An active landslide automatic identification method based on spatial guidance deep abstraction network

Abstract

The application belongs to the field of active landslide identification, and specifically discloses an active landslide automatic identification method based on a space-guided deep abstract network, which comprises the following steps: acquiring a slope unit data set, each slope unit containing disaster-causing attributes and spatial coordinate features; using a space high-frequency mapping module to perform Fourier position coding on the spatial coordinate features to obtain a high-dimensional spatial feature matrix; using a dynamic local feature selection module to input the high-dimensional spatial feature matrix into a super network to generate a dynamic sparse mask, and performing element-wise weighted fusion of the disaster-causing attribute features and the dynamic sparse mask to obtain mask disaster-causing attribute features; using a space-guided residual block to extract the mask disaster-causing attribute features and introducing the high-dimensional spatial feature matrix for spatial constraint to obtain fusion features; and inputting the fusion features into a classification head to output the active landslide identification result of each slope unit. The application can effectively improve the active landslide automatic identification precision under a wide-area complex geological environment and reduce the false alarm rate and the missed identification rate.

Description

Technical Field

[0001] This application belongs to the field of active landslide identification, and more specifically, relates to an automatic identification method for active landslides based on spatially guided deep abstraction networks. Background Technology

[0002] Landslides are the most common and extremely dangerous type of geological hazard, which can cause significant economic losses and pose a huge risk to human settlements. Active landslides are those that show obvious signs of movement but have not yet experienced overall instability. Since sudden destructive landslides often evolve from potential active landslides, early identification of these dangerous slopes can effectively avoid or mitigate catastrophic consequences.

[0003] In recent years, temporal synthetic aperture radar interferometry (InSAR) technology has become an important means of identifying active landslides due to its ability to observe minute surface deformations over all weather conditions and over a wide area. However, traditional identification relies on "InSAR deformation + manual visual interpretation," a method that is time-consuming, labor-intensive, and highly subjective. With the development of artificial intelligence technology, automatic landslide identification algorithms based on deep learning or machine learning (such as SVM, CNN, U-Net, CART, etc.) have gradually become a research hotspot. However, existing automatic identification methods still have the following shortcomings: on the one hand, the surface deformation information extracted by InSAR is easily affected by errors caused by non-landslide factors (such as groundwater extraction, freeze-thaw cycles, etc.), leading to high false alarm rates; on the other hand, when dealing with the spatial location of slope units, mainstream deep learning algorithms, due to the inherent bias of the model towards learning low-frequency smooth functions, i.e., spectral bias, struggle to accurately capture the complex topographical boundaries of landslides, resulting in high false alarm and missed identification rates in complex geological environments.

[0004] Therefore, how to improve the accuracy of automatic identification of active landslides in a wide range of complex geological environments in order to reduce the false alarm rate and the missed identification rate is an urgent problem to be solved. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this application is to provide an automatic identification method for active landslides based on spatially guided deep abstraction networks, which can improve the accuracy of automatic identification of active landslides in a wide range of complex geological environments and reduce the false alarm rate and missed identification rate.

[0006] To achieve the above objectives, in a first aspect, this application provides an automatic identification method for active landslides based on spatially guided deep abstraction networks, comprising the following steps: S10 provides a spatially guided deep abstraction network and obtains a slope unit dataset, wherein each slope unit in the slope unit dataset contains disaster-causing attribute features and spatial coordinate features; the spatially guided deep abstraction network includes a spatial high-frequency mapping module, a dynamic local feature selection module, a spatially guided residual block, and a classification head; S20, The spatial high-frequency mapping module is used to perform Fourier position encoding on the spatial coordinate features, mapping the low-dimensional spatial coordinates to the high-dimensional frequency domain space to obtain a high-dimensional spatial feature matrix. S30, the high-dimensional spatial feature matrix is ​​input into the hypernetwork using the dynamic local feature selection module to generate a dynamic sparse mask, and the disaster-causing attribute features are fused with the dynamic sparse mask element-wise to obtain the mask disaster-causing attribute features. S40, the spatially guided residual block is used to extract features from the catastrophic attribute features of the mask, and the high-dimensional spatial feature matrix is ​​introduced into the feature extraction process for spatial constraints to obtain fused features; S50, input the fused features into the classification head, and output the active landslide identification result for each slope unit.

[0007] As a further preferred embodiment, the specific process of the Fourier position encoding is as follows: The spatial coordinate features are expanded in the last dimension, with a preset number of frequency bands. According to the formula Calculate the frequency basis matrix ; The spatial coordinate features are subjected to sine and cosine transforms using the frequency basis matrix, preserving the original low-dimensional input. The transform results are then concatenated with the spatial coordinate features along the channel dimension and flattened to obtain a high-dimensional spatial feature matrix. The dimension of the high-dimensional spatial feature matrix Satisfy the formula .

[0008] As a further preferred embodiment, the hypernetwork comprises two linear transformation layers, wherein the first fully connected layer maps the input dimension to 64 dimensions and performs nonlinear activation through a modified linear unit, and the second fully connected layer maps the feature dimension from 64 dimensions to... The dimension reshapes the output continuous tensor into a three-dimensional dynamic primitive weight tensor. ,in K The number of independent attention sub-basis sets. D The dimension of the disaster-causing attribute features is the input. The dynamic sparse mask is generated by: using Activation function on the three-dimensional dynamic original weight tensor The last dimension is sparsified, and the sparse activation function is constrained based on the 1.5-order Tsallis entropy, and its mathematical expression is: ,in This indicates retrieving the maximum value. express A probability simplex space of dimension 1. Tsallis-1.5 entropy is the core physical mechanism of dynamic sparsity. This represents the probability distribution vector assigned to each feature factor, i.e., all probability allocation schemes. Its transpose represents the relative importance of each local disaster-causing factor at its current spatial location. The unnormalized feature scores output by the network's front-end layer represent preliminary "evaluations" of the importance of each feature factor at the current specific spatial location. The activation function has geometric projection and truncation properties, which are used to set the weights of non-dominant factors that are irrelevant to the current spatial location and do not satisfy the local landslide disaster mechanism to 0, thereby generating an absolutely sparse spatially guided local feature selection mask matrix for each sample. .

[0009] As a further preferred embodiment, the specific process of the element-wise weighted fusion is as follows: the disaster-causing attribute feature vector is expanded in the second dimension by increasing its dimension, and the dynamic sparse mask is multiplied element-wise with the expanded disaster-causing attribute feature tensor using the tensor automatic broadcasting mechanism.

[0010] As a further preferred embodiment, the spatially guided residual block comprises two cascaded spatially guided abstraction layers and an explicit downsampling shortcut branch. The spatially guided abstraction layer includes a one-dimensional grouped convolutional layer, a virtual mini-batch normalization layer, and a gated linear unit. The gated linear unit is used to perform nonlinear selection filtering and aggregation on the features after grouped convolution and virtual mini-batch normalization.

[0011] As a further preferred embodiment, the downsampling shortcut branch introduces Dropout perturbation into the input features and integrates them into the high-dimensional spatial feature matrix for independent abstraction. The output of the spatial guided abstraction layer is added to the output of the downsampling shortcut branch as a residual, and then activated by a leaky modified linear unit to obtain the fused features.

[0012] As a further preferred embodiment, the disaster-causing attribute features include statistical values ​​of stratigraphic lithology, distance from fault, elevation, slope, aspect, planar curvature, profile curvature, topographic relief, distance from river, normalized vegetation index, annual average rainfall, distance from road, land use type, and annual average surface deformation rate in the east-west direction; the discrete features in the slope unit dataset are converted into integer indices through label encoding, the continuous features in the slope unit dataset are standardized using Z-scores based on the mean and standard deviation of the training set, and the spatial coordinate features are constrained to the range of negative 1 to positive 1 through mapping normalization.

[0013] As a further preferred embodiment, the classification head includes a Dropout layer and a two-layer fully connected classification head. The fused features are processed by the Dropout layer and then input into the two-layer fully connected classification head, and combined with the Sigmoid activation function to output the active landslide identification result for each slope unit.

[0014] As a further preferred embodiment, the specific method for obtaining the slope unit dataset is as follows: using digital elevation model data, the slope units are divided using the surface curvature watershed method. Through zonal statistics, the statistical values ​​of stratigraphic lithology, distance from fault, elevation, slope, aspect, plane curvature, profile curvature, topographic relief, distance from river, normalized vegetation index, annual average rainfall, distance from road, land use type, and east-west annual average surface deformation rate are summarized into the slope unit vector data to generate the slope unit dataset.

[0015] Secondly, this application provides an automatic identification system for active landslides based on spatially guided deep abstraction networks, used to implement the method described in any one of the above, comprising: The data acquisition module is used to acquire a slope unit dataset, wherein each slope unit in the slope unit dataset contains disaster-causing attribute features and spatial coordinate features; The spatial high-frequency mapping module is used to perform Fourier position encoding on the spatial coordinate features, mapping the low-dimensional spatial coordinates to the high-dimensional frequency domain space to obtain a high-dimensional spatial feature matrix. The dynamic local feature selection module is used to input the high-dimensional spatial feature matrix into the hypernetwork to generate a dynamic sparse mask, and to perform element-wise weighted fusion of the catastrophic attribute features and the dynamic sparse mask to output the catastrophic attribute features of the mask. A spatially guided residual block is used to extract features from the catastrophic attribute features of the mask, and the high-dimensional spatial feature matrix is ​​introduced into the feature extraction process for spatial constraints, and the fused features are output. A classification head is used to classify the fused features and output the active landslide identification result for each slope unit.

[0016] This application has the following effects: (1) A spatial high-frequency mapping mechanism is proposed to solve the problem of landslide boundary ambiguity caused by "spectral bias". To address the "spectral bias" defect that traditional deep neural networks easily fall into when learning low-dimensional, low-frequency spatial features, they tend to learn low-frequency smoothing functions. This application introduces a high-dimensional random Fourier position coding technique to project the spatial location of the low-dimensional slope unit center onto high-dimensional, high-frequency spatial features. This mapping mechanism effectively breaks the model's tendency to over-smooth complex terrain features and significantly improves the model's boundary recognition accuracy for active landslides in a wide-area, complex mountainous geological environment.

[0017] (2) Constructing an absolutely sparse feature selection framework with "space-data" dual-stream coupling. Addressing the strong spatial heterogeneity of landslide induction mechanisms across a wide area (e.g., topography dominating in high-altitude areas, while rainfall or human activities dominate in river valleys), this application breaks away from the limitations of existing mainstream algorithms that overly rely on a globally unified static feature weight matrix. It uses a hypernetwork as the spatial guidance core to construct a "space-data" dual-stream coupling architecture. This mechanism uses high-frequency spatial signals as prior guidance to dynamically generate a unique feature selection mask for each local spatial location, and combines it with... By employing sparse activation functions with geometric projection properties, the sparsity and isolation of feature weights are achieved from a mathematical perspective. This breakthrough not only amplifies the main disaster-causing factors in the current region according to local conditions, but also accurately eliminates redundant features that have no significant impact on the formation of local landslides. This isolates the interference of environmental background noise caused by non-landslide factors such as groundwater extraction or engineering construction settlement on classification decisions, greatly reducing the false alarm rate and the missed identification rate.

[0018] (3) Establishing a multi-source collaborative identification mechanism based on the landslide disaster mechanism. Considering that the pixel-level segmentation method commonly used in traditional computer vision models would severely disrupt the inherent geological and physical continuity of the slope and lack practical engineering significance, this application abandons simple pixel-level identification and adopts "slope units" with clear three-dimensional real landform physical boundaries extracted based on the surface curvature watershed algorithm as the basic identification unit. Under this physical entity framework, this application uses time-series interferometric synthetic aperture radar (InSAR) and other technologies to extract the effective surface deformation and multi-source heterogeneous features covering disaster-inducing factors such as strata lithology and topography, as well as disaster-causing factors such as rainfall and human engineering activities, and deeply integrates them into the slope unit. This multi-dimensional collaborative mechanism completely breaks the limitations of single deformation monitoring, provides a comprehensive dataset that conforms to the geological evolution law for deep neural networks, and improves the identification accuracy of potential active landslides output by the algorithm. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall technical flow of the automatic identification method for active landslides provided in the embodiments of this application; Figure 2 This is a schematic diagram of the ascending and descending rails and the annual average deformation rate after two-dimensional decomposition provided in the embodiments of this application; wherein, (a) is the annual average deformation rate of the ascending rail, (b) is the annual average deformation rate of the descending rail, (c) is the annual average deformation rate in the east-west direction, and (d) is the annual average deformation rate in the vertical direction. Figure 3 The images show a comparison of the identification results of five algorithms provided in this application for active landslides in Hualong County; where (a)-(e) are the identification results of SGDANet, U-Net, CNN, SVM and CART algorithms, respectively. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0021] This application proposes an automatic identification method for active landslides based on Spatially Guided Deep Abstract Network (SGDANet). Its technical features are as follows: (1) Constructing a comprehensive set of identification indicators: Based on the development and evolution mechanism of landslides, a comprehensive set of identification indicators is proposed, which covers geological environment, rainfall, engineering activities and surface deformation factors. (2) Introducing Fourier position coding to enhance high-frequency geomorphic boundary perception: A Fourier position coding mechanism is introduced to map low-dimensional original spatial features to a high-dimensional high-frequency feature space. Through this technical feature, the network can effectively extract and retain high-frequency spatial feature signals of complex terrain, thereby improving the problem of missed identification caused by boundary smoothing in complex mountainous and steep terrain by mainstream deep learning models. (3) Based on Dynamic sparse feature selection using operators and hypernetworks: This application utilizes hypernetworks to receive high-dimensional, high-frequency spatial features and dynamically generate feature weights, while simultaneously employing... The operator generates an absolutely sparse mask. Using this sparse mask, the model can adaptively adjust the weight proportions of various disaster-causing factors according to the specific spatial location of different slope units, and precisely constrain the weights of irrelevant disaster-causing features at the current location to zero. This mechanism achieves dynamic feature selection and local noise reduction, improving the model's generalization adaptability in complex geological environments.

[0022] Reference Figure 1 As shown, this implementation scheme provides a new method for automatic identification of active landslides—the SGDANet algorithm, which specifically includes the following steps: Step 1) Effective Surface Deformation Extraction. In this application, synthetic aperture radar (SAR) data from the European Space Agency's (ESA) Sentinel-1 satellite is first used to measure the line-of-sight (LOS) deformation displacement and rate of the landslide surface using Small Baseline Set (SBAS-InSAR) technology. During this process, Advanced Land Observing Satellite (ALOS) DEM data with a resolution of 12.5 m is introduced to eliminate topographic effects and perform geometric correction. By screening short baseline interferometric pairs and setting a coherence coefficient threshold, filtering, phase unwrapping, and removal of orbital errors, atmospheric delay, and topographic errors are performed sequentially to finally obtain high-quality annual average LOS deformation rates of the ascending and descending slopes. Given that a single LOS-direction deformation cannot accurately reflect the deformation characteristics of the real surface, and that synthetic aperture radar is extremely insensitive to north-south deformation components, this study ignores the contribution of north-south displacement and uses observations from two different line-of-sight directions (ascending and descending orbits) to decompose the LOS deformation in two dimensions (2D), extracting the vertical and east-west deformation components (defining vertical upward as positive and eastward as positive). The satellite line-of-sight deformation can be expressed as the sum of the projections of the vertical deformation and the east-west deformation onto the LOS direction. Based on this, the decomposition equation is established as shown in equation (1): (1) In the formula: and These are the LOS directional deformations observed during satellite ascent and descent, respectively. and These represent the actual deformations in the vertical and east-west directions, respectively; The incident angle of the radar satellite; The radar line-of-sight (LOS) azimuth (0° with true north as the base, rotated clockwise) can be obtained from the relationship between the satellite's flight direction angle and imaging geometry; subscript and These correspond to the ascending and descending orbit parameters, respectively. By solving equation (1) simultaneously, accurate vertical and east-west effective two-dimensional surface deformation information is finally extracted. In this study, east-west surface deformation is selected as the effective deformation information.

[0023] Step 2) Constructing an active landslide identification index set. The identification index set is constructed using multi-source data: (1) Extracting surface deformation features using Sentinel-1A SAR imagery. (2) Calculating the Normalized Difference Vegetation Index (NDVI) using Sentinel-2 multispectral imagery. (3) Generating topographic factors, including elevation, slope, aspect, and curvature, using Advanced Land Observing Satellite (ALOS) Digital Elevation Model (DEM) data. (4) Constructing two types of geological factors: stratigraphic lithology and fault distance, using geological maps. (5) Establishing factors such as road distance and river distance using geographic data. (6) Generating land use type factors reflecting human engineering activities using land cover data. (7) Generating a precipitation factor based on annual average precipitation using Climate Hazards Group Infrared Precipitation and Station Data (CHIRPS) precipitation data.

[0024] Step 3) Slope Unit Segmentation. First, using Advanced Land Observation Satellite (ALOS) DEM data with a resolution of 12.5m, slope units were divided using the surface curvature watershed method. Then, through regional statistical analysis, 14 evaluation indicators, including stratigraphic lithology, distance from faults, elevation, slope, aspect, planar curvature, profile curvature, topographic relief, distance from rivers, Normalized Difference Vegetation Index (NDVI), average annual rainfall, distance from roads, land use type, and average annual surface deformation rate in the east-west direction, were statistically analyzed into slope unit vector data. This generated dataset was used for model training, validation, and testing based on the SGDANet algorithm. The regional statistical types and categories of the indicator set are shown in Table 1.

[0025] Table 1. Statistical Types and Categories of Multi-Source Data Partitions

[0026] In Table 1: MEAN is the average value of all pixels within the ramp cell; MAJORITY is the value with the highest frequency among all pixels within the ramp cell; MAXIMUM is the maximum value of all pixels within the ramp cell; RANGE is the difference between the maximum and minimum values ​​of all pixels within the ramp cell; STD is the standard deviation of all pixels within the ramp cell.

[0027] Step 4) Dataset Splitting. To ensure stable gradient propagation and accurate adaptation to the underlying architecture, this implementation scheme divides the dataset into three types of features: discrete features, continuous features, and spatial coordinate features. To prevent information leakage during model training, the complete dataset is pre-split into independent training, validation, and test sets in a 7:1:2 ratio, and all subsequent statistical benchmarks for scaling are only allowed to be extracted from the training set. Discrete features are converted into integer indices using label encoding; continuous features utilize the mean of the training set. with standard deviation Z-score standardization To address the spatial coordinate characteristics and avoid phase aliasing caused by sine and cosine transforms, as well as to meet the periodic input requirements of Fourier position coding, this application employs a limited mapping normalization. It is strictly constrained to the core interval [-1,1].

[0028] Step 5) Dynamic Feature Selection. The processed slope cell dataset is input into SGDANet, which consists of a "hazardous attribute stream" and a "spatial information stream." The core of the SGDANet algorithm mainly consists of a spatial high-frequency mapping module, a dynamic local feature selection module, and a spatially guided residual block. Dynamic feature selection is performed by the spatial high-frequency mapping module and the dynamic local feature selection module, and its specific processing is as follows: Step 5.1) Spatial High-Frequency Mapping Module. This study introduces Fourier Position Encoding to project low-dimensional, low-frequency spatial features onto high-dimensional, high-frequency spatial features. The process is as follows: Input the center geographic coordinate matrix of the current batch of slope units. The coordinate matrix is ​​expanded in its last dimension and then input into the Fourier high-frequency position encoder. The encoder has a preset number of frequency bands. for (default (8), calculated according to formula (2) to The frequency base matrix grows exponentially : (2) The spatial features of the input are expanded using sine and cosine periodic transformations based on this frequency basis, while preserving the original low-dimensional input to maintain global position awareness. Finally, a high-dimensional spatial feature matrix is ​​output by fully connected concatenation and flattening along the channel dimension. Because it includes the original coordinate terms and the sine and cosine transform terms corresponding to each coordinate, the output spatial feature dimension... Strictly satisfy formula (3): (3) Step 5.2) Based on hypernetwork and A Spatially Guided Locality (SDR) module for sparse activation functions. In each forward propagation layer of the catastrophic attribute feature extraction flow, this study constructs a spatially guided locality module, utilizing a hypernetwork to dynamically generate sample-specific feature masks based on spatial location. The data flow and computation logic are as follows: a. Hypernetwork weight generation: The module receives the high-dimensional spatial feature matrix output from step 5.1. Nonlinear mapping is achieved through a supernetwork consisting of explicit fully connected layers. This supernetwork contains two linear transformation layers: the first fully connected layer transforms the dimension from... The feature dimension is mapped to 64 dimensions and in-situ nonlinear activation is performed by introducing a rectified linear unit (ReLU); the second fully connected layer reprojects the feature dimension from 64 dimensions to 64 dimensions. Dimension, among which The number of independent attention sub-basis sets constructed for extracting local information (in this embodiment, it is reflected as the number of independent feature extraction branches). The input slope unit's hazard-causing attribute feature dimension. The output continuous tensor is reshaped into a three-dimensional dynamic original weight tensor. .

[0029] b. Absolutely Sparse Mask Calculation: Traditional Softmax activation functions, when calculating attention weights, can only infinitely approximate the weights of irrelevant factors to zero, but cannot truly reduce them to zero. This leads to continuous interference from environmental noise caused by non-slippage factors in classification decisions. Therefore, this application abandons Softmax and innovatively adopts... Activation function acts on The last dimension (disaster-causing attribute feature dimension) The constraint is imposed by introducing a 1.5-order Tsallis entropy, as shown in equation (4): (4) In the formula: express A probability simplex space of dimension 1. The entropy is Tsallis-1.5. Because... It possesses powerful geometric projection and truncation characteristics, enabling it to set the weights of non-dominant factors that are unrelated to the current spatial location and do not satisfy the local landslide disaster mechanism. This generates an absolutely sparse spatially guided local feature selection mask matrix for each sample. .

[0030] c. Physical factor masking interception: This intercepts the input slope unit disaster-causing attribute feature vector. In the second dimension, dimensional expansion is performed, transforming it into... Using the tensor automatic broadcasting mechanism, the sparse mask is multiplied element-wise with the expanded feature matrix, as shown in formula (5): (5) Where: the output mask catastrophic attribute feature tensor The dimension is This step achieves physical interception at the mathematical level: under the dynamic scheduling of spatial signals, irrelevant geological background noise is directly filtered to zero, while key disaster-causing factors are significantly amplified.

[0031] Step 6) Active landslide identification. Features filtered by the dynamic local feature selection module need further extraction of higher-order nonlinear semantics, and it is also necessary to prevent the original physical spatial constraints from being forgotten during deep network iterations. Therefore, this application designs a Spatially Guided Basic Block containing a Spatially Guided Abstract Layer. Its specific feature aggregation and transfer mechanism is as follows: a. Spatial guided abstraction layer includes grouped convolution and gated aggregation: After flattening, the input is a one-dimensional grouped convolutional layer (Conv1d, number of groups is...). This approach expands the feature dimensions while maintaining the independence of each sub-base set. After mitigating the degradation of model generalization ability caused by large-batch training through a virtual mini-batch normalization layer (Ghost BatchNorm), a gated linear unit (GLU) is introduced for non-linear selection filtering and aggregation, outputting an abstract feature vector. .

[0032] b. Spatial Guided Residual Blocks: A backbone path is constructed containing two concatenated abstraction layers, and an explicit downsampling shortcut branch is designed. The shortcut branch introduces Dropout perturbation to the original features and then re-incorporates spatial features. Perform independent abstraction. Add the residuals of the main path and shortcut branches, and activate them with a leaky modified linear unit (LeakyReLU) to continuously strengthen the spatial physical constraints in the deep stack.

[0033] Finally, the fused features from the deep output of multiple spatially guided residual blocks are fed into the Dropout layer and then through a two-layer high-dimensional fully connected classification head. Combined with the Sigmoid activation function, the value of each slope unit (0 or 1) is output, thereby realizing the automatic identification of active landslides in a wide-area urban area.

[0034] The effect of implementing this application is as follows: 1. This application uses Sentinel-1 SAR data and SBAS-InSAR technology to extract the annual average deformation rate of the ascending and descending orbits. The LOS deformation rate of the ascending and descending orbits is decomposed into vertical and east-west directions in two dimensions. The annual average deformation rates of the ascending and descending orbits, as well as the vertical and east-west directions, are shown below. Figure 2 As shown.

[0035] 2. This application conducts automatic identification of active landslides in Hualong County, Qinghai Province. The identified active landslides are active deformation zones within the boundaries of potential landslide hazards. The SGDANet algorithm proposed in this application significantly outperforms existing mainstream computer vision and machine learning algorithms (U-Net, CNN, SVM, and CART) in identifying active landslides in a wide-area, complex environment. In the test set evaluation, the model of this application achieved an F1-score of 0.8743, an accuracy of 99.52%, a precision of 89.89%, an AUC of 0.9971, and a Kappa coefficient of 0.8719. In particular, in terms of recall, which represents the ability to identify disasters, SGDANet achieves significant improvements of 10.35%, 10.84%, 25.92%, and 26.21% compared to U-Net, CNN, SVM, and CART, respectively. Meanwhile, this application effectively overcomes the interference of background noise in complex environments, achieving the lowest levels of false alarms and missed detections, with a false alarm rate of only 0.10% and a missed detection rate of only 0.22%. Compared with the comparison algorithm, the false alarm rate is reduced by 0.12% to 0.27%, and the missed detection rate is reduced by 0.29% to 0.69%, demonstrating good performance. The accuracy evaluation is shown in Table 2, and the false alarm and missed detection results are shown in Table 3.

[0036] Table 2. Accuracy comparison of different algorithms on the test set

[0037] Table 3. False alarm and false negative rates of different algorithms on the test set.

[0038] 3. The results of active landslide identification in Hualong County, Qinghai Province are as follows: Figure 3 As shown, analysis reveals that SGDANet produces fewer false alarms and missed detections of slope units compared to U-Net, CNN, SVM, and CART models. In particular, SGDANet significantly outperforms other models in terms of missed detections. Therefore, the SGDANet algorithm demonstrates better overall performance in active landslide identification.

[0039] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. An automatic identification method for active landslides based on spatially guided deep abstraction networks, characterized in that, Includes the following steps: S10 provides a spatially guided deep abstraction network and obtains a slope unit dataset, wherein each slope unit in the slope unit dataset contains disaster-causing attribute features and spatial coordinate features; the spatially guided deep abstraction network includes a spatial high-frequency mapping module, a dynamic local feature selection module, a spatially guided residual block, and a classification head; S20, The spatial high-frequency mapping module is used to perform Fourier position encoding on the spatial coordinate features, mapping the low-dimensional spatial coordinates to the high-dimensional frequency domain space to obtain a high-dimensional spatial feature matrix. S30, the high-dimensional spatial feature matrix is ​​input into the hypernetwork using the dynamic local feature selection module to generate a dynamic sparse mask, and the disaster-causing attribute features are fused with the dynamic sparse mask element-wise to obtain the mask disaster-causing attribute features. S40, the spatially guided residual block is used to extract features from the catastrophic attribute features of the mask, and the high-dimensional spatial feature matrix is ​​introduced into the feature extraction process for spatial constraints to obtain fused features; S50, input the fused features into the classification head, and output the active landslide identification result for each slope unit.

2. The automatic identification method for active landslides based on spatially guided deep abstraction networks as described in claim 1, characterized in that, The specific process of the Fourier position encoding is as follows: The spatial coordinate features are expanded in the last dimension, with a preset number of frequency bands. According to the formula Calculate the frequency basis matrix ; The spatial coordinate features are subjected to sine and cosine transforms using the frequency basis matrix, preserving the original low-dimensional input. The transform results are then concatenated with the spatial coordinate features along the channel dimension and flattened to obtain a high-dimensional spatial feature matrix. The dimension of the high-dimensional spatial feature matrix Satisfy the formula .

3. The automatic identification method for active landslides based on spatially guided deep abstraction networks as described in claim 1, characterized in that, The hypernetwork comprises two linear transformation layers. The first fully connected layer maps the input dimension to 64 dimensions and performs non-linear activation through modified linear units. The second fully connected layer maps the feature dimension from 64 dimensions to... The dimension reshapes the output continuous tensor into a three-dimensional dynamic primitive weight tensor. ,in K The number of independent attention sub-basis sets. D The dimension of the disaster-causing attribute features is the input. The dynamic sparse mask is generated by: using Activation function on the three-dimensional dynamic original weight tensor The last dimension is sparsified, and the sparse activation function is constrained based on the 1.5-order Tsallis entropy, and its mathematical expression is: ,in This indicates retrieving the maximum value. express A probability simplex space of dimension 1. Tsallis-1.5 entropy is the core physical mechanism of dynamic sparsity. This represents the probability distribution vector assigned to each feature factor, i.e., all probability allocation schemes. Its transpose represents the relative importance of each local disaster-causing factor at its current spatial location. The unnormalized feature scores output by the network's front-end layer represent preliminary "evaluations" of the importance of each feature factor at the current specific spatial location. The activation function has geometric projection and truncation properties, which are used to set the weights of non-dominant factors that are irrelevant to the current spatial location and do not satisfy the local landslide disaster mechanism to 0, thereby generating an absolutely sparse spatially guided local feature selection mask matrix for each sample. .

4. The method for automatic identification of active landslides based on spatially guided deep abstraction networks as described in claim 1, characterized in that, The specific process of the element-wise weighted fusion is as follows: the disaster-causing attribute feature vector is expanded in the second dimension, and the dynamic sparse mask is multiplied element-wise with the expanded disaster-causing attribute feature tensor using the tensor automatic broadcasting mechanism.

5. The method for automatic identification of active landslides based on spatially guided deep abstraction networks as described in claim 1, characterized in that, The spatially guided residual block comprises two cascaded spatially guided abstraction layers and an explicit downsampling shortcut branch. The spatially guided abstraction layer includes a one-dimensional grouped convolutional layer, a virtual mini-batch normalization layer, and a gated linear unit. The gated linear unit is used to perform non-linear selection filtering and aggregation on the features after grouped convolution and virtual mini-batch normalization.

6. The automatic identification method for active landslides based on spatially guided deep abstraction networks as described in claim 5, characterized in that, The downsampling shortcut branch introduces Dropout perturbation into the input features and integrates them into the high-dimensional spatial feature matrix for independent abstraction. The output of the spatial guided abstraction layer is added to the output of the downsampling shortcut branch, and the residuals are then activated by a leaky modified linear unit to obtain the fused features.

7. The method for automatic identification of active landslides based on spatially guided deep abstraction networks as described in claim 1, characterized in that, The disaster-causing attribute features include statistical values ​​of stratigraphic lithology, distance from fault, elevation, slope, aspect, planar curvature, profile curvature, topographic relief, distance from river, normalized vegetation index, annual average rainfall, distance from road, land use type, and east-west trending annual average surface deformation rate. The discrete features in the slope unit dataset are converted into integer indices through label encoding. The continuous features in the slope unit dataset are standardized using Z-scores based on the mean and standard deviation of the training set. The spatial coordinate features are constrained to the range of -1 to +1 through mapping normalization.

8. The automatic identification method for active landslides based on spatially guided deep abstraction networks as described in claim 1, characterized in that, The classification head includes a Dropout layer and a two-layer fully connected classification head. The fused features are processed by the Dropout layer and then input into the two-layer fully connected classification head. The active landslide identification result of each slope unit is output by combining the Sigmoid activation function.

9. The method for automatic identification of active landslides based on spatially guided deep abstraction networks as described in claim 1, characterized in that, The specific method for obtaining the slope unit dataset is as follows: using digital elevation model data, slope units are divided using the surface curvature watershed method. Through zonal statistics, the statistical values ​​of stratigraphic lithology, distance from fault, elevation, slope, aspect, plane curvature, profile curvature, topographic relief, distance from river, normalized vegetation index, annual average rainfall, distance from road, land use type, and east-west annual average surface deformation rate are summarized into the slope unit vector data to generate the slope unit dataset.

10. An automatic identification system for active landslides based on spatially guided deep abstraction networks, characterized in that, To implement the method of any one of claims 1 to 9, comprising: The data acquisition module is used to acquire a slope unit dataset, wherein each slope unit in the slope unit dataset contains disaster-causing attribute features and spatial coordinate features; The spatial high-frequency mapping module is used to perform Fourier position encoding on the spatial coordinate features, mapping the low-dimensional spatial coordinates to the high-dimensional frequency domain space to obtain a high-dimensional spatial feature matrix. The dynamic local feature selection module is used to input the high-dimensional spatial feature matrix into the hypernetwork to generate a dynamic sparse mask, and to perform element-wise weighted fusion of the catastrophic attribute features and the dynamic sparse mask to output the catastrophic attribute features of the mask. A spatially guided residual block is used to extract features from the catastrophic attribute features of the mask, and the high-dimensional spatial feature matrix is ​​introduced into the feature extraction process for spatial constraints, and the fused features are output. A classification head is used to classify the fused features and output the active landslide identification result for each slope unit.

Images