Landslide identification method and device based on multi-dimensional feature aggregation and computer equipment

Abstract

The application relates to a landslide identification method and device based on multi-dimensional feature aggregation and computer equipment. The method comprises the following steps: acquiring remote sensing image data and basic geographic data of a target area; based on the remote sensing image data and the basic geographic data, spectral, terrain and geological multi-dimensional features reflecting landslide development environment are extracted and fused to construct a multi-dimensional feature cube; a landslide identification model is constructed, the landslide identification model comprises a backbone network of an encoder and a decoder, wherein the encoder is used to realize multiplexing and fusion of cross-level features in the multi-dimensional feature cube; sample data containing landslide labeling information is used to train and optimize the landslide identification model; and according to the multi-dimensional feature cube, the landslide identification model after training and optimization is used to output a landslide spatial distribution identification result of the target area. The method can improve the precision and efficiency of automatic landslide identification and is easy to deploy in practice.

Description

Technical Field

[0001] This application relates to the field of landslide identification technology, and in particular to a landslide identification method, apparatus, computer equipment, computer-readable storage medium, and computer program product based on multidimensional feature aggregation. Background Technology

[0002] Landslides, a typical geological hazard formed by the sliding of slope rock and soil along weak surfaces, require accurate identification and mapping for disaster prevention and mitigation. Early methods relied primarily on professional field surveys and manual remote sensing interpretation. While reliable, these methods were inefficient, costly, and subjective, making them unsuitable for large-scale rapid surveys and emergency responses. With technological advancements, automated identification methods based on machine learning have been applied. However, these methods rely solely on pixel spectral information, lacking utilization of spatial features such as landslide morphology and texture. This leads to "salt-and-pepper noise" in the results and difficulty in distinguishing landslides from high-reflectivity features like bare land. Subsequent object-oriented methods, while considering spatial features, heavily depend on image segmentation quality, and the lack of a universally applicable segmentation scale results in unstable performance in complex, large-scale scenarios.

[0003] In recent years, deep learning technology, especially convolutional neural networks (CNNs), has opened up new avenues for landslide identification by automatically extracting deep features through end-to-end learning. However, existing models, in pursuit of accuracy, often continuously deepen the network and increase the number of parameters, neglecting the optimization of feature transfer efficiency and computational cost. This leads to two prominent problems: first, high model complexity and parameter redundancy result in enormous consumption of training and inference computational resources, posing difficulties in practical deployment; second, the lack of efficient feature reuse mechanisms makes gradient vanishing or information loss prone to occur when the network is deepened, limiting further improvement in feature representation capabilities. Therefore, the current core challenge lies in how to improve the practicality and deployment feasibility of landslide identification models by constructing a lightweight network architecture and an efficient feature hierarchical interaction mechanism, while ensuring recognition accuracy, thereby achieving full reuse of cross-layer features and optimization of model parameters.

[0004] Therefore, there is an urgent need for a landslide identification method, device, computer equipment, computer-readable storage medium, and computer program product based on multi-dimensional feature aggregation, which can improve the accuracy and efficiency of automated landslide identification and is easy to deploy in practice. Summary of the Invention

[0005] Therefore, it is necessary to provide a landslide identification method, device, computer equipment, computer-readable storage medium, and computer program product based on multi-dimensional feature aggregation that can improve the accuracy and efficiency of automated landslide identification and is easy to deploy in practice, in order to address the above-mentioned technical problems.

[0006] Firstly, this application provides a landslide identification method based on multidimensional feature aggregation, including:

[0007] Acquire remote sensing image data and basic geographic data of the target area. Based on the remote sensing image data and the basic geographic data, extract and fuse spectral, topographic and geological multi-dimensional features that reflect the landslide development environment to construct a multi-dimensional feature cube.

[0008] A landslide identification model is constructed, which includes a backbone network of encoder and decoder, wherein the encoder is used to realize the reuse and fusion of cross-level features in the multi-dimensional feature cube;

[0009] The landslide identification model is trained and optimized using sample data containing landslide annotation information.

[0010] Based on the multidimensional feature cube, the landslide recognition model, trained and optimized, outputs the spatial distribution recognition results of landslides in the target area.

[0011] In one embodiment, the process of constructing the multidimensional feature cube includes:

[0012] The remote sensing image data is subjected to radiometric calibration, atmospheric correction, and spatial registration.

[0013] The calculation includes spectral index characteristics of the normalized vegetation index and the modified normalized water index;

[0014] Based on digital elevation model data, topographic features including elevation, slope, aspect, plane curvature, and distance from the hydrological network are extracted.

[0015] Based on geological map data, geological environmental characteristics including lithology and distance from tectonic faults are extracted;

[0016] The spectral index features, topographic features, and geological environment features are layered and channels are merged to generate multidimensional feature cube data.

[0017] In one embodiment, before layering and merging the spectral index features, topographic features, and geological environment features, the method further includes:

[0018] Calculate the Pearson correlation coefficient matrix between the spectral index features, the topographic features, and the geological environment features, and remove redundant features according to a preset correlation threshold;

[0019] A variance inflation factor analysis is performed on the feature set after removing redundant features, and features are removed again from the feature set based on the variance inflation factor.

[0020] In one embodiment, the encoder includes at least one densely connected block, wherein the input of each convolutional layer in the densely connected block is the concatenation of the output feature maps of all preceding convolutional layers in the channel dimension;

[0021] The encoder also includes a downsampling transition layer disposed between the densely connected blocks for downsampling in the spatial dimension and compressing the number of feature channels;

[0022] The decoder section includes an upsampling transition layer corresponding to the downsampling transition layer, which is used to restore the spatial resolution of the feature map and fuse feature information from the corresponding level of the encoder through skip connection paths.

[0023] In one embodiment, training and optimizing the landslide identification model includes:

[0024] The landslide identification model is trained iteratively using an end-to-end approach with forward and backward propagation.

[0025] During model training, random dropout layers are performed between the densely connected blocks to prevent overfitting;

[0026] After the model training is completed, the model weight parameters with the best recognition performance are selected based on the performance metrics of the independent validation set or test set to generate the fully trained landslide recognition model.

[0027] In one embodiment, the method further includes:

[0028] The regions corresponding to independent test samples in the landslide spatial distribution identification results are compared pixel by pixel with the real landslide distribution data verified in the field to construct a confusion matrix.

[0029] Based on the confusion matrix, multiple quantitative evaluation indicators, including overall precision, accuracy, recall, F1 score, Kappa coefficient, and average intersection-union ratio, are calculated to assess the recognition performance and reliability of the landslide identification model.

[0030] Secondly, this application also provides a landslide identification device based on multidimensional feature aggregation, comprising:

[0031] The acquisition module is used to acquire remote sensing image data and basic geographic data of the target area. Based on the remote sensing image data and the basic geographic data, it extracts and fuses spectral, topographic and geological multi-dimensional features that reflect the landslide development environment to construct a multi-dimensional feature cube.

[0032] A construction module is used to construct a landslide recognition model, which includes a backbone network of encoder and decoder. The encoder is used to realize the reuse and fusion of cross-level features in the multi-dimensional feature cube.

[0033] The training and optimization module is used to train and optimize the landslide identification model using sample data containing landslide annotation information.

[0034] The identification module is used to output the spatial distribution identification results of landslides in the target area based on the multi-dimensional feature cube and the trained and optimized landslide identification model.

[0035] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0036] Acquire remote sensing image data and basic geographic data of the target area. Based on the remote sensing image data and the basic geographic data, extract and fuse spectral, topographic and geological multi-dimensional features that reflect the landslide development environment to construct a multi-dimensional feature cube.

[0037] A landslide identification model is constructed, which includes a backbone network of encoder and decoder, wherein the encoder is used to realize the reuse and fusion of cross-level features in the multi-dimensional feature cube;

[0038] The landslide identification model is trained and optimized using sample data containing landslide annotation information.

[0039] Based on the multidimensional feature cube, the landslide recognition model, trained and optimized, outputs the spatial distribution recognition results of landslides in the target area.

[0040] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0041] Acquire remote sensing image data and basic geographic data of the target area. Based on the remote sensing image data and the basic geographic data, extract and fuse spectral, topographic and geological multi-dimensional features that reflect the landslide development environment to construct a multi-dimensional feature cube.

[0042] A landslide identification model is constructed, which includes a backbone network of encoder and decoder, wherein the encoder is used to realize the reuse and fusion of cross-level features in the multi-dimensional feature cube;

[0043] The landslide identification model is trained and optimized using sample data containing landslide annotation information.

[0044] Based on the multidimensional feature cube, the landslide recognition model, trained and optimized, outputs the spatial distribution recognition results of landslides in the target area.

[0045] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0046] Acquire remote sensing image data and basic geographic data of the target area. Based on the remote sensing image data and the basic geographic data, extract and fuse spectral, topographic and geological multi-dimensional features that reflect the landslide development environment to construct a multi-dimensional feature cube.

[0047] A landslide identification model is constructed, which includes a backbone network of encoder and decoder, wherein the encoder is used to realize the reuse and fusion of cross-level features in the multi-dimensional feature cube;

[0048] The landslide identification model is trained and optimized using sample data containing landslide annotation information.

[0049] Based on the multidimensional feature cube, the landslide recognition model, trained and optimized, outputs the spatial distribution recognition results of landslides in the target area.

[0050] The aforementioned landslide identification method, apparatus, computer equipment, computer-readable storage medium, and computer program product based on multi-dimensional feature aggregation construct an information-rich multi-dimensional feature cube by fusing spectral, topographic, and geological features from remote sensing imagery and geographic environmental data. This provides the model with a comprehensive and three-dimensional representation of the landslide development environment, overcoming the shortcomings of insufficient information from a single data source or single type of feature. The constructed landslide identification model integrates a feature extraction module based on a dense connection mechanism in its encoder, enabling feature reuse and short-circuit connections between layers in the network. This ensures effective gradient propagation in deep networks, alleviates the gradient vanishing problem, and significantly improves feature utilization efficiency and model training stability. The end-to-end training method allows the model to directly learn from the original multi-dimensional feature data and output landslide identification results, avoiding complex manual feature engineering and error accumulation from multi-stage processing, thus improving the overall efficiency and practicality of landslide identification. In summary, combining multi-dimensional feature fusion with a densely connected network architecture forms a complete intelligent landslide identification solution. This solution not only fully utilizes the complementary information provided by multi-source data but also ensures that the model maintains strong feature extraction capabilities while possessing good training characteristics and parameter efficiency through innovative network structure design. Ultimately, it can achieve high-precision and high-efficiency automatic landslide identification and mapping in complex environments, meeting the practical application needs of large-scale and rapid emergency response. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is an application environment diagram of a landslide identification method based on multidimensional feature aggregation in one embodiment;

[0053] Figure 2 This is a flowchart illustrating a landslide identification method based on multidimensional feature aggregation in one embodiment;

[0054] Figure 3 This is a flowchart illustrating a landslide identification method based on multidimensional feature aggregation in another embodiment;

[0055] Figure 4 This is a schematic diagram illustrating the fabrication of a landslide sample in one embodiment;

[0056] Figure 5 This is a framework diagram of a multidimensional feature aggregation model for a landslide identification method in one embodiment;

[0057] Figure 6 This is a flowchart illustrating the overall process of a landslide identification method based on multidimensional feature aggregation in one embodiment.

[0058] Figure 7 This is a structural block diagram of a landslide identification device based on multidimensional feature aggregation in one embodiment;

[0059] Figure 8 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0060] 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.

[0061] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0062] The landslide identification method based on multidimensional feature aggregation provided in this application can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104, or it can be located in the cloud or on another network server.

[0063] Server 104 acquires remote sensing image data and basic geographic data of the target area through terminal 102. Based on the remote sensing image data and basic geographic data, server 104 extracts and fuses spectral, topographic, and geological multi-dimensional features reflecting the landslide development environment to construct a multi-dimensional feature cube. It then constructs a landslide identification model, which includes a backbone network of encoder and decoder. The encoder is used to realize the reuse and fusion of cross-level features in the multi-dimensional feature cube. The landslide identification model is trained and optimized using sample data containing landslide annotation information. Based on the multi-dimensional feature cube, the trained and optimized landslide identification model outputs the spatial distribution identification results of landslides in the target area.

[0064] The terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle systems, and projection devices. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted displays. Head-mounted displays can be virtual reality (VR) devices, augmented reality (AR) devices, and smart glasses. The server 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0065] In one exemplary embodiment, such as Figure 2 As shown, a landslide identification method based on multidimensional feature aggregation is provided, and this method is applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps S202 to S208. Wherein:

[0066] Step S202: Obtain remote sensing image data and basic geographic data of the target area. Based on the remote sensing image data and basic geographic data, extract and fuse multi-dimensional spectral, topographic and geological features that reflect the landslide development environment to construct a multi-dimensional feature cube.

[0067] Specifically, remote sensing imagery data mainly refers to optical satellite imagery (such as Landsat 8OLI), used to provide spectral reflectance information of the Earth's surface; basic geographic data includes digital elevation models (DEMs, such as ASTERGDEM), geological maps, topographic maps, etc., used to provide background environmental information such as elevation, lithology, and geological structure. Based on the above data, landslide-causing factors of different dimensions are extracted respectively:

[0068] Remote sensing indices such as the Normalized Difference Vegetation Index (NDVI) and the Modified Normalized Difference Water Index (MNDWI) are calculated to indirectly reflect state information closely related to landslide activity, such as surface vegetation cover and soil moisture. Based on the Digital Earth Model (DEM), spatial analysis using a Geographic Information System (GIS) extracts factors such as elevation, slope, aspect, curvature, and distance from the river. These factors directly control slope stability, surface hydrological pathways, and stress distribution. Based on geological maps, factors such as lithology and distance from faults are extracted. Lithology determines the mechanical properties of the soil and rock mass, while faults are often geologically weak zones and tectonic activity areas; together, they constitute the geological basis for landslide occurrence.

[0069] All extracted single-factor layers (each layer can be viewed as a two-dimensional matrix) undergo rigorous spatial registration to ensure that each geographic coordinate point is aligned across different layers. Subsequently, these layers are overlaid along the channel dimension to form a three-dimensional data structure, namely a multidimensional feature cube. For example... Figure 3 As shown, the two spatial dimensions (rows and columns) of this cube correspond to the pixel grid of the geographic region, while each slice of the third dimension (channel) corresponds to a specific disaster-causing factor (such as NDVI, slope, lithology coding, etc.). For example, if 10 factors are extracted, the constructed cube is a three-dimensional array of height × width × 10.

[0070] Step S204: Construct a landslide recognition model. The landslide recognition model includes a backbone network of encoder and decoder. The encoder is used to realize the reuse and fusion of cross-level features in the multi-dimensional feature cube.

[0071] Specifically, an MFAN (Multi-dimensional Feature Aggregation Network) architecture is constructed, which mainly consists of Dense Block (DB), Transition Down (TD), and Transition Up (TU). The DB module adopts a dense connection mechanism, the network downsampling path is implemented through the TD module, and the network upsampling path is implemented through the TU module.

[0072] The encoder consists of multiple stages. Within each stage, the input first passes through a densely connected block. The input to each convolutional operation is not only the output of the previous layer, but also a concatenation of the output feature maps of all preceding layers within that block along the channel dimension. This enables direct reuse of features from shallow to deep layers. For example, edge information from the first layer and texture information from the third layer can be directly used as part of the input to the fifth convolution. This greatly enhances gradient flow, alleviates the gradient vanishing problem in deep networks, and allows the network to fully utilize features of different granularities extracted from each layer.

[0073] When restoring spatial details, the decoder can not only utilize the most abstract and semantically strongest features from the end of the encoder, but also fuse shallow features from the early stages of the encoder that have higher spatial resolution and contain more geometric and textural details. This ensures the accuracy of the final landslide boundary recognition and avoids the loss of spatial information caused by multiple downsampling.

[0074] Step S206: Use sample data containing landslide annotation information to train and optimize the landslide identification model.

[0075] Specifically, the sample data consists of three parts: a multidimensional feature cube as input, a binary landslide label map of the corresponding region as a supervision signal (typically, landslide areas are labeled as 1 and non-landslide areas as 0), and a test set for independent evaluation. The multidimensional feature cube is input into the model, where features are extracted and fused sequentially through densely connected blocks of the encoder, then the resolution is restored by the decoder, and finally, the output layer generates a prediction probability map of the same size as the input. The predicted probability map output by the model is compared pixel-by-pixel with the real landslide label map, and a scalar loss is calculated using a pre-defined loss function (e.g., the cross-entropy loss function for binary classification tasks).

[0076] Using the backpropagation algorithm, the calculated scalar loss is propagated back layer by layer from the model's output layer to the input layer, calculating the gradient of the loss function with respect to each trainable parameter of the model. These gradients indicate the direction and magnitude in which each parameter should be adjusted to reduce the loss. Subsequently, an optimization algorithm (such as Adam, a variant of stochastic gradient descent) is used to fine-tune all the weights and biases of the model based on this gradient information.

[0077] Step S208: Based on the multi-dimensional feature cube, the landslide recognition model, trained and optimized, is used to output the spatial distribution recognition result of landslides in the target area.

[0078] Specifically, for the region to be identified, the process strictly follows the same procedure as the training phase: acquiring multi-source data and constructing corresponding multi-dimensional feature cubes. These multi-dimensional feature cubes are then used as input, loaded with the optimal model weights obtained during previous training. The model operates in a forward propagation manner: data flows through an encoder-decoder network. Densely connected blocks in the encoder reuse and deeply abstract the features, while the decoder combines skip connections to recover spatial details and upsample. Finally, the model's output layer (typically a 1×1 convolution followed by a sigmoid activation function) calculates an independent probability value for each spatial pixel in the input cube, representing the confidence level that the pixel belongs to the landslide category. The landslide probability values ​​of all pixels collectively form a landslide spatial distribution probability map that perfectly matches the spatial range of the input region. To obtain the final binarized identification result (i.e., clearly distinguishing between landslide and non-landslide areas), a threshold (e.g., 0.5) can be set for this probability map. Pixels with probability values ​​higher than the threshold are classified as "landslides," while those with values ​​lower are classified as "non-landslides," thus generating an intuitive thematic map of landslide identification results.

[0079] The aforementioned landslide identification method based on multidimensional feature aggregation integrates spectral, topographic, and geological features from remote sensing images and geographic environmental data to construct an information-rich multidimensional feature cube. This provides the model with a comprehensive and three-dimensional representation of the landslide development environment, overcoming the limitations of insufficient information from a single data source or single type of feature. The constructed landslide identification model integrates a feature extraction module based on a dense connection mechanism in its encoder, enabling feature reuse and short-circuit connections between layers in the network. This ensures effective gradient propagation in deep networks, alleviates gradient vanishing problems, and significantly improves feature utilization efficiency and model training stability. The end-to-end training approach allows the model to directly learn from the original multidimensional feature data and output landslide identification results, avoiding complex manual feature engineering and error accumulation from multi-stage processing, thus improving the overall efficiency and practicality of landslide identification. In summary, combining multi-dimensional feature fusion with a densely connected network architecture forms a complete intelligent landslide identification solution. This solution not only fully utilizes the complementary information provided by multi-source data but also ensures that the model maintains strong feature extraction capabilities while possessing good training characteristics and parameter efficiency through innovative network structure design. Ultimately, it can achieve high-precision and high-efficiency automatic landslide identification and mapping in complex environments, meeting the practical application needs of large-scale and rapid emergency response.

[0080] In one exemplary embodiment, such as Figure 3 As shown, the process of constructing a multidimensional feature cube includes:

[0081] Step S302: Perform radiometric calibration, atmospheric correction, and spatial registration on the remote sensing image data;

[0082] Step S304: Calculate the spectral index characteristics including the normalized vegetation index and the corrected normalized water index;

[0083] Step S306: Based on the digital elevation model data, extract topographic features including elevation, slope, aspect, plane curvature, and distance from the hydrological network;

[0084] Step S308: Based on geological map data, extract geological environmental characteristics including lithology and distance from tectonic faults;

[0085] Step S310 involves overlaying and merging the spectral index features, topographic features, and geological environment features into layers to generate multidimensional feature cube data.

[0086] Specifically, such as Figure 4 As shown, spectral index factors, such as the Normalized Difference Vegetation Index (NDVI) and the Modified Normalized Water Index (NDWI), are calculated based on each band of remote sensing imagery. Topographic and hydrological factors, such as elevation, slope, aspect, plane curvature, and distance from the hydrological network, are extracted based on the Global Digital Elevation Model (DEM). Geological environmental characteristics, including lithological type and distance from tectonic faults, are extracted based on geological map data.

[0087] After completing spatial registration and normalization (such as Min-Max standardization), all the extracted single-factor layers (each layer is a two-dimensional raster matrix, representing a feature channel) are overlaid and merged. This is equivalent to treating each layer as a channel and stacking all layers on the third dimension (channel dimension) to generate a three-dimensional data structure, namely a multi-dimensional feature cube.

[0088] In this embodiment, by constructing a multidimensional feature cube, each spatial pixel no longer contains only a few spectral band values, but is represented by a high-dimensional feature vector that integrates multiple attributes such as spectrum, topography, and geology. This provides a comprehensive and structured input for subsequent deep learning models to learn the complex nonlinear relationship between landslides and complex environments, and is a key data foundation for achieving high-precision and automated landslide identification.

[0089] In an exemplary embodiment, before layering and merging spectral index features, topographic features, and geological environment features, the method further includes:

[0090] Calculate the Pearson correlation coefficient matrix between spectral index features, topographic features, and geological environment features, and remove redundant features based on a preset correlation threshold;

[0091] Variance inflation factor analysis is performed on the feature set after removing redundant features. Based on the variance inflation factor, features are removed from the feature set again.

[0092] Specifically, the extracted landslide-causing factors were resampled and standardized. The Pearson Correlation Coefficient (PCC) was calculated using the Python platform. PCC measures the linear correlation between two variables X and Y, with values ​​ranging from -1 to 1. Intuitively, this linear correlation indicates whether Y increases or decreases simultaneously as X increases. A positive PCC value indicates a positive correlation, and vice versa. When the two variables are distributed along a straight line, the PCC value is 1 or -1; when there is no linear relationship between the two variables, the PCC value is 0; values ​​from 0 to 1 indicate increasingly stronger correlations. If the correlation index between factors exceeds 0.7, it indicates a strong correlation between factors, requiring selective removal.

[0093] ;

[0094] in, Indicates the number of samples; , They represent The standard deviation and mean of the sample; , They represent The standard deviation and mean of the sample are given; cov(X,Y) represents... and The covariance results.

[0095] Further multicollinearity diagnosis was performed on the extracted landslide causative factors, specifically using the variance inflation factor (VIF) and tolerance (TOL) indices for quantitative evaluation. The VIF value of each factor was calculated using SPSS software; when VIF > 10, severe multicollinearity was identified, and the corresponding factor was removed. Simultaneously, combined with Pearson correlation analysis results, a set of highly independent causative factors was comprehensively selected, ultimately constructing a landslide sample library with a clear structure and low data redundancy, providing a high-quality data foundation for subsequent model training.

[0096] The preprocessed landslide-causing factors are further input into the ENVI 5.6 platform, and its layer overlay tool is used to construct h×w×c dimensional multidimensional cube data, where h is the image height, w is the image width, and c is the total number of channels after overlay. Specifically, this refers to the number of landslide-causing factors used in this invention, as detailed below. Figure 4As shown. The multidimensional cube data is then divided into multiple image patches of the same size, and these patches are further divided into training and testing sets according to a preset 7:3 ratio. The training set data is primarily used to train the network, while the testing set data is used to evaluate the landslide recognition accuracy of the optimal model.

[0097] In this embodiment, the statistical screening process described above, involving two steps, directly reduces the complexity of the model by decreasing the number of input features, thus accelerating the training process. Simultaneously, it eliminates multicollinearity, leading to more stable model parameter estimation and reducing the risk of overfitting. Furthermore, it removes internal noise and redundant information from the subsequently constructed multidimensional feature cube, ensuring the quality of the data input to the deep learning model.

[0098] In one exemplary embodiment, the encoder includes at least one densely connected block, wherein the input of each convolutional layer in the densely connected block is the concatenation result of the output feature maps of all preceding convolutional layers in the channel dimension;

[0099] The encoder also includes downsampling transition layers positioned between densely connected blocks to downsample and compress the number of feature channels in the spatial dimension;

[0100] The decoder section includes an upsampling transition layer corresponding to the downsampling transition layer, which is used to restore the spatial resolution of the feature map and fuse feature information from the corresponding level of the encoder through skip connection paths.

[0101] Specifically, landslide data from 70% of the training set are input into the model in batches for training to obtain the optimal landslide identification model, such as... Figure 5 As shown, the specific training process is as follows:

[0102] The preprocessed multidimensional cubic data is input into the network, where basic features are first extracted through an initial convolutional layer. The data then enters densely connected blocks, each employing a dense connection mechanism. The input to each layer is composed of feature maps from all preceding layers concatenated along the channel dimension. Nonlinear transformations are then performed sequentially through batch normalization, ReLU activation, and 3×3 convolutions, achieving efficient feature reuse and direct gradient propagation.

[0103] The network achieves feature compression through downsampling transition modules, each of which sequentially includes 1×1 convolutions and 2×2 max pooling operations. The 1×1 convolutions reduce the number of feature maps by setting a compression parameter θ, thus controlling the model complexity; the max pooling operation gradually reduces the spatial resolution of the feature maps, expanding the receptive field while preserving important features, forming a hierarchical feature representation.

[0104] The decoding path gradually restores spatial resolution through upsampling transition modules. Each module uses a 3×3 convolution for feature reshaping, followed by a 2×2 upsampling operation to double the feature map size. During upsampling, skip connections fuse feature maps from corresponding layers in the encoding path with the decoded features, fully utilizing shallow spatial details to improve boundary localization accuracy. In the final layer of the network, a 1×1 convolution maps features to the number of channels corresponding to the number of classes, and a sigmoid activation function generates a slippage probability for each pixel, completing the end-to-end semantic segmentation task.

[0105] A gradient-based optimization algorithm was used to train the entire network end-to-end, and the model parameters were adjusted through backpropagation. The performance on the validation set was monitored during training, and the model weights that performed best on the test set were finally selected to obtain the final model that can be used for landslide identification across the entire region.

[0106] In this embodiment, efficient feature extraction and reuse are achieved through dense connections in the encoder, multi-scale feature representation is constructed through a symmetrical downsampling-upsampling structure between the encoder and decoder, and complementary fusion of multi-level features is achieved through skip connections. These three elements work together to enable the model to simultaneously capture both the subtle local features and global contextual information of landslides, thereby achieving high-precision, clear-boundary automatic landslide identification and mapping in complex remote sensing scenarios.

[0107] In one exemplary embodiment, training and optimizing the landslide identification model includes:

[0108] An end-to-end approach was used to iteratively train the landslide identification model using forward and backward propagation.

[0109] During model training, random dropout layers are performed between densely connected blocks to prevent overfitting;

[0110] After the model training is completed, the model weight parameters with the best recognition performance are selected based on the performance metrics of the independent validation set or test set to generate a fully trained landslide recognition model.

[0111] Specifically, the end-to-end approach means the model's input is the raw, processed multi-dimensional feature cube, and the output is a direct landslide probability map. Forward propagation involves inputting a batch of training data into the model, where the data flows through densely connected blocks and downsampling layers in the encoder, then through upsampling layers and skip connections in the decoder, ultimately generating the prediction result. Using a predefined loss function (such as binary cross-entropy loss for pixel-level binary classification tasks), the difference between the model's predicted map and the actual landslide annotation map is calculated; this difference is quantized as a scalar loss value. The backpropagation algorithm calculates the gradient of the loss value relative to each trainable parameter (weights and biases) of the model. The gradient indicates the direction and magnitude by which each parameter should be adjusted to reduce the loss. Subsequently, an optimization algorithm (such as Adam) is used to update all parameters based on the gradient information.

[0112] Introducing Dropout layers between densely connected blocks is a key regularization technique to prevent model overfitting. During the forward propagation of each training iteration, the Dropout layer randomly drops (i.e. temporarily sets its output to zero) a portion of the neurons in the layer with a pre-set probability (such as 0.2 or 0.5).

[0113] Throughout the training process, not only are parameters optimized using the training set, but model performance is also periodically evaluated on a separate validation set (or a reserved test set). The training process generates model weight snapshots for multiple epochs. After training is complete, instead of simply selecting the weights for the final epoch, the model weight parameters corresponding to the epoch that performs best on independent data are chosen based on the comprehensive performance metrics on the validation / test set.

[0114] In this embodiment, end-to-end training ensures the efficiency of the process, Dropout regularization improves the generalization ability of the model, and model selection based on the validation set ensures the optimization of the final model performance.

[0115] In one exemplary embodiment, the method further includes:

[0116] The regions corresponding to independent test samples in the landslide spatial distribution identification results are compared pixel by pixel with the real landslide distribution data verified in the field to construct a confusion matrix.

[0117] Based on the confusion matrix, multiple quantitative evaluation indicators, including overall precision, accuracy, recall, F1 score, Kappa coefficient, and average crossover ratio, are calculated to assess the recognition performance and reliability of the landslide identification model.

[0118] Specifically, such as Figure 6As shown, the trained landslide recognition model is used to predict the area to be identified, generating a landslide recognition result map. The landslide recognition model trained on the test set is then tested. The test set results are compared pixel-by-pixel with the labels of landslide samples validated in the field. An accuracy index is calculated based on the confusion matrix. The predicted category (landslide / non-landslide) and the true category of each pixel are statistically analyzed and filled into the four cells of a 2×2 confusion matrix.

[0119] Based on the confusion matrix, a series of complementary statistical indicators are calculated. The following six are commonly used evaluation indicators:

[0120] 1) Overall accuracy: This indicates the overall prediction accuracy of the model across all pixels in both landslide and non-landslide areas, comprehensively reflecting the model's recognition performance.

[0121]

[0122] 2) Precision: This represents the proportion of correctly predicted landslide samples out of all samples predicted as landslides, and evaluates the accuracy of the model's predictions for landslide pixels.

[0123]

[0124] 3) Recall: This represents the proportion of correctly classified landslide samples to all real landslide samples, reflecting whether the model comprehensively detects landslide pixels.

[0125]

[0126] 4) F1 score: The harmonic mean of precision and recall, which avoids biased evaluation of the model.

[0127]

[0128] 5) Kappa coefficient: This is a metric for evaluating the performance of a classifier. It measures the degree of deviation between the actual classification result and the random guess result. The higher the value, the more stable the model.

[0129]

[0130]

[0131] 6) Mean Intersection over Union (MIU): An important metric for evaluating the overall performance of a segmentation model. A higher MIU indicates that the model can more accurately predict the boundaries between landslide and non-landslide areas.

[0132]

[0133] In this context, TP (True Positive) represents the number of landslide pixels accurately identified in the study area, TN (True Negative) represents the number of non-landslide pixels correctly excluded, FP (False Positive) represents the number of non-landslide pixels misidentified as landslides, and FN (False Negative) represents the number of pixels incorrectly identified as other types of landslides.

[0134] In this embodiment, a feature cube is constructed by fusing multidimensional features of spectrum, topography and geology, and end-to-end training and prediction are performed using a densely connected encoder-decoder network, which effectively achieves high-precision and clear-boundary automated landslide identification.

[0135] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0136] Based on the same inventive concept, this application also provides a landslide identification device based on multidimensional feature aggregation for implementing the landslide identification method based on multidimensional feature aggregation described above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the landslide identification device based on multidimensional feature aggregation provided below can be found in the limitations of the landslide identification method based on multidimensional feature aggregation described above, and will not be repeated here.

[0137] In one exemplary embodiment, such as Figure 7 As shown, a landslide identification device based on multidimensional feature aggregation is provided, comprising:

[0138] The acquisition module 702 is used to acquire remote sensing image data and basic geographic data of the target area. Based on the remote sensing image data and basic geographic data, it extracts and fuses spectral, topographic and geological multi-dimensional features that reflect the landslide development environment to construct a multi-dimensional feature cube.

[0139] Module 704 is used to build a landslide recognition model. The landslide recognition model contains a backbone network of encoder and decoder. The encoder is used to realize the reuse and fusion of cross-level features in the multi-dimensional feature cube.

[0140] The training and optimization module 706 is used to train and optimize the landslide recognition model using sample data containing landslide annotation information.

[0141] The recognition module 708 is used to output the spatial distribution recognition results of landslides in the target area based on the multi-dimensional feature cube and the trained and optimized landslide recognition model.

[0142] In one embodiment, the construction module 704 is further configured to perform radiometric calibration, atmospheric correction, and spatial registration on remote sensing image data; calculate spectral index features including normalized vegetation index and modified normalized water index; extract topographic features including elevation, slope, aspect, plane curvature, and distance from hydrological network based on digital elevation model data; extract geological environmental features including lithology and distance from tectonic faults based on geological map data; and perform layer overlay and channel merging of spectral index features, topographic features, and geological environmental features to generate multidimensional feature cube data.

[0143] In one embodiment, the construction module 704 is further configured to calculate the Pearson correlation coefficient matrix between spectral index features, topographic features and geological environment features, remove redundant features according to a preset correlation threshold, perform variance inflation factor analysis on the feature set after removing redundant features, and remove features from the feature set again according to the variance inflation factor.

[0144] In one embodiment, the encoder includes at least one densely connected block, where the input of each convolutional layer in the densely connected block is the concatenation of the output feature maps of all preceding convolutional layers in the channel dimension; the encoder also includes downsampling transition layers disposed between the densely connected blocks for downsampling in the spatial dimension and compressing the number of feature channels; the decoder includes an upsampling transition layer corresponding to the downsampling transition layer for restoring the spatial resolution of the feature maps and fusing feature information from the corresponding level of the encoder through skip connection paths.

[0145] In one embodiment, the training optimization module 706 is specifically used to train the landslide recognition model iteratively using an end-to-end approach with forward and backward propagation; during the model training process, random dropout layers are performed between densely connected blocks to prevent overfitting; after the model training is completed, the model weight parameters with the best recognition performance are selected based on the performance metrics of the independent validation set or test set to generate a fully trained landslide recognition model.

[0146] In one embodiment, the evaluation module is used to compare the regions corresponding to independent test samples in the landslide spatial distribution identification results with real landslide distribution data verified in the field on a pixel-by-pixel basis to construct a confusion matrix; based on the confusion matrix, it calculates multiple quantitative evaluation indicators, including overall precision, accuracy, recall, F1 score, Kappa coefficient, and average intersection-union ratio, to evaluate the recognition performance and reliability of the landslide identification model.

[0147] Each module in the aforementioned landslide identification device based on multidimensional feature aggregation can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0148] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 8 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs in the non-volatile storage media to run. The database stores remote sensing image data and basic geographic data. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a landslide identification method based on multi-dimensional feature aggregation.

[0149] Those skilled in the art will understand that Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0150] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described above.

[0151] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0152] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the method described above.

[0153] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0154] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0155] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0156] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A landslide identification method based on multidimensional feature aggregation, characterized in that, The method includes: Acquire remote sensing image data and basic geographic data of the target area. Based on the remote sensing image data and the basic geographic data, extract and fuse spectral, topographic and geological multi-dimensional features that reflect the landslide development environment to construct a multi-dimensional feature cube. A landslide identification model is constructed, which includes a backbone network of encoder and decoder, wherein the encoder is used to realize the reuse and fusion of cross-level features in the multi-dimensional feature cube; The landslide identification model is trained and optimized using sample data containing landslide annotation information. Based on the multidimensional feature cube, the landslide recognition model, trained and optimized, outputs the spatial distribution recognition results of landslides in the target area.

2. The method according to claim 1, characterized in that, The process of constructing the multidimensional feature cube includes: The remote sensing image data is subjected to radiometric calibration, atmospheric correction, and spatial registration. The calculation includes spectral index characteristics of the normalized vegetation index and the modified normalized water index; Based on digital elevation model data, topographic features including elevation, slope, aspect, plane curvature, and distance from the hydrological network are extracted. Based on geological map data, geological environmental characteristics including lithology and distance from tectonic faults are extracted; The spectral index features, topographic features, and geological environment features are layered and channels are merged to generate multidimensional feature cube data.

3. The method according to claim 2, characterized in that, Before layering and merging the spectral index features, topographic features, and geological environment features, the method further includes: Calculate the Pearson correlation coefficient matrix between the spectral index features, the topographic features, and the geological environment features, and remove redundant features according to a preset correlation threshold; A variance inflation factor analysis is performed on the feature set after removing redundant features, and features are removed again from the feature set based on the variance inflation factor.

4. The method according to claim 1, characterized in that, The encoder includes at least one densely connected block, wherein the input of each convolutional layer in the densely connected block is the concatenation result of the output feature maps of all preceding convolutional layers in the channel dimension; The encoder also includes a downsampling transition layer disposed between the densely connected blocks for downsampling in the spatial dimension and compressing the number of feature channels; The decoder section includes an upsampling transition layer corresponding to the downsampling transition layer, which is used to restore the spatial resolution of the feature map and fuse feature information from the corresponding level of the encoder through skip connection paths.

5. The method according to claim 4, characterized in that, The process of training and optimizing the landslide identification model includes: The landslide identification model is trained iteratively using an end-to-end approach with forward and backward propagation. During model training, random dropout layers are performed between the densely connected blocks to prevent overfitting; After the model training is completed, the model weight parameters with the best recognition performance are selected based on the performance metrics of the independent validation set or test set to generate the fully trained landslide recognition model.

6. The method according to claim 1, characterized in that, The method further includes: The regions corresponding to independent test samples in the landslide spatial distribution identification results are compared pixel by pixel with the real landslide distribution data verified in the field to construct a confusion matrix. Based on the confusion matrix, multiple quantitative evaluation indicators, including overall precision, accuracy, recall, F1 score, Kappa coefficient, and average intersection-union ratio, are calculated to assess the recognition performance and reliability of the landslide identification model.

7. A landslide identification device based on multidimensional feature aggregation, characterized in that, The device includes: The acquisition module is used to acquire remote sensing image data and basic geographic data of the target area. Based on the remote sensing image data and the basic geographic data, it extracts and fuses spectral, topographic and geological multi-dimensional features that reflect the landslide development environment to construct a multi-dimensional feature cube. A construction module is used to construct a landslide recognition model, which includes a backbone network of encoder and decoder. The encoder is used to realize the reuse and fusion of cross-level features in the multi-dimensional feature cube. The training and optimization module is used to train and optimize the landslide identification model using sample data containing landslide annotation information. The identification module is used to output the spatial distribution identification results of landslides in the target area based on the multi-dimensional feature cube and the trained and optimized landslide identification model.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images