Intelligent landslide identification method and system based on hierarchical frame selection and boundary feature fusion

Abstract

The invention provides an intelligent landslide identification method and system based on hierarchical frame selection and boundary feature fusion, and the method comprises the steps: obtaining a target remote sensing image data set, carrying out the hierarchical frame selection of the target remote sensing image data set, obtaining a hierarchical frame selection region set, carrying out the boundary feature extraction of the hierarchical frame selection region set, and obtaining a boundary feature extraction result; obtaining a boundary feature set of each hierarchical framed area, inputting the boundary feature sets into a pre-trained boundary feature fusion model for feature fusion processing to generate a fused boundary feature set, performing feature comparison processing with a preset landslide morphological feature library based on the fused boundary feature set to generate a landslide identification feature set, and performing landslide identification on the landslide identification feature set. And performing region labeling processing on the target remote sensing image data set according to the landslide recognition feature set to generate a landslide region recognition result. According to the invention, through topology reconstruction of hierarchical frame selection and feature fusion, a landslide identification link with a self-interpretation capability is formed, and the problem of misjudgment accumulation of the model caused by feature black box transmission is avoided.

Description

Technical Field

[0001] The present invention relates to the field of slope identification, and more particularly, to a landslide intelligent identification method and system based on hierarchical box selection and boundary feature fusion. Background Art

[0002] In the field of geological disaster monitoring, landslide identification technology, as one of the core application directions of remote sensing image analysis, is mainly used to identify potential landslide areas through surface deformation characteristics and terrain evolution laws. After extracting image features in the prior art, it relies on static threshold segmentation or isolated boundary detection algorithms to generate preliminary identification areas, and linearly matches and verifies them through a preset landslide morphological feature library. However, due to the fixed segmentation scale in traditional methods, it is difficult to adapt to the dynamic changes of multi-scale geological features in complex terrain, resulting in insufficient sensitivity in identifying the boundaries of incipient landslides and slow-moving landslides; the simple boundary feature extraction mechanism severs the correlation between spatial distribution, geometric morphology, and surface coverage features, causing the misjudgment rate in the transition area between the landslide body and the surrounding terrain to rise; the static feature library matching mode cannot fuse the dynamic evolution law of geological conditions in real time, making the identification results vulnerable to the influence of seasonal changes in surface coverage and artificial engineering interference. Summary of the Invention

[0003] The present invention provides a landslide intelligent identification method and system based on hierarchical box selection and boundary feature fusion.

[0004] In a first aspect, an embodiment of the present invention provides a landslide intelligent identification method based on hierarchical box selection and boundary feature fusion, including the following steps: Obtain a target remote sensing image dataset, and perform hierarchical box selection processing on the target remote sensing image dataset to obtain a set of hierarchical box selection regions; each of the hierarchical box selection regions corresponds to a box selection level identifier; Perform boundary feature extraction processing on the set of hierarchical box selection regions to obtain a boundary feature set for each of the hierarchical box selection regions; the boundary feature set includes spatial distribution features, geometric morphology features, and texture correlation features; Input the boundary feature set into a pre-trained boundary feature fusion model for feature fusion processing to generate a fused boundary feature set; the fused boundary feature set is used to represent the boundary feature correlation between different box selection levels; Perform feature comparison processing based on the fused boundary feature set and a preset landslide morphological feature library to generate a landslide identification feature set; the landslide identification feature set includes geological deformation features, slope correlation features, and surface coverage features of the target region; Perform region annotation processing on the target remote sensing image dataset according to the landslide identification feature set to generate a landslide region identification result.

[0005] Second aspect, an embodiment of the present invention provides a computer system, including: A memory storing a computer program; A processor for loading the computer program to implement the landslide intelligent recognition method based on hierarchical bounding box selection and boundary feature fusion as described above.

[0006] The landslide intelligent recognition method based on hierarchical bounding box selection and boundary feature fusion provided by the present invention constructs a remote sensing image processing system with spatial priority identification through a dynamic hierarchical bounding box selection mechanism, effectively solving the problems of incomplete capture of terrain features and blurred boundaries caused by a single segmentation scale in traditional landslide recognition methods. Based on the dynamic fusion architecture of the three-dimensional boundary feature set, for the first time, the non-linear collaborative calculation of the spatial distribution topological relationship, geometric form evolution law, and surface texture correlation features is realized, breaking through the feature fault limitation caused by the isolated analysis of a single boundary dimension in the prior art. Through the dynamic adjustment mechanism of the cross-hierarchical feature association weight, an adaptive matching channel for multi-scale terrain features is established, enabling potential landslide bodies at different development stages in complex geological environments to obtain accurate mapping. The dynamic interaction mechanism between the feature backtracking verification architecture and the landslide morphological feature library realizes the two-way coupling verification of geographical space features and geological evolution laws without relying on a historical landslide sample library, significantly improving the recognition sensitivity of the boundaries of initial landslides and hidden landslides. This method forms a landslide recognition link with self-explanatory ability through the topological reconstruction of hierarchical bounding box selection and feature fusion, avoiding the problem of cumulative misjudgment caused by the feature black box transmission in existing deep learning models, and ensuring the reliability and traceability of geological disaster recognition while reducing the dependence on manual annotation. Description of the Drawings

[0007] Figure 1 is a flowchart of a landslide intelligent recognition method based on hierarchical bounding box selection and boundary feature fusion provided by an embodiment of the present invention.

[0008] Figure 2 is a schematic diagram of the composition of a computer system provided by an embodiment of the present invention. Detailed Embodiments

[0009] Please refer to Figure 1 , which is a flowchart of a landslide intelligent recognition method based on hierarchical bounding box selection and boundary feature fusion provided by an embodiment of the present invention. This method is executed by a computer system and includes the following steps: Step S100: Obtain a target remote sensing image dataset, and perform hierarchical bounding box selection processing on the target remote sensing image dataset to obtain a set of hierarchical bounding box regions; each hierarchical bounding box region corresponds to a bounding box level identifier.

[0010] The target remote sensing image dataset refers to a collection of remote sensing image data containing the geographical information of the target area for intelligent landslide identification. These image data can be obtained through remote sensing devices such as satellites and drones, and contain rich surface information such as terrain, landform, and vegetation cover. Hierarchical bounding box processing is to perform multi-scale and multi-level regional division on the target remote sensing image dataset to screen out areas where landslides may exist. The hierarchical bounding box area set is a collection of a series of areas obtained after hierarchical bounding box processing, and each area has its geographical scope and characteristics. The bounding box hierarchy identifier is an identifier assigned to each hierarchical bounding box area, which is used to indicate the spatial division priority of the area at different segmentation scales.

[0011] As an implementation manner, in step S100, obtain the target remote sensing image dataset and perform hierarchical bounding box processing on the target remote sensing image dataset to obtain a hierarchical bounding box area set, which may specifically include the following steps S110 to S140: Step S110: Perform multi-scale segmentation processing on the target remote sensing image dataset to obtain a plurality of initial segmentation areas; each initial segmentation area corresponds to a segmentation scale identifier.

[0012] Multi-scale segmentation processing is to segment the target remote sensing image dataset according to different scales to obtain areas of different sizes and characteristics. Multi-scale segmentation can better adapt to landslide areas of different scales and characteristics, and improve the accuracy of landslide identification. The initial segmentation area is the basic area unit obtained after multi-scale segmentation processing, and each area has relatively independent geographical features and attributes. The segmentation scale identifier is an identifier assigned to each initial segmentation area, which is used to indicate at which segmentation scale the area is obtained, reflecting the size and characteristics of the area.

[0013] As an implementation manner, in step S110, perform multi-scale segmentation processing on the target remote sensing image dataset to obtain a plurality of initial segmentation areas, which may specifically include the following steps S111 to S114: Step S111: Obtain the image resolution and surface cover type distribution information of the target remote sensing image dataset.

[0014] The image resolution refers to the actual ground distance represented by each pixel in the remote sensing image, which determines the detail representation ability of the image. The higher the resolution, the richer the detail information in the image. The surface cover type distribution information refers to the distribution of different surface cover types (such as vegetation, bare land, water bodies, etc.) in the target area, and these information can be obtained through image classification and recognition technologies. Obtaining the image resolution and surface cover type distribution information can provide an important basis for the subsequent selection of segmentation scales.

[0015] Specifically, the image resolution can be obtained by reading the metadata of the image through remote sensing image processing software. For the distribution information of land cover types, classification algorithms based on machine learning, such as Support Vector Machine (SVM), Random Forest (RF), etc., can be used to classify the remote sensing image to obtain the distribution of different land cover types. For example, when using the SVM algorithm, first select a certain number of training samples, label the land cover type for each sample, and then input these samples into the SVM model for training to obtain a classification model. Finally, input the target remote sensing image into the trained model for classification prediction to obtain the distribution information of land cover types.

[0016] Step S112: Determine the segmentation scale range according to the image resolution, and select multiple candidate segmentation scales within the segmentation scale range based on the land cover type distribution information.

[0017] The segmentation scale range is an interval of a series of possible segmentation scales determined according to the image resolution, which limits the size range of the segmented areas. The land cover type distribution information plays an important role in selecting candidate segmentation scales. Different land cover types have different sensitivities to the segmentation scale, so it is necessary to select appropriate segmentation scales according to the characteristics of land cover types. Candidate segmentation scales are a set of possible scales selected from within the segmentation scale range.

[0018] As an implementation, step S112, determining the segmentation scale range according to the image resolution and selecting multiple candidate segmentation scales within the segmentation scale range based on the land cover type distribution information, may specifically include the following steps S1121~S1126: Step S1121: Extract the size of the smallest recognizable unit included in the image resolution, and determine the minimum segmentation parameter and the maximum segmentation parameter of the segmentation scale range according to the size of the smallest recognizable unit.

[0019] The size of the smallest recognizable unit refers to the size of the smallest geographical unit that can be clearly recognized at the image resolution, and it is the basis for determining the segmentation scale range. The minimum segmentation parameter and the maximum segmentation parameter are the lower limit and the upper limit of the segmentation scale range respectively, which determine the minimum and maximum sizes of the segmented areas.

[0020] Specifically, the size of the smallest recognizable unit can be determined by analyzing the image resolution. For example, if the image resolution is 1 meter / pixel, then the size of the smallest recognizable unit may be 1 meter × 1 meter. Then, according to experience or experimental data, determine the minimum segmentation parameter and the maximum segmentation parameter. For example, the minimum segmentation parameter is 2 times the size of the smallest recognizable unit, and the maximum segmentation parameter is 10 times the size of the smallest recognizable unit, then the segmentation scale range is 2 meters × 2 meters to 10 meters × 10 meters.

[0021] Step S1122: Assign segmentation parameter adjustment weights to each land cover type according to the area proportion and spatial distribution continuity of different land cover types included in the land cover type distribution information; the segmentation parameter adjustment weights are used to characterize the differences in the sensitivity of different land cover types to the segmentation scale.

[0022] The area proportion refers to the area ratio of different land cover types in the target area, and the spatial distribution continuity refers to the degree of continuity of the same land cover type in space. The segmentation parameter adjustment weight is a weight value assigned to each land cover type according to the area proportion and spatial distribution continuity, and is used to adjust the segmentation scale. Different land cover types have different sensitivities to the segmentation scale. For example, the vegetation-covered area may require a smaller segmentation scale to accurately identify its boundaries and features, while the water body area may have a lower sensitivity to the segmentation scale.

[0023] Specifically, the area proportion and spatial distribution continuity of different land cover types can be obtained through statistical analysis of the land cover type distribution information. Then, according to experience or experimental data, assign segmentation parameter adjustment weights to each land cover type. For example, for the vegetation-covered area, due to its complex boundaries and features, assign a higher segmentation parameter adjustment weight; for the water body area, assign a lower segmentation parameter adjustment weight.

[0024] Step S1123: Perform dynamic range expansion processing on the minimum segmentation parameter and the maximum segmentation parameter based on the segmentation parameter adjustment weights to generate an expanded segmentation scale range.

[0025] The dynamic range expansion processing adjusts the minimum segmentation parameter and the maximum segmentation parameter according to the segmentation parameter adjustment weights to expand the selection range of the segmentation scale. The expanded segmentation scale range can better adapt to the segmentation requirements of different land cover types.

[0026] Specifically, the minimum segmentation parameter and the maximum segmentation parameter can be weighted and adjusted according to the segmentation parameter adjustment weights. For example, for a certain land cover type, its segmentation parameter adjustment weight is 1.2. Then, when expanding the segmentation scale range, the minimum segmentation parameter and the maximum segmentation parameter can be multiplied by 1.2 respectively to obtain the expanded minimum segmentation parameter and maximum segmentation parameter.

[0027] Step S1124: Generate an initial candidate segmentation parameter set according to the segmentation parameter interval step size within the expanded segmentation scale range.

[0028] The segmentation parameter interval step size refers to the difference between two adjacent segmentation parameters within the expanded segmentation scale range. The initial candidate segmentation parameter set is a set of a series of possible segmentation parameters generated according to the segmentation parameter interval step size.

[0029] Specifically, starting from the expanded minimum segmentation parameter, the segmentation parameters are incremented in sequence according to the segmentation parameter interval step size until the expanded maximum segmentation parameter is reached, and these segmentation parameters form the initial candidate segmentation parameter set. For example, if the expanded segmentation scale range is from 2.4 m × 2.4 m to 12 m × 12 m, and the segmentation parameter interval step size is 0.6 m, then the initial candidate segmentation parameter set is {2.4 m × 2.4 m, 3 m × 3 m, 3.6 m × 3.6 m, …, 12 m × 12 m}.

[0030] Step S1125: Based on the type spatial aggregation degree in the land cover type distribution information, perform clustering and screening processing on the segmentation parameters in the initial candidate segmentation parameter set, remove the segmentation parameters whose spatial aggregation degree is lower than the preset type aggregation threshold, and generate an optimized candidate segmentation parameter set.

[0031] The type spatial aggregation degree refers to the degree of aggregation of the same land cover type in space, which reflects whether the distribution of this type is concentrated. The clustering and screening processing is to screen the initial candidate segmentation parameter set according to the type spatial aggregation degree, and remove the segmentation parameters that do not suit the distribution characteristics of the land cover type. The preset type aggregation threshold is a preset threshold used to judge whether the segmentation parameter meets the spatial aggregation degree requirement. The optimized candidate segmentation parameter set is a segmentation parameter set that is more suitable for the distribution characteristics of the land cover type after clustering and screening processing.

[0032] Specifically, a clustering analysis algorithm such as the K-Means algorithm can be used to cluster the segmentation parameters in the initial candidate segmentation parameter set. First, calculate the spatial aggregation degree of the land cover type corresponding to each segmentation parameter, and then remove the segmentation parameters whose spatial aggregation degree is lower than the preset type aggregation threshold to obtain the optimized candidate segmentation parameter set. For example, if the preset type aggregation threshold is 0.8, and for a certain segmentation parameter, the spatial aggregation degree of the land cover type corresponding to it is 0.7, then this segmentation parameter will be removed.

[0033] Step S1126: According to the segmentation parameters in the optimized candidate segmentation parameter set, combined with the type boundary clarity in the land cover type distribution information, perform boundary adaptability verification processing on each segmentation parameter, retain the segmentation parameters that meet the boundary consistency condition, and generate multiple candidate segmentation scales; the boundary consistency condition is used to ensure that the segmentation scale corresponding to the segmentation parameter can completely cover the continuous area of the same land cover type and the distribution of boundary turning points conforms to the preset terrain feature constraint rules.

[0034] The type boundary clarity refers to the clarity of the boundary between the same land cover type and other types. The boundary adaptability verification process verifies the segmentation parameters in the optimized candidate segmentation parameter set according to the type boundary clarity to determine whether they can accurately segment the boundary of the land cover type. The boundary consistency condition is a judgment criterion used to ensure that the segmentation scale corresponding to the segmentation parameters can completely cover the continuous area of the same land cover type and the distribution of the boundary turning points conforms to the preset terrain feature constraint rules. The candidate segmentation scales are a set of scales that are obtained after the boundary adaptability verification process and are suitable for segmentation.

[0035] Specifically, an edge detection algorithm, such as the Canny edge detection algorithm, can be used to perform edge detection on the segmentation results corresponding to each segmentation parameter to obtain the distribution of the boundary turning points. Then, according to the preset terrain feature constraint rules, it is judged whether the distribution of the boundary turning points meets the requirements. For example, if the preset terrain feature constraint rules require that the angles of the boundary turning points are within the target range, then the angles of the boundary turning points can be calculated, the segmentation parameters that do not meet the requirements can be removed, and the segmentation parameters that meet the boundary consistency condition can be retained to generate multiple candidate segmentation scales.

[0036] Step S113: Perform image segmentation processing on the target remote sensing image dataset under each candidate segmentation scale to generate initial segmentation regions; the image segmentation processing includes superpixel segmentation processing and edge detection processing.

[0037] The image segmentation processing is to segment the target remote sensing image dataset according to the candidate segmentation scales to obtain the initial segmentation regions. The superpixel segmentation processing is to segment the image into a series of superpixels with similar features, and these superpixels can be used as the basic units of the initial segmentation regions. The edge detection processing is to detect the boundaries between different regions in the image to determine the boundaries of the initial segmentation regions.

[0038] Specifically, the simple linear iterative clustering (SLIC) algorithm can be used for the superpixel segmentation processing. This algorithm first initializes the pixels in the image, and then iteratively updates the labels of the pixels to divide the pixels into different superpixels. The Canny edge detection algorithm can be used for the edge detection processing. This algorithm detects the edge information in the image by calculating the gradient values of the image. For example, for a certain candidate segmentation scale, the target remote sensing image dataset is input into the SLIC algorithm to obtain the superpixel segmentation result, and then the superpixel segmentation result is input into the Canny edge detection algorithm to obtain the boundary information of the initial segmentation regions.

[0039] Step S114: Perform scale alignment processing on the initial segmentation regions generated under different candidate segmentation scales to obtain multiple initial segmentation regions; each initial segmentation region corresponds to a segmentation scale identifier and the corresponding segmentation parameter set.

[0040] The scale alignment process is to uniformly process the initial segmentation regions generated under different candidate segmentation scales so that they have the same spatial coordinates and scales. The segmentation scale identifier is used to indicate under which segmentation scale the initial segmentation region is generated, and the segmentation parameter set contains the specific segmentation parameters under that segmentation scale.

[0041] Specifically, the nearest neighbor interpolation algorithm can be used to perform scale alignment on the initial segmentation regions under different candidate segmentation scales. This algorithm fills the target pixel by finding the nearest pixel value, thereby achieving scale alignment. For example, for the initial segmentation regions generated under two different candidate segmentation scales, their spatial coordinates are unified, and then the nearest neighbor interpolation algorithm is used to adjust the scale of one region to be the same as that of the other region, obtaining the initial segmentation regions after scale alignment. Each initial segmentation region is assigned a segmentation scale identifier and the corresponding segmentation parameter set for subsequent processing.

[0042] Step S120: Perform region merging processing on multiple initial segmentation regions according to the segmentation scale identifier to generate a merged region set; each merged region in the merged region set includes at least two initial segmentation regions.

[0043] The region merging process is to merge the initial segmentation regions with similar characteristics to reduce the number of regions and improve the integrity and coherence of the regions. The merged region set is a set of a series of regions obtained after the region merging process, and each merged region contains at least two initial segmentation regions. Specifically, a rule-based merging algorithm can be used for region merging according to the segmentation scale identifier and the characteristics of the initial segmentation regions (such as terrain, vegetation coverage, etc.). For example, for the initial segmentation regions with the same segmentation scale identifier and adjacent to each other, if their terrain characteristics and vegetation coverage types are similar, then they are merged into one merged region. The terrain similarity and vegetation coverage similarity between two initial segmentation regions can be calculated, and when the similarity is higher than the preset merging threshold, the merging operation is performed.

[0044] Step S130: Perform region screening processing on the merged region set to obtain a candidate region set; each candidate region in the candidate region set meets the preset landslide candidate conditions; the landslide candidate conditions include the region area threshold, the regional terrain continuity, and the regional texture consistency.

[0045] The regional screening process is to screen out the regions that meet the landslide candidate conditions from the merged region set, so as to reduce the workload of subsequent processing. The candidate region set is a set of regions that may have landslides obtained after the regional screening process. The landslide candidate conditions are a set of criteria used to determine whether a region may have a landslide, including the regional area threshold, the regional terrain continuity, and the regional texture consistency. Specifically, the regional area threshold can be set according to experience or experimental data. For example, the regional area threshold is set to 100 square meters, and the merged regions with an area smaller than this threshold are removed. The regional terrain continuity can be judged by analyzing the terrain undulation and slope change of the region. By using terrain analysis algorithms, such as slope calculation algorithms and aspect calculation algorithms, the slope and aspect within the region are calculated. When the changes in slope and aspect are within the target range, the regional terrain is considered continuous. The regional texture consistency can be judged by calculating the texture features of the region, such as the gray-level co-occurrence matrix (GLCM) features. When the texture features of the region are within the target range, the regional texture is considered consistent. The merged regions that meet the landslide candidate conditions are screened out to obtain the candidate region set.

[0046] Step S140: Perform hierarchical division processing on the candidate region set according to the distribution density and regional overlap degree of the candidate region set to generate a hierarchical bounding region set; each hierarchical bounding region corresponds to a bounding level identifier, and the bounding level identifier is used to indicate the spatial division priority of the hierarchical bounding region at different segmentation scales.

[0047] The hierarchical division processing is to divide the candidate regions into different levels according to the distribution density and regional overlap degree of the candidate region set, so as to achieve multi-scale spatial division. The hierarchical bounding region set is a set of regions with different levels obtained after the hierarchical division processing. The bounding level identifier is used to indicate the spatial division priority of the hierarchical bounding region at different segmentation scales, and the regions with higher priority are considered first in subsequent processing.

[0048] Specifically, a density clustering algorithm, such as the DBSCAN algorithm, can be used to perform clustering analysis on the candidate region set, and the distribution density of the regions can be obtained according to the clustering results. The regional overlap degree can be determined by calculating the overlapping area between candidate regions. According to the distribution density and regional overlap degree, the candidate regions are divided into different levels. For example, the candidate regions with high distribution density and large regional overlap degree are divided into higher levels, and the candidate regions with low distribution density and small regional overlap degree are divided into lower levels. Each hierarchical bounding region is assigned a bounding level identifier for subsequent processing.

[0049] Step S200: Perform boundary feature extraction processing on the hierarchical bounding region set to obtain a boundary feature set for each hierarchical bounding region; the boundary feature set includes spatial distribution features, geometric shape features, and texture correlation features.

[0050] The boundary feature extraction process extracts the boundary features of each hierarchical bounding box region from the set of hierarchical bounding box regions for subsequent feature fusion and landslide identification. The boundary feature set is a set of feature sets that describe the boundary features of hierarchical bounding box regions obtained after the boundary feature extraction process, including spatial distribution features, geometric shape features, and texture correlation features.

[0051] The spatial distribution features describe the spatial distribution of the hierarchical bounding box regions, such as the position, direction, and range of the regions. The geometric shape features describe the geometric shapes of the hierarchical bounding box regions, such as perimeter, area, and shape index. The texture correlation features describe the texture features of the hierarchical bounding box regions, such as the roughness and contrast of the texture.

[0052] Specifically, the spatial distribution features can be obtained by calculating the center point coordinates of the hierarchical bounding box region, the position and size of the circumscribed rectangle, etc. The geometric shape features can be obtained by calculating geometric parameters such as the perimeter, area, and shape index of the region. The texture correlation features can be obtained by calculating the gray-level co-occurrence matrix (GLCM) features of the region, such as contrast, correlation, and energy. For example, for a certain hierarchical bounding box region, a boundary detection algorithm is used to detect its boundary, and then the center point coordinates of the boundary are calculated as part of the spatial distribution features, the perimeter and area of the boundary are calculated as part of the geometric shape features, and the GLCM features of the boundary region are calculated as part of the texture correlation features to obtain the boundary feature set of the hierarchical bounding box region.

[0053] Step S300: Input the boundary feature set into a pre-trained boundary feature fusion model for feature fusion processing to generate a fused boundary feature set; the fused boundary feature set is used to represent the boundary feature correlation between different bounding box levels.

[0054] Feature fusion processing is to fuse different features in the boundary feature set to generate a fused boundary feature set that can represent the boundary feature correlation between different bounding box levels. The pre-trained boundary feature fusion model is a pre-trained model for feature fusion, which can learn the correlation and interaction between different features, thereby generating more effective fused features.

[0055] Specifically, the pre-trained boundary feature fusion model can adopt a deep learning model, such as a convolutional neural network (CNN) or a recurrent neural network (RNN). Taking CNN as an example, the model includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The input layer receives the boundary feature set as input. The convolutional layer extracts local information of the features through convolutional kernels. The pooling layer reduces the dimensionality of the features. The fully connected layer integrates the extracted features. The output layer outputs the fused boundary feature set. In the training phase, a large amount of historical data is used to train the model and adjust its parameters so that it can learn the associations and interactions between different features. In the inference phase, the boundary feature set is input into the trained model to obtain the fused boundary feature set.

[0056] As an implementation, the boundary feature fusion model is trained through the following steps S301 to S3010: Step S301: Obtain a historical remote sensing image sample set and the corresponding labeled landslide area set; perform hierarchical bounding box processing on the historical remote sensing image sample set to obtain a sample hierarchical bounding box area set; perform boundary feature extraction processing on the sample hierarchical bounding box area set to obtain a sample boundary feature set.

[0057] The historical remote sensing image sample set is a set containing historical remote sensing image data, which can come from different times and locations and contains rich landslide information. The labeled landslide area set is a set obtained by labeling the landslide areas in the historical remote sensing image sample set, and the labeling information can include the location, scope, etc. of the landslide area. The sample hierarchical bounding box area set is a set of a series of areas obtained by performing hierarchical bounding box processing on the historical remote sensing image sample set, and each area corresponds to a bounding box level identifier. The sample boundary feature set is a set of features that describe the boundary features of the sample hierarchical bounding box areas obtained by performing boundary feature extraction processing on the sample hierarchical bounding box area set.

[0058] Specifically, the historical remote sensing image sample set can be downloaded through a remote sensing data platform or a related database. The method of performing hierarchical bounding box processing on the historical remote sensing image sample set is the same as the method described in step S100. The method of performing boundary feature extraction processing on the sample hierarchical bounding box area set is the same as the method described in step S200. For example, 100 historical remote sensing images are downloaded from the remote sensing data platform as the historical remote sensing image sample set, hierarchical bounding box processing is performed on each image to obtain a sample hierarchical bounding box area set, and then boundary feature extraction processing is performed on each sample hierarchical bounding box area to obtain a sample boundary feature set.

[0059] Step S302: Initialize the feature fusion network structure; the feature fusion network structure includes a multi-level feature association module, a dynamic weight generation module, and a non-linear mapping module.

[0060] The feature fusion network structure is a network architecture for feature fusion, which includes multiple modules, each with different functions. The multi-level feature association module is used to learn the feature association relationships between different-level bounding regions, the dynamic weight generation module is used to generate the dynamic weights of different features, and the non-linear mapping module is used to perform non-linear transformation on the features to enhance the expression ability of the features.

[0061] Specifically, the multi-level feature association module can adopt an attention mechanism, such as a multi-head attention mechanism, to learn the feature association relationships between different-level bounding regions. The dynamic weight generation module can adopt a fully connected layer to generate the dynamic weights of different features according to the input features. The non-linear mapping module can adopt an activation function, such as the ReLU function, to perform non-linear transformation on the features. For example, when initializing the feature fusion network structure, set the number of heads of the multi-level feature association module to 8, the number of neurons in the fully connected layer of the dynamic weight generation module to 64, and the non-linear mapping module adopts the ReLU activation function.

[0062] Step S303: Input the sample boundary feature set into the multi-level feature association module in the feature fusion network structure to generate a sample first association feature sequence.

[0063] The multi-level feature association module receives the sample boundary feature set as input and generates a sample first association feature sequence by learning the feature association relationships between different-level bounding regions. The sample first association feature sequence is a set of feature sequences that describe the feature association relationships between different-level bounding regions.

[0064] Specifically, input the sample boundary feature set into the multi-level feature association module. The multi-head attention mechanism of the multi-level feature association module processes the input features, calculates the attention scores between different-level bounding regions, and then performs weighted summation on the features according to the attention scores to obtain the sample first association feature sequence. For example, for the boundary features of each-level bounding region in the sample boundary feature set, the multi-level feature association module calculates the attention scores between its boundary features and the boundary features of other-level bounding regions, performs weighted summation on the features with higher attention scores to obtain the first association feature corresponding to this-level bounding region, and forms the sample first association feature sequence with the first association features of all-level bounding regions.

[0065] Step S304: Input the sample first association feature sequence into the dynamic weight generation module to generate a sample dynamic weight adjustment coefficient in combination with the sample geometric shape features and texture association features.

[0066] The dynamic weight generation module receives the sample first associated feature sequence as input, combines the sample geometric shape features and texture association features, and generates the sample dynamic weight adjustment coefficients. The sample dynamic weight adjustment coefficients are a set of coefficients used to adjust the weights of different features during the feature fusion process.

[0067] Specifically, the sample first associated feature sequence, the sample geometric shape features, and the texture association features are input into the fully connected layer of the dynamic weight generation module. The fully connected layer performs a linear transformation on the input features, and then maps the output values to the range [0,1] through an activation function (such as the Sigmoid function) to obtain the sample dynamic weight adjustment coefficients. For example, the input feature dimension of the fully connected layer is 128, and the output feature dimension is 1. After performing a linear transformation on the input features, the output values are mapped to the range [0,1] through the Sigmoid function to obtain the sample dynamic weight adjustment coefficients.

[0068] Step S305: Perform weighted aggregation processing on the sample texture association features based on the sample dynamic weight adjustment coefficients to generate the sample initial fusion texture features.

[0069] The weighted aggregation processing is to perform weighted summation on the sample texture association features according to the sample dynamic weight adjustment coefficients to generate the sample initial fusion texture features. The sample initial fusion texture features are a set of features that fuse the texture association features of different hierarchical bounding box regions.

[0070] Specifically, the sample dynamic weight adjustment coefficients are multiplied element by element with the sample texture association features, and then the results after multiplication are summed to obtain the sample initial fusion texture features.

[0071] Step S306: Perform feature splicing processing on the sample initial fusion texture features and the sample first associated feature sequence to generate the predicted fusion feature set.

[0072] The feature splicing processing is to splice the sample initial fusion texture features and the sample first associated feature sequence to generate the predicted fusion feature set. The predicted fusion feature set is a set of features that fuse the sample initial fusion texture features and the sample first associated feature sequence.

[0073] Specifically, the sample initial fusion texture features and the sample first associated feature sequence are spliced in the feature dimension to obtain the predicted fusion feature set. For example, the dimension of the sample initial fusion texture features is 3, and the dimension of the sample first associated feature sequence is 10. They are spliced in the feature dimension to obtain the predicted fusion feature set with a dimension of 13.

[0074] Step S307: Extract the standard spatial distribution features, standard geometric shape features, and standard texture association features in the labeled landslide area set, and construct the standard fusion feature template.

[0075] The standard spatial distribution features, standard geometric shape features, and standard texture correlation features are a set of standard features extracted from the labeled landslide area set to describe the characteristics of the landslide area. The standard fusion feature template is a template constructed based on these standard features and is used to evaluate the accuracy of the predicted fusion feature set.

[0076] Specifically, the standard spatial distribution features can be extracted by calculating the center point coordinates of the labeled landslide area, the position and size of the circumscribed rectangle, etc. The standard geometric shape features can be extracted by calculating geometric parameters such as the perimeter, area, and shape index of the labeled landslide area. The standard texture correlation features can be extracted by calculating the gray-level co-occurrence matrix (GLCM) features of the labeled landslide area, such as contrast, correlation, energy, etc. These standard features are fused to construct the standard fusion feature template. For example, for each landslide area in the labeled landslide area set, calculate its center point coordinates, perimeter, area, GLCM features, etc., and fuse these features to construct the standard fusion feature template.

[0077] Step S308: Calculate the distribution difference degree, shape matching degree, and texture consistency index between the predicted fusion feature set and the standard fusion feature template.

[0078] The distribution difference degree refers to the degree of difference in the spatial distribution between the predicted fusion feature set and the standard fusion feature template. The shape matching degree refers to the degree of matching in the geometric shape between the predicted fusion feature set and the standard fusion feature template. The texture consistency index refers to the degree of consistency in the texture features between the predicted fusion feature set and the standard fusion feature template.

[0079] Specifically, the distribution difference degree can be calculated using distance measurement methods such as Euclidean distance or Manhattan distance to calculate the distance between the predicted fusion feature set and the standard fusion feature template in terms of spatial distribution features. The shape matching degree can be calculated using a shape matching algorithm, such as the Hausdorff distance algorithm, to calculate the distance between the predicted fusion feature set and the standard fusion feature template in terms of geometric shape features. The texture consistency index can be calculated using similarity measurement methods such as correlation coefficient or cosine similarity to calculate the similarity between the predicted fusion feature set and the standard fusion feature template in terms of texture correlation features. For example, use Euclidean distance to calculate the distance between the predicted fusion feature set and the standard fusion feature template in terms of spatial distribution features, use the Hausdorff distance algorithm to calculate their distance in terms of geometric shape features, and use cosine similarity to calculate their similarity in terms of texture correlation features.

[0080] Step S309: Generate a model optimization loss value according to the weighted combination result of the distribution difference degree, shape matching degree, and texture consistency index.

[0081] The model optimization loss value is a value calculated based on the weighted combination results of the distribution difference degree, the shape matching degree, and the texture consistency index, and is used to evaluate the difference degree between the prediction result of the model and the standard fusion feature template.

[0082] Specifically, different weights are assigned to the distribution difference degree, the shape matching degree, and the texture consistency index. For example, the weight of the distribution difference degree is 0.3, the weight of the shape matching degree is 0.3, and the weight of the texture consistency index is 0.4. Then, the distribution difference degree, the shape matching degree, and the texture consistency index are respectively multiplied by their corresponding weights, and the results are added together to obtain the model optimization loss value. For example, if the distribution difference degree is 0.2, the shape matching degree is 0.3, and the texture consistency index is 0.8, the model optimization loss value = 0.3×0.2 + 0.3×0.3 + 0.4×(1 - 0.8) = 0.19.

[0083] Step S3010: Based on the model optimization loss value, jointly and iteratively optimize the network parameters of the multi-level feature association module, the dynamic weight generation module, and the non-linear mapping module until the feature deviation between the predicted fusion feature set and the standard fusion feature template is less than the preset optimization threshold, and obtain the trained boundary feature fusion model.

[0084] The joint iterative optimization is to update the network parameters of the multi-level feature association module, the dynamic weight generation module, and the non-linear mapping module according to the model optimization loss value to improve the prediction accuracy of the model. The preset optimization threshold is a pre-set threshold used to determine whether the model converges. When the feature deviation between the predicted fusion feature set and the standard fusion feature template is less than the preset optimization threshold, it is considered that the model converges and the training is completed.

[0085] Specifically, the gradient descent algorithm, such as the Stochastic Gradient Descent (SGD) algorithm or the Adaptive Moment Estimation (Adam) algorithm, is used to calculate the gradients of the network parameters of the multi-level feature association module, the dynamic weight generation module, and the non-linear mapping module according to the model optimization loss value, and then update the network parameters according to the gradients. This process is continuously repeated until the feature deviation between the predicted fusion feature set and the standard fusion feature template is less than the preset optimization threshold. For example, the preset optimization threshold is 0.01, and the Adam algorithm is used to update the network parameters. When the feature deviation between the predicted fusion feature set and the standard fusion feature template is less than 0.01, it is considered that the model converges and the trained boundary feature fusion model is obtained.

[0086] As an implementation manner, in step S300, the boundary feature set is input into the pre-trained boundary feature fusion model for feature fusion processing to generate a fused boundary feature set, which may specifically include the following steps S310~S360: Step S310: Perform multi-level feature correlation analysis on the spatial distribution features in the boundary feature set to generate a first correlation feature sequence; the first correlation feature sequence is used to characterize the spatial distribution correlation strength between different bounding levels.

[0087] The multi-level feature correlation analysis analyzes the spatial distribution features in the boundary feature set to learn the spatial distribution correlation strength between different bounding levels. The first correlation feature sequence is a set of feature sequences that describe the spatial distribution correlation strength between different bounding levels.

[0088] Specifically, an attention mechanism, such as a multi-head attention mechanism, is used to process the spatial distribution features in the boundary feature set. The multi-head attention mechanism calculates the attention scores of the spatial distribution features between different bounding levels, and then weights and sums the features according to the attention scores to obtain the first correlation feature sequence. For example, for the spatial distribution features of each bounding level in the boundary feature set, the multi-head attention mechanism calculates the attention scores between its spatial distribution features and those of other bounding levels, weights and sums the features with higher attention scores to obtain the first correlation feature corresponding to this bounding level, and combines the first correlation features of all bounding levels to form the first correlation feature sequence.

[0089] Step S320: Perform morphological similarity measurement on the geometric shape features in the boundary feature set to generate a geometric shape similarity map; the geometric shape similarity map is used to indicate the geometric shape difference degree between different bounding levels.

[0090] The morphological similarity measurement compares the geometric shape features in the boundary feature set to calculate the geometric shape difference degree between different bounding levels. The geometric shape similarity map is a map used to visually display the geometric shape difference degree between different bounding levels.

[0091] Specifically, a shape matching algorithm, such as the Hausdorff distance algorithm, is used to calculate the distances between the geometric shape features of different bounding levels. The calculated distances are formed into a matrix, and this matrix is the geometric shape similarity map. For example, for the geometric shape features of each bounding level in the boundary feature set, the Hausdorff distance algorithm is used to calculate the distances between its geometric shape features and those of other bounding levels, and these distances are formed into a matrix. Each element in the matrix represents the geometric shape difference degree between two bounding levels.

[0092] Step S330: Perform cross-feature matching based on the first correlation feature sequence and the geometric shape similarity map to generate a dynamic weight adjustment coefficient; the dynamic weight adjustment coefficient is used to balance the influence weights of the spatial distribution and geometric shape of different bounding levels on the fusion result.

[0093] Cross - feature matching is to match the first associated feature sequence with the geometric - shape similarity map to generate dynamic weight adjustment coefficients. The dynamic weight adjustment coefficients are a set of coefficients used to adjust the weights of the spatial distribution and geometric shape at different box - selection levels during the feature - fusion process.

[0094] Specifically, the first associated feature sequence and the geometric - shape similarity map are input into a fully - connected layer. The fully - connected layer performs a linear transformation on the input features, and then maps the output values to the range [0, 1] through an activation function (such as the Sigmoid function) to obtain the dynamic weight adjustment coefficients. For example, the input - feature dimension of the fully - connected layer is the dimension of the first associated feature sequence plus the dimension of the geometric - shape similarity map, and the output - feature dimension is 1. After performing a linear transformation on the input features, the output values are mapped to the range [0, 1] through the Sigmoid function to obtain the dynamic weight adjustment coefficients.

[0095] Step S340: Perform an adaptive weighted aggregation process on the texture - associated features in the boundary - feature set according to the dynamic weight adjustment coefficients to generate initial fused texture features.

[0096] The adaptive weighted aggregation process is to perform a weighted sum on the texture - associated features in the boundary - feature set according to the dynamic weight adjustment coefficients to generate initial fused texture features. The initial fused texture features are a set of features that fuse the texture - associated features at different box - selection levels. Specifically, the dynamic weight adjustment coefficients are multiplied element - by - element with the texture - associated features in the boundary - feature set, and then the results of the multiplication are summed to obtain the initial fused texture features.

[0097] Step S350: Perform a feature - concatenation process on the initial fused texture features and the first associated feature sequence to generate cross - level concatenated features.

[0098] The feature - concatenation process is to concatenate the initial fused texture features and the first associated feature sequence to generate cross - level concatenated features. The cross - level concatenated features are a set of features that fuse the initial fused texture features and the first associated feature sequence. Specifically, the initial fused texture features and the first associated feature sequence are concatenated in the feature dimension to obtain the cross - level concatenated features.

[0099] Step S360: Perform a non - linear mapping process on the cross - level concatenated features to eliminate feature redundancy and enhance the cross - level feature - expression consistency, generating a fused - boundary - feature set; each feature in the fused - boundary - feature set contains the fused - expression information of spatial distribution, geometric shape, and texture association.

[0100] Nonlinear mapping processing performs a nonlinear transformation on the cross-level concatenated features to eliminate feature redundancy and enhance the consistency of cross-level feature representation. The fused boundary feature set is a set of features that integrates spatial distribution, geometric shape, and texture correlation information after nonlinear mapping processing.

[0101] Specifically, an activation function, such as the ReLU function, is used to perform a nonlinear transformation on the cross-level concatenated features. The ReLU function sets negative values in the input features to 0 and retains positive values, thereby eliminating feature redundancy. After being processed by the ReLU function, the consistency of feature representation is enhanced, generating a fused boundary feature set. For example, for the cross-level concatenated features [1, -2, 3, 4, -5, 6], after being processed by the ReLU function, the fused boundary feature set [1, 0, 3, 4, 0, 6] is obtained.

[0102] Step S400: Perform feature comparison processing based on the fused boundary feature set and a preset landslide morphology feature library to generate a landslide recognition feature set; the landslide recognition feature set includes geological deformation features, slope correlation features, and surface coverage features of the target area.

[0103] Feature comparison processing is to compare the fused boundary feature set with a preset landslide morphology feature library to find features that match the landslide morphology features. The landslide morphology feature library is a database containing various landslide morphology features, which can be constructed through historical data and expert knowledge. The landslide recognition feature set is a set of features obtained after feature comparison processing, describing the possible landslides in the target area, including geological deformation features, slope correlation features, and surface coverage features.

[0104] Specifically, calculate the similarity between each feature in the fused boundary feature set and the features in the landslide morphology feature library. When the similarity is higher than the preset matching threshold, it is considered that the feature matches the landslide morphology feature. The matching features are composed into a landslide recognition feature set. For example, the preset matching threshold is 0.8. For a certain feature in the fused boundary feature set, calculate its similarity with the features in the landslide morphology feature library to be 0.9, and consider that the feature matches the landslide morphology feature and add it to the landslide recognition feature set.

[0105] As an implementation, step S400, performing feature comparison processing based on the fused boundary feature set and a preset landslide morphology feature library to generate a landslide recognition feature set, may specifically include the following steps S410~S450: Step S410: Obtain a standard landslide morphology feature set from the landslide morphology feature library; the standard landslide morphology feature set includes geological deformation features, slope correlation features, and surface coverage features of typical landslide areas.

[0106] The standard landslide morphological feature set is a set of standard features extracted from the landslide morphological feature library to describe the characteristics of typical landslide areas. These features can be obtained through the analysis and summary of a large amount of historical landslide data.

[0107] Specifically, the features of typical landslide areas related to the current target area are screened from the landslide morphological feature library to form the standard landslide morphological feature set. For example, if the current target area is a mountainous area, the geological deformation features, slope correlation features, and surface coverage features of typical landslide areas in the mountainous area are screened from the landslide morphological feature library to form the standard landslide morphological feature set.

[0108] Step S420: Perform feature analysis processing on each fusion boundary feature in the fusion boundary feature set to extract the target geological deformation parameters, target slope parameters, and target surface coverage parameters.

[0109] Feature analysis processing is to analyze each fusion boundary feature in the fusion boundary feature set to extract the target geological deformation parameters, target slope parameters, and target surface coverage parameters contained therein. The target geological deformation parameters, target slope parameters, and target surface coverage parameters are parameters describing the geological deformation, slope, and surface coverage conditions of the target area.

[0110] Specifically, a feature extraction algorithm, such as the principal component analysis (PCA) algorithm or the independent component analysis (ICA) algorithm, is used to perform dimensionality reduction processing on the fusion boundary features to extract the main features therein. Then, based on these main features, the target geological deformation parameters, target slope parameters, and target surface coverage parameters are parsed. For example, using the PCA algorithm to perform dimensionality reduction processing on the fusion boundary features to obtain the main features, and based on these main features, the target geological deformation parameter is the ground settlement amount, the target slope parameter is the average slope, and the target surface coverage parameter is the vegetation coverage rate.

[0111] Step S430: Calculate the similarity between the target geological deformation parameters, target slope parameters, and target surface coverage parameters and the corresponding parameters in the standard landslide morphological feature set respectively to generate the geological deformation similarity, slope correlation similarity, and surface coverage similarity.

[0112] Similarity calculation is to compare the target geological deformation parameters, target slope parameters, and target surface coverage parameters with the corresponding parameters in the standard landslide morphological feature set respectively to calculate the degree of similarity between them. The geological deformation similarity, slope correlation similarity, and surface coverage similarity are indicators respectively describing the degree of similarity between the geological deformation, slope, and surface coverage conditions of the target area and the standard landslide morphological feature set.

[0113] Specifically, a similarity measurement method, such as cosine similarity or Euclidean distance, is used to calculate the similarity between the target geological deformation parameters and the geological deformation parameters in the standard landslide morphological feature set, obtaining the geological deformation similarity. Similarly, the similarity between the target slope parameters and the slope parameters in the standard landslide morphological feature set is calculated to obtain the slope correlation similarity. The similarity between the target surface coverage parameters and the surface coverage parameters in the standard landslide morphological feature set is calculated to obtain the surface coverage similarity. For example, when using cosine similarity to calculate the similarity between the target geological deformation parameters and the geological deformation parameters in the standard landslide morphological feature set, if the similarity is 0.8, the geological deformation similarity is considered to be 0.8.

[0114] Step S440: Determine the landslide recognition confidence corresponding to the fusion boundary feature according to the weighted sum result of the geological deformation similarity, the slope correlation similarity, and the surface coverage similarity.

[0115] The landslide recognition confidence is a probability value used to judge whether there is a landslide in the target area corresponding to the fusion boundary feature. According to the weighted sum result of the geological deformation similarity, the slope correlation similarity, and the surface coverage similarity, the landslide recognition confidence can be obtained.

[0116] Specifically, different weights are assigned to the geological deformation similarity, the slope correlation similarity, and the surface coverage similarity. For example, the weight of the geological deformation similarity is 0.4, the weight of the slope correlation similarity is 0.3, and the weight of the surface coverage similarity is 0.3. Then, the geological deformation similarity, the slope correlation similarity, and the surface coverage similarity are multiplied by their corresponding weights respectively, and the results are added to obtain the landslide recognition confidence. For example, the geological deformation similarity is 0.8, the slope correlation similarity is 0.7, and the surface coverage similarity is 0.6. The landslide recognition confidence = 0.4×0.8 + 0.3×0.7 + 0.3×0.6 = 0.71.

[0117] Step S450: Determine the feature set of the target area corresponding to the fusion boundary feature with a landslide recognition confidence greater than the preset confidence threshold as the landslide recognition feature set.

[0118] The preset confidence threshold is a preset threshold used to judge whether there is a landslide in the target area corresponding to the fusion boundary feature. When the landslide recognition confidence is greater than the preset confidence threshold, it is considered that the possibility of a landslide in the target area corresponding to the fusion boundary feature is relatively high, and the feature set of its corresponding target area is determined as the landslide recognition feature set.

[0119] Specifically, traverse each fusion boundary feature in the fusion boundary feature set, and calculate its corresponding landslide recognition confidence. Filter out the fusion boundary features whose landslide recognition confidence is greater than the preset confidence threshold, and form a landslide recognition feature set with the feature sets of the corresponding target regions. For example, if the preset confidence threshold is 0.7, and for a certain fusion boundary feature, its corresponding landslide recognition confidence is 0.8, then add the feature set of the target region corresponding to this fusion boundary feature to the landslide recognition feature set.

[0120] As an implementation manner, in step S450, determine the feature set of the target region corresponding to the fusion boundary feature whose landslide recognition confidence is greater than the preset confidence threshold as the landslide recognition feature set, which may specifically include the following steps S451 to S455: Step S451: Obtain the box selection level identifier and spatial position information corresponding to the fusion boundary feature set.

[0121] The box selection level identifier is used to indicate the spatial division priority of the hierarchical box selection area corresponding to the fusion boundary feature at different segmentation scales. The spatial position information is used to describe the geographical location of the target region corresponding to the fusion boundary feature.

[0122] Specifically, extract the box selection level identifier and spatial position information corresponding to each fusion boundary feature from the fusion boundary feature set. For example, for a certain fusion boundary feature in the fusion boundary feature set, its corresponding box selection level identifier is "level 3", and the spatial position information is "longitude 110°, latitude 30°".

[0123] Step S452: Determine the spatial division priority of the target region according to the box selection level identifier, and perform spatial clustering processing on the target region according to the spatial position information to generate a set of candidate landslide regions.

[0124] The spatial division priority is used to determine the dependency relationship of different target regions in the processing order. The spatial clustering processing is to cluster the target regions with similar spatial positions to generate a set of candidate landslide regions.

[0125] Specifically, determine the spatial division priority of each target area according to the frame selection level identifier, and process the target areas with higher priority first. Then, adopt a spatial clustering algorithm, such as the DBSCAN algorithm, to cluster the target areas according to the spatial position information. The DBSCAN algorithm clusters the target areas with similar spatial positions into one category based on the distance and density between the target areas, generating a set of candidate landslide areas. For example, for the set of target areas {Area 1, Area 2, Area 3, Area 4}, according to the frame selection level identifier, it is determined that the spatial division priority of Area 1 is the highest, and it is processed first. Then, use the DBSCAN algorithm to cluster Area 1, Area 2, Area 3, and Area 4. For example, Area 1 and Area 2 are close in spatial position and are clustered into one category, and Area 3 and Area 4 are close in spatial position and are clustered into one category, generating a set of candidate landslide areas {{Area 1, Area 2}, {Area 3, Area 4}}.

[0126] As an implementation, in step S452, determine the spatial division priority of the target area according to the frame selection level identifier, and perform spatial clustering processing on the target area according to the spatial position information to generate a set of candidate landslide areas, which may specifically include the following steps S4521 to S4528: Step S4521: Extract the hierarchical division rules corresponding to the frame selection level identifier. The hierarchical division rules include the decreasing relationship of the spatial coverage range between different frame selection levels and the hierarchical weight allocation strategy.

[0127] The hierarchical division rules are determined according to the frame selection level identifier, which clarifies the spatial coverage range of different frame selection levels and the weights of each level. The decreasing relationship of the spatial coverage range reflects that as the frame selection level increases, the corresponding spatial range will gradually decrease, which reflects the spatial division process from macro to micro. The hierarchical weight allocation strategy assigns different importance to different levels and is used to reflect the priority of each level in subsequent processing.

[0128] Specifically, the frame selection level identifier can be matched with a pre-set hierarchical division rule table to extract the corresponding rules. For example, if the frame selection level identifier is "Level 2", the spatial coverage range decreasing ratio between this level and other levels (such as Level 1 and Level 3) can be found in the hierarchical division rule table, as well as the weight coefficient of Level 2 in the overall processing. For example, the hierarchical division rule table stipulates that the spatial coverage range of Level 2 is 50% of that of Level 1, and the hierarchical weight of Level 2 is 0.3.

[0129] Step S4522: Determine the spatial division priority corresponding to each frame selection level according to the decreasing relationship of the spatial coverage range and the hierarchical weight allocation strategy; the spatial division priority is used to indicate the dependency relationship of different frame selection levels in the processing order.

[0130] The spatial division priority is jointly determined by the decreasing relationship of spatial coverage and the hierarchical weight allocation strategy, which determines the order of processing for different selection levels. The level with a higher priority will be processed first, which helps to ensure that the target area can be accurately determined step by step from the macro to the micro in the overall processing process.

[0131] Specifically, for each selection level, a comprehensive priority index can be calculated based on its spatial coverage and hierarchical weight. For example, the weighted average method can be used to perform a weighted calculation on the decreasing ratio of spatial coverage and the hierarchical weight. For example, the decreasing ratio of the spatial coverage of level 2 is 0.5, and the hierarchical weight is 0.3. The decreasing ratio of the spatial coverage of level 3 is 0.2, and the hierarchical weight is 0.5. Then the comprehensive priority index of level 2 is 0.5×0.3 = 0.15, and the comprehensive priority index of level 3 is 0.2×0.5 = 0.1. It can be seen from this that the spatial division priority of level 2 is higher than that of level 3.

[0132] Step S4523: Perform hierarchical sorting on the target area based on the spatial division priority to generate a sorted target area sequence. The hierarchical sorting process sorts the target areas according to the spatial division priority, with the target areas with higher priority ranked in the front and the target areas with lower priority ranked in the back, forming an ordered sequence. Such sorting helps the subsequent processing to proceed in a reasonable order, improving the processing efficiency and accuracy.

[0133] During specific operations, a sorting algorithm such as the quicksort algorithm can be used. For a set containing multiple target areas, the target areas are sorted according to the spatial division priority of the selection level to which each target area belongs. For example, there is a target area A belonging to level 2, a target area B belonging to level 3, and a target area C belonging to level 1. According to the previously calculated priorities, the sorted target area sequence is [target area C, target area A, target area B].

[0134] Step S4524: Perform spatial proximity analysis on the sorted target area sequence according to the coordinate distribution characteristics and regional density characteristics in the spatial position information to generate a set of adjacent area clusters.

[0135] Spatial proximity analysis is to find adjacent target areas in space by analyzing the coordinate distribution characteristics and regional density characteristics of the target areas and cluster them. The set of adjacent area clusters is a set of a series of area clusters obtained after spatial proximity analysis. The target areas within each area cluster are adjacent in space.

[0136] Specifically, the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm can be used for spatial proximity analysis. This algorithm first defines a neighborhood radius and a minimum number of points threshold. For each target region in the sorted target region sequence, with it as the center, the number of points is counted within the neighborhood radius. If the number of points exceeds the minimum number of points threshold, these adjacent target regions are clustered into a cluster. For example, the neighborhood radius is set to 10 meters and the minimum number of points threshold is set to 3. For target region D, there are target regions E and F within its 10-meter neighborhood and the minimum number of points threshold is met, then target regions D, E, and F will be clustered into a neighboring region cluster.

[0137] Step S4525: Perform hierarchical consistency verification processing on the set of neighboring region clusters, and filter out the region clusters that meet the hierarchical division rules as the preliminary candidate region clusters.

[0138] The hierarchical consistency verification processing is to check whether each region cluster in the set of neighboring region clusters meets the hierarchical division rules. Only the region clusters that meet the rules can be selected as the preliminary candidate region clusters, which helps to ensure the rationality of the regions in the hierarchical structure for subsequent processing.

[0139] Specifically, the hierarchical consistency verification can be carried out from two aspects. On the one hand, check whether the selected levels of the target regions within the region cluster conform to the decreasing relationship of the spatial coverage range and the hierarchical weight allocation strategy. On the other hand, check whether the hierarchical attributes of the entire region cluster match the hierarchical division rules. For example, the target regions within a region cluster belong to level 2 and level 3 respectively, but the number of target regions at level 2 is too large and does not conform to the spatial coverage range and weight ratio of this level in the hierarchical division rules, then this region cluster will be excluded.

[0140] Step S4526: Based on the spatial distribution consistency of the geological deformation characteristics and slope correlation characteristics of the preliminary candidate region clusters, perform region merging processing on the preliminary candidate region clusters to generate merged candidate region clusters.

[0141] The region merging processing is to merge the region clusters with similar characteristics according to the spatial distribution consistency of the geological deformation characteristics and slope correlation characteristics of the preliminary candidate region clusters. The spatial distribution consistency of the geological deformation characteristics and slope correlation characteristics reflects the similarity of the regions in terms of geology and terrain. Merging the region clusters with similar characteristics can reduce redundancy and improve the efficiency of subsequent processing.

[0142] Specifically, the similarity of geological deformation features and the similarity of slope correlation features between the preliminary candidate region clusters can be calculated. The cosine similarity algorithm or the Euclidean distance algorithm can be used to calculate the similarity. When the similarity of geological deformation features and the similarity of slope correlation features of two region clusters are both higher than the preset similarity threshold, these two region clusters are merged. For example, the preset similarity threshold is 0.8, the similarity of geological deformation features between region cluster G and region cluster H is 0.85, and the similarity of slope correlation features is 0.9. Then region cluster G and region cluster H are merged into a merged candidate region cluster.

[0143] Step S4527: According to the spatial boundary superposition degree and the surface coverage continuity of the merged candidate region clusters, perform boundary optimization processing on the merged candidate region clusters to eliminate the boundary conflict regions and generate smooth candidate regions.

[0144] The boundary optimization processing is to adjust the boundaries of the merged candidate region clusters to eliminate the boundary conflict regions and ensure the continuity of the surface coverage. The spatial boundary superposition degree reflects the overlapping situation of the boundaries between the merged candidate region clusters, and the surface coverage continuity reflects the coherence degree of the regions in terms of surface coverage types.

[0145] Specifically, when operating, a boundary smoothing algorithm such as the Laplace smoothing algorithm can be used. This algorithm smooths the boundary points of the merged candidate region clusters to reduce the abrupt changes of the boundaries. For the boundary conflict regions, they can be adjusted according to the spatial boundary superposition degree, and the overlapping parts are reasonably divided. At the same time, check the surface coverage continuity and correct the regions with inconsistent surface coverage types. For example, there is a boundary overlap between the merged candidate region cluster I and the merged candidate region cluster J. The boundary is smoothed by the Laplace smoothing algorithm, and the overlapping part is divided between the two region clusters according to a set ratio according to the spatial boundary superposition degree. At the same time, the regions with inconsistent surface coverage types are corrected to generate smooth candidate regions.

[0146] Step S4528: Perform feature matching verification processing on the smooth candidate regions and the landslide recognition feature set, and retain the regions whose matching degrees meet the preset landslide form constraint conditions to generate a set of candidate landslide regions.

[0147] The feature matching verification processing is to compare the smooth candidate regions with the landslide recognition feature set to check whether the features of the smooth candidate regions match the landslide recognition features. The preset landslide form constraint condition is a preset standard used to judge whether the smooth candidate regions have the characteristics of a landslide.

[0148] Specifically, the matching degree between the smoothed candidate region and each feature in the landslide recognition feature set can be calculated. The calculation of the matching degree can adopt a feature similarity calculation method, such as the cosine similarity calculation based on feature vectors. When the matching degree between the smoothed candidate region and the landslide recognition feature is higher than the preset landslide morphology constraint condition, it is considered that there may be a landslide in this region, and it is retained as a candidate landslide region. For example, the preset landslide morphology constraint condition is that the matching degree is greater than 0.7, and the matching degree between the smoothed candidate region K and a certain feature in the landslide recognition feature set is 0.8, then the smoothed candidate region K is added to the candidate landslide region set.

[0149] Step S453: Perform region merging verification processing on each candidate landslide region in the candidate landslide region set to obtain a verified merged region; the region merging verification processing includes region overlap degree verification and geological continuity verification.

[0150] The region merging verification processing is to further verify the candidate landslide regions in the candidate landslide region set to ensure that the merged region is reasonable in space and continuous geologically. The region overlap degree verification is to check the overlap situation between candidate landslide regions, and the geological continuity verification is to check whether the merged region is coherent in geological features.

[0151] Specifically, for every two candidate landslide regions in the candidate landslide region set, calculate their region overlap degree. The region overlap degree can be obtained by calculating the ratio of the overlapping area of the two regions to the total area. If the region overlap degree is higher than the preset overlap degree threshold, then consider merging these two regions. At the same time, perform geological continuity verification, and it can be judged whether the geology of the merged region is continuous by analyzing geological information such as geological deformation characteristics and slope correlation characteristics. For example, the preset overlap degree threshold is 0.3, the region overlap degree between the candidate landslide region L and the candidate landslide region M is 0.4, and the geological continuity verification passes, then the candidate landslide region L and the candidate landslide region M are merged into a verified merged region.

[0152] Step S454: Perform a secondary comparison process between the feature set corresponding to the verified merged region and the standard landslide morphology feature set to generate a secondary comparison similarity.

[0153] The secondary comparison process is to compare the feature set corresponding to the verified merged region with the standard landslide morphology feature set again to more accurately judge whether the verified merged region is a landslide region. The secondary comparison similarity is an index reflecting the similarity degree between the verified merged region and the standard landslide morphology feature set.

[0154] During specific operations, a similarity calculation method similar to that in step S430 can be adopted. Calculate the similarity between the geological deformation characteristics, slope correlation characteristics, surface coverage characteristics, etc. of the verification merged area and the corresponding characteristics in the standard landslide form feature set respectively, and then comprehensively calculate these similarities to obtain the secondary comparison similarity. For example, using the weighted average method, different weights are assigned to the geological deformation feature similarity, slope correlation similarity, and surface coverage similarity to calculate the secondary comparison similarity.

[0155] Step S455: If the secondary comparison similarity is greater than the preset secondary threshold, add the feature set corresponding to the verification merged area to the landslide recognition feature set.

[0156] The preset secondary threshold is a pre-set standard used to determine whether the verification merged area truly conforms to the characteristics of a landslide. When the secondary comparison similarity is greater than the preset secondary threshold, it is considered that the verification merged area is a landslide area, and the corresponding feature set is added to the landslide recognition feature set.

[0157] Specifically, compare the calculated secondary comparison similarity with the preset secondary threshold. For example, the preset secondary threshold is 0.7, and the secondary comparison similarity of the verification merged area N is 0.8. Then, add the feature set corresponding to the verification merged area N to the landslide recognition feature set.

[0158] Step S500: Perform regional annotation processing on the target remote sensing image dataset according to the landslide recognition feature set to generate a landslide area recognition result.

[0159] Regional annotation processing is to mark the areas where landslides may exist on the target remote sensing image dataset according to the landslide recognition feature set, thereby generating a landslide area recognition result. Through regional annotation, it can be intuitively shown which parts of the target area may have landslides, providing an important basis for subsequent geological disaster prevention and control.

[0160] As an implementation method, step S500, performing regional annotation processing on the target remote sensing image dataset according to the landslide recognition feature set to generate a landslide area recognition result, may specifically include the following steps S510 to S570: Step S510: Conduct a spatial continuity analysis on the geological deformation characteristics in the landslide recognition feature set to generate a deformation continuous area distribution map; the deformation continuous area distribution map is used to identify the target areas with consistent surface deformation trends and coherent spatial distributions.

[0161] Spatial continuity analysis is to analyze the geological deformation characteristics in the landslide recognition feature set to find the areas with consistent surface deformation trends and coherent spatial distributions. The deformation continuous area distribution map is a visual result that graphically shows these target areas with continuous deformation characteristics.

[0162] Specifically, a spatial clustering algorithm, such as the DBSCAN algorithm, can be used to cluster the geological deformation features. This algorithm clusters regions with similar deformation trends and adjacent in space into clusters according to the spatial distribution and similarity of the geological deformation features. For each cluster, it is marked on the target remote sensing image dataset to form a distribution map of continuous deformation regions. For example, in a target remote sensing image dataset of a mountainous area, the DBSCAN algorithm is used to cluster the geological deformation features, and it is found that the surface deformation trends of several regions are consistent and adjacent in space. These regions are marked on the image to generate a distribution map of continuous deformation regions.

[0163] Step S520: Perform an overlay analysis on the distribution map of continuous deformation regions and the slope-related features to extract candidate deformation regions where the slope change overlaps with the continuous deformation regions.

[0164] The overlay analysis overlaps the distribution map of continuous deformation regions with the slope-related features to find the overlapping parts between the slope change and the continuous deformation regions. These overlapping parts are the candidate deformation regions. The overlap between the slope change and the continuous deformation regions may imply that landslides are more likely to occur in this area.

[0165] In specific operations, the overlay analysis function of geographic information system (GIS) software can be used. Import the distribution map of continuous deformation regions and the slope-related feature layer into the GIS software for overlay operation. The software will automatically identify the regions where the slope change overlaps with the continuous deformation regions and extract them as candidate deformation regions. For example, in the GIS software, when the distribution map of continuous deformation regions and the slope-related feature layer are overlaid, it is found that there are some regions with both continuous surface deformation and obvious slope change, and these regions are extracted as candidate deformation regions.

[0166] Step S530: Perform texture pattern matching processing based on the texture-related features of the candidate deformation regions to generate a set of texture consistent regions; the regions in the set of texture consistent regions satisfy the preset landslide texture evolution law.

[0167] The texture pattern matching processing finds the regions that satisfy the preset landslide texture evolution law by analyzing the texture-related features of the candidate deformation regions. The preset landslide texture evolution law is the texture change characteristics summarized based on historical landslide data and geological knowledge.

[0168] Specifically, a template matching algorithm can be used for texture pattern matching. First, a texture template is constructed according to the preset landslide texture evolution law. Then, the texture correlation features of the candidate deformation region are matched with the texture template. When the matching degree between the texture features of the candidate deformation region and the texture template is higher than the preset matching threshold, it is considered that this region satisfies the landslide texture evolution law, and it is added to the texture consistency region set. For example, if the preset matching threshold is 0.7 and the matching degree between the texture features of the candidate deformation region P and the texture template is 0.8, then the candidate deformation region P is added to the texture consistency region set.

[0169] Step S540: Perform boundary closure degree verification processing on the texture consistency region set, and screen out the regions with the boundary turning point density lower than the preset closure threshold as the closure candidate regions.

[0170] The boundary closure degree verification processing is to check the boundary closure of each region in the texture consistency region set. The boundary turning point density is an index to measure the boundary closure degree, which represents the ratio of the number of turning points on the boundary to the boundary length. The preset closure threshold is a preset standard for judging whether the boundary of the region is closed.

[0171] Specifically, the boundary turning point density of each region in the texture consistency region set can be calculated. By analyzing the boundary of the region, the number of turning points on the boundary is counted, and the boundary length is calculated, so as to obtain the boundary turning point density. The regions with the boundary turning point density lower than the preset closure threshold are screened out as the closure candidate regions. For example, if the preset closure threshold is 0.1 and the boundary turning point density of the texture consistency region Q is 0.08, then the texture consistency region Q is used as the closure candidate region.

[0172] Step S550: Perform coverage type matching processing on the spatial range of the closure candidate region and the surface coverage features in the target remote sensing image dataset, exclude the interference regions with the coverage type not matching the landslide surface type, and generate the initial landslide annotation region.

[0173] The coverage type matching processing is to compare the spatial range of the closure candidate region with the surface coverage features in the target remote sensing image dataset, find out the regions with the coverage type not matching the landslide surface type, and exclude them. The landslide surface type is the surface coverage type related to landslides summarized according to geological knowledge and historical data.

[0174] Specifically, the spatial range of the closed candidate region can be compared with the land cover classification layer in the target remote sensing image dataset. For each closed candidate region, check whether its coverage type matches the landslide surface type. If not, exclude this region. For example, the landslide surface type is usually bare land or sparse vegetation. If the coverage type of a certain closed candidate region is dense forest, then exclude this region, and finally generate the initial landslide annotation region.

[0175] Step S560: Perform multi-scale context consistency correction processing on the initial landslide annotation region, fuse the boundary feature correlations at different bounding levels, and generate the corrected landslide annotation region.

[0176] The multi-scale context consistency correction processing corrects the initial landslide annotation region by comprehensively considering the boundary feature correlations at different bounding levels to improve the accuracy and consistency of the annotation. The boundary features at different bounding levels may contain landslide information at different scales. Fusing these features can more comprehensively describe the landslide area.

[0177] As an implementation manner, in step S560, performing multi-scale context consistency correction processing on the initial landslide annotation region, fusing the boundary feature correlations at different bounding levels, and generating the corrected landslide annotation region may specifically include the following steps S561 to S568: Step S561: Extract the bounding level identifier and the boundary feature correlation parameter corresponding to the initial landslide annotation region.

[0178] The bounding level identifier is used to indicate the bounding level to which the initial landslide annotation region belongs, and the boundary feature correlation parameter describes the boundary feature correlation relationship between this region and other bounding levels.

[0179] Specifically, extract the bounding level identifier and the boundary feature correlation parameter from the relevant data of the initial landslide annotation region. For example, by querying the database or data file, obtain that the bounding level identifier corresponding to the initial landslide annotation region is "level 3", and its boundary feature correlation parameters with other levels, such as the boundary overlap ratio with level 2.

[0180] Step S562: Obtain the boundary feature set of the adjacent level in the level bounding region set according to the bounding level identifier.

[0181] The adjacent level refers to the level adjacent to the bounding level to which the initial landslide annotation region belongs. Obtaining the boundary feature set of the adjacent level can provide more information for subsequent context correlation analysis.

[0182] During specific operations, based on the selected hierarchical identifier, search for adjacent hierarchies in the set of hierarchical selection areas. For example, if the selected hierarchy of the initial landslide annotation area is "Hierarchy 3", then search for the boundary feature sets of Hierarchy 2 and Hierarchy 4. For each hierarchical selection area of Hierarchy 2 and Hierarchy 4, extract its boundary features to form the boundary feature set of the adjacent hierarchies.

[0183] Step S563: Perform context correlation analysis on the boundary feature sets of adjacent hierarchies to generate cross-hierarchical boundary constraint conditions; the cross-hierarchical boundary constraint conditions are used to limit the spatial offset amplitude of boundary features at different hierarchies.

[0184] Context correlation analysis is to analyze the boundary feature sets of adjacent hierarchies, find the correlation relationships between boundary features at different hierarchies, and thus generate cross-hierarchical boundary constraint conditions. The cross-hierarchical boundary constraint conditions can ensure the spatial consistency of boundary features at different hierarchies and avoid excessive offsets. Specifically, machine learning algorithms such as support vector machines (SVM) can be used to analyze the boundary feature sets of adjacent hierarchies. This algorithm learns the relationships between boundary features and establishes a classification model to judge whether the spatial offset of boundary features at different hierarchies is reasonable. According to the output of the model, cross-hierarchical boundary constraint conditions are generated. For example, by analyzing the boundary feature sets of Hierarchy 2 and Hierarchy 3 using the SVM algorithm, it is found that the offset of the boundary features of Hierarchy 3 relative to the boundary features of Hierarchy 2 in space cannot exceed a set distance, and this distance is used as the cross-hierarchical boundary constraint condition.

[0185] Step S564: Dynamically adjust the boundary of the initial landslide annotation area based on the cross-hierarchical boundary constraint conditions to generate an adjusted boundary area.

[0186] The dynamic adjustment process is to adjust the boundary of the initial landslide annotation area according to the cross-hierarchical boundary constraint conditions to make it meet the spatial consistency requirements of boundary features at different hierarchies.

[0187] During specific operations, for each boundary point of the initial landslide annotation area, check whether it meets the cross-hierarchical boundary constraint conditions. If not, adjust the position of the boundary point according to the constraint conditions. For example, according to the cross-hierarchical boundary constraint conditions, the offset of the boundary features of Hierarchy 3 relative to the boundary features of Hierarchy 2 cannot exceed 5 meters. For a certain boundary point of the initial landslide annotation area, it is found that its offset exceeds 5 meters, and the boundary point is moved in the direction that meets the constraint conditions to generate an adjusted boundary area.

[0188] Step S565: Perform geological continuity verification processing on the adjusted boundary area to detect the distribution of terrain mutation points and fault zones inside the area.

[0189] The geological continuity verification process is to examine the geological conditions within the adjusted boundary area and identify the distribution of terrain mutation points and fault zones. The existence of terrain mutation points and fault zones may affect the stability of the landslide area, so detection is required.

[0190] Specifically, existing terrain analysis algorithms such as slope analysis and curvature analysis can be used to analyze the adjusted boundary area. Slope analysis can identify areas with significant changes in terrain slope, and curvature analysis can detect the degree of terrain curvature. Through these analyses, the locations of terrain mutation points and fault zones can be found. For example, in the adjusted boundary area, if a sudden increase in slope is found in a certain area through slope analysis, that area is marked as a terrain mutation point.

[0191] Step S566: Perform regional segmentation processing on the adjusted boundary area according to the distribution of terrain mutation points and fault zones to generate multiple sub-areas.

[0192] Regional segmentation processing is to divide the adjusted boundary area into multiple sub-areas according to the distribution of terrain mutation points and fault zones. Such segmentation can more accurately describe the geological structure of the landslide area and provide more detailed information for subsequent processing.

[0193] During specific operation, using the terrain mutation points and fault zones as boundaries, the adjusted boundary area is divided into multiple parts. For example, in the adjusted boundary area, if a fault zone is found to pass through, the area is divided into two sub-areas along the fault zone.

[0194] Step S567: Perform secondary matching processing on the surface coverage characteristics and texture correlation characteristics of each sub-area independently, and select the sub-areas with a matching degree higher than the preset consistency sub-threshold as effective landslide sub-areas.

[0195] Secondary matching processing is to perform re-matching on the surface coverage characteristics and texture correlation characteristics of each sub-area to check whether they conform to the landslide characteristics. The preset consistency sub-threshold is a pre-set standard used to determine whether a sub-area is an effective landslide sub-area.

[0196] Specifically, the matching degree between the surface coverage characteristics and texture correlation characteristics of each sub-area and the landslide characteristics can be calculated. The calculation of the matching degree can adopt feature similarity calculation methods such as cosine similarity calculation based on feature vectors. When the matching degree of a sub-area is higher than the preset consistency sub-threshold, that sub-area is considered an effective landslide sub-area. For example, the preset consistency sub-threshold is 0.6, and the matching degree between the surface coverage characteristics and texture correlation characteristics of sub-area R and the landslide characteristics is 0.7, then sub-area R is taken as an effective landslide sub-area.

[0197] Step S568: Topologically merge the spatial ranges of the effective landslide sub-regions, eliminate the overlaps and gaps, and generate a corrected landslide annotation region.

[0198] Topological merging processing is to merge the spatial ranges of the effective landslide sub-regions, eliminate the overlaps and gaps, and make the corrected landslide annotation region more continuous and reasonable in space.

[0199] During specific operations, topological operation methods, such as union operation, can be used to merge the spatial ranges of the effective landslide sub-regions. For the overlapping parts, only one copy is retained; for the gap parts, they are filled. For example, there are two effective landslide sub-regions S and T, and their spatial ranges partially overlap. Through the union operation, they are merged into a continuous region, the overlapping parts are eliminated, and a corrected landslide annotation region is generated.

[0200] Step S570: Generate a landslide region identification result according to the spatial coordinate mapping relationship between the corrected landslide annotation region and the target remote sensing image dataset.

[0201] The spatial coordinate mapping relationship refers to the coordinate correspondence between the corrected landslide annotation region and the target remote sensing image dataset. Through this mapping relationship, the corrected landslide annotation region can be accurately marked on the target remote sensing image dataset to generate the final landslide region identification result.

[0202] Specifically, according to the spatial coordinates of the corrected landslide annotation region, find the corresponding positions on the target remote sensing image dataset and mark the regions. For example, the coordinates of the corrected landslide annotation region are from (x1, y1) to (x2, y2). Find the corresponding pixel range on the target remote sensing image dataset and mark this pixel range as the landslide region to generate the landslide region identification result.

[0203] As an implementation manner, after generating the landslide region identification result, the method provided by the embodiment of the present invention may further include the following steps S580 to S5130: Step S580: Perform feature backtracking verification processing on the target regions in the landslide region identification result, and extract the corresponding fusion boundary feature set and landslide identification feature set of the target regions.

[0204] Feature backtracking verification processing is to recheck the target regions in the landslide region identification result and extract the corresponding fusion boundary feature set and landslide identification feature set. This helps to further confirm whether the target regions are truly landslide regions and improve the accuracy of the identification result.

[0205] Specifically, for each target area in the landslide area recognition result, search for its corresponding fusion boundary feature set and landslide recognition feature set in the relevant data recorded in the previous processing. For example, by querying a database or a data file, obtain the fusion boundary feature set and landslide recognition feature set corresponding to target area A.

[0206] Step S590: Perform a secondary dynamic matching process on the fusion boundary feature set and the standard landslide morphology feature set in the landslide morphology feature library to generate a regional matching confidence set.

[0207] The secondary dynamic matching process is to compare the fusion boundary feature set with the standard landslide morphology feature set again, calculate the matching degree between the fusion boundary features of each target area and the standard landslide morphology features, so as to generate a regional matching confidence set. The regional matching confidence set reflects the similarity degree between each target area and the standard landslide morphology features.

[0208] Specifically, when operating, a method similar to the previous feature comparison process can be used to calculate the similarity between the fusion boundary feature set and the standard landslide morphology feature set. For example, use the cosine similarity calculation method based on feature vectors. For the fusion boundary features of each target area, calculate the cosine similarity between it and each feature in the standard landslide morphology feature set. These similarities are comprehensively calculated to obtain the regional matching confidence of this target area. For all target areas, the calculated regional matching confidences are composed into a regional matching confidence set.

[0209] Step S5100: Based on the regional matching confidence set, perform a screening process on the target areas in the landslide area recognition result, remove the misjudged areas with a confidence lower than the preset backtracking threshold, and generate an optimized landslide area set.

[0210] The screening process is to screen the target areas in the landslide area recognition result according to the regional matching confidence set, consider the target areas with a confidence lower than the preset backtracking threshold as misjudged areas, and remove them, so as to generate an optimized landslide area set. The preset backtracking threshold is a pre-set standard for judging whether a target area is a real landslide area.

[0211] Specifically, compare each regional matching confidence in the regional matching confidence set with the preset backtracking threshold. For example, the preset backtracking threshold is 0.7. For target area B, its regional matching confidence is 0.6. Then remove target area B from the landslide area recognition result. The remaining target areas form an optimized landslide area set.

[0212] Step S5110: Analyze the spatial topological relationship of adjacent areas in the optimized landslide area set, and extract the regional boundary overlap degree and geological continuity parameters.

[0213] Spatial topological relationship analysis is to analyze and optimize the spatial position relationship of adjacent regions in the landslide area set, and extract the regional boundary overlap degree and geological continuity parameters. The regional boundary overlap degree reflects the overlap of adjacent regional boundaries, and the geological continuity parameter describes the coherence degree of adjacent regions in geological characteristics.

[0214] In specific operations, the spatial analysis function of Geographic Information System (GIS) software can be used to analyze adjacent regions in the optimized landslide area set. For each pair of adjacent regions, calculate the ratio of the overlapping area of their boundaries to the total area to obtain the regional boundary overlap degree. At the same time, analyze geological information such as the geological deformation characteristics and slope correlation characteristics of adjacent regions, and calculate the geological continuity parameters. For example, by analyzing adjacent regions C and D through GIS software, the calculated regional boundary overlap degree is 0.2 and the geological continuity parameter is 0.8.

[0215] Step S5120: Perform regional merging processing on the optimized landslide area set according to the regional boundary overlap degree and geological continuity parameters to generate the merged landslide area.

[0216] Regional merging processing is to merge adjacent regions with a high overlap degree and geological continuity according to the regional boundary overlap degree and geological continuity parameters to generate the merged landslide area. Such merging can make the division of the landslide area more reasonable and reduce unnecessary segmentation.

[0217] Specifically, for each pair of adjacent regions in the optimized landslide area set, compare the regional boundary overlap degree and geological continuity parameters with the preset merging thresholds. For example, the preset regional boundary overlap degree merging threshold is 0.1, and the geological continuity parameter merging threshold is 0.7. If the regional boundary overlap degree of adjacent regions E and F is 0.2 and the geological continuity parameter is 0.8, meeting the merging threshold conditions, then merge regions E and F into a merged landslide area.

[0218] Step S5130: Perform coverage type cross-validation processing on the landslide identification feature set of the merged landslide area and the surface coverage features of the target remote sensing image dataset, exclude the regions with coverage type conflicts, and generate the final landslide area identification result.

[0219] The coverage type cross - validation process is to compare the landslide recognition feature set of the merged landslide area with the surface coverage features of the target remote sensing image dataset, check whether the coverage type of the merged landslide area matches the landslide surface type, exclude the areas with coverage type conflicts, and generate the final landslide area recognition result. Specifically, for each merged landslide area, the surface coverage features in its landslide recognition feature set are compared with the surface coverage classification layer of the target remote sensing image dataset. If it is found that the coverage type does not match the landslide surface type, that area is excluded. For example, the landslide surface type is usually bare land or sparse vegetation, and if the coverage type of a certain merged landslide area is water body, that area is excluded. The remaining merged landslide areas are composed of the final landslide area recognition result.

[0220] In summary, the landslide intelligent recognition method based on hierarchical box selection and boundary feature fusion proposed by the present invention can accurately identify the landslide area through a series of steps such as hierarchical box selection processing, boundary feature extraction and fusion, feature comparison, and region annotation of the target remote sensing image dataset, providing effective technical support for the prevention and control of geological disasters. In each step, specific technical means and algorithms are adopted to ensure the feasibility and accuracy of the method. At the same time, through multiple verifications and corrections, the reliability of the landslide area recognition result is further improved.

[0221] It can be understood that in the above introductions of each embodiment of the present invention, various algorithms involved, such as the Euclidean distance algorithm, terrain analysis algorithm, DBSCAN algorithm, etc., can be obtained from relevant contents in the prior art. For the sake of saving space, they are not elaborated in the embodiments of the present invention. In addition, those skilled in the art can make detailed supplements according to the common general knowledge in the art when implementing the solution of the present invention. For example, according to the general knowledge in the art, normalization can be used to eliminate the dimension conflict before feature fusion, interpolation can be used to eliminate the dimension difference, historical data, experience or business scenario requirements can be combined for reasonable setting of thresholds, the model can be trained based on the general model training method, the number of layers in the model structure can be set according to actual needs, the activation function can be selected, etc. The present invention will no longer give redundant introductions to the overly detailed implementation process.

[0222] Please refer to Figure 2 , Figure 2Schematic structural diagram of a computer system provided by an embodiment of the present invention. The computer system at least includes a processor 101, a communication interface 102, and a memory 103. Among them, the processor 101, the communication interface 102, and the memory 103 can be connected through a bus or other means. Among them, the processor 101 (or central processing unit (CPU)) is the computing core and control core of the computer system, which can parse various instructions in the computer system and process various data in the computer system. The communication interface 102 can optionally include a standard wired interface, a wireless interface (such as WI-FI, a mobile communication interface, etc.), and can be used to send and receive data under the control of the processor 101; the communication interface 102 can also be used for the transmission and interaction of internal data in the computer system. The memory 103 (Memory) is a memory device in the computer system, used to store programs and data. It can be understood that the memory 103 here can include both the built-in memory of the computer system and, of course, the extended memory supported by the computer system. The memory 103 provides a storage space, and the operating system of the computer system is stored in this storage space, which is not limited in the present invention. In one embodiment, the processor 101 executes the landslide intelligent recognition method based on hierarchical box selection and boundary feature fusion provided above in the embodiments of the present invention by running the computer program in the memory 103.

Claims

1. A landslide intelligent recognition method based on hierarchical box selection and boundary feature fusion, characterized in that Including the following steps: Obtain a target remote sensing image dataset, and perform hierarchical bounding box selection processing on the target remote sensing image dataset to obtain a set of hierarchical bounding box regions; each of the hierarchical bounding box regions corresponds to a bounding box level identifier; Perform boundary feature extraction processing on the set of hierarchical bounding box regions to obtain a boundary feature set for each of the hierarchical bounding box regions; the boundary feature set includes spatial distribution features, geometric shape features, and texture correlation features; Input the boundary feature set into a pre-trained boundary feature fusion model for feature fusion processing to generate a fused boundary feature set; the fused boundary feature set is used to characterize the boundary feature correlation between different bounding box levels; Perform feature comparison processing based on the fused boundary feature set and a preset landslide form feature library to generate a landslide recognition feature set; the landslide recognition feature set includes geological deformation features, slope correlation features, and surface cover features of the target area; Perform area annotation processing on the target remote sensing image dataset according to the landslide recognition feature set to generate a landslide area recognition result.

2. The method according to claim 1, wherein The obtaining of the target remote sensing image dataset and the performing of hierarchical bounding box selection processing on the target remote sensing image dataset to obtain a set of hierarchical bounding box regions includes: Perform multi-scale segmentation processing on the target remote sensing image dataset to obtain a plurality of initial segmentation regions; each initial segmentation region corresponds to a segmentation scale identifier; Perform region merging processing on the plurality of initial segmentation regions according to the segmentation scale identifier to generate a set of merged regions; each merged region in the set of merged regions includes at least two initial segmentation regions; Perform region screening processing on the set of merged regions to obtain a set of candidate regions; each candidate region in the set of candidate regions meets a preset landslide candidate condition; the landslide candidate condition includes a region area threshold, regional terrain continuity, and regional texture consistency; Perform hierarchical division processing on the set of candidate regions according to the distribution density and regional overlap degree of the set of candidate regions to generate the set of hierarchical bounding box regions; each hierarchical bounding box region corresponds to a bounding box level identifier, and the bounding box level identifier is used to indicate the spatial division priority of the hierarchical bounding box region at different segmentation scales.

3. The method according to claim 2, wherein The performing of multi-scale segmentation processing on the target remote sensing image dataset to obtain a plurality of initial segmentation regions includes: Obtain the image resolution and surface cover type distribution information of the target remote sensing image dataset; Determine the segmentation scale range according to the image resolution, and select a plurality of candidate segmentation scales within the segmentation scale range based on the surface cover type distribution information; Perform image segmentation processing on the target remote sensing image dataset at each candidate segmentation scale to generate initial segmentation regions; the image segmentation processing includes superpixel segmentation processing and edge detection processing; Perform scale alignment processing on the initial segmentation regions generated at different candidate segmentation scales to obtain the plurality of initial segmentation regions; each initial segmentation region corresponds to a segmentation scale identifier and a corresponding set of segmentation parameters.

4. The method according to claim 1, wherein Inputting the boundary feature set into a pre-trained boundary feature fusion model for feature fusion processing to generate a fused boundary feature set includes: Performing multi-level feature correlation analysis on the spatial distribution features in the boundary feature set to generate a first correlation feature sequence; the first correlation feature sequence is used to characterize the spatial distribution correlation strength between different box selection levels; Performing morphological similarity measurement on the geometric shape features in the boundary feature set to generate a geometric shape similarity map; the geometric shape similarity map is used to indicate the geometric shape difference degree of different box selection levels; Performing cross-feature matching based on the first correlation feature sequence and the geometric shape similarity map to generate a dynamic weight adjustment coefficient; the dynamic weight adjustment coefficient is used to balance the influence weights of the spatial distribution and geometric shape of different box selection levels on the fusion result; Performing adaptive weighted aggregation processing on the texture correlation features in the boundary feature set according to the dynamic weight adjustment coefficient to generate an initial fused texture feature; Performing feature splicing processing on the initial fused texture feature and the first correlation feature sequence to generate a cross-level splicing feature; Performing non-linear mapping processing on the cross-level splicing feature to eliminate feature redundancy and enhance the cross-level feature expression consistency to generate the fused boundary feature set.

5. The method according to claim 4, characterized in that, The boundary feature fusion model is trained through the following steps: Obtaining a historical remote sensing image sample set and a corresponding labeled landslide area set; performing hierarchical box selection processing on the historical remote sensing image sample set to obtain a sample hierarchical box selection area set; performing boundary feature extraction processing on the sample hierarchical box selection area set to obtain a sample boundary feature set; Initializing a feature fusion network structure; the feature fusion network structure includes a multi-level feature correlation module, a dynamic weight generation module, and a non-linear mapping module; Inputting the sample boundary feature set into the multi-level feature correlation module in the feature fusion network structure to generate a sample first correlation feature sequence; Inputting the sample first correlation feature sequence into the dynamic weight generation module, and combining the sample geometric shape features and texture correlation features to generate a sample dynamic weight adjustment coefficient; Performing weighted aggregation processing on the sample texture correlation features based on the sample dynamic weight adjustment coefficient to generate a sample initial fused texture feature; Inputting the sample initial fused texture feature and the sample first correlation feature sequence into the non-linear mapping module to generate a predicted fusion feature set; Extracting the standard spatial distribution features, standard geometric shape features, and standard texture correlation features in the labeled landslide area set to construct a standard fusion feature template; Calculating the distribution difference degree, shape matching degree, and texture consistency index between the predicted fusion feature set and the standard fusion feature template; Generating a model optimization loss value according to the weighted combination result of the distribution difference degree, shape matching degree, and texture consistency index; Based on the optimized loss value of the model, jointly iteratively optimize the network parameters of the multi-level feature association module, dynamic weight generation module, and non-linear mapping module until the feature deviation between the predicted fusion feature set and the standard fusion feature template is less than a preset optimization threshold, and obtain a trained boundary feature fusion model.

6. The method according to claim 1, wherein, Performing feature comparison processing on the fusion boundary feature set and a preset landslide morphology feature library to generate a landslide recognition feature set, including: Obtain a standard landslide morphology feature set from the landslide morphology feature library; the standard landslide morphology feature set includes geological deformation features, slope correlation features, and surface coverage features of typical landslide areas; Perform feature analysis processing on each fusion boundary feature in the fusion boundary feature set to extract target geological deformation parameters, target slope parameters, and target surface coverage parameters; Calculate the similarity between the target geological deformation parameters, target slope parameters, and target surface coverage parameters and the corresponding parameters in the standard landslide morphology feature set respectively to generate geological deformation similarity, slope correlation similarity, and surface coverage similarity; Determine the landslide recognition confidence corresponding to the fusion boundary feature according to the weighted sum result of the geological deformation similarity, slope correlation similarity, and surface coverage similarity; Determine the feature set of the target area corresponding to the fusion boundary feature with the landslide recognition confidence greater than the preset confidence threshold as the landslide recognition feature set.

7. The method according to claim 6, characterized in that, The determining the feature set of the target area corresponding to the fusion boundary feature with the landslide recognition confidence greater than the preset confidence threshold as the landslide recognition feature set includes: Obtain the box selection level identifier and spatial position information corresponding to the fusion boundary feature set; Determine the spatial division priority of the target area according to the box selection level identifier, and perform spatial clustering processing on the target area according to the spatial position information to generate a candidate landslide area set; Perform region merging and verification processing on each candidate landslide area in the candidate landslide area set to obtain a verified merged area; the region merging and verification processing includes region overlap degree verification and geological continuity verification; Perform secondary comparison processing on the feature set corresponding to the verified merged area and the standard landslide morphology feature set to generate a secondary comparison similarity; If the secondary comparison similarity is greater than the preset secondary threshold, add the feature set corresponding to the verified merged area to the landslide recognition feature set.

8. The method according to claim 1, wherein The performing region annotation processing on the target remote sensing image data set according to the landslide recognition feature set to generate a landslide area recognition result includes: Perform spatial continuity analysis on the geological deformation features in the landslide recognition feature set to generate a deformation continuous area distribution map; the deformation continuous area distribution map is used to identify the target area with consistent surface deformation trend and continuous spatial distribution; Perform overlay analysis on the deformation continuous area distribution map and the slope correlation features to extract candidate deformation areas where the slope change overlaps with the deformation continuous area; Perform texture pattern matching processing based on the texture correlation features of the candidate deformation regions to generate a set of texture consistency regions; the regions in the set of texture consistency regions satisfy the preset landslide texture evolution law; Perform boundary closure degree verification processing on the set of texture consistency regions, and screen out the regions with the boundary turning point density lower than the preset closure threshold as the closure candidate regions; Perform coverage type matching processing on the spatial range of the closure candidate regions and the surface coverage features in the target remote sensing image dataset, exclude the interference regions with the coverage type not matching the landslide surface type, and generate the initial landslide annotation regions; Perform multi-scale context consistency correction processing on the initial landslide annotation regions, fuse the boundary feature correlations at different box selection levels, and generate the corrected landslide annotation regions; Generate the landslide region recognition result according to the spatial coordinate mapping relationship between the corrected landslide annotation regions and the target remote sensing image dataset.

9. The method according to claim 8, wherein The performing multi-scale context consistency correction processing on the initial landslide annotation regions, fusing the boundary feature correlations at different box selection levels, and generating the corrected landslide annotation regions includes: Extract the box selection level identifier and the boundary feature correlation parameters corresponding to the initial landslide annotation regions; Obtain the boundary feature sets of adjacent levels in the hierarchical box selection region set according to the box selection level identifier; Perform context correlation analysis on the boundary feature sets of the adjacent levels to generate cross-level boundary constraint conditions; the cross-level boundary constraint conditions are used to limit the spatial offset amplitude of the boundary features at different levels; Perform dynamic adjustment processing on the boundaries of the initial landslide annotation regions based on the cross-level boundary constraint conditions to generate the adjusted boundary regions; Perform geological continuity verification processing on the adjusted boundary regions to detect the distribution of terrain mutation points and fault zones inside the regions; Perform region segmentation processing on the adjusted boundary regions according to the distribution of the terrain mutation points and fault zones to generate multiple sub-regions; Perform secondary matching processing on the surface coverage features and texture correlation features of each sub-region independently, and screen out the sub-regions with the matching degree higher than the preset consistency sub-threshold as the effective landslide sub-regions; Perform topological merging processing on the spatial ranges of the effective landslide sub-regions to eliminate overlaps and gaps, and generate the corrected landslide annotation regions.

10. A computer system, characterized in that, Including: A memory, in which a computer program is stored; A processor, configured to load the computer program to implement the landslide intelligent recognition method based on hierarchical box selection and boundary feature fusion according to any one of claims 1-9.

Images