch B;

[0045] Feature map of GCF² module after gated fusion Introducing channel attention mechanism The output feature map of the GCF² module is obtained. ;

[0046] GCF² module output feature map Represented as:

[0047] .

[0048] Furthermore, in S4, the EdgeECI module outputs a feature map to the GCF² module. Perform depthwise separable convolution The feature map obtained after depthwise separable convolution is obtained. ;

[0049] Feature map after depthwise separable convolution Represented as:

[0050]

[0051] By analyzing the feature map after depthwise separable convolution... Perform point convolution operation to obtain the feature map after point convolution operation. ;

[0052] Feature map after point convolution operation Represented as:

[0053]

[0054] Feature maps after point convolution operation Employing an efficient channel spatial attention mechanism The feature map is obtained after processing by the efficient channel spatial attention module. ;

[0055] Feature map processed by efficient channel spatial attention mechanism Represented as:

[0056]

[0057] Feature maps processed by an efficient channel spatial attention mechanism Employing edge-guided mechanism Weighting is performed to obtain the weighted feature map obtained through the edge-guided mechanism. The final output feature map is obtained by integration. That is, the landslide detection results;

[0058] Weighted feature map guided by edge mechanism Represented as:

[0059]

[0060] The final output feature map after fusion Represented as:

[0061] .

[0062] The beneficial effects of the present application are as follows: the core of the present application is to construct structure-aware geometric coding, multi-scale context and frequency domain feature extractor in parallel, and introduce foreground coding and learnable gating mechanism in the fusion module to realize adaptive weighting and coupling of features; through the geometric dominant gate, the frequency domain and the context information are selectively compensated and noise suppressed under the guidance of geometry, thereby enhancing the perception and discrimination ability of landslide features under complex terrain, hidden small scale and noise interference;

[0063] The present application proposes a combination of ghost convolution and lightweight multi-scale pyramid pooling PGF Block to extract rich semantic features across scales, and the PGF Block module expands the receptive field without increasing parameters, and improves the ability of the encoder to model spatial dependencies in large and fragmented landslide areas.

[0064] The present application designs a geometry-context-frequency fusion module GCF², which combines spatial edges and frequency domain features to strengthen the delineation of boundaries and internal textures, and a foreground enhancement branch to further emphasize semantics specific to landslides and suppress background noise, thereby achieving clearer segmentation in low-contrast and visually confusing areas.

[0065] The present application is based on the EdgeECI interaction module of the learnable edge guidance and efficient channel spatial attention mechanism, which introduces learnable interaction between channels and spatial clues, significantly improves the detection accuracy of subtle and small-scale landslides by enhancing fine-grained boundary information and restoring key details.

[0066] To verify the effectiveness and superiority of the GCF-Net model proposed in the present application in the landslide remote sensing image segmentation task, I carried out comparative experiments with five representative semantic segmentation methods, including U-Net, UNet++, AttUNet, SwinUNet and TransUNet, all models use the same training settings and do not use pre-trained weights to ensure fairness.

[0067] The experimental results show that the landslide detection model based on the GCF-Net architecture performs well in overall precision and key indicators, the overall accuracy (OA) of the model reaches 96.83%, which is about 1.93% and 2.03% higher than the traditional U-Net and UNet++, respectively, and the F1 score is 1.65% higher than the optimal comparative model TransUNet, which reflects a good balance between precision and recall rate, and the improvement is mainly due to the fusion of frequency domain, geometry and context information in the encoder bridge stage, which enables the model to achieve global consistency and fine identification in complex terrain and small landslides.

[0068] In terms of model complexity, the number of parameters, floating point operations and model size of the model are significantly lower than those of the comparative method, achieving a reduction in computational and storage burden while maintaining high accuracy, providing conditions for small memory device deployment. Meanwhile, the GCF2 module performs best when accessing middle-level features, balancing foreground enhancement and background suppression. Shallow layers are affected by noise, and deep layers are affected by boundary precision due to loss of details, indicating that middle-level feature access significantly optimizes performance.

[0069] In addition, the learnable edge guiding mechanism further improves the ability to capture small landslide areas and boundaries. Under moderate fusion weights, the model achieves optimal performance in OA, AA, MIoU, recall rate and F1, verifying its adaptive learning ability for boundary details and small targets, while avoiding negative interference that may be caused by excessive guidance.

[0070] In summary, the model is superior to mainstream segmentation methods in terms of accuracy, stability, boundary perception ability and model deployment efficiency, especially in complex terrain and small landslide areas, demonstrating its practical application value and promotion potential in landslide remote sensing image segmentation tasks. BRIEF DESCRIPTION OF DRAWINGS

[0071] The drawings described herein are used to provide a further understanding of the present application, form a part of the present application, and the illustrative embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation of the present application. In the drawings:

[0072] Figure 1 Flowchart of a landslide detection method fusing geometric, contextual and frequency domain multi-source features;

[0073] Figure 2 Structure diagram of a landslide detection model based on GCF-Net architecture;

[0074] Figure 3 Structure diagram of the EdgeECI module of the learnable edge guiding and efficient channel spatial attention mechanism;

[0075] Figure 4 Structure diagram of the GCF2 module fusing geometric-context-frequency domain;

[0076] Figure 5 Flowchart of an embodiment of a landslide detection method fusing geometric-context-frequency domain multi-source features. DETAILED DESCRIPTION

[0077] In order to make the technical solutions and advantages of the embodiments of the present application clearer and more apparent, the exemplary embodiments of the present application are further described in detail below with reference to the drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than an exhaustive enumeration of all embodiments. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0078] Reference Figures 1-5 The landslide detection method of the present embodiment fuses geometry, context and frequency domain multi-source features, and specifically includes the following steps:

[0079] S1. Construct a landslide detection model based on the GCF-Net architecture, which includes an encoder and a decoder, and input the image to be detected into the landslide detection model;

[0080] S2. Through the ghost pyramid feature fusion module in the encoder, use multi-scale pyramid pooling and ghost convolution mechanism to capture feature information of different scales in the image to be detected;

[0081] S3. Through the GCF² module in the encoder, further fuse the feature information of different scales in the connection bridge, and input the output feature map of the GCF² module to the decoder, which is used to restore image details and help the decoder to generate high-quality segmentation results;

[0082] S4. Through the EdgeECI module in the decoder, perform depth separable convolution and point convolution operations on the output feature map of the GCF² module step by step, and use the efficient channel-spatial attention mechanism and the edge-guided mechanism to process the operation results to obtain the final output feature map after fusion.

[0083] Specifically, referring to Figure 2 The proposed GCF-Net network architecture adopts an overall U-shaped structure of encoder and decoder, and its core is the fusion of geometry-context-frequency domain features to improve the fine-grained recognition and boundary recovery ability of landslide areas. Overall, the GCF-Net network architecture mainly consists of two parts: the encoder backbone and the decoder backbone, and three types of key modules designed therebetween: the ghost pyramid feature fusion module (PGF Block), the geometry-context-frequency domain fusion module (GCF²) and the edge detail enhancement module (EdgeECI). The core task of the encoder backbone is to extract high-level semantic features from the input image. Through step-by-step down-sampling, the encoder not only reduces the spatial resolution, but also effectively captures the key information in the image while enhancing the semantic expression ability of the feature map, providing high-quality semantic features for subsequent feature fusion and recovery.

[0084] First, in the encoder section, the GCF-Net network architecture extracts multi-scale spatial features stepwise through a four-level downsampling structure. Each level of the encoding unit consists of a Ghost Pyramid Feature Fusion Block (PGF Block) and a max pooling layer (MP2D). The PGF Block is a lightweight ghost feature generation unit. Using a GM (Maximum Generator) layer as the core, parallel feature channels are constructed at different scales to effectively improve feature diversity and representational ability. Subsequently, a max pooling layer (MP2D) downsamples the features, further expanding the receptive field and compressing redundant spatial information, thereby achieving hierarchical feature encoding from local texture to global semantics. This is achieved through a four-layer PGF (Programmable Generation Function). By employing block and pooling operations, the encoder can extract multi-scale representations rich in geometric and contextual information while preserving key terrain features, providing high-quality prior features for subsequent cross-modal fusion and decoding reconstruction. Subsequently, in the bridging phase, the model introduces a geometry-context-frequency domain fusion module (GCF²) to achieve collaborative modeling of cross-domain information. The GCF² module constructs a unified multimodal feature space from three perspectives: geometric feature modeling, context dependency capture, and frequency domain information enhancement. The geometric modeling branch captures the boundary structure of landslide morphology through spatial gradient-based structural encoding; the context modeling branch utilizes inter-channel relationship matrices to achieve global dependency awareness, improving the model's understanding of long-range semantics; and the frequency domain enhancement branch extracts high- and low-frequency feature components through fast Fourier transform and adaptively fuses them with spatial features to suppress texture noise and enhance significant boundary responses. During the fusion phase, the three branches dynamically adjust learnable weights to achieve information complementarity and adaptive feature selection, thus achieving a balance between semantic consistency and structural detail, providing the decoder with more discriminative fused features. (Features); Finally, in the decoder section, the model employs a symmetrical upsampling structure to progressively restore spatial resolution and introduces the EdgeECI module to improve the reconstruction accuracy of boundary regions. Specifically, the feature map after each level of upsampling is skipped with the encoded features of the corresponding layer, fusing shallow spatial details with high-level semantic information. The EdgeECI module, through gradient-guided edge attention mechanism and detail consistency constraints, focuses on enhancing the feature response of landslide edge regions, effectively suppressing blurring transitions and boundary drift problems during the decoding process. Finally, the feature maps after progressive decoding are mapped to pixel-level prediction results through a 1×1 convolutional layer (Conv1×1). This enables precise segmentation of the landslide area from the background;

[0085] Overall, when the input image is After the encoder After processing, the feature map is output. ,Right now ,in, For feature map height, For feature map Image width, For feature map The number of channels;

[0086] decoder The main task is to progressively restore the spatial resolution of the image and generate the final segmentation result. In the decoder, the feature maps are upsampled. Gradually recover, and obtain the feature map after one upsampling operation. After each upsampling, the decoder fuses the features passed from the encoder and the connecting bridge to better recover the detailed information of the target and obtain the final output feature map. ;

[0087] Feature map after one upsampling operation Represented as:

[0088]

[0089] Final output feature map Represented as:

[0090]

[0091] Finally, the final segmentation result is generated through the last convolutional layer of the decoder. ;

[0092] Final segmentation result Represented as:

[0093]

[0094] Among them, the characteristics transmitted by the connecting bridge are: .

[0095] Furthermore, in S2, the ghost pyramid feature fusion module utilizes a ghost convolution mechanism. For the input image to be detected The process is performed to obtain the processed feature map. Multi-scale pyramid pooling processes the feature maps at each scale. Perform average pooling operation Post-fusion features yield pyramid pooling output. Perform convolution operation on it The feature map after the convolution operation is obtained. ;

[0096] Processed feature map Represented as:

[0097]

[0098] wherein, , , , respectively represent the image height, image width and channel number, represents the intermediate channel number in the ghost pyramid feature fusion module;

[0099] Pyramid Pooling output is represented as:

[0100]

[0101] wherein, represents the window size of the average pooling operation, represents the channel number of the convolution operation output in the ghost pyramid feature fusion module;

[0102] Feature map after convolution operation is represented as:

[0103] .

[0104] Specifically, in the landslide image segmentation task, the key to improving the segmentation accuracy lies in efficiently extracting and fusing features from different scales, especially when dealing with complex landslide images, the extraction of multi-scale features is crucial; however, traditional methods are often limited in application due to high computational cost, in order to break through this bottleneck, the present application proposes a ghost pyramid feature fusion module, which combines ghost convolution and pyramid pooling technology, effectively balancing the computational complexity while ensuring the feature expression ability.

[0105] The PGF Block mainly acts on the encoder part, responsible for multi-scale feature extraction and fusion, by obtaining context information in different scales in parallel, the PGF Block enhances the network's learning ability for multi-scale features, while reducing the computational overhead through an efficient fusion mechanism, referring to Figure 2 , the left side shows the structure of the PGF Block, which combines two innovative technologies, ghost convolution (GhostModule) and pyramid pooling (LightPPM), ghost convolution significantly reduces computational complexity by generating multiple low-dimensional feature maps, while pyramid pooling extracts context information at multiple scales, thereby improving feature expression ability.

[0106] In the preliminary spatial feature extraction stage of the encoder, the PGF Block can effectively improve the diversity and accuracy of feature representation while ensuring computational efficiency, this design not only strengthens the learning of multi-scale features, but also makes the entire network more efficient in the landslide image segmentation task.

[0107] Further, in the S3, the GCF2 module includes a branch A and a branch B, the branch A processes the feature map after the convolution operation through a frequency domain enhancement and an autonomous learning position encoding , to obtain a branch A output ;

[0108] The frequency domain enhancement process is represented as:

[0109]

[0110] wherein, represents the feature map of the frequency domain enhancement output, i.e., the frequency domain feature, is a frequency domain high-frequency mask, is a 1x1 convolution operation;

[0111] The autonomous learning position encoding process is represented as:

[0112]

[0113] wherein, represents the position encoding, is an encoding result of the frequency domain feature;

[0114] The branch A output is represented as:

[0115]

[0116] wherein, represents group normalization;

[0117] The branch B extracts the context information of the feature map after the convolution operation through a multi-scale convolution operation, to obtain feature maps , , under different convolution kernels, fuse the feature maps through a fusion convolution layer and a concatenation operation, and obtain a fusion feature map ;

[0118] The feature maps , , under different convolution kernels are represented as:

[0119]

[0120] wherein, represents a 3x3 convolution operation, represents a 5x5 convolution operation;

[0121] The fusion feature map is represented as:

[0122]

[0123] According to the feature map after the convolution operation foreground coding , the feature map after the convolution operation and the fusion feature map contrast enhancement , get the enhanced feature map ;

[0124]

[0125] In order to efficiently fuse the features of branch A and branch B, the GCF2 module introduces a gating mechanism, which fuses the enhanced feature map and branch A output through splicing operation, and then weights the fusion through gating mechanism, to get the gated fusion feature map ;

[0126] The gated fusion feature map is represented as:

[0127]

[0128] wherein, indicates the gating weight of branch A, indicates the gating weight of branch B;

[0129] In order to further enhance the representation ability of the feature, the GCF2 module introduces a channel attention mechanism on the gated fusion feature map , to get the GCF2 module output feature map ;

[0130] The GCF2 module output feature map is represented as:

[0131] .

[0132] Specifically, high-frequency detail information has a decisive impact on the segmentation accuracy of remote sensing images, especially in the processing of complex terrain. How to extract high-frequency features such as edges and textures in the image is the key to improving the segmentation effect. Traditional feature extraction methods often tend to acquire low-frequency information and ignore the expression of high-frequency features, resulting in loss of details and missed detection of small targets. In order to solve the above problems, this invention proposes the GCF² module, which introduces techniques such as frequency domain enhancement, position and foreground coding and contrast enhancement to highlight the acquisition of high-frequency information, suppress background interference, and improve the network's ability to understand spatial structure.

[0133] refer to Figure 3 The core idea of ​​the GCF² module is to highlight high-frequency information in images through frequency domain enhancement methods, thereby improving the network's sensitivity to detailed regions (such as edges and textures). Simultaneously, it utilizes learnable positional encoding to provide spatial location information for feature maps, enhancing the network's ability to model the spatial structure of images. The introduction of spatial location information not only helps the network capture spatial relationships more accurately but also makes the segmentation results more precise. Finally, through foreground encoding, the semantic representation of foreground targets is adaptively strengthened, achieving decoupling between foreground and background features. This process effectively suppresses background noise, deepens the fusion of foreground and contextual features, and improves target perception capabilities in complex scenes. By combining frequency domain, geometric information, and contextual features, the GCF² module can capture richer detailed features, further improving the model's performance.

[0134] The gating mechanism designed in step S3 adaptively learns the fusion weights of the features from the two branches, enabling the model to intelligently determine the contribution of each branch to the final output; channel attention mechanism. The channel attention mechanism can dynamically adjust the weights of each channel based on its importance, allowing the network to focus more on key features and improve the model's performance. This allows the GCF² module to significantly improve the model's ability to capture high-frequency details and enhance its understanding of the image's spatial structure, ultimately improving segmentation accuracy.

[0135] Furthermore, in S4, the EdgeECI module outputs a feature map to the GCF² module. Perform depthwise separable convolution The feature map obtained after depthwise separable convolution is obtained. ;

[0136] Feature map after depthwise separable convolution Represented as:

[0137]

[0138] By analyzing the feature map after depthwise separable convolution... Point convolution operation is performed , to obtain a feature map after point convolution operation ;

[0139] Feature map after point convolution operation is expressed as:

[0140]

[0141] The feature map after point convolution operation An efficient channel and spatial attention mechanism containing channel attention CA and spatial attention SA is adopted , to obtain a feature map after processing by the efficient channel and spatial attention mechanism ;

[0142] Feature map after processing by the efficient channel and spatial attention mechanism is expressed as:

[0143]

[0144] The feature map after processing by the efficient channel and spatial attention mechanism An edge guiding mechanism is adopted Weighting is performed to obtain a weighted feature map after the edge guiding mechanism , and a fused final output feature map is obtained , i.e., a landslide detection result;

[0145] Weighted feature map after the edge guiding mechanism is expressed as:

[0146]

[0147] Fused final output feature map is expressed as:

[0148] .

[0149] Specifically, in the segmentation of complex terrain images such as landslides, the recovery of image details often involves accurate capture of edges and small targets. Traditional convolutional neural networks (CNNs) often have limitations in processing such details, especially in the process of recovering small targets and details. In order to overcome this defect, the present application proposes an EdgeECI module, which enhances the network's ability to recover detailed features by combining a learnable edge guiding mechanism with an efficient channel and spatial attention mechanism.

[0150] The EdgeECI module is the core unit of the decoder part, which aims to accurately recover the detailed areas in the image, especially the edges and small targets, by working together. The efficient channel and spatial attention mechanism The feature map is weighted through the channel and space attention mechanism, thereby optimizing feature fusion under multi-scale and enhancing the expression ability of global information; subsequently, the edge guiding mechanism generates a spatial weighting map based on edge information to highlight the detail area, ensuring that the network pays more attention to the micro features during the restoration process. Through this synergistic mechanism, the EdgeECI module not only enhances the detail restoration capability of the decoder, but also effectively fuses information with the encoder through the connection bridge, thereby improving the accuracy of the entire network.

[0151] Although the present application has been described in terms of limited embodiments, those skilled in the art, given the benefit of the above description, will appreciate that other embodiments are contemplated within the scope of the application described herein. Furthermore, it should be noted that the language used in the specification has primarily been chosen for readability and instructional purposes and can not have been chosen to delineate or circumscribe the subject application. Accordingly, numerous modifications and adaptations thereof are enabled as are apparent to those skilled in the art without departing from the scope and spirit of the appended claims. In terms of the scope of the present application, the disclosure made of the present application is illustrative rather than restrictive, and the scope of the present application is defined by the appended claims.

Claims

1. A landslide detection method integrating geometric, contextual, and frequency domain multi-source features, characterized in that, Includes the following steps: S1. Construct a landslide detection model based on the GCF-Net architecture, which includes an encoder and a decoder, and input the image to be detected to the landslide detection model; S2. By using the ghost pyramid feature fusion module in the encoder, multi-scale pyramid pooling and ghost convolution mechanisms are employed to capture feature information at different scales in the image to be detected. S3. Through the GCF² module in the encoder, feature information of different scales is further fused in the connecting bridge, and the resulting GCF² module output feature map is input into the decoder; S4. The EdgeECI module in the decoder performs depthwise separable convolution and point convolution operations on the output feature map of the GCF² module step by step. The results of the operation are processed by an efficient channel spatial attention mechanism and an edge-guided mechanism to obtain the fused final output feature map.

2. The landslide detection method integrating geometric, contextual, and frequency domain multi-source features according to claim 1, characterized in that, In S2, the ghost convolution mechanism is used. For the input image to be detected The process is performed to obtain the processed feature map. Multi-scale pyramid pooling processes the feature maps at each scale. Perform average pooling operation Post-fusion features yield pyramid pooling output. Perform convolution operation on it The feature map after the convolution operation is obtained. ; Processed feature map Represented as: in, , , , These represent the image height, image width, and number of channels, respectively. This indicates the number of intermediate channels in the Ghost Pyramid feature fusion module; Pyramid pooling output Represented as: in, This indicates the window size for the average pooling operation. This represents the number of channels output by the convolution operation in the Ghost Pyramid feature fusion module; Feature map after convolution operation Represented as: 。 3. The landslide detection method integrating geometric, contextual, and frequency domain multi-source features according to claim 2, characterized in that, In S3, the GCF² module includes branch A and branch B. Branch A processes the feature map after the convolution operation through frequency domain enhancement and self-learned position encoding. The output of branch A is obtained. ; The frequency domain enhancement process is represented as: in, The feature map representing the frequency domain enhanced output is called the frequency domain feature. For high-frequency masks in the frequency domain, This is a 1×1 convolution operation; The self-learning position encoding process is represented as: in, Indicates position code, This represents the encoding result of the frequency domain features; Branch A output Represented as: in, Indicates group normalization; Branch B extracts the feature map after the convolution operation through multi-scale convolution operations. Contextual information is used to obtain feature maps under different convolutional kernels. , , fused convolutional layers and connection The operation yields a fused feature map. ; Feature maps under different convolution kernels , , Represented as: in, This represents a 3×3 convolution operation. This represents a 5×5 convolution operation; Fusion Feature Map Represented as: Based on the feature map after convolution operation Foreground encoding The feature map after convolution operation and fusion feature map Enhance contrast The enhanced feature map is obtained. ; Enhanced feature map and branch A output The feature maps are fused by stitching and then weighted by a gating mechanism to obtain the gated fused feature maps. ; Feature map after gating fusion Represented as: in, This represents the gating weight of branch A. This represents the gating weight of branch B; Feature map of GCF² module after gated fusion Introducing channel attention mechanism The output feature map of the GCF² module is obtained. ; GCF² module output feature map Represented as: 。 4. The landslide detection method integrating geometric, contextual, and frequency domain multi-source features according to claim 3, characterized in that, In S4, the EdgeECI module outputs a feature map to the GCF² module. Perform depthwise separable convolution The feature map obtained after depthwise separable convolution is obtained. ; Feature map after depthwise separable convolution Represented as: By analyzing the feature map after depthwise separable convolution... Perform point convolution operation The feature map obtained after point convolution operation is obtained. ; Feature map after point convolution operation Represented as: Feature maps after point convolution operation Employing an efficient channel spatial attention mechanism The feature map is obtained after processing by the efficient channel spatial attention module. ; Feature map processed by efficient channel spatial attention mechanism Represented as: Feature maps processed by an efficient channel spatial attention mechanism Employing edge-guided mechanism Weighting is performed to obtain the weighted feature map obtained through the edge-guided mechanism. The final output feature map is obtained by integration. That is, the landslide detection results; Weighted feature map guided by edge mechanism Represented as: The final output feature map after fusion Represented as: 。

Images