Landslide Identification Method, Device and Storage Medium Based on Multi-Path Feature Fusion

Abstract

The present invention provides a landslide recognition method, device and storage medium based on multi-path feature fusion, which relates to the field of geological disaster monitoring and early warning. The method includes: obtaining a landslide image data set, performing preprocessing, dividing the data set, constructing a multi-path landslide recognition model integrating a dual attention mechanism, training the multi-path landslide recognition model using the data set, outputting the multi-path landslide recognition model after the training is completed, inputting the landslide image data to be recognized into the trained multi-path landslide recognition model, and outputting the landslide recognition result. The device and storage medium are used to implement the method. The beneficial effects of the present invention are: a multi-path landslide recognition model is proposed, realizing deep interaction of different types of landslide image data at the feature level, maintaining high-resolution feature information, significantly improving the accuracy of landslide recognition; reducing the calculation cost, significantly improving the real-time performance of landslide recognition, and realizing real-time monitoring and early warning.

Description

Technical Field

[0001] The present invention relates to the field of geological disaster monitoring and early warning, and particularly to a landslide identification method, device and storage medium based on multi-path feature fusion. Background Art

[0002] With the development of technology, modern geological disaster monitoring methods have gradually introduced advanced technical means, including remote sensing technology, sensor networks and big data analysis. These methods have made remarkable progress in improving monitoring efficiency and early warning accuracy.

[0003] Remote sensing technology: Using satellite images and unmanned aerial vehicle technology, large areas can be monitored. Remote sensing technology can provide high-resolution surface images to help identify early signs of geological disasters. Defects: The acquisition frequency of remote sensing data is low, making it difficult to achieve real-time monitoring; data processing is complex and requires a large amount of computing resources.

[0004] Sensor network: Sensors such as rain gauges, displacement gauges, and groundwater level gauges are arranged in areas prone to geological disasters to monitor changes in environmental parameters in real time. Defects: The cost of sensor deployment is high and the maintenance difficulty is large; the data of a single sensor is easily affected by noise interference, which affects the accuracy of early warning.

[0005] Big data analysis: By collecting and analyzing a large amount of historical data, a geological disaster risk assessment model is constructed to improve the scientificity and accuracy of early warning. Defects: Model training depends on a large amount of high-quality data, and data acquisition is difficult; the calculation cost of complex models is high and the real-time performance is poor.

[0006] Traditional single encoder network architecture: In the field of geological disaster monitoring, traditional landslide identification methods usually adopt a single encoder network architecture. These methods rely on classical deep learning and machine learning models such as UNE, FCN, DeeplabV3, multi-layer perceptron (MLP), support vector machine (SVM) and artificial neural network (ANN). These traditional models have played an important role in landslide identification, but there are also certain limitations. Defects: After being trained on a specific dataset, the models of the traditional single encoder network architecture may be difficult to adapt to new environments or geological conditions; due to the high computational complexity of the models, the real-time processing ability is weak, making it difficult to respond quickly in sudden landslide disasters. Summary of the Invention

[0007] The purpose of the present invention is to: In order to overcome the deficiencies of multi-source data fusion, high computational cost and poor real-time performance existing in the traditional single encoder network architecture in landslide intelligent identification, the present invention provides a landslide identification method, device and storage medium based on multi-path feature fusion. A landslide identification method based on multi-path feature fusion mainly includes the following steps:

[0008] S1. Obtain the landslide image dataset and perform preprocessing;

[0009] S2. Divide the dataset;

[0010] S3. Construct a multi-path landslide recognition model integrating a dual attention mechanism;

[0011] S4. Use the dataset to train the multi-path landslide recognition model, and output the multi-path landslide recognition model after training;

[0012] S5. Input the landslide image data to be recognized into the trained multi-path landslide recognition model, and output the landslide recognition result.

[0013] Furthermore, the multi-path landslide recognition model includes a main encoder module and a secondary encoder module containing stacked encoders, and a decoder module containing stacked decoders. The convolutional block attention mechanism module is introduced into the composition structures of both the encoder and the decoder;

[0014] It also includes that the main encoder module and the secondary encoder module are connected through a feature-aware self-attention mechanism gate, and are connected to the decoder module through the deepest encoder and skip connections.

[0015] Furthermore, the landslide image dataset includes landslide images and their terrain factor images. The preprocessing steps are as follows: perform image enhancement and normalization on the landslide images and their terrain factor images respectively, and then match and bind them one by one.

[0016] Furthermore, the convolutional block attention mechanism module is a sequential integration of the channel and spatial attention mechanisms. The mathematical representation of the process is as follows:

[0017]

[0018]

[0019] Among them, represents the input feature map of the convolutional block attention mechanism module, , respectively represent the integration results of the channel and spatial attention mechanisms, also represents the output feature map of the convolutional block attention mechanism module; and respectively represent the channel and spatial attention functions; represents element-wise multiplication.

[0020] Furthermore, the main encoder module and the secondary encoder module respectively process the landslide images and the terrain factor images. The input images go through a 1x1 convolution, and then are encoded by the stacked encoders. The main encoder module has one more encoder than the secondary encoder module;

[0021] The decoder module adopts the decoder architecture of the U-net model, including the same number of stacked decoders as the encoders in the main encoder module, decodes the encoding result of the deepest encoder in the main encoder module, and finally outputs the landslide recognition result map through upsampling and convolution operations.

[0022] Furthermore, the encoder consists of a residual connection block, a convolutional block attention mechanism module, and a convolutional module in sequence;

[0023] Furthermore, the decoder consists of an upsampling module, a splicing module, a residual structure module, and a convolutional block attention mechanism module in sequence.

[0024] Furthermore, the feature-aware self-attention mechanism fuses the encoded feature maps of the encoder layers in the secondary encoder module with the encoded feature maps of the same-level encoders in the main encoder module;

[0025] The specific working process is as follows: A query matrix and a sum matrix are generated from the output feature map of the encoder layer in the main encoder module through a 1x1 convolutional layer. At the same time, a key matrix is generated from the output feature map of the encoder layer in the secondary encoder module, and then the self-attention mechanism weighted feature is calculated. Finally, the weighted feature is fused with the output feature map of the encoder in the main encoder module to obtain a feature map with self-attention adjustment, which is used as the input of the next-level encoder in the main encoder module.

[0026] Furthermore, the skip connection is implemented by a dilated spatial pyramid pooling module, which is connected to the output of the encoder in the main encoder module and the splicing module of the decoder at the same level in the decoder module at both ends, retains the output of each encoder layer, and passes it to the decoder layer of the same level;

[0027] The formula expression of the working process of the dilated spatial pyramid pooling module is as follows:

[0028]

[0029] Among them, is the input feature map, represents the dilated convolution operation with a dilation rate of , is the corresponding weight, and * represents the convolution operation, is the number of types of dilated convolutions.

[0030] Furthermore, step S4 is specifically as follows:

[0031] S41. Set the iteration period and the maximum number of training epochs;

[0032] S42. Select the landslide images and their topographic factor images in the training set, and input them into the main encoder module and the secondary encoder module of the multi-path landslide recognition model respectively. The recognition result map is output through the decoder;

[0033] S43. Calculate the loss function of the model, specifically a weighted combination of the cross-entropy loss function and the Dice loss function, and optimize the model parameters through the backpropagation algorithm and gradient descent;

[0034] S44. Repeat steps S42 - S43. Retain the model once for each iteration cycle. Input the data in the validation set into the updated model to evaluate the model performance;

[0035] S45. End the training when the maximum number of training epochs is reached, and select the output of the model with the optimal performance as the trained multi-path landslide recognition model.

[0036] A storage medium stores instructions and data for implementing a landslide recognition method based on multi-path feature fusion.

[0037] A computer device includes: a processor and the storage medium; the processor loads and executes the instructions and data in the storage medium for implementing a landslide recognition method based on multi-path feature fusion.

[0038] The beneficial effects brought by the technical solution provided by the present invention are as follows: The present invention proposes a multi-path landslide recognition model, which includes encoder modules with multiple paths and introduces a dual attention mechanism to achieve deep interaction of different types of landslide image data at the feature level. In addition, the encoder module and the decoder module are connected through skip connections to maintain high-resolution feature information, significantly improving the accuracy of landslide recognition; the parallel operation of the multi-path encoder module reduces the computational cost, and the fast decoding ability of the U-net selected by the decoder module significantly improves the real-time performance of landslide recognition, realizing real-time monitoring and early warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The present invention will be further described below in conjunction with the drawings and embodiments. In the drawings:

[0040] Figure 1 is a flowchart of a landslide recognition method based on multi-path feature fusion in an embodiment of the present invention;

[0041] Figure 2 is a schematic diagram of the network structure of a multi-path landslide recognition model in an embodiment of the present invention;

[0042] Figure 3 is a schematic diagram of the structure of a convolutional attention module in an embodiment of the present invention;

[0043] Figure 4It is a schematic structural diagram of the residual structure block in the embodiment of the present invention;

[0044] Figure 5 It is a schematic structural diagram of the feature-aware self-attention mechanism gate in the embodiment of the present invention;

[0045] Figure 6 It is a schematic structural diagram of the atrous spatial pyramid pooling module in the embodiment of the present invention;

[0046] Figure 7 It is a schematic diagram of the landslide recognition effect of each model in the comparative experiment in the embodiment of the present invention;

[0047] Figure 8 It is a schematic diagram of the operation of the hardware device in the embodiment of the present invention. Specific embodiments

[0048] In order to have a clearer understanding of the technical features, objectives, and effects of the present invention, the specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0049] The embodiment of the present invention provides a landslide recognition method, device, and storage medium based on multi-path feature fusion.

[0050] Please refer to Figure 1 , Figure 1 It is a flowchart of a landslide recognition method based on multi-path feature fusion in the embodiment of the present invention, specifically including the following steps:

[0051] First step, obtain the landslide image dataset and perform preprocessing.

[0052] The landslide image dataset includes landslide images and their terrain factor images. The preprocessing steps are: perform image enhancement and normalization on the landslide images and their terrain factor images respectively, and then match and bind them one by one.

[0053] Second step, divide the dataset.

[0054] Third step, construct a multi-path landslide recognition model integrating a dual attention mechanism, as shown in Figure 2 , specifically:

[0055] It includes a main encoder module and a secondary encoder module containing stacked encoders, and a decoder module containing stacked decoders. The convolutional block attention mechanism module is introduced into the composition structures of both the encoder and the decoder.

[0056] The main encoder module and the secondary encoder module process the landslide images and the terrain factor images respectively. The input image passes through a 1x1 convolution and then is encoded by the stacked encoders. The main encoder module has one more encoder than the secondary encoder module.

[0057] The decoder module adopts the decoder architecture of the U-net model, which contains the same number of stacked decoders as the encoders in the main encoder module. It decodes the encoding results of the deepest encoder in the main encoder module, and finally outputs the landslide recognition result map through upsampling and convolution operations.

[0058] The composition structure of the encoder is, in sequence, a residual connection block, a convolutional block attention mechanism module, and a convolutional module; the composition structure of the decoder is, in sequence, an upsampling module, a splicing module, a residual structure module, and a convolutional block attention mechanism module.

[0059] In the encoding and decoding stages, a convolutional attention module (CBAM) is introduced. The specific structure is as Figure 3 shown. It processes the feature map with channel attention and spatial attention to enhance the expression ability of important features. The convolutional block attention mechanism module is the sequential integration of the channel and spatial attention mechanisms. The specific steps are as follows:

[0060] Channel attention: Global average pooling and global max pooling are performed on the input feature map to obtain two different feature maps, which are processed through a shared fully connected layer and element-wise weighted.

[0061] Spatial attention: Max pooling and average pooling in the channel dimension are performed on the input feature map to obtain two different feature maps, which are processed through a convolutional layer and element-wise weighted.

[0062] The mathematical representation of the process is as follows:

[0063]

[0064]

[0065] Among them, represents the input feature map of the convolutional block attention mechanism module, , respectively represent the integration results of the channel and spatial attention mechanisms, also represents the output feature map of the convolutional block attention mechanism module; and respectively represent the channel and spatial attention functions; represents element-wise multiplication.

[0066] The residual connection technology is adopted, and a residual structure block (Residual Block) is introduced in the encoder and decoder layers. The structure is as Figure 4 shown. This connection method can effectively alleviate the problem of gradient disappearance in deep neural networks and retain high-resolution feature information at the same time.

[0067] Its wide application in the model aims to solve the problem of gradient disappearance in deep networks and improve landslide recognition performance. Through residual connections, the input features are directly passed to subsequent layers, simplifying the training and enhancing feature propagation ( Figure 3 )). For each residual structure block, its output is defined by the following formula:

[0068]

[0069] where, represents the output of the residual structure block; represents the input features, represents the weighted operation and non-linear activation function in the residual block. This structure enables the network to learn the residual mapping between the input and output, helping to capture the subtle feature changes in landslide recognition, thereby improving the recognition accuracy.

[0070] It also includes that the main encoder module and the secondary encoder module are connected through the feature-aware self-attention mechanism gate, and are connected to the decoder module through the deepest encoder and skip connections. The present invention realizes the dual attention mechanism through the feature-aware self-attention mechanism gate and the convolutional block attention mechanism module.

[0071] The structure of the feature-aware self-attention mechanism gate is as shown in Figure 5 . The encoded feature map is fused with the encoded feature map of the encoder in the same layer in the main encoder module. The specific working process is as follows: A query matrix and a sum matrix are generated from the encoder output feature map in the main encoder module through a 1x1 convolutional layer. At the same time, a key matrix is generated from the encoder output feature map in the secondary encoder module, and then the weighted features of the self-attention mechanism are calculated. Specifically: First, the similarity between the query matrix (Q) and the key matrix (K) is calculated through matrix multiplication to obtain the attention score; Secondly, the obtained score matrix is normalized and transformed into attention weights through the sigmoid function; Finally, the attention weights and the sum matrix (V) are weighted and summed to obtain the weighted features; The mathematical representation of the self-attention weights is:

[0072]

[0073] where, represents the self-attention weight function; is the activation function; is the dimension of the key matrix, is used to scale the dot product to prevent the gradient from being too small.

[0074] The weighted features are fused with the feature maps output by the encoder in the main encoder module to obtain feature maps with self-attention adjustment, which serve as the input for the next layer of the encoder in the main encoder module. The feature maps output by the self-attention mechanism contain rich context information, significantly improving the model's sensitivity to landslide features (such as terrain changes, vegetation cover differences, etc.). This enables the model to focus on key features and deeply learn the high-level representations of landslides.

[0075] The skip connections are implemented by the Atrous Spatial Pyramid Pooling (ASPP) module. The two ends are respectively connected to the output of the encoder in the main encoder module and the splicing module of the decoder at the same level in the decoder module.

[0076] The specific structure of the Atrous Spatial Pyramid Pooling is as Figure 6 shown. By using atrous convolutions with different dilation rates, it effectively captures landslide features at different scales. It can provide a larger receptive field without increasing the computational cost, thereby capturing more extensive context information and enhancing the model's adaptability to complex environments.

[0077] The formula for the working process of the Atrous Spatial Pyramid Pooling module is as follows:

[0078]

[0079] Among them, is the input feature map, represents the atrous convolution operation with a dilation rate of , is the corresponding weight, * represents the convolution operation, is the number of types of atrous convolutions.

[0080] Fourthly, use the dataset to train the multi-path landslide recognition model, and output the multi-path landslide recognition model after training. Specifically:

[0081] Step 1, set the number of epochs and the maximum number of training rounds.

[0082] Step 2, select the landslide images and their terrain factor images in the training set, and input them into the main encoder module and the secondary encoder module of the multi-path landslide recognition model respectively, and output the recognition result map through the decoder.

[0083] Step 3, calculate the loss function of the model, specifically a weighted combination of the cross-entropy loss function and the Dice loss function, and optimize the model parameters through the backpropagation algorithm and gradient descent;

[0084] Step 4: Repeat Steps 2 and 3. Retain the model once an iteration cycle is completed. Input the data in the validation set into the updated model and evaluate the model performance.

[0085] Step 5: End the training when the maximum number of training epochs is reached. Select the model with the optimal performance as the trained multi-path landslide recognition model for output.

[0086] Step 5: Input the landslide image data to be recognized into the trained multi-path landslide recognition model and output the landslide recognition result.

[0087] To verify the performance of the proposed multi-path feature fusion neural network in landslide recognition, four classic semantic segmentation models, including FCN, DeeplabV3, Unet, and ResUnet, were selected as comparison objects. All models were trained on the same training dataset and evaluated on the same test dataset. The training dataset includes various landslide samples to ensure that the model can learn the characteristics of different types of landslides. The test dataset contains various actual landslide scenarios to evaluate the performance of the model in real applications. The specific landslide recognition effects of each model are as Figure 7 shown. The accuracy of the recognition results is summarized in Table 1. Three typical landslide images (labeled a, b, and c respectively) in the test set were selected for detailed comparison.

[0088] This comparative analysis reveals the significant advantages of the multi-path feature fusion neural network in accurate landslide recognition compared to the FCN, DeeplabV3, Unet, and ResUnet models. Specifically, there are certain limitations in accurately identifying the landslide area for other models, manifested as misclassification and omission of the landslide area.

[0089] By comprehensively comparing the performance of these five models in multiple accuracy metrics, it can be clearly seen that the multi-path feature fusion neural network not only has obvious advantages in the accuracy of identifying the landslide area, but also performs well in capturing the comprehensiveness and consistency of landslides.

[0090] Table 1 Comparison of recognition results of different network models

[0091]

[0092] Verification of the effect of the dual-encoding path:

[0093] In order to further verify the effectiveness of the multi-path feature fusion neural network, four groups of experiments were designed, using different input sample methods for verification. The first group of experiments used only optical remote sensing images as input, and the results showed that the precision was 0.856, but the recall was low, which was 0.778. The second group of experiments used only terrain factor images as input. The results showed that the precision and recall of this method were 0.602 and 0.553, respectively, and it was almost impossible to effectively identify landslides. The third group of experiments merged the optical remote sensing images with the terrain factor images and input them into the main encoder. The results showed that the precision was 0.848 and the recall was 0.790. Although it has improved, it has not reached the optimal level. Finally, the fourth group of experiments used optical remote sensing images and terrain factor images to input the main encoder and auxiliary encoder respectively. The results showed that the precision, recall, F1 score and average intersection-over-union ratio all reached 0.866, 0.831, 0.848 and 0.860. These results verify the significant advantages of the multi-path input sample method in the landslide identification task, indicating that the model can better integrate and utilize information from different sources to improve the accuracy and comprehensiveness of identification.

[0094] Table 2 Comparison of results of different input sample methods

[0095]

[0096] Dual attention mechanism ablation verification:

[0097] In order to verify the impact of the attention mechanism on the performance of the model, an ablation experiment was conducted to test the effects of the convolutional block attention mechanism (CBAM) and the feature-aware self-attention mechanism (Self-attention Gate). The results show that the addition of CBAM improves the model's precision from 0.857 to 0.864, the recall from 0.805 to 0.811, and the F1 score and average intersection-over-union ratio increase to 0.837 and 0.851, respectively. After the introduction of the Self-attention Gate, the recall rate of the model is significantly improved to 0.827, while the precision and F1 score are also improved to 0.853 and 0.840, respectively. When the two attention mechanisms are applied simultaneously, all evaluation indicators of the model reach the optimal level, with the precision, recall, F1 score and average intersection-over-union ratio being 0.866, 0.831, 0.848 and 0.860, respectively. These results indicate that the convolutional block attention mechanism and feature-aware self-attention mechanism play an important role in improving model performance, helping the model to more effectively capture the characteristics of the landslide area and improve recognition accuracy and comprehensiveness.

[0098] Table 3 Attention mechanism ablation experiment results

[0099]

[0100] Please refer to Figure 8 , Figure 8 which is a schematic diagram of the operation of the hardware device according to an embodiment of the present invention. The hardware device specifically includes: a computer device 401, a processor 402, and a storage medium 403.

[0101] A computer device 401: The computer device 401 implements the method for landslide recognition based on multi-path feature fusion.

[0102] A processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the method for landslide recognition based on multi-path feature fusion.

[0103] A storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the method for landslide recognition based on multi-path feature fusion.

[0104] The beneficial effects of the present invention are as follows: The present invention proposes a multi-path landslide recognition model, which includes encoder modules of multiple paths and introduces a dual attention mechanism to achieve deep interaction of different types of landslide image data at the feature level. In addition, the encoder module and the decoder module are connected through skip connections to maintain high-resolution feature information, significantly improving the accuracy of landslide recognition. By running the multi-path encoder modules in parallel, the computational cost is reduced, and the fast decoding ability of the U-net selected for the decoder module significantly improves the real-time performance of landslide recognition, realizing real-time monitoring and early warning.

[0105] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. However, such modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A landslide identification method based on multi-path feature fusion, characterized in that: The specific steps include: S1, obtain landslide image dataset and perform preprocessing; S2, divide the data set; S3. Construct a multi-path landslide recognition model integrating dual attention mechanism, specifically: Including a main encoder module and a sub-encoder module including stacked encoders and a decoder module including stacked decoders, and the composition structures of the encoder and the decoder both introduce a convolutional block attention mechanism module; It also includes that the main encoder module is connected to the sub-encoder module through the feature-aware self-attention mechanism gate, and is connected to the decoder module through the deepest encoder and skip connection; S4, using the data set to train the multi-path landslide recognition model, and outputting the multi-path landslide recognition model after the training is completed; S5, inputting the landslide image data to be identified into the trained multi-path landslide identification model, and outputting the landslide identification result; The main encoder module and the sub-encoder module process the landslide image and the terrain factor image respectively. The input image undergoes a 1x1 convolution and is then encoded by a stacked encoder. The main encoder module has one more encoder than the sub-encoder module. The decoder module adopts the decoder architecture of the U-net model, including the same number of stacked decoders as the encoders in the main encoder module, and decodes the encoding results of the deepest encoder of the main encoder module step by step, and finally outputs the landslide recognition result map through upsampling and convolution operations; The encoder structure consists of a residual connection block, a convolution block attention mechanism module and a convolution module in sequence; The decoder consists of an upsampling module, a splicing module, a residual structure module, and a convolutional block attention mechanism module. The feature-aware self-attention mechanism gate fuses the encoded feature map of the encoder layer in the secondary encoder module with the encoded feature map of the same layer encoder in the primary encoder module; The specific working process is as follows: the query matrix and the sum matrix are generated from the output feature map of the encoder layer in the main encoder module through a 1x1 convolution layer, and the key matrix is ​​generated from the output feature map of the encoder layer in the secondary encoder module, and then the weighted features of the self-attention mechanism are calculated. Finally, the weighted features are fused with the encoder output feature map in the main encoder module to obtain the feature map with self-attention adjustment as the input of the next layer of encoder in the main encoder module; The jump connection is implemented by a dilated spatial pyramid pooling module, and the two ends are respectively connected to the output of the encoder in the main encoder module and the splicing module of the decoder at the same level in the decoder module, and the output of each encoder layer is retained and passed to the decoder layer at the same level; The formula of the working process of the atrous space pyramid pooling module is as follows: Among them, X is the input feature map, Indicates that the expansion rate is d i The dilated convolution operation, W i is the corresponding weight, * represents the convolution operation, and N is the number of types of dilated convolution.

2. A landslide identification method based on multi-path feature fusion as claimed in claim 1, characterized in that: The landslide image data set includes landslide images and terrain factor images thereof. The preprocessing steps are: performing image enhancement and normalization on the landslide images and terrain factor images thereof respectively, and then matching and binding them one by one.

3. A landslide identification method based on multi-path feature fusion as claimed in claim 1, characterized in that: The convolutional block attention mechanism module is a sequential integration of channel and spatial attention mechanisms. The mathematical representation of the process is as follows: F′=M c (F)⊙F F″=M s (F′)⊙F′ Among them, F represents the input feature map of the convolutional block attention mechanism module, F′ and F″ represent the integrated results of the channel and spatial attention mechanisms respectively, and F″ also represents the output feature map of the convolutional block attention mechanism module; M c (·) and M s (·) denote channel and spatial attention functions respectively; ⊙ denotes element-wise multiplication.

4. A landslide identification method based on multi-path feature fusion as claimed in claim 1, characterized in that: Step S4 is specifically as follows: S41, setting the iteration period and the maximum number of training rounds; S42, selecting a landslide image and a terrain factor image in a training set, inputting them into a main encoder module and a sub-encoder module of a multi-path landslide recognition model respectively, and outputting a recognition result image through a decoder; S43, calculating the loss function of the model, specifically a weighted combination of a cross entropy loss function and a Dice loss function, and optimizing the model parameters through a back propagation algorithm and gradient descent; S44, repeating steps S42-S43, retaining the model once after reaching one iteration cycle, inputting the validation set data into the updated model, and evaluating the model performance; S45, when the maximum number of training rounds is reached, the training is terminated, and the model output with the best performance is selected as the trained multi-path landslide recognition model.

5. A storage medium, characterized in that: The storage medium stores instructions and data for implementing a landslide identification method based on multi-path feature fusion as described in any one of claims 1 to 4.

6. A computer device, characterized in that: include: Processor and storage medium; the processor loads and executes instructions and data in the storage medium to implement a landslide identification method based on multi-path feature fusion as described in any one of claims 1 to 4.

Images