Landslide segmentation model based on multi-loss function fusion

Abstract

The invention discloses a landslide segmentation model based on multi-loss function fusion, particularly relates to the field of landslide monitoring, is used for solving the problems of precise segmentation of a landslide area and insufficient identification precision in a complex scene, realizes precise segmentation of a landslide image through multi-level feature extraction and adaptive optimization, and has remarkable beneficial effects. The detail identification of the landslide area is enhanced by utilizing self-adaptive channel response and multi-scale feature fusion, so that features of different scales are fully expressed; weighted combination and residual connection of global and local features ensure continuous transmission and integration of high-level and low-level features, and detail and global information are completely reserved. The spatial resolution is recovered through up-sampling and channel splicing, and it is ensured that a segmentation result is consistent with an original image; the multi-level nonlinear loss function improves the robustness of the model under the condition that the samples are unbalanced, samples difficult to classify are effectively concerned, and the model shows a high-precision segmentation effect in a complex scene.

Description

Technical Field

[0001] This invention relates to the field of landslide monitoring, and more specifically, to a landslide segmentation model based on multi-loss function fusion. Background Technology

[0002] With the increasing frequency of geological disasters, landslides have become a serious threat to infrastructure and the safety of residents in mountainous areas. Landslide monitoring and early warning have gradually become key tasks in disaster prevention and control. To achieve effective identification and dynamic monitoring of landslide areas, automated segmentation of landslide areas using remote sensing imagery has become an important technical means. Remote sensing imagery has advantages in acquiring large-scale, real-time data, providing strong support for landslide early warning, disaster assessment, and emergency response. Against this backdrop, deep learning-based image segmentation models have demonstrated high accuracy and efficiency in landslide detection, making landslide segmentation a hot topic in the field of geological disaster research and application.

[0003] Current landslide image segmentation methods face numerous technical bottlenecks in practical applications, primarily manifested in insufficient segmentation accuracy, limited feature representation, and poor adaptability to complex scenes. Firstly, because the feature differences between the landslide area and its surrounding background in images are subtle, models often make identification errors when extracting landslide edges and complex terrain features. For example, surface cover such as vegetation and rocks have highly similar spectral characteristics to the landslide area, and traditional segmentation models often confuse these subtle differences, affecting the clarity of segmentation boundaries and overall accuracy. Furthermore, the shape and size of landslide areas vary significantly, ranging from small-scale localized landslides to large-scale mountain landslides, resulting in a lack of stable feature extraction capabilities when dealing with landslide areas of different sizes, leading to insufficient reliability of segmentation results.

[0004] Secondly, due to the extreme imbalance in landslide sample data—landslide areas typically occupy a very small portion of the imagery—the model struggles to acquire sufficient learning signals for small-scale landslides. Traditional loss functions, such as cross-entropy loss, tend to significantly increase the model's focus on background areas when dealing with such severely imbalanced data, while neglecting to learn about landslide areas. This results in low accuracy in landslide area identification and frequent missed detections. Particularly, the model fails to effectively capture the initial, small cracks or micro-slip areas of a landslide, limiting its application value in landslide early warning and monitoring.

[0005] Finally, existing landslide segmentation models suffer from significant shortcomings in robustness to complex scenarios. The formation and development of landslide areas are typically influenced by multiple factors, such as topographic relief, geological structure, and meteorological conditions, resulting in complex and variable morphological representations of landslide areas in images. Standard segmentation models are poorly adapted to these complex backgrounds. Furthermore, difficult-to-classify samples (such as landslide edges and boundaries) lack targeted processing under existing loss function systems, easily leading to blurred boundaries or incomplete segmentation. Especially in some extreme scenarios (such as strong light, shadows, occlusion, etc.), detailed information about landslide areas is difficult to preserve. These problems collectively affect the refined representation and global consistency of landslide image segmentation, making it difficult for segmentation models to meet the high-precision requirements of actual monitoring and early warning.

[0006] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0007] To overcome the aforementioned shortcomings of existing technologies, embodiments of this invention provide a landslide segmentation model based on multi-loss function fusion. Through multi-level feature extraction and fusion, and adaptive optimization of complex loss functions, high-precision segmentation of landslide images is achieved, demonstrating significant beneficial effects. First, adaptive channel response and multi-scale feature fusion effectively enhance the model's ability to identify details in landslide areas, ensuring that landslide features at different scales are fully represented. Second, based on weighted combination and residual connections of global and local features, continuous transmission and integration of high and low-level features are constructed, enabling the complete preservation of edge details and global structure in landslide images, improving the model's spatial consistency and detail accuracy in landslide areas. Furthermore, upsampling and channel stitching restore the spatial resolution of the feature map, ensuring accurate matching between the landslide segmentation map and the original image, laying a solid foundation for segmentation accuracy. Finally, employing multi-level nonlinear loss functions, through adaptive fusion of cross-entropy, Dice, and Focal losses, improves the model's robustness under imbalanced sample conditions, strengthens its focus on difficult-to-classify samples, and enables the model to maintain high-precision segmentation performance even in complex terrain and extreme scenarios. This invention effectively improves the reliability and application value of landslide segmentation models, and is applicable to scenarios such as landslide monitoring, disaster early warning and emergency response, in order to solve the problems mentioned in the background technology.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] S1: Acquire the landslide image to be segmented and perform initial convolution processing to extract channel response feature maps.

[0010] S2 performs dilated convolution on the channel response feature maps at different dilation rates to generate multi-scale feature fusion images.

[0011] S3 generates a weighted feature map based on global and local features from a multi-scale feature fusion image.

[0012] S4 combines low-level and high-level features of the weighted feature map through residual connections and skip connections to obtain a combined feature map with continuous information transmission.

[0013] S5, after upsampling the combined feature map, splices high and low level features by channel to generate a segmentation feature map restored to the original spatial dimension.

[0014] S6 generates an optimized loss function for segmentation feature maps based on a weighted combination of cross-entropy loss, Dice loss, and Focal loss.

[0015] In a preferred embodiment, step S1 includes the following:

[0016] S1.1, for the input landslide image Initial convolution operations are performed using a 3×3 convolution kernel K to generate preliminary feature maps. Where C ' The output channel number is denoted by H, W, and C, where H, W, and C represent the image height, width, and number of channels, respectively.

[0017] S1.2, To enhance the characteristic response of the landslide area, a dynamic channel weight adjustment mechanism is adopted; specifically, a weight function W is constructed. c Feature enhancement for each channel is based on dynamically allocating response intensity according to image content. The calculation steps are as follows:

[0018] Calculate the global average pooling of the input image to generate the channel-level average feature matrix.

[0019] A weight function matrix is ​​generated based on the channel-level average feature matrix, and a nonlinear transformation is applied to obtain the weighted response. A custom activation function g(x) is used to satisfy g(x) = α1x + β1x. 2 Where α1 and β1 are coefficients, dynamically adjusted through parameter tuning; weighting function W c : The obtained weight function is applied to each channel of the initial feature map to enhance the features along the channel dimension. The expression is: F' i,j,k =F i,j,k ·W c The adjusted weights are applied to the feature maps of each channel, making the features of the landslide area more prominent.

[0020] S1.3, based on the feature map, the feature map is normalized by batch normalization to eliminate bias and accelerate convergence; by calculating the mean and standard deviation of the current batch, the feature map is adjusted one by one by channel so that the mean of the feature map in different channels is close to zero and the standard deviation is close to 1.

[0021] S1.4, ReLU activation is applied to the normalized feature map to introduce nonlinearity and generate the final output feature map.

[0022] In a preferred embodiment, step S2 includes the following:

[0023] S2.1, For the output feature map generated in step S1, set different dilation rates and perform dilated convolution processing on the landslide image features with different receptive fields.

[0024] S2.2, based on the output of each dilation rate convolutional layer, expands and superimposes the spatial dimension of the feature map to form feature maps at different scales.

[0025] S2.3, Multi-scale feature maps are stitched together along the channel dimension to overlay features from each scale onto the same fused feature map, thus obtaining a multi-scale feature fusion image.

[0026] In a preferred embodiment, step S3 includes the following:

[0027] S3.1 For the multi-scale feature fusion image generated in step S2, first use 1×1 convolution to compress it to obtain the global feature map of the image.

[0028] S3.2, the global feature map is subjected to global average pooling to generate a single global feature matrix, where each element of the matrix represents the global response intensity of the corresponding channel.

[0029] S3.3 directly uses 1×1 convolution on the multi-scale fused feature map to obtain the local feature map.

[0030] S3.4, add the local feature map and the global feature matrix to generate an intermediate feature map containing both global and local information.

[0031] S3.5, perform another 1×1 convolution operation on the intermediate feature map to generate a weighted feature map.

[0032] S3.6 Multiply the weighted feature map and the local feature map point by point to generate the final weighted feature map, thereby enhancing the significance of the landslide area features.

[0033] In a preferred embodiment, step S4 includes the following:

[0034] S4.1 In the process of encoding the weighted feature map, in order to ensure the continuous transmission of features, a 1×1 convolution structure is added after each 3×3 convolutional layer; after the 3×3 convolution extracts spatial features, the number of channels is adjusted by the 1×1 convolution to keep the feature map compact in the channel dimension; then, the residual connection mechanism is applied to add the input feature map and the output feature map after convolution layer by layer to maintain the continuity of information flow.

[0035] S4.2 In the decoding stage, in order to preserve the detailed information of the image, the low-level features in the coding layer that have undergone residual connection processing are directly connected to the high-level features in the decoding layer. In this skip connection mechanism, the low-level features of the coding layer are directly passed to the decoding layer, ensuring that the decoding layer can obtain detailed information from the low-level features.

[0036] S4.3, after point-by-point combination of low-level and high-level features, the generated combined feature map F combined It has the ability to express complete information about the landslide area.

[0037] In a preferred embodiment, step S5 includes the following:

[0038] S5.1, In the decoding stage, bilinear interpolation is used to analyze the combined feature map F. combined A 2x upsampling process is performed to gradually restore the spatial resolution of the feature map; bilinear interpolation ensures the continuity of the upsampled feature map by linearly weighting adjacent pixels, resulting in the upsampled feature map F. upsampled .

[0039] S5.2, upsample the feature map F upsampled The second feature map F at the corresponding spatial location of the encoding stage encoded Channel splicing is performed to form a combined feature map F containing high-level and low-level information. concat .

[0040] S5.3, to avoid redundancy in the number of channels in the concatenated feature map, a 1×1 convolution operation is used to concatenate the combined feature map F. concat Channel compression is performed; the compressed feature map after convolution is represented as F. compressed .

[0041] S5.4, After channel compression, to ensure that spatial details are represented in the feature map, the compressed feature map F is... compressed Residual structures are applied for refinement; residual refinement is performed by adding the output of convolution compression to the original input features through pointwise addition operations, so as to recover detailed information in the segmentation feature map; the final segmentation feature map is restored to the original spatial dimension.

[0042] In a preferred embodiment, step S6 includes the following:

[0043] Cross-entropy loss is used to calculate the classification probability of landslide regions and non-landslide regions in the segmentation feature map; let the true class label be y∈{0,1}, and the predicted probability be p∈[0,1]. The expression for calculating cross-entropy is as follows: L CE =-∑ylog(p)+(1-y)log(1-p).

[0044] The overlap between the segmentation mask and the ground truth mask is calculated using the Dice loss function. The Dice loss is defined as follows: The Dice loss is calculated by determining the intersection ratio between the segmented regions and the ground truth labels.

[0045] The response of the Focal loss is multiplied by the dynamic adjustment coefficient of the Dice loss, forming an enhanced Focal loss term. The multiplication result of the Focal loss is defined as follows:

[0046] Using cross-entropy loss as the initial term of the excitation, and combining it with the adjustment factor of Dice loss and the enhanced Focal loss, the total loss function L is constructed. total The total loss function is expressed using a hierarchically weighted nonlinear combination, as shown in the following formula:

[0047] The technical effects and advantages of the landslide segmentation model based on multi-loss function fusion in this invention are as follows:

[0048] This invention achieves high-precision segmentation of landslide images through multi-level feature extraction and fusion, and adaptive optimization of complex loss functions, yielding significant benefits. First, adaptive channel response and multi-scale feature fusion effectively enhance the model's ability to identify details in landslide areas, ensuring comprehensive representation of landslide features at different scales. Second, based on weighted combinations of global and local features and residual connections, continuous transmission and integration of high- and low-level features are constructed, enabling the complete preservation of edge details and global structure in landslide images, improving the model's spatial consistency and detail accuracy in landslide areas. Furthermore, upsampling and channel stitching restore the spatial resolution of the feature map, ensuring accurate matching between the landslide segmentation map and the original image, laying a solid foundation for segmentation accuracy. Finally, employing multi-level nonlinear loss functions, through adaptive fusion of cross-entropy, Dice, and Focal losses, improves the model's robustness under imbalanced sample conditions, strengthens its focus on difficult-to-classify samples, and enables the model to maintain high-precision segmentation even in complex terrain and extreme scenarios. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the landslide segmentation model based on multi-loss function fusion of the present invention. Detailed Implementation

[0050] 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.

[0051] Example 1: Figure 1 This invention presents a landslide segmentation model based on multi-loss function fusion, comprising:

[0052] S1: Acquire the landslide image to be segmented and perform initial convolution processing to extract channel response feature maps.

[0053] S2 performs dilated convolution on the channel response feature maps at different dilation rates to generate multi-scale feature fusion images.

[0054] S3 generates a weighted feature map based on multi-scale feature fusion images, using global and local features.

[0055] S4 combines low-level and high-level features of the weighted feature map through residual connections and skip connections to obtain a combined feature map with continuous information transmission.

[0056] S5, after upsampling the combined feature map, splices high and low level features by channel to generate a segmentation feature map restored to the original spatial dimension.

[0057] S6 generates an optimized loss function for segmentation feature maps based on a weighted combination of cross-entropy loss, Dice loss, and Focal loss.

[0058] In landslide image segmentation, traditional methods struggle to accurately extract key features of landslide areas in both spatial and channel dimensions due to the complexity and diversity of landslide region features and the intricate background. Therefore, this paper proposes a feature enhancement method for landslide images using initial convolution and adaptive channel weight adjustment. This enhances the response accuracy to landslide areas in subsequent processing steps and ensures accurate model focusing on the landslide region within the feature space.

[0059] Step S1 includes the following:

[0060] S1.1, for the input landslide image (Where H, W, and C represent the image height, width, and number of channels, respectively) Initial convolution operations are performed using a 3×3 convolution kernel K to generate preliminary feature maps. Where C 'This represents the number of output channels. The expression for convolution is as follows: F i,j,k =∑ m,n I i+m,j+n,c ·K m,n +b k ;where F i,j,k b represents the output feature value at position (i,j) on channel k; k is the convolution bias; m, n are the convolution kernel sizes.

[0061] The role of the convolution kernel is to obtain the local spatial features of the landslide image. Through layer-by-layer spatial weighted calculation, a multi-channel preliminary feature map is output for subsequent feature refinement.

[0062] S1.2, To enhance the characteristic response of the landslide area, a dynamic channel weight adjustment mechanism is adopted. Specifically, a weight function W is constructed. c Feature enhancement for each channel is essentially based on dynamically allocating response intensity according to the image content. The calculation steps are as follows:

[0063] Calculate the global average pooling GAP of the input image to generate the channel-level average feature matrix.

[0064] A weight function matrix is ​​generated based on the channel-level average feature matrix, and a nonlinear transformation is applied to obtain the weighted adjusted response. A custom activation function g(x) is used to satisfy g(x) = α1x + β1x. 2 , where α1 and β1 are coefficients, dynamically adjusted through parameter tuning. Weighting function W c :

[0065] The resulting weight function can be applied to each channel of the initial feature map, enhancing the features along the channel dimension. The expression is: F' i,j,k =F i,j,k ·W c .

[0066] The adjusted weights are further applied to the feature maps of each channel, making the features of the landslide area more prominent, thereby highlighting the key areas in the image.

[0067] Step S1.3, based on the feature maps, performs batch normalization to normalize the feature maps, eliminating bias and accelerating convergence. By calculating the mean and standard deviation of the current batch, the feature maps are adjusted channel by channel, making the mean of the feature maps close to zero and the standard deviation close to 1 in different channels. The purpose of this step is to eliminate bias in the feature distribution, improving the stability and convergence speed of the training process.

[0068] S1.4, ReLU activation is applied to the normalized feature map to introduce non-linearity and generate the final output feature map. The purpose of ReLU activation is to set negative features to zero while keeping positive features unchanged, thus ensuring the model's outstanding performance on positive features.

[0069] Step S1 constructs a highly recognizable landslide area feature map through refined initial convolution and dynamic adjustment of channel weights. Batch normalization and nonlinear activation are then used to standardize and saliency the features, enhancing the representation of the landslide area in the feature space. This ultimately provides a precise feature foundation for subsequent multi-scale information fusion and accurate segmentation of landslide areas in complex backgrounds.

[0070] In landslide image segmentation, to accurately extract multi-scale features of the landslide area and overcome the problem of insufficient capture of complex terrain information by single-scale convolution, dilated convolution operations are performed on the feature maps at different dilation rates to achieve effective fusion of multi-scale features. This step ensures that the model can capture detailed features and global background information in the image while maintaining spatial resolution, laying the foundation for accurate segmentation.

[0071] Step S2 includes the following:

[0072] S2.1, For the output feature map generated in step S1, different dilation rates (3, 6, 9) are set to perform dilated convolution processing on the landslide image features with different receptive fields. Dilated convolution expands the receptive field of the feature map by increasing the number of holes inside the convolution kernel during calculation, while maintaining the spatial resolution of the landslide image features. This operation can effectively capture multi-scale feature details of the landslide area, forming feature maps at different scales.

[0073] In step S2.2, based on the output of each dilation rate convolutional layer, the spatial dimension of the feature maps is expanded and superimposed to form feature maps at different scales. These feature maps are named F3, F6, and F9, corresponding to the convolutional outputs with dilation rates of 3, 6, and 9, respectively. Feature maps at different scales have independent receptive fields, containing rich information from local to global perspectives.

[0074] S2.3, Multi-scale feature maps are concatenated along the channel dimension, stacking features from each scale onto a single fused feature map F. fuse In this process, a multi-scale feature fusion image containing rich multi-scale information is obtained. The stitching operation can be expressed as: F fuse = [F3, F6, F9]; This stitching operation integrates features at different scales into a multi-channel feature map, so that the fused feature map not only has a large receptive field, but can also present the complexity of the landslide area at different scales.

[0075] Step S2 achieves the fusion of multi-scale feature maps through dilated convolution operations with multiple dilation rates, and generates a fused feature map containing rich spatial context by combining splicing and dynamically weighted adaptive processing, providing key multi-scale information support for fine-grained segmentation of landslide areas.

[0076] In landslide image segmentation, to further highlight the feature details of landslide areas, a weighted feature map is generated based on multi-scale feature fusion images by combining global and local features. This enables the model to achieve higher saliency within the landslide area. This step aims to enhance the depth of multi-scale information representation, thereby improving the model's ability to distinguish landslide areas in complex backgrounds.

[0077] Step S3 includes the following:

[0078] S3.1, For the multi-scale feature fusion image generated in step S2, first use a 1×1 convolution to compress it to obtain the global feature map F of the image. global The purpose of 1×1 convolution is to linearly combine only the channel dimensions without changing the spatial resolution, and to extract the overall features of the entire image into compact global information.

[0079] S3.2, the global feature map is subjected to global average pooling to generate a single global feature matrix G, where each element g of the matrix... k This represents the global response intensity for the corresponding channel k. The global feature matrix is ​​represented as: Where H ' and W ' The height and width of the feature map are used to compress spatial information through global pooling to obtain the channel-level global response value.

[0080] S3.3, directly apply a 1×1 convolution to the multi-scale fused feature map to obtain the local feature map F. local This convolution operation only adjusts the channel dimension, allowing each location in the local feature map to retain its fine-grained spatial information, thus enhancing feature saliency in subsequent steps.

[0081] S3.4, add the local feature map and the global feature matrix to generate an intermediate feature map F containing both global and local information. mid To ensure effective information transmission and adaptive fusion between different channels, a nonlinear combination function can be used for correction: F mid =F local +Φ(G); whereΦ(x)=α2x+β2x 3This is to enhance the nonlinear response of the feature map while preserving global information, making it more focused on the feature representation of complex regions. α2 controls the linear portion of the feature map, directly retaining the original information of the feature map at a basic ratio, ensuring the preservation of global features in the overall feature map. Its function is to directly amplify or reduce the input feature map, thereby maintaining a basic response within the global information of the feature map. β2 controls the nonlinear enhancement portion of the feature map, used to weight and enhance the nonlinear feature representation of landslide regions in the feature map. Through x 3 The nonlinear enhancement makes complex patterns in the feature map more prominent, thereby highlighting key regions in the fusion of local and global features.

[0082] S3.5, perform another 1×1 convolution operation on the intermediate feature map to generate the weighted feature map F. weight This 1×1 convolution is used for subsequent point-by-point weighting. This further compresses spatially redundant information, retaining only the weight information that is most effective in representing the characteristics of the landslide area.

[0083] S3.6, multiply the weighted feature map and the local feature map point by point to generate the final weighted feature map F. final This enhances the salience of landslide area features. The final feature map is calculated using the formula: F final =F weight ⊙F local ; where ⊙ represents point-by-point multiplication, which makes the response of each channel in the weighted feature map effectively amplify only the landslide area, ensuring the saliency of the image feature map in the landslide area.

[0084] Step S3 generates a more distinctive weighted feature map by weighted combination of global and local features, which enhances the expression of key features of the landslide area and provides in-depth feature support for accurate identification of landslide areas in images.

[0085] In landslide image segmentation using deep networks, the layer-by-layer transmission of information can easily lead to feature loss, especially detailed information about the landslide area. To address this issue, a combination of residual connections and skip connections is used to effectively integrate low-level and high-level features across different network layers. This ensures that both detailed and global information in the feature maps are fully preserved, supporting refined landslide area identification.

[0086] Step S4 includes the following:

[0087] In S4.1, during the encoding of the weighted feature map, to ensure continuous feature propagation, a 1×1 convolutional structure is added after each 3×3 convolutional layer. After the 3×3 convolution extracts spatial features, the number of channels is adjusted through 1×1 convolutions to keep the feature map compact in the channel dimension. Then, a residual connection mechanism is applied to add the input feature map to the convolutionally processed output feature map layer by layer to maintain the continuity of the information flow. The residual connection formula is as follows: F residual =F in +Ψ(F 3×3 (F in )); where F residual It is an output of the encoding stage, containing the input feature map F. in The result of adding the feature maps processed by 3×3 convolution and 1×1 convolution; F in For the input feature map, F 3×3 Ψ represents a 3×3 convolution operation, and Ψ represents a 1×1 convolution operation used for channel matching followed by addition. Residual connections not only effectively transfer feature information but also reduce the vanishing gradient phenomenon in deep networks.

[0088] S4.2, In the decoding stage, to preserve image detail, low-level features processed by residual connections in the coding layer are directly connected to high-level features in the decoding layer. In this skip connection mechanism, low-level features from the coding layer are directly passed to the decoding layer, ensuring that the decoding layer can obtain rich detail information from these low-level features. The specific calculation is as follows: F combined =F residual ⊙F high Among them, F high For the high-level feature map of the decoding layer, F residual This is the low-level feature map after the residual connection of the coding layer. ⊙ represents the pointwise multiplication operation to achieve the fusion of features at different levels and ensure a balanced representation of details and global information in the segmentation map.

[0089] S4.3, after point-by-point combination of low-level and high-level features, the generated combined feature map F combined It possesses the ability to represent complete information about landslide areas. To further enhance the adaptability of the combined feature map, a custom nonlinear function Ω(x) = γx + δx is used. 2 The combined feature map is adaptively adjusted to enhance the feature's detail representation. The non-linear combination formula is as follows: Here, γ and δ are adjustment coefficients, which increase the expressive flexibility of the combined feature map through nonlinear processing, so that it maintains significant feature response to landslide areas of different complexities.

[0090] Step S4 maintains the continuity of features in the network through residual connections, and in the decoding stage, it uses skip connections to adaptively combine low-level and high-level features to generate a combined feature map that balances detailed and global information, providing key feature representation support for accurate segmentation of landslide areas.

[0091] In the segmentation processing of landslide images, to ensure that the segmentation feature map is restored to its original spatial dimension and possesses multi-level information representation, upsampling and channel stitching are used to combine high-level features from the decoding stage with low-level features from the encoding stage, thereby forming a detailed and complete feature representation of the landslide area. This step ensures the consistency of the feature map in terms of spatial resolution and hierarchical detail, providing crucial support for segmentation accuracy.

[0092] Step S5 includes the following:

[0093] S5.1, In the decoding stage, bilinear interpolation is used to analyze the combined feature map F. combined A 2x upsampling process is performed to gradually restore the spatial resolution of the feature map. Bilinear interpolation, through linear weighting of adjacent pixels, ensures the continuity of the upsampled feature map, resulting in the upsampled feature map F. upsampled .

[0094] S5.2, upsample the feature map F upsampled The second feature map F at the corresponding spatial location of the encoding stage encoded Channel splicing is performed to form a combined feature map F containing high-level and low-level information. concat The concatenation operation, by superimposing feature maps along the channel dimension, imbues the concatenated feature map with more hierarchical information, thereby improving segmentation accuracy. Channel concatenation is expressed as follows: F concat =F upsampled ,F encoded This operation combines high- and low-level features along the channel dimension, laying the foundation for subsequent feature refinement.

[0095] Low-level feature map F encoded This refers to feature maps generated during the encoding phase (i.e., the downsampling part of the network) that have not undergone multiple convolutions, typically originating from shallow layers of the network. Because these feature maps are located in the initial layers of the network's encoding, they contain relatively raw spatial details but lack deep semantic understanding. Therefore, F... encoded This refers to the feature map generated at the end of each coding block. It is usually concatenated with the upsampled high-level feature map through skip connections during the decoding stage to maintain the network's sensitivity to image details and enhance its ability to restore details of landslide areas.

[0096] S5.3, to avoid redundancy in the number of channels in the concatenated feature map, a 1×1 convolution operation is used to concatenate the combined feature map F. concat Channel compression is performed. The 1×1 convolution linearly combines the channels, reducing the number of channels while retaining effective information, thus keeping the feature map computationally compact. The compressed feature map after convolution is represented as F. compressed .

[0097] S5.4, After channel compression, to ensure that spatial details are represented in the feature map, the compressed feature map F is... compressed Residual structures are applied for refinement. Residual refinement uses pointwise addition to add the output of the convolutional compression to the original input features, thereby recovering detailed information from the segmentation feature map. The refined segmentation feature map F... refined Expressed as: F refined =F compressed +Λ(F compressed ); where Λ(x) is a user-defined nonlinear refinement function, using the nonlinear combination form Λ(x)=(α4x+β4x) 2 ), to enhance the expressive effect of the feature map; α4 and β4 are adjustment coefficients.

[0098] The final segmentation feature map restores the original spatial dimension and retains rich segmentation features, providing a high-precision representation of the landslide area and providing sufficient support for subsequent landslide area identification and extraction.

[0099] Step S5 achieves the organic fusion of high and low level information through bilinear interpolation upsampling, channel splicing, convolutional compression and residual refinement, so that the generated segmentation feature map not only has the original spatial dimension, but also retains rich details of the landslide area, providing a solid spatial and semantic foundation for the final landslide segmentation.

[0100] In landslide image segmentation, to improve the accuracy and robustness of landslide region identification, a multi-level nonlinear fusion loss function is constructed. Based on cross-entropy loss, Dice loss, and Focal loss, a hierarchical adaptive adjustment is performed on the weight allocation for class discrimination, consistency optimization, and hard-to-classify samples. This design aims to enhance the fine representation of landslide region edges and details, thereby meeting the segmentation requirements in complex landslide scenarios.

[0101] Step S6 includes the following:

[0102] Cross-entropy loss is used to calculate the classification probability of landslide areas and non-landslide areas in a segmentation feature map. Assuming the true class label is y∈{0,1} and the predicted probability is p∈[0,1], the cross-entropy loss function measures the classification accuracy between landslide areas and background areas. The expression for calculating cross-entropy is as follows: LCE =-∑ylog(p)+(1-y)log(1-p);

[0103] This step ensures that the segmentation feature map distinguishes between landslide areas and background areas, and lays the foundation for subsequent feature consistency adjustments.

[0104] To adjust the overlap consistency between the landslide area and the background area, the Dice loss function is used to calculate the degree of overlap between the segmentation mask and the ground truth mask. Dice loss is suitable for addressing the problem of imbalanced positive and negative samples, resulting in higher consistency in the segmentation results. The Dice loss is defined as: Among them, Dice loss improves the overall segmentation effect of landslide areas by calculating the intersection ratio between segmented regions and real labels.

[0105] To enhance the focus on hard-to-classify samples, the response of the Focal loss is multiplied by the dynamically adjusted coefficient of the Dice loss, forming an enhanced Focal loss term. The Focal part L in the adaptive loss function... Focal The square root function is used to amplify the impact of hard-to-classify samples on model optimization. The doubling effect of the Focal loss is defined as follows:

[0106] Using cross-entropy loss as the initial term of the excitation, and combining it with the adjustment factor of Dice loss and the enhanced Focal loss, the total loss function L is constructed. total The total loss function is expressed using a hierarchically weighted nonlinear combination, as shown in the following formula:

[0107] In this formula, the cross-entropy loss L CE The effect of Dice loss gradually diminishes. Dice loss, as the reciprocal factor, improves regional consistency, while Focal loss is multiplied after being adjusted by Dice weights to ensure higher sensitivity to detailed regions in complex landslide scenarios.

[0108] Step S6 constructs an adaptive loss function with dynamic adjustment capabilities through a hierarchical combination of cross-entropy, Dice, and Focal losses, thereby improving the accuracy of landslide area segmentation and optimizing sample consistency, providing sufficient support for the stable performance of the model under varying terrain conditions.

[0109] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0110] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0111] It should be noted that, in this document, the use of relational terms such as "first" and "second" is merely to distinguish one entity or operation from another, and does not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0112] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A landslide segmentation model based on multi-loss function fusion, characterized in that, Including the following steps: S1, acquire the landslide image to be segmented and perform initial convolution processing to extract channel response feature maps; S2, dilated convolution is performed on the channel response feature map at different dilation rates to generate a multi-scale feature fusion image; S3 generates a weighted feature map based on global and local features from a multi-scale feature fusion image. S4, the weighted feature map is combined with low-level and high-level features through residual connections and skip connections to obtain a combined feature map with continuous information transmission; S5, after upsampling the combined feature map, splice high and low level features by channel to generate a segmentation feature map restored to the original spatial dimension; S6 generates an optimized loss function for segmentation feature maps based on a weighted combination of cross-entropy loss, Dice loss, and Focal loss.

2. The landslide segmentation model based on multi-loss function fusion according to claim 1, characterized in that: Step S1 includes the following: S1.1, for the input landslide image Initial convolution operations are performed using a 3×3 convolution kernel K to generate preliminary feature maps. Where C ' The output channel number is denoted by H, W, and C, where H, W, and C represent the image height, width, and number of channels, respectively. S1.2, To enhance the characteristic response of the landslide area, a dynamic channel weight adjustment mechanism is adopted; specifically, a weight function W is constructed. c Feature enhancement for each channel is based on dynamically allocating response intensity according to image content. The calculation steps are as follows: Calculate the global average pooling of the input image to generate the channel-level average feature matrix. A weight function matrix is ​​generated based on the channel-level average feature matrix, and a nonlinear transformation is applied to obtain the weighted response. A custom activation function g(x) is used to satisfy g(x) = α1x + β1x. 2 Where α1 and β1 are coefficients, dynamically adjusted through parameter tuning; weighting function W c : The obtained weight function is applied to each channel of the initial feature map to enhance the features along the channel dimension. The expression is: F' i,j,k =F i,j,k ·W c The adjusted weights are applied to the feature maps of each channel, making the features of the landslide area more prominent. S1.3, Based on the feature map, the feature map is normalized by batch normalization to eliminate bias and accelerate convergence; By calculating the mean and standard deviation of the current batch, the feature map is adjusted channel by channel so that the mean of the feature map in different channels is close to zero and the standard deviation is close to 1. S1.4, ReLU activation is applied to the normalized feature map to introduce nonlinearity and generate the final output feature map.

3. The landslide segmentation model based on multi-loss function fusion according to claim 2, characterized in that: Step S2 includes the following: S2.1, For the output feature map generated in step S1, set different dilation rates and perform dilated convolution processing on the landslide image features with different receptive fields; S2.2, based on the output of each dilation rate convolutional layer, the spatial dimension of the feature map is expanded and superimposed to form feature maps at different scales; S2.3, Multi-scale feature maps are stitched together along the channel dimension to overlay features from each scale onto the same fused feature map, thus obtaining a multi-scale feature fusion image.

4. The landslide segmentation model based on multi-loss function fusion according to claim 3, characterized in that: Step S3 includes the following: S3.1 For the multi-scale feature fusion image generated in step S2, first use 1×1 convolution to compress it to obtain the global feature map of the image; S3.2, the global feature map is subjected to global average pooling to generate a single global feature matrix, where each element of the matrix represents the global response intensity of the corresponding channel; S3.3, directly use 1×1 convolution on the multi-scale fused feature map to obtain the local feature map; S3.4, add the local feature map and the global feature matrix to generate an intermediate feature map containing both global and local information; S3.5, perform another 1×1 convolution operation on the intermediate feature map to generate a weighted feature map; S3.6 Multiply the weighted feature map and the local feature map point by point to generate the final weighted feature map, thereby enhancing the significance of the landslide area features.

5. The landslide segmentation model based on multi-loss function fusion according to claim 4, characterized in that: Step S4 includes the following: S4.1 In the process of encoding the weighted feature map, to ensure the continuous transmission of features, a 1×1 convolutional structure is added after each 3×3 convolutional layer; after the 3×3 convolution extracts spatial features, the number of channels is adjusted by the 1×1 convolution to keep the feature map compact in the channel dimension; then, the residual connection mechanism is applied to add the input feature map and the output feature map after convolution layer by layer to maintain the continuity of information flow; S4.2 In the decoding stage, in order to preserve the detailed information of the image, the low-level features in the coding layer that have undergone residual connection processing are directly connected to the high-level features in the decoding layer; in this skip connection mechanism, the low-level features of the coding layer are directly passed to the decoding layer, ensuring that the decoding layer can obtain detailed information from the low-level features; S4.3, after point-by-point combination of low-level and high-level features, the generated combined feature map F combined It has the ability to express complete information about the landslide area.

6. The landslide segmentation model based on multi-loss function fusion according to claim 5, characterized in that: Step S5 includes the following: S5.1, In the decoding stage, bilinear interpolation is used to analyze the combined feature map F. combined A 2x upsampling process is performed to gradually restore the spatial resolution of the feature map; bilinear interpolation ensures the continuity of the upsampled feature map by linearly weighting adjacent pixels, resulting in the upsampled feature map F. upsampled ; S5.2, upsample the feature map F upsampled The second feature map F at the corresponding spatial location of the encoding stage encoded Channel splicing is performed to form a combined feature map F containing high-level and low-level information. concat ; S5.3, to avoid redundancy in the number of channels in the concatenated feature map, a 1×1 convolution operation is used to concatenate the combined feature map F. concat Channel compression is performed; the compressed feature map after convolution is represented as F. compressed ; S5.4, After channel compression, to ensure that spatial details are represented in the feature map, the compressed feature map F is... compressed Residual structures are applied for refinement; residual refinement is performed by adding the output of convolution compression to the original input features through pointwise addition operations, so as to recover detailed information in the segmentation feature map; the final segmentation feature map is restored to the original spatial dimension.

7. The landslide segmentation model based on multi-loss function fusion according to claim 6, characterized in that: Step S6 includes the following: Cross-entropy loss is used to calculate the classification probability of landslide regions and non-landslide regions in the segmentation feature map; let the true class label be y∈{0,1}, and the predicted probability be p∈[0,1]. The expression for calculating cross-entropy is as follows: L CE =-∑ylog(p)+(1-y)log(1-p); The overlap between the segmentation mask and the ground truth mask is calculated using the Dice loss function. The Dice loss is defined as follows: The Dice loss is calculated by measuring the intersection ratio between the segmented region and the ground truth label. The response of the Focal loss is multiplied by the dynamic adjustment coefficient of the Dice loss, forming an enhanced Focal loss term. The multiplication result of the Focal loss is defined as follows: Using cross-entropy loss as the initial term of the excitation, and combining it with the adjustment factor of Dice loss and the enhanced Focal loss, the total loss function L is constructed. total The total loss function is expressed using a hierarchically weighted nonlinear combination, as shown in the following formula:

Images