A machine vision-based in-service slope inspection method and system

Abstract

The application provides a kind of in-service side slope inspection method and system based on machine vision, by acquiring the continuous acquisition visible light image data and synchronous acquisition infrared thermal imaging data of in-service side slope, and three-dimensional point cloud data, composite data set is constructed, multidimensional information extraction processing is carried out, the texture change sequence of in-service side slope surface and the temperature gradient distribution corresponding to infrared thermal imaging data are obtained. Input the two into the pre-trained transfer learning model, combine the side slope instability case data in the historical disaster database to optimize the model, output the instability risk probability distribution of in-service side slope. The instability risk probability distribution is matched with the real-time acquired meteorological data in time and space, the influence weight of rainfall and temperature change on the stability of in-service side slope is quantified through the meteorological and geological coupling model, the deformation prediction curve of in-service side slope is updated, based on the matching analysis with the disaster threshold library in the historical disaster database, the side slope maintenance decision table is generated.

Description

Technical Field

[0001] This invention relates to the field of machine vision technology, and in particular to a machine vision-based method and system for inspecting in-service slopes. Background Technology

[0002] In-service slope inspection is a crucial aspect of ensuring the safety of geotechnical engineering projects. Continuous monitoring and analysis of slope surface conditions can promptly identify potential instability risks and support maintenance decisions. Currently, machine vision technology has been gradually applied to slope inspection, using collected slope image data and feature extraction algorithms to assess surface conditions. Some solutions attempt to integrate multi-source sensor data to improve the comprehensiveness of analysis; however, there is still room for improvement in the depth of data correlation and the guarantee of spatiotemporal consistency. Furthermore, in terms of the scenario adaptability of risk assessment models and the quantification of the dynamic impact of external environmental factors, existing technologies struggle to fully integrate historical disaster case experience with real-time monitoring data to achieve accurate predictions. This results in insufficient sensitivity in capturing early signs of slope instability, and the timeliness and reliability of deformation trend prediction need to be improved. Summary of the Invention

[0003] In view of this, the present invention provides a method and system for inspecting in-service slopes based on machine vision. The technical solution of the embodiments of the present invention is implemented as follows:

[0004] On one hand, embodiments of the present invention provide a machine vision-based method for inspecting in-service slopes. The method includes: acquiring continuously acquired visible light image data and synchronously acquired infrared thermal imaging data of the in-service slope, as well as three-dimensional point cloud data with the same acquisition timestamp as the visible light image data and the infrared thermal imaging data, constructing a composite dataset with spatiotemporal alignment; performing multi-dimensional information extraction processing on the composite dataset to obtain a texture change sequence of the in-service slope surface and a temperature gradient distribution corresponding to the infrared thermal imaging data; and inputting the texture change sequence and the temperature gradient distribution into a pre-training process. The transfer learning model is optimized by combining slope instability case data from a historical disaster database, and outputs the instability risk probability distribution of the in-service slope. The instability risk probability distribution is spatiotemporally matched with real-time meteorological data, and the influence weights of rainfall and temperature changes on the stability of the in-service slope are quantified through a meteorological-geological coupling model to update the deformation prediction curve of the in-service slope. Based on the deformation prediction curve and the disaster threshold library in the historical disaster database, a slope maintenance decision table containing displacement critical values, risk levels, and maintenance priorities is generated.

[0005] On the other hand, embodiments of the present invention provide a computer system including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the program to implement the steps in the above-described method.

[0006] This invention provides a machine vision-based method for inspecting in-service slopes. By constructing a composite dataset with spatiotemporal alignment, it achieves deep collaboration between visible light image data, infrared thermal imaging data, and 3D point cloud data in both temporal and spatial dimensions. This transforms multi-source heterogeneous data from independent entities into a mutually corroborating organic whole, enabling the simultaneous capture of the microscopic texture evolution, spatial differences in temperature fields, and macroscopic geometric features of the slope surface. This provides multi-scale, multi-physics field information support for slope instability risk assessment. The texture change sequence and temperature gradient distribution obtained by multi-dimensional information extraction from the composite dataset, through time series analysis and spatial gradient calculation, accurately characterize the dynamic texture evolution trend and 3D temperature field gradient direction of the slope surface. This method can capture gradual deformation precursors and structural surface temperature anomalies that are difficult to detect in traditional inspections, providing key feature-level evidence for early instability risk identification. Texture change sequences and temperature gradient distributions are input into a pre-trained transfer learning model and optimized using historical disaster case data. The transfer learning algorithm transfers feature patterns from historical slope instability cases to the current inspection scenario, enabling the model to quickly adapt to specific slope geological conditions and instability patterns. This improves the adaptability of the instability risk probability distribution to complex geological environments and enhances the scenario-specificity and reliability of risk prediction. A meteorological-geological coupling model quantifies the impact weights of rainfall and temperature changes on slope stability, dynamically linking real-time meteorological data with slope geological features. This achieves a quantitative expression of the impact of meteorological factors on slope stability, allowing the deformation prediction curve to adjust in real-time according to meteorological conditions, improving the prediction accuracy of dynamic slope instability processes. A slope maintenance decision table generated based on the deformation prediction curve and a historical disaster threshold database transforms abstract prediction data into structured decision information containing displacement critical values, risk levels, and maintenance priorities. This directly connects to engineering maintenance needs, providing technical support for the precise allocation of maintenance resources and the dynamic adjustment of risk control strategies, thus improving the intelligence and refinement of in-service slope safety management. Attached Figure Description

[0007] Figure 1 This is a schematic diagram illustrating the implementation process of an in-service slope inspection method based on machine vision, provided in an embodiment of the present invention.

[0008] Figure 2 This is a schematic diagram of the hardware entity of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0009] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The described embodiments should not be regarded as limitations on the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0010] This invention provides a machine vision-based method for inspecting in-service slopes, which can be executed by a computer system's processor. The computer system can refer to devices with data processing capabilities, such as servers, laptops, tablets, and desktop computers.

[0011] Figure 1 This is a schematic diagram illustrating the implementation process of a machine vision-based in-service slope inspection method according to an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes:

[0012] Step S100: Acquire continuously acquired visible light image data and synchronously acquired infrared thermal imaging data of the in-service slope, as well as three-dimensional point cloud data with the same acquisition timestamp as the visible light image data and infrared thermal imaging data, and construct a composite dataset with spatiotemporal alignment.

[0013] Slopes in service refer to slopes that are currently in actual use. Continuous acquisition of visible light image data involves continuously collecting visible light images of the slope over a continuous period of time. This data can visually present the texture, color, shape, and other appearance characteristics of the slope surface. Simultaneously acquired infrared thermal imaging data, obtained at the same time as the visible light image data, uses infrared thermal imaging technology to acquire the temperature distribution of the slope surface. Different temperature distributions may indicate structural changes or potential hazards within the slope. Three-dimensional point cloud data, acquired through technologies such as 3D laser scanning, records the coordinate information of various points on the slope surface in three-dimensional space. It shares the same acquisition timestamp as the visible light image data and infrared thermal imaging data, meaning that all three types of data were collected simultaneously from the same slope area at the same time.

[0014] When constructing a composite dataset with spatiotemporal alignment, the three types of collected data are first calibrated with timestamps. Then, these data are integrated according to a set format. For example, they can be stored in a database with timestamps and spatial coordinates as indexes to facilitate subsequent queries and processing.

[0015] Step S200: Perform multi-dimensional information extraction processing on the composite dataset to obtain the texture change sequence of the in-service slope surface and the temperature gradient distribution corresponding to the infrared thermal imaging data.

[0016] In one implementation, step S200 may specifically include the following steps S210 to S260:

[0017] Step S210: Construct a spatiotemporal index for the continuously acquired visible light image data, synchronously acquired infrared thermal imaging data, and 3D point cloud data contained in the composite dataset. Add a unified timestamp and absolute coordinate label to each set of data to generate a 3D data volume with spatiotemporal identifiers. Each data unit in the 3D data volume contains the visible light pixel value, infrared temperature value, and point cloud 3D coordinates at the corresponding time coordinate.

[0018] Spatiotemporal indexes are constructed to facilitate rapid querying and location of data within composite datasets. Uniform timestamps assign the same format and standard time identifier to all data, ensuring temporal comparability between different data types. Absolute coordinate labels assign absolute location information in three-dimensional space to each data point, giving the data a clear spatial location. A three-dimensional data volume with spatiotemporal identifiers is a data structure integrating temporal and spatial information. Each data unit contains visible light pixel values, infrared temperature values, and point cloud three-dimensional coordinates at a specific time and spatial location, comprehensively reflecting the characteristics of in-service slopes at that moment and location.

[0019] When constructing a spatiotemporal index, spatial indexing algorithms, such as quadtrees and octrees, can be used to divide the three-dimensional space into different regions, with each region corresponding to an index node. Then, the data is assigned to the appropriate index node based on its absolute coordinates. For adding timestamps, a unified time format, such as the ISO 8601 standard, can be used to convert the acquisition time to this format and add it to the data. When generating a three-dimensional data volume, the processed data can be arranged in temporal and spatial order to form a three-dimensional data matrix, where each element corresponds to a data unit.

[0020] Step S220: Perform multimodal spatial registration on the continuously acquired visible light image data and the synchronously acquired infrared thermal imaging data in the 3D data volume. Establish the coordinate mapping relationship between visible light pixels and infrared pixels through feature point matching. Combine the geometric constraints of the 3D point cloud data to correct the registration deviation and generate a spatially consistent registration data pair. The spatial resolution of the registration data pair is adapted to the sampling density of the 3D point cloud data.

[0021] Multimodal spatial registration aligns data from different modalities (such as visible light image data and infrared thermal imaging data) spatially, ensuring they accurately correspond to the same physical location. Feature point matching involves finding points with similar features in visible light and infrared thermal imaging, such as corner points and edge points, and then establishing correspondences between these feature points to determine the coordinate mapping between visible light and infrared pixels. Geometric constraints on 3D point cloud data utilize the geometric shape information of the slope surface provided by the 3D point cloud data to constrain the registration process and correct potential registration deviations. Spatially consistent registered data pairs are those where, after registration processing, the visible light image data and infrared thermal imaging data have a good spatial correspondence, and their spatial resolution is compatible with the sampling density of the 3D point cloud data, thus ensuring the accuracy and consistency of subsequent analysis.

[0022] When performing multimodal spatial registration, feature-based registration algorithms such as SIFT (Scale Invariant Feature Transform) and SURF (Accelerated Robust Feature Transform) can be used. First, feature points are extracted from visible light and infrared thermal images. Then, feature points are matched using feature descriptors. During the matching process, some incorrect matching points may occur. In such cases, geometric constraints of the 3D point cloud data can be used for filtering and correction. For example, the spatial positions of feature points are calculated based on the 3D point cloud data, and then it is determined whether the spatial distance between matching points meets the geometric constraints. If not, the matching points are discarded. Finally, the coordinate mapping relationship between visible light pixels and infrared pixels is established based on the matching results, generating registration data pairs.

[0023] Step S230: Dynamic window texture extraction is performed on the continuously acquired visible light image data in the registration data pair. The image is traversed with a preset spatial window, and the amplitude and direction of the gray value change within the window are calculated. The texture feature evolution of the same window in the continuous acquisition period is tracked in combination with the timestamp information to generate an initial texture dynamic feature set. The initial texture dynamic feature set contains the texture change direction and change amplitude of each window.

[0024] Dynamic window texture extraction is a method for extracting texture features from image data. It involves sliding a pre-sized spatial window across the image and analyzing the texture features within each window. The pre-sized spatial window is a rectangular area of ​​fixed size, set according to actual needs and image resolution. The magnitude and direction of grayscale value changes represent the degree and direction of pixel grayscale value changes in different directions within the window, reflecting the texture characteristics within the window. Combining timestamp information to track the evolution of texture features within the same window over consecutive acquisition periods allows for comparison and analysis of texture features within the same window at different time points to understand how the texture changes over time. The initial dynamic texture feature set is a dataset containing the direction and magnitude of texture changes for each window, providing a foundation for subsequent texture analysis.

[0025] Step S240: Calculate the neighborhood temperature difference of the synchronously acquired infrared thermal imaging data in the registration data pair. Taking each pixel as the center, calculate the temperature difference between it and the surrounding neighboring pixels. Combine the elevation information of the three-dimensional point cloud data to correct the influence of slope aspect on temperature distribution and generate a preliminary temperature gradient field. The spatial range of the preliminary temperature gradient field covers the entire surface of the in-service slope.

[0026] The surrounding neighbor pixels are pixels within a certain range adjacent to the center pixel. This range can be set according to actual needs, such as 3×3 or 5×5. The elevation information of the 3D point cloud data reflects the topographic undulations of the slope surface. The slope aspect is the orientation of the slope; different slope aspects receive varying degrees of sunlight and heat radiation, thus affecting temperature distribution. Combining elevation information with slope aspect correction to eliminate interference from slope aspect factors in temperature gradient calculation is to ensure the temperature gradient field more accurately reflects temperature changes within the slope. The preliminary temperature gradient field is a dataset describing the temperature gradient distribution on the surface of an in-service slope, spatially covering the entire slope surface.

[0027] When calculating neighborhood temperature differences, for each pixel, its surrounding pixels are traversed, and the temperature difference between them is calculated. A subtraction operation can be used, subtracting the temperatures of the neighboring pixels from the temperature of the center pixel to obtain the temperature difference. When correcting the influence of slope aspect on temperature distribution using elevation information from 3D point cloud data, the slope aspect of each pixel's location can first be calculated based on the elevation information. Then, the temperature at that location can be corrected according to the slope aspect and illumination model. For example, a table of temperature correction coefficients under different slope aspects and illumination conditions can be established by searching historical meteorological and illumination data. Based on the current pixel's slope aspect and illumination conditions, the corresponding correction coefficient can be found in the table to correct the temperature. Finally, the temperature gradient information of all pixels is integrated into a preliminary temperature gradient field.

[0028] Step S250: Spatial topological association is performed between the initial texture dynamic feature set and the 3D point cloud data in the 3D data volume. Based on the topological structure of the point cloud data, the slope surface is divided into multiple connected regions. The texture changes of all dynamic windows in each connected region are tracked. Through the synergistic analysis of texture changes in connected regions, a texture change sequence with spatial correlation is generated. The texture change sequence contains the overall texture evolution trend of each connected region in a continuous time period.

[0029] In one implementation, step S250 may specifically include the following steps S251 to S256:

[0030] Step S251: Construct a topology for the three-dimensional point cloud data in the three-dimensional data volume, calculate the adjacency relationship between each point in the point cloud, and generate a three-dimensional topology map of the slope surface. The nodes of the three-dimensional topology map are the points in the point cloud data, and the edges represent the spatial proximity relationship between the points.

[0031] Topology construction establishes connections and spatial relationships between points in 3D point cloud data. Adjacency relationships are the connections between adjacent points in the point cloud, determined by calculating spatial distances and positional relationships. A 3D topology graph is a graphical representation of a 3D spatial structure, where nodes represent points in the point cloud data, and edges represent spatial proximity relationships between points. Spatial indexing and graph theory algorithms can be used for topology construction. First, spatial indexing algorithms, such as KD-trees and octrees, are used to spatially partition the 3D point cloud data, improving the efficiency of distance calculation between points. Then, for each point, its spatial distance to its surrounding neighbors is calculated; if the distance is less than a certain threshold, the two points are considered to be adjacent. Based on these adjacency relationships, a 3D topology graph is constructed, with points as nodes and adjacency relationships as edges. During the construction process, graph data structures, such as adjacency matrices or adjacency lists, can be used to store the topology graph information.

[0032] Step S252: Perform region growth division based on the three-dimensional topology map, determine the region growth seed point based on the curvature change, aggregate spatially adjacent points with curvature differences less than a preset threshold into the same connected region, and generate the connected region division result of the slope surface. Each connected region contains a set of point cloud data points with similar geometric features.

[0033] When performing region growth and partitioning, the curvature value of each point in the 3D point cloud data is first calculated using curvature calculation algorithms such as Principal Component Analysis (PCA). Then, seed points for region growth are determined based on curvature changes; for example, points with significant curvature changes can be selected as seed points. Starting from the seed point, its neighboring points are traversed, and the curvature difference between the neighboring points and the seed point is calculated. When the curvature difference is less than a preset threshold and the points are spatially adjacent, the neighboring point is added to the current connected region. This process is repeated until no neighboring points meet the criteria. Finally, the entire 3D point cloud data is traversed to generate the connected region partitioning results for the slope surface.

[0034] Step S253: Extract the spatial coordinates of each dynamic window in the initial texture dynamic feature set, assign the dynamic window to the corresponding connected region according to the connected region partitioning result, and establish a region-window mapping relationship. Each connected region contains at least one dynamic window.

[0035] When extracting the spatial coordinates of a dynamic window, it can be converted into three-dimensional spatial coordinates based on the window's position in the visible light image and the image's georeferenced information. When assigning the dynamic window to a corresponding connected region, spatial indexing and a point-region inclusion algorithm can be used. First, spatial indexing is used to quickly locate the approximate region where the window is located. Then, it is determined whether the window's spatial coordinates lie within a connected region; if so, the window is assigned to that region. Finally, a region-window mapping relationship is established, storing the corresponding information of connected regions and dynamic windows in a data structure, such as a dictionary or table.

[0036] Step S254: Perform time series alignment on the texture changes of all dynamic windows within each connected region to ensure that the texture change timestamps of each window are consistent. Calculate the average and standard deviation of the texture change amplitude of all windows within the region to generate regional texture change statistics.

[0037] When performing time series alignment, methods such as linear interpolation or spline interpolation can be used to calculate the texture change value at a unified time point based on the texture change timestamps and data values ​​of the windows. For each connected region, all dynamic windows within it are traversed, and their texture change data are time-series aligned. Then, the mean and standard deviation of the texture change amplitude of all windows within the region are calculated. The mean is calculated by adding the texture change amplitudes of all windows and then dividing by the number of windows. The standard deviation is calculated by first squared the difference between the texture change amplitude of each window and the mean, then averaging these squared values, and finally taking the square root. Finally, the mean and standard deviation of the texture change for each connected region are recorded in the region texture change statistics.

[0038] Step S255: Analyze the evolution trend of regional texture change statistics over time. By calculating the rate of change of statistics within adjacent time periods, determine the acceleration or deceleration characteristics of regional texture change. Combine the regional geometric features to determine whether the changes are synergistic. Synergy is manifested in the consistent direction of texture change of most windows within the region.

[0039] Analyzing the evolution trend of regional texture change statistics over time involves observing the numerical changes of regional texture change statistics (such as mean and standard deviation) at different time points to understand the development trend of texture changes within the region. The rate of change of statistics within adjacent time periods is the ratio of the difference in statistics between two adjacent time points to the statistics at the previous time point. Calculating the rate of change can determine whether regional texture changes are accelerating or decelerating. Regional geometric features are the geometric attributes of connected regions, such as shape, size, and slope. Combining regional geometric features with the analysis of whether changes exhibit synergy is crucial because different geometric features can affect the consistency of texture changes. Synergy is manifested in the consistent direction of texture changes in most windows within the region; that is, the texture changes in most windows show similarity in direction.

[0040] When analyzing the evolution trend of regional texture change statistics over time, a curve showing the change of statistics over time can be plotted. The slope and trend of the curve can be observed to determine the changing trend. When calculating the rate of change of statistics within adjacent time periods, the mean and standard deviation are calculated separately. If the rate of change is positive and gradually increases, it indicates that the regional texture change is accelerating; if the rate of change is positive but gradually decreases, it indicates that the regional texture change is decelerating; if the rate of change is negative, it indicates that the regional texture change is reversing. When determining whether the changes are synergistic in conjunction with regional geometric features, the distribution of windows within the region and the influence of geometric features on texture changes can be analyzed. For example, if windows within the region are evenly distributed and have similar geometric features, and the texture change direction of most windows is consistent, it indicates that the changes are synergistic.

[0041] Step S256: Arrange the collaborative texture change trends of each connected region in chronological order, and combine them with regional spatial coordinate information to generate a spatially correlated texture change sequence covering the entire surface of the in-service slope. The texture change sequence contains the overall texture evolution trend of each connected region in a continuous time period.

[0042] When arranging the collaborative texture change trends in chronological order, time-sorting algorithms such as bubble sort and quicksort can be used to sort the texture change data of each connected region according to its timestamp. When combining regional spatial coordinate information, the sorted texture change data can be integrated with the spatial coordinate information of the connected regions and stored in a data structure, such as a multidimensional array or a database. Finally, a spatially correlated texture change sequence covering the entire surface of the in-service slope is generated based on the integrated data. This sequence can be used for subsequent slope stability analysis and monitoring.

[0043] Step S260: Perform spatial structural surface constraint processing on the preliminary temperature gradient field, combine the slope structural surface information extracted from the three-dimensional point cloud data, correct the abnormal values ​​of the temperature gradient near the structural surface, calculate the angle relationship between the gradient direction and the orientation of the structural surface, and convert the planar temperature gradient field into a three-dimensional temperature gradient distribution.

[0044] Spatial structural surface constraint processing utilizes the structural surface information of slopes to constrain and correct the initial temperature gradient field, thereby improving its accuracy. In this process, slope structural surface information is first extracted from 3D point cloud data using structural surface extraction algorithms, such as those based on curvature and normal direction. Then, the temperature gradient values ​​near the structural surfaces in the initial temperature gradient field are corrected based on the characteristics of the structural surfaces and the temperature conduction model. For example, if the structural surface is an insulation layer, the temperature gradient near it may change abruptly; in this case, the temperature gradient can be adjusted according to the characteristics of the insulation layer. When calculating the angle between the gradient direction and the orientation of the structural surface, vector operations can be used to represent both the temperature gradient direction and the orientation of the structural surface as vectors, and then the angle between them can be calculated. Finally, based on the angle relationship and the corrected temperature gradient values, the planar temperature gradient field is converted into a 3D temperature gradient distribution, extending the temperature gradient information into 3D space.

[0045] Step S300: Input the texture change sequence and temperature gradient distribution into the pre-trained transfer learning model, combine it with slope instability case data from the historical disaster database to optimize the transfer learning model, and output the probability distribution of instability risk of in-service slopes.

[0046] In one implementation, step S300 may specifically include the following steps S310 to S360:

[0047] Step S310: Perform cross-timescale feature enhancement on the texture change sequence, decompose the texture change trend of different time periods into multi-resolution, extract the features of short-term rapid change and long-term slow change, and generate a multi-scale texture feature set. The feature dimension of the multi-scale texture feature set is adapted to the input layer dimension of the pre-trained transfer learning model.

[0048] Cross-timescale feature enhancement involves extracting and enhancing features from texture change sequences at different timescales to better capture the diverse characteristics of texture changes. Multi-resolution decomposition is a method that decomposes signals or data at different resolutions. Through multi-resolution decomposition, texture change trends can be decomposed into components of different frequencies, where short-term rapid changes correspond to high-frequency components, and long-term slow changes correspond to low-frequency components. Short-term rapid changes reflect drastic changes in texture over a short period, possibly related to sudden events or short-term disturbances on the slope; long-term slow changes reflect gradual changes in texture over a longer period, possibly related to long-term geological evolution or environmental factors on the slope.

[0049] When performing cross-timescale feature enhancement, multi-resolution analysis methods such as wavelet decomposition can be used to decompose the texture change sequence. Wavelet decomposition can decompose the signal into wavelet coefficients at different scales and locations. By selecting different wavelet bases and decomposition levels, texture features at different resolutions can be obtained. For the high-frequency and low-frequency components obtained from the decomposition, feature extraction is performed separately, such as calculating statistical features like mean, variance, and energy. Finally, the extracted texture features at different scales are integrated into a multi-scale texture feature set, and the feature dimensions are adjusted to match the input layer dimension of the pre-trained transfer learning model.

[0050] Step S320: Spatial feature encoding of temperature gradient distribution is performed, and the three-dimensional temperature gradient data is converted into a multi-channel feature map. Each channel corresponds to a temperature gradient component in a different direction. The gradient direction information is enhanced by the correlation analysis between channels, and a temperature gradient feature map is generated. The size of the temperature gradient feature map is consistent with the spatial dimension of the multi-scale texture feature set.

[0051] When performing spatial feature encoding, the three-dimensional temperature gradient data can be separated according to different directional components, and each component can be stored in a channel to form a multi-channel feature map. In the correlation analysis between channels, methods such as correlation coefficients and mutual information can be used to calculate the correlation between different channels. Based on the correlation analysis results, channels can be weighted or fused to enhance gradient direction information. Finally, the size of the temperature gradient feature map is adjusted to match the spatial dimension of the multi-scale texture feature set; interpolation or downsampling methods can be used for size adjustment.

[0052] Step S330: Perform cross-modal feature fusion between the multi-scale texture feature set and the temperature gradient feature map, construct a cross-modal attention mechanism, calculate the mutual information between texture features and temperature features, assign feature fusion weights according to the mutual information magnitude, and generate a fused feature tensor. The fused feature tensor contains the collaborative correlation information between texture and temperature.

[0053] When performing cross-modal feature fusion, a cross-modal attention mechanism is first constructed. This can be achieved using a convolutional neural network (CNN)-based attention module, such as the Squeeze-and-Excitation (SE) module, which automatically learns the importance of features from different modalities. Next, the mutual information between texture and temperature features is calculated using the mutual information calculation formula from information theory. Weights are then assigned to each feature based on the magnitude of the mutual information; for example, the mutual information can be normalized to obtain weight values. Finally, the multi-scale texture feature set and the temperature gradient feature map are weighted and summed according to their respective weights to generate a fused feature tensor.

[0054] In one implementation, step S330 may specifically include the following steps S331 to S336:

[0055] Step S331: Expand the channel dimension of the multi-scale texture feature set, map the texture features of different time scales to independent channels, and generate a multi-channel texture feature map. The number of channels in the multi-channel texture feature map is the same as the number of time scales.

[0056] When expanding the channel dimension, first determine the number of different time scales in the multi-scale texture feature set. Assuming the multi-scale texture feature set contains texture features at three time scales: short-term, medium-term, and long-term, then the number of channels is 3. For each time scale texture feature, add it as an independent channel to the feature map. This process can be achieved using matrix operations, such as concatenating the texture feature matrices of each time scale according to the channel dimension.

[0057] Step S332: Enhance the spatial dimension of the temperature gradient feature map by calculating the spatial distribution histogram of the gradient direction, encoding the gradient direction information into spatial features, and generating an enhanced temperature feature map. The spatial resolution of the enhanced temperature feature map is consistent with that of the multi-channel texture feature map.

[0058] When enhancing spatial dimensions, the temperature gradient feature map is first divided into multiple small spatial regions. For each small region, the gradient direction of all pixels within it is calculated, and the frequency of occurrence of different gradient directions is counted, thus obtaining a histogram of the spatial distribution of gradient directions for that region. The gradient directions can be divided into several discrete intervals, and the number of pixels with gradient directions within each interval is counted. Then, the histogram of the spatial distribution of gradient directions for each region is used as a new feature vector for that region. Finally, these new feature vectors are combined to form the enhanced temperature feature map. To ensure that the spatial resolution of the enhanced temperature feature map is consistent with that of the multi-channel texture feature map, interpolation or downsampling methods can be used to adjust the size of the enhanced temperature feature map.

[0059] Step S333: Calculate the mutual information of the multi-channel texture feature map and the enhanced temperature feature map at each spatial location. The mutual information value represents the degree of correlation between the two features at that spatial location. Generate a mutual information weight map, where each pixel value in the mutual information weight map is the mutual information value at the corresponding location.

[0060] By calculating the mutual information at each spatial location, the dependency between texture and temperature features at that location can be understood. A larger mutual information value indicates a stronger correlation between the two features at that location; a smaller mutual information value indicates a weaker correlation. The mutual information weight map is a graph with the same spatial size as the multi-channel texture feature map and the enhanced temperature feature map, where each pixel value corresponds to a mutual information value at a spatial location. This graph visually displays the distribution of the correlation between the two features across the entire space.

[0061] When calculating mutual information, for each spatial location (i,j) of the multi-channel texture feature map and the enhanced temperature feature map, the eigenvectors of the multi-channel texture feature map and the enhanced temperature feature map at that location are treated as two random variables. Discretization methods can be used to discretize the eigenvector values, for example, dividing the eigenvalues ​​into several intervals and then counting the frequency of occurrence of the eigenvalues ​​within each interval. The mutual information value at that location is calculated according to the mutual information calculation formula in information theory. By traversing all spatial locations of the entire feature map, the mutual information value at each location is obtained, thus generating a mutual information weight map.

[0062] Step S334: Based on the mutual information weight map, the multi-channel texture feature map and the enhanced temperature feature map are weighted and summed. The positions with higher weights are given more feature contributions to generate a preliminary fused feature map. The number of channels in the preliminary fused feature map is the sum of the number of channels in the two feature maps.

[0063] The multi-channel texture feature map and the enhanced temperature feature map are weighted and summed based on the mutual information weight map. This ensures that at locations with high mutual information values ​​(i.e., where the two features are closely related), the two features contribute more to the fusion result, while at locations with low mutual information values, the contribution of the features is relatively smaller. This highlights the synergistic effect of the two features at closely related locations, improving the quality of the fused features.

[0064] Step S335: Perform feature interaction modeling on the preliminary fused feature map. By calculating the feature correlation between different channels, construct channel attention weights, enhance the feature expression of important channels, suppress the interference of secondary channels, and generate channel-enhanced fused features.

[0065] Feature interaction modeling analyzes the feature correlations between different channels in the preliminary fused feature map to uncover potential information and synergistic effects between channels. Channel attention weights are weights used to measure the importance of each channel. By constructing channel attention weights, the model can focus more on the feature information of important channels while suppressing the interference of secondary channels, thereby improving the quality and expressive power of the fused features.

[0066] When modeling feature interactions, channel attention mechanisms, such as the Squeeze-and-Excitation (SE) module, can be used. First, a global average pooling operation is performed on the preliminary fused feature map, compressing the feature map of each channel into a scalar, resulting in a global feature representation for each channel. Then, a fully connected layer performs a non-linear transformation on these global feature representations to generate channel attention weights. The input dimension of the fully connected layer is the number of channels in the preliminary fused feature map, and the output dimension is also the number of channels. Next, the channel attention weights are multiplied element-wise with each channel of the preliminary fused feature map, enhancing the feature representation of important channels and suppressing interference from less important channels. Finally, the channel-enhanced fused feature is obtained.

[0067] Step S336: Integrate the channel-enhanced fusion features with spatial and channel dimensions, reshape them into a three-dimensional tensor, and generate a fusion feature tensor containing texture-temperature co-processing information.

[0068] The integration of spatial and channel dimensions involves reorganizing and arranging the channel-enhanced fusion features in both spatial and channel dimensions, transforming them into a three-dimensional tensor. A three-dimensional tensor is a data structure with three dimensions, which, in this step, better represents the collaborative information of texture and temperature features, facilitating subsequent input into the transfer learning model for analysis and processing. The fusion feature tensor containing texture-temperature collaborative information is the final result obtained after integrating spatial and channel dimensions. It combines information from multi-channel texture feature maps and enhanced temperature feature maps, and highlights important feature information through a channel attention mechanism, enabling a more comprehensive and accurate reflection of the characteristics of in-service slopes.

[0069] Step S340: Extract slope instability case data from the historical disaster database, standardize the texture changes and temperature gradient features in the case data to make their feature distribution match the distribution of the fused feature tensor, and establish a case feature library containing feature patterns corresponding to different instability types.

[0070] The texture variations and temperature gradient features in the case data are standardized to ensure their distribution matches the distribution of the fused feature tensor. This guarantees that data from different sources can be compared and analyzed at the same scale during subsequent feature matching and model tuning. The case feature library is a database containing feature patterns corresponding to different instability types, providing historical case feature references for the transfer learning model and helping the model learn the feature patterns of different instability types.

[0071] When extracting slope instability case data, case data relevant to currently in-service slopes is selected based on the database's storage structure and query interface. For the selected case data, texture variations and temperature gradient features are standardized. Common standardization methods, such as Z-score standardization, can be used. Then, the standardized case data is classified according to instability type, and feature patterns for each instability type are extracted to establish a case feature library.

[0072] Step S350: Associate and match the fused feature tensor with the case features in the case feature library. Through feature pattern similarity analysis, add case similarity labels to the fused feature tensor. Input the fused feature tensor with similarity labels into the pre-trained transfer learning model. Adjust the parameters of the model's feature extraction layer and output layer so that the model learns the feature association between the current slope and historical cases.

[0073] When performing association matching, similarity metrics such as Euclidean distance and cosine similarity can be used. For each case feature in the fused feature tensor and the case feature library, the similarity between them is calculated. The case with the highest similarity is selected as the most similar case, and a similarity label for that case is added to the fused feature tensor. After inputting the fused feature tensor with similarity labels into the pre-trained transfer learning model, the model parameters are fine-tuned. First, some parameters of the lower-level feature extraction layers are frozen, and only the parameters of the top-level output layer are adjusted to allow the model to initially adapt to the new feature distribution. Then, some parameters of the lower-level feature extraction layers are gradually unfrozen, layer by layer from the top down. After each layer is unfrozen, the parameters are fine-tuned so that the lower-level feature extraction can capture the features of the current slope. Through multiple rounds of layer-by-layer unfreezing and parameter fine-tuning, the model's feature extraction layers can simultaneously learn the general features of historical cases and the features of the current slope, achieving feature association learning between the current slope and historical cases.

[0074] In one implementation, step S350 may specifically include the following steps S351 to S356:

[0075] Step S351: Perform feature pattern clustering on the case features in the case feature library, group cases with similar feature distributions into one class, extract the feature center of each class of cases, and generate a case feature pattern library, where each feature pattern represents a class of unstable features.

[0076] When performing feature pattern clustering, clustering algorithms such as K-Means can be used. First, determine the number of clusters K, which can be determined based on historical experience or experimental methods. Then, randomly initialize K cluster centers. For each case feature in the case feature library, calculate its distance to each cluster center and assign it to the cluster containing the nearest cluster center. Next, recalculate the feature centers for each category and update the positions of the cluster centers. Repeat the above assignment and update process until the cluster centers no longer change significantly or a preset number of iterations is reached. Finally, extract the feature centers for each case category to generate a case feature pattern library.

[0077] Step S352: Calculate the similarity between the fused feature tensor and each feature pattern in the case feature pattern library, and generate a similarity distribution vector. Each element of the similarity distribution vector corresponds to the similarity value of a feature pattern.

[0078] When calculating similarity, similarity measurement methods can be used, such as Euclidean distance, cosine similarity, etc., without any specific limitations.

[0079] Step S353: Based on the feature pattern corresponding to the maximum similarity value in the similarity distribution vector, add case similarity labels to the fused feature tensor. The label content includes the instability type of the case and the degree of similarity.

[0080] Case similarity labels are added to the fused feature tensor based on the feature pattern corresponding to the maximum similarity value in the similarity distribution vector because the feature pattern corresponding to the maximum similarity value represents the instability type most similar to the characteristics of the currently in-service slope. When adding case similarity labels, the maximum similarity value and its corresponding feature pattern index j are first found from the similarity distribution vector. Based on feature pattern index j, the corresponding feature pattern is retrieved from the case feature pattern library, and the instability type represented by this feature pattern is determined. Similarity information can be directly represented using the maximum similarity value. The instability type and similarity information are combined to form a case similarity label, which is then added to the fused feature tensor.

[0081] Step S354: Input the fused feature tensor with similarity labels into the pre-trained transfer learning model, freeze the parameters of the model's bottom feature extraction layer, and only adjust the parameters of the top output layer to allow the model to initially adapt to the new feature distribution.

[0082] Inputting a fused feature tensor with similarity labels into a pre-trained transfer learning model allows the model to learn the features of currently in-service slopes using knowledge from historical cases. Freezing the parameters of the model's lower-level feature extraction layers and adjusting only the parameters of the top-level output layer is a fine-tuning strategy. The lower-level feature extraction layers typically learn general features such as edges and textures, which have a certain degree of universality across different tasks. Therefore, freezing these parameters avoids over-adjusting on new data, preventing the model from forgetting previously learned knowledge. Adjusting only the parameters of the top-level output layer allows the model to quickly adapt to new feature distributions and initially learn the feature associations between the current slope and historical cases.

[0083] When tuning the model, the fused feature tensor with similarity labels is first input into the pre-trained transfer learning model. Then, by setting the trainable attribute of the model parameters, the parameters of the lower-level feature extraction layers are set to non-trainable, i.e., these parameters are frozen. Only the parameters of the top-level output layer are allowed to be updated. Optimization algorithms, such as stochastic gradient descent (SGD) and Adam, are used to train the parameters of the top-level output layer. During training, the loss value is calculated based on the model's loss function, and the parameters of the top-level output layer are updated based on the loss value, allowing the model to initially adapt to the new feature distribution.

[0084] Step S355: After the model has initially adapted, gradually unfreeze some parameters of the bottom feature extraction layer, unfreezing layer by layer from the top layer down. After each layer is unfrozen, fine-tune the parameters so that the bottom feature extraction can capture the features of the current slope.

[0085] Unfreezing layers sequentially from top to bottom is done because the features at the top are usually more relevant to the specific task and output. Unfreezing the top layer first allows the model to better adjust the extraction of output-related features based on its initial adaptation. Fine-tuning the parameters after each unfreezing layer allows the lower feature extraction layers to gradually adapt to the current slope characteristics, avoiding model instability or overfitting caused by unfreezing too many parameters at once. During fine-tuning, a small learning rate is set to avoid excessive parameter updates. The loss value is calculated based on the model's loss function, and the parameters of the unfrozen layer are updated according to the loss value. This unfreezing and fine-tuning process is repeated until the preset number of unfrozen layers is reached or other stopping conditions are met.

[0086] Step S356: Through multiple rounds of layer-by-layer unfreezing and parameter fine-tuning, the feature extraction layer of the pre-trained transfer learning model can simultaneously learn the general features of historical cases and the features of the current slope, thereby realizing the feature association learning between the current slope and historical cases.

[0087] In the multi-round layer-by-layer unfreezing and parameter fine-tuning process, each round follows the method described in step S355 above. As the number of rounds increases, the model's feature extraction layer gradually adapts to the characteristics of the current slope while retaining the general features of historical cases. After each round of fine-tuning, the model's performance is evaluated, such as by using methods like cross-validation to calculate metrics like accuracy and recall. Based on the evaluation results, fine-tuning parameters, such as the learning rate and the number of unfrozen layers, are adjusted. Through multiple iterations, the model's feature extraction layer can effectively learn both the general features of historical cases and the features of the current slope, achieving feature association learning between the current slope and historical cases.

[0088] Step S360: After multiple rounds of iterative optimization, the fused feature tensor is input into the optimized pre-trained transfer learning model, which outputs the instability probability of each region of the in-service slope in the future time period. The outlier values ​​of the instability probability of isolated regions are corrected by combining the slope spatial topology, and an instability risk probability distribution covering the entire in-service slope is generated.

[0089] Multi-round iterative optimization involves multiple parameter adjustments and model training steps in S350, enabling the pre-trained transfer learning model to better learn the feature associations between the current slope and historical cases, thus improving model performance. The fused feature tensor is input into the optimized model, which can then output the instability probability of each region of the in-service slope within a future time period based on the learned feature associations. The instability probability of each region of the in-service slope is an estimated value of the probability of instability occurring in each region within a future period. The slope spatial topology describes the spatial connectivity and positional relationships between different regions of the slope. Combining the slope spatial topology with the correction of outlier instability probability values ​​for isolated regions is crucial because the instability probability of isolated regions may be abnormal due to data noise or other factors. By considering their spatial relationships with surrounding areas, the estimation of instability probability can be made more reasonable. The instability risk probability distribution covering the entire in-service slope describes the distribution of instability probability in different regions of the in-service slope, visually demonstrating the instability risk situation in different areas of the slope and providing important basis for subsequent slope maintenance decisions.

[0090] In one implementation, step S360 may specifically include the following steps S361 to S366:

[0091] Step S361: Input the fused feature tensor into the optimized pre-trained transfer learning model. The model outputs the instability probability of each spatial location on the surface of the in-service slope in multiple future time periods, generating a preliminary instability probability map. The spatial resolution of the preliminary instability probability map is consistent with the spatial dimension of the fused feature tensor.

[0092] The fused feature tensor is input into the optimized pre-trained transfer learning model because the optimized model has learned the feature associations between the current in-service slope and historical cases, enabling it to predict the instability probability of different areas of the slope based on the information in the fused feature tensor. The model outputs the instability probability of each spatial location on the in-service slope surface over multiple future time periods. For each spatial point on the slope surface, the model provides an estimate of the probability of instability at different future time points. The preliminary instability probability map is a graph representing these instability probabilities in image form, with a spatial resolution consistent with the spatial dimension of the fused feature tensor. This ensures that the instability probability information accurately corresponds to the actual spatial location of the slope, providing an accurate data foundation for subsequent analysis and correction.

[0093] Step S362: Perform spatial topology analysis on the preliminary instability probability map. Based on the results of connected region division, calculate the average value and distribution range of instability probability in each connected region, and identify whether there are isolated high instability probability points in the connected regions. High instability probability points are points whose instability risk probability exceeds the preset multiple of the region average.

[0094] Spatial topology analysis analyzes the preliminary instability probability map in terms of spatial structure to understand the distribution of instability probability across different regions. The connected region division result, obtained in step S250, divides the slope surface into multiple connected regions. Analysis based on this result allows for better consideration of the spatial correlation between different regions of the slope. Calculating the average instability probability and its distribution range within each connected region reveals the overall level and dispersion of instability probability in that region. Identifying isolated high-instability-probability points within connected regions is crucial because these points may have abnormally high instability probabilities due to data noise, special geological conditions, or other factors, requiring further verification and correction. High-instability-probability points are those whose instability risk probability exceeds a preset multiple of the regional average. This preset multiple can be set according to actual conditions, for example, 2 or 3 times.

[0095] Step S363: Perform neighborhood consistency verification on the identified isolated high instability probability points, compare the difference between the point and the average instability probability of its connected region, and when the difference exceeds the preset range, it is judged as an outlier and corrected based on the neighborhood instability probability.

[0096] Neighborhood consistency verification checks whether the instability probability of an isolated high-probability point is consistent with that of its surrounding neighboring points. Comparing the point's probability with the average instability probability of its connected regions helps determine if the point's probability deviates significantly from the overall situation of the region. When the difference exceeds a preset range, the point is considered an outlier, as its instability probability may be due to data errors or other abnormal factors, inconsistent with the actual situation in the surrounding area. Correction based on neighborhood instability probability aims to make the point's probability more consistent with the situation in its surrounding area, improving the accuracy of the instability risk probability distribution.

[0097] Step S364: After correcting outliers, calculate the rate of change of the instability probability of each connected region over time, analyze the accelerating trend of the rate of change, compare the rate of change with the rate of change before instability in historical cases, and adjust the temporal distribution of the instability probability of connected regions.

[0098] Calculating the rate of change of instability probability over time for each connected region is to understand how the instability probability of that region changes over time. An accelerating trend in the rate of change can reflect whether the instability probability is increasing at an accelerating or decelerating pace, or whether there are other trends. Comparing the rate of change with the pre-instability rate of change in historical cases is because the pre-instability rate of change in historical cases can serve as a reference to help determine whether there is a potential risk of instability in the current connected region. Adjusting the temporal distribution of instability probability for connected regions involves adjusting the instability probability of connected regions at different future time points based on the comparison results, making the prediction of instability probability more accurate and reasonable.

[0099] When calculating the rate of change of instability probability over time, for each connected region, its instability probability value is obtained at different time periods. The difference in instability probability between adjacent time periods is calculated and then divided by the time interval to obtain the rate of change of instability probability. When analyzing the accelerating trend of the rate of change, the rate of change of the rate of change (i.e., the second derivative) can be calculated. If the second derivative is positive, it indicates that the instability probability is increasing at an accelerating rate; if the second derivative is negative, it indicates that the instability probability is increasing at a decelerating rate. The rate of change of instability probability of the current connected region is compared with the rate of change before instability in historical cases, and the instability probability value of the connected region at different future time points is adjusted according to the comparison results. For example, if the current rate of change is close to the rate of change before instability in historical cases and shows an accelerating trend, the instability probability value of the connected region at future time points is appropriately increased.

[0100] Step S365: Based on the spatial relationship of each connected region, calculate the correlation of the instability probability of adjacent connected regions. Through the instability risk transmission analysis between connected regions, adjust the instability probability value of the edge region to enhance the spatial coherence of the overall distribution.

[0101] The spatial relationships between connected regions describe their adjacency and connectivity in space. Calculating the correlation of instability probabilities between adjacent connected regions aims to understand the degree of mutual influence between their instability probabilities. The analysis of instability risk transmission between connected regions considers the potential impact of instability in one connected region on its neighboring regions; this impact can be reflected through the transmission of instability risk. Adjusting the instability probability values ​​of edge regions is necessary because the interaction between edge regions and their neighboring regions is more pronounced. By considering the influence of neighboring regions and adjusting the instability probability values ​​of edge regions, the spatial distribution of instability risk probabilities can be made more coherent and reasonable.

[0102] When calculating the correlation between the instability probabilities of adjacent connected regions, correlation analysis methods, such as the Pearson correlation coefficient, can be used. For two adjacent connected regions, the correlation between their instability probability values ​​over the same time period is calculated. Based on the correlation analysis results, an analysis of the transmission of instability risk between connected regions is performed. For example, if the instability probabilities of two adjacent regions are highly correlated, and the instability probability of one region increases, it can be considered that the instability risk of that region may be transmitted to the adjacent region, and the instability probability value of the edge region of the adjacent region should be appropriately increased. For the edge region, its instability probability value is adjusted according to its correlation with the adjacent region and the transmission of instability risk. A weighted average method can be used to calculate the adjusted instability probability value of the edge region based on the instability probability values ​​and correlation coefficients of the adjacent regions.

[0103] Step S366: Piece together the corrected instability probabilities of all connected regions according to the spatial topology to generate an instability risk probability distribution covering the entire surface of the in-service slope. The instability risk probability distribution includes the instability probability value of each spatial location in future time periods.

[0104] The modified instability probability of all connected regions is pieced together according to the spatial topology. This involves accurately combining the instability probability values ​​of each connected region based on the spatial topology of the slope, forming a complete instability risk probability distribution. The spatial topology describes the spatial location and connectivity of each connected region; piecing together the values ​​according to this structure ensures the spatial continuity and accuracy of the instability probability values. The instability risk probability distribution covering the entire surface of the in-service slope provides a comprehensive description of the instability probability distribution in all areas of the in-service slope, including the instability probability value for each spatial location in future time periods, providing detailed and accurate information for subsequent slope maintenance decisions.

[0105] Step S400: The probability distribution of instability risk is spatiotemporally matched with the real-time meteorological data. The influence weights of rainfall and temperature changes on the stability of the in-service slope are quantified through the meteorological-geological coupling model, and the deformation prediction curve of the in-service slope is updated.

[0106] In one implementation, step S400 may specifically include the following steps S410 to S460:

[0107] Step S410: Construct a time series from the real-time acquired meteorological data, organize meteorological elements such as rainfall and temperature changes in chronological order, add a timestamp to each meteorological element, and generate a meteorological time series. The time granularity of the meteorological time series is adapted to the time period of the instability risk probability distribution.

[0108] Real-time meteorological data is the latest data on meteorological elements such as rainfall and temperature changes obtained from meteorological monitoring equipment or meteorological databases. Time series construction involves organizing and arranging this meteorological data in chronological order to form an ordered sequence. Organizing meteorological elements such as rainfall and temperature changes in chronological order means arranging the data according to the order of their collection time to clearly show how these elements change over time. Precise timestamps are accurate time identifiers added to each meteorological element data point, accurate to the specific date and time, used to accurately pinpoint the time of data collection. A meteorological time series is a sequence containing meteorological element data arranged in chronological order and their corresponding timestamps. Its time granularity is adapted to the time period of the instability risk probability distribution, meaning that the time intervals and time ranges of the meteorological time series must match the time period corresponding to the instability risk probability distribution to ensure accurate subsequent spatiotemporal matching operations.

[0109] Step S420: Extract the time dimension of the instability risk probability distribution, obtain all the time period information it contains, determine the start and end times of each time period, generate a risk time axis, and the time range of the risk time axis covers the time range of the meteorological time series.

[0110] The instability risk probability distribution describes the distribution of the likelihood of instability in different areas of an in-service slope over different time periods. Time dimension extraction involves extracting time-related information from this distribution. Obtaining all time period information included in the distribution involves identifying all time periods covered by the instability risk probability distribution; these time periods may be continuous or discrete. Determining the start and end times of each time period is crucial for clarifying its specific time range. The risk time axis is a time-based sequence that records the start and end times of each time period in the instability risk probability distribution. Its time range covers the time range of meteorological time series, ensuring that meteorological data can be fully mapped to the time periods of the instability risk probability distribution during spatiotemporal matching.

[0111] When extracting the time dimension, the data structure of the instability risk probability distribution is first analyzed to identify time-related information. If the instability risk probability distribution is stored as a matrix or tensor, one dimension may represent time, from which time period information can be extracted. For each time period, its start and end times are determined based on the data storage method and relevant identifiers. This time period information can be organized into an ordered list to generate a risk time axis. When generating the risk time axis, it is essential to ensure that its time range covers the time range of the meteorological time series. If the time range of the instability risk probability distribution is smaller than that of the meteorological time series, the instability risk probability distribution can be expanded to match its time range. If the time range of the instability risk probability distribution is larger than that of the meteorological time series, the portion overlapping with the meteorological time series can be used as the risk time axis.

[0112] Step S430: Align the meteorological time series with the risk time axis so that each meteorological element value corresponds to the corresponding time period of the instability risk probability distribution. Use time interpolation to compensate for the time granularity difference between meteorological data and risk data, and generate spatiotemporally aligned data pairs. Each spatiotemporally aligned data pair contains meteorological elements for one time period and the instability risk probability for the corresponding period.

[0113] In one implementation, step S430 may specifically include the following steps S431 to S436:

[0114] Step S431: Parse the metadata of the meteorological time series, obtain the start time, end time and collection interval of the meteorological data, and determine the time base of the meteorological time axis.

[0115] Metadata for meteorological time series data describes the basic information of the meteorological time series, including the start and end times of data collection and the collection intervals. Parsing the metadata of a meteorological time series involves extracting this metadata information from the data file or related information. Obtaining the start and end times and collection intervals of meteorological data collection is to clarify the time range and time intervals of the meteorological data collection, thereby determining the time reference for the meteorological time axis. The time reference for the meteorological time axis serves as the starting point and time interval standard, providing the foundation for subsequent time node alignment and time interpolation processing.

[0116] When parsing metadata for meteorological time series data, if the time series is stored in file format, the metadata may be contained in the file's header information or metadata fields. File reading tools and parsing algorithms can be used to extract information such as the start and end times of data collection and the collection intervals from the file. The start time of data collection is used as the starting point of the meteorological time axis, and the collection intervals are used as the time interval standards for the time axis to determine the time base of the meteorological time axis.

[0117] Step S432: Analyze the time period division rules of the risk time axis, determine the start and end times of each time period, and generate a time period division table. Each table entry contains a period number and the corresponding time range.

[0118] When analyzing the time period division rules of the risk timeline, it is necessary to consider the data storage method and related identifiers of the risk timeline. If the risk timeline is stored as an array or list, it may contain records of the start and end times of each time period, from which the time period information can be directly extracted. If the risk timeline division rules are determined by a certain algorithm or formula, then this algorithm or formula needs to be analyzed to calculate the start and end times of each time period. The period number and corresponding time range of each time period are then compiled into a table to generate a time period division table. Each row of the table represents a time period, containing the period number and the corresponding start and end times.

[0119] Step S433: Based on the time period division table, the meteorological time series is periodically extracted, meteorological element data within each time period is extracted, the cumulative value or average value of meteorological elements within the period is calculated, and a periodic meteorological element set is generated. Each periodic meteorological element set contains the meteorological element statistics for one time period.

[0120] When extracting time periods, each time period in the time period division table is traversed, and meteorological element data belonging to that time period is selected from the meteorological time series based on its start and end times. For the selected meteorological element data, if the meteorological element is rainfall, its cumulative value is usually calculated, which is the sum of all rainfall data within that time period; if the meteorological element is temperature change, its average value is usually calculated, which is the average of all temperature change data within that time period. The statistical values ​​of meteorological elements for each time period are then compiled into a set to generate a periodic meteorological element set.

[0121] Step S434: When the collection interval of the meteorological time series is inconsistent with the time period of the risk time axis, the meteorological element data is resampled using the time interpolation method so that the time node of the resampled meteorological data is consistent with the time period node of the risk time axis.

[0122] When resampling using time interpolation methods, the first step is to determine the sampling interval of the meteorological time series and the time period of the risk time axis. If the sampling interval of the meteorological time series is shorter than the time period of the risk time axis, it indicates that the time resolution of the meteorological data is high, and downsampling is required. If the sampling interval of the meteorological time series is longer than the time period of the risk time axis, it indicates that the time resolution of the meteorological data is low, and upsampling is required. Common time interpolation methods include linear interpolation and spline interpolation.

[0123] Step S435: Perform time integrity verification on the interpolated periodic meteorological element set to check for any missing meteorological data in the time period. If missing data is found, perform trend filling based on the meteorological data of adjacent periods to ensure that each time period has a corresponding meteorological element value.

[0124] When performing time integrity verification, each time period in the time period division table is traversed, and the meteorological element set for each time period is checked to see if there are corresponding meteorological element statistical values ​​for that time period. If meteorological data is found to be missing for a certain time period, the changing trend of the meteorological element is analyzed based on the meteorological element statistical values ​​of the preceding and following periods. For example, if the meteorological element is rainfall, and the rainfall in the preceding and following periods shows an increasing trend, the rainfall in the missing period can be estimated based on the magnitude of the increase. Trend filling methods such as linear extrapolation or exponential extrapolation can be used. The filled meteorological element values ​​are then added to the meteorological element set for each time period to ensure that there are corresponding meteorological element values ​​for each time period.

[0125] Step S436: Combine the set of periodic meteorological elements for each time period with the probability value of instability risk for that period to generate spatiotemporally aligned data pairs. Each spatiotemporally aligned data pair includes a time period identifier, statistical values ​​of meteorological elements, and the probability of instability risk for the corresponding period.

[0126] When generating spatiotemporally aligned data pairs, each time period in the time period division table is traversed. The statistical values ​​of meteorological elements for that time period are obtained from the periodic meteorological element set, and the instability risk probability values ​​for that time period are obtained from the instability risk probability distribution. The time period identifier, the meteorological element statistical value, and the corresponding instability risk probability value are combined into a single data pair. Data structures such as lists, tuples, or dictionaries can be used to store the spatiotemporally aligned data pairs. For example, a dictionary can be used, where the key is the time period identifier and the values ​​are lists containing meteorological element statistical values ​​and instability risk probability values.

[0127] Step S440: Extract geological features from the 3D point cloud data in the 3D data volume, analyze the geological attributes of the slope such as lithological distribution, structural surface density, and weathering degree, and generate a geological vulnerability feature set. Each feature in the geological vulnerability feature set corresponds to a quantitative description of the geological attribute of the slope.

[0128] When extracting geological features, the 3D point cloud data is first preprocessed, such as through filtering and noise reduction, to improve data quality. For lithology distribution analysis, the characteristics of geological survey data and 3D point cloud data can be combined to use classification algorithms to divide regions of different lithologies. For example, machine learning-based classification algorithms, such as Support Vector Machine (SVM) or Random Forest, can be used to classify point cloud data into different lithology categories based on geometric features and reflection intensity, thereby obtaining the lithology distribution. For calculating the density of structural surfaces, structural surface features such as planes and cracks in the 3D point cloud data can be identified, and the number of structural surfaces per unit volume can be counted to obtain the density. For assessing the degree of weathering, a weathering assessment model can be established based on the surface roughness, reflection intensity, and other characteristics of the point cloud data, combined with geological knowledge and experience, to quantify the degree of weathering of slopes. The quantified results of geological attributes such as lithology distribution, structural surface density, and weathering degree are compiled into a geological vulnerability feature set.

[0129] Step S450: Input the spatiotemporally aligned data pairs and the geological vulnerability feature set into the meteorological-geological coupling model, analyze the interaction between meteorological elements and geological vulnerability features, calculate the degree of infiltration impact of rainfall on different lithological regions, the weakening effect of temperature changes on the strength of structural surfaces, and generate an influence relationship matrix. The rows of the influence relationship matrix represent meteorological elements, the columns represent geological vulnerability features, and the elements represent the intensity of the interaction between the two.

[0130] The spatiotemporally aligned data pairs contain statistical values ​​of meteorological elements for each time period and the corresponding probability of instability risk. The geological vulnerability feature set contains a quantitative description of various geological attributes of the slope. Inputting these into the meteorological-geological coupling model is to comprehensively consider the impact of meteorological and geological factors on slope stability. The meteorological-geological coupling model is a model capable of simulating the interaction between meteorological elements and geological vulnerability features. Through this model, the interaction between meteorological elements and geological vulnerability features can be analyzed to understand how meteorological factors affect slope stability by influencing geological conditions. The calculation of the impact of rainfall on the permeability of different lithological regions is because different lithologies have different permeabilities; the amount and duration of rainfall affect the permeability of water in different lithological regions, thus affecting slope stability. The weakening effect of temperature changes on the strength of structural surfaces is that temperature changes alter the physical and mechanical properties of structural surfaces, reducing their strength and thus affecting slope stability. The influence relationship matrix is ​​a matrix used to describe the intensity of the interaction between meteorological elements and geological vulnerability features. Its rows represent meteorological elements, columns represent geological vulnerability features, and matrix elements represent the intensity of the interaction between the two. By generating the influence relationship matrix, the interaction between meteorological factors and geological factors can be displayed intuitively.

[0131] In one implementation, step S450 may specifically include the following steps S451 to S456:

[0132] Step S451: Decompose the meteorological element statistics in the spatiotemporal aligned data pair into elements, decompose the rainfall into single rainfall derivative elements and consecutive rainfall days, decompose the temperature change into daily temperature difference derivative elements and cumulative temperature change derivative elements, and generate a refined meteorological element set.

[0133] Decomposing the statistical values ​​of meteorological elements in spatiotemporally aligned data pairs allows for a more detailed analysis of their impact on slope stability. The influence of rainfall and temperature changes on slope stability depends not only on their overall statistical values ​​but also on the specific characteristics of these changes. Decomposing rainfall into single-event rainfall derivatives and the number of consecutive rainfall days is crucial because the magnitude of a single rainfall event and the number of consecutive rainfall days have different effects on slope water infiltration and soil saturation. Single-event rainfall derivatives can include maximum single rainfall and average single rainfall, reflecting the intensity of a single rainfall event. The number of consecutive rainfall days reflects the duration of rainfall; prolonged continuous rainfall can keep the slope soil saturated for extended periods, increasing the risk of instability. Decomposing temperature changes into diurnal temperature range derivatives and cumulative temperature change derivatives is important because the magnitude of the diurnal temperature range affects the thermal expansion and contraction of rocks, influencing the strength of structural surfaces; cumulative temperature change reflects the overall trend of temperature changes over a period of time, potentially affecting the degree of rock weathering and the physical properties of the soil. A refined meteorological element set is a collection of elements derived from the decomposed meteorological elements. By generating a refined meteorological element set, the impact of meteorological elements on slope stability can be considered more comprehensively.

[0134] Step S452: Standardize each geological vulnerability feature in the geological vulnerability feature set to generate a standardized geological feature set. Each feature value in the standardized geological feature set reflects the relative strength of the geological attribute.

[0135] When performing standardization, general standardization methods such as Z-score standardization or Min-Max standardization can be used, without specific limitations. The standardized geological vulnerability characteristics are then compiled into a standardized geological feature set.

[0136] Step S453: Construct interaction terms between the refined meteorological element set and the standardized geological feature set, calculate the product term of each meteorological derivative element and each geological feature, and generate an interaction feature set, where each interaction feature represents a meteorological-geological interaction.

[0137] Interaction term construction involves combining a refined set of meteorological elements and a standardized set of geological features to calculate their interactions. The product term of each meteorological derivative element and each geological feature is calculated because the interaction between meteorological elements and geological features can be represented by their product. For example, the product of a single rainfall event derivative element and a lithological distribution feature reflects the impact of a single rainfall event on different lithological regions; the product of a diurnal temperature range derivative element and a structural surface density feature reflects the impact of diurnal temperature range on regions with different structural surface densities. The interaction feature set is a collection of features encompassing all meteorological-geological interactions, with each interaction feature representing a specific meteorological-geological interaction. By generating the interaction feature set, the interactions between meteorological and geological factors can be comprehensively considered.

[0138] Step S454: Analyze the correlation between the probability values ​​of instability risk in the interaction feature set and the spatiotemporally aligned data pair. Through correlation analysis, determine the direction and degree of influence of each interaction feature on instability risk, and generate the interaction influence coefficient. A positive coefficient indicates that it promotes instability, and a negative coefficient indicates that it inhibits instability.

[0139] When conducting correlation analysis, correlation analysis methods such as Pearson correlation coefficient or Spearman correlation coefficient can be used. For each interactive feature in the interactive feature set, calculate its correlation coefficient with the instability risk probability value in the spatiotemporally aligned data pair. The correlation coefficient ranges from -1 to 1, with the absolute value closer to 1 indicating a stronger correlation. Based on the sign and magnitude of the correlation coefficient, determine the direction and extent of each interactive feature's influence on instability risk. Organize the interaction influence coefficients of each interactive feature into a coefficient set.

[0140] Step S455: Based on the interaction influence coefficient, extract the coefficients of the interaction characteristics related to rainfall as the degree of influence of rainfall on the permeability of different lithological regions, and extract the coefficients of the interaction characteristics related to temperature change as the weakening effect of temperature change on the strength of the structural surface.

[0141] The interaction coefficient is used because it reflects the degree and direction of each meteorological-geological interaction's impact on slope instability risk. Coefficients of rainfall-related interaction features are extracted as the degree of rainfall's permeability impact on different lithological regions because rainfall-related interaction features (such as the interaction between single rainfall-derived elements and lithological distribution characteristics) are closely related to rainfall's permeability impact on different lithological regions, and their interaction coefficients can quantify the degree of this permeability impact. Coefficients of temperature-related interaction features are extracted as the weakening effect of temperature changes on structural surface strength because temperature-related interaction features (such as the interaction between diurnal temperature range-derived elements and structural surface density characteristics) are related to the impact of temperature changes on structural surface strength, and their interaction coefficients can reflect the degree of this weakening effect. In extracting coefficients, coefficients of rainfall-related interaction features are selected from the set of interaction coefficients, and these coefficients are compiled into a set as the degree of rainfall's permeability impact on different lithological regions. Similarly, coefficients of temperature-related interaction features are selected and compiled into a set as the weakening effect of temperature changes on structural surface strength.

[0142] Step S456: Classify and organize the degree of infiltration influence and weakening effect according to meteorological elements and geological characteristics, and construct an influence relationship matrix with rows representing meteorological elements and columns representing geological characteristics. The matrix elements are the influence intensity of the corresponding meteorological-geological interaction.

[0143] When constructing the influence relationship matrix, the rows and columns are first determined. Rows represent meteorological elements, such as rainfall and temperature changes; columns represent geological features, such as lithology distribution, structural surface density, and weathering degree. Based on the previously extracted data on the permeability influence of rainfall on different lithological regions and the weakening effect of temperature changes on structural surface strength, the corresponding interaction coefficients are filled into the corresponding positions in the matrix. For example, the element where rainfall and lithology distribution intersect is the interaction coefficient corresponding to the permeability influence of rainfall on the lithology distribution region; the element where temperature changes and structural surface density intersect is the interaction coefficient corresponding to the weakening effect of temperature changes on the structural surface density region. After filling all the interaction coefficients into the matrix, the influence relationship matrix with rows representing meteorological elements and columns representing geological features is obtained.

[0144] Step S460: Based on the influence relationship matrix, quantify the influence weights of rainfall and temperature changes on the stability of in-service slopes. The weight values ​​reflect the degree of contribution of meteorological elements to the slope instability risk. Introduce the influence weights into the time evolution model of the probability distribution of instability risk, adjust the model parameters in combination with historical deformation monitoring data, and update the deformation prediction curve of in-service slopes.

[0145] In one implementation, step S460 may specifically include the following steps S461 to S466:

[0146] Step S461: Perform row summation on the influence relationship matrix, calculate the total influence intensity of each meteorological element, and generate the total influence value of the meteorological elements.

[0147] Row summation is performed on the influence matrix because each row represents a meteorological element, and each element in the row represents the intensity of the interaction between that meteorological element and different geological features. Row summation yields the total influence intensity of each meteorological element on all geological features. Calculating the total influence intensity of each meteorological element (rainfall, temperature change) is to comprehensively consider its impact on slope stability under different geological conditions. The total influence intensity is the sum of the absolute values ​​of all interaction coefficients for that element. This is because interaction coefficients can be positive or negative; taking the absolute value eliminates the influence of direction and only considers the degree of influence. The total influence value of meteorological elements is a set containing the total influence intensity of each meteorological element. By generating the total influence value of meteorological elements, the magnitude of the impact of different meteorological elements on slope stability can be intuitively compared.

[0148] Step S462: Normalize the total impact value of meteorological elements so that the sum of the total impact values ​​of rainfall and temperature change is 1, and obtain the impact weight of rainfall and the impact weight of temperature change. The larger the weight value, the more significant the impact of the meteorological element on slope stability.

[0149] Normalizing the total impact value of meteorological elements is to make the influence intensity of different meteorological elements comparable and to intuitively reflect the relative contribution of each meteorological element to slope stability in the form of weights. Making the sum of the total impact values ​​of rainfall and temperature change equal to 1 is a common normalization method. This method converts the total impact value into weighted values; the larger the weighted value, the more significant the impact of that meteorological element on slope stability. The weights for rainfall and temperature change represent the relative importance of these two meteorological elements in influencing slope stability, respectively.

[0150] Step S463: Introduce the weights of rainfall and temperature change into the time evolution model of the probability distribution of instability risk, and adjust the time decay coefficient of meteorological factors in the model. The larger the weight, the smaller the decay coefficient, and the longer the duration of the influence of meteorological factors.

[0151] When introducing weights and adjusting the attenuation coefficient, it is assumed that the influence term of meteorological factors in the time evolution model of the instability risk probability distribution is I(t), which is a function of time t, and can be expressed as the influence term of rainfall I. rain (t) and the effect of temperature change I temp The combination of (t), i.e., I(t) = w rain I rain (t)+wtemp I temp (t). For the rainfall impact term I rain (t), whose time decay coefficient is γ rain It can generally be expressed in the form of exponential decay, such as I rain (t)=I rain (0)e -γraint , where I rain (0) represents the initial rainfall amount. Similarly, the temperature change factor I... temp (t) has a time decay coefficient γ temp I temp (t)=I temp (0)e -γtempt Based on the weighting of rainfall amount w rain The influence of temperature change on weight w temp Adjust the attenuation coefficient when w rain When it is large, decrease γ rain The value of w makes the impact of rainfall persist over a longer period of time; when w temp When it is large, decrease γ temp The value of allows the effects of temperature changes to last longer.

[0152] Step S464: Collect historical deformation monitoring data of in-service slopes, extract slope deformation under different meteorological conditions in history, establish historical meteorological-deformation relationship, and input the relationship as a constraint into the time evolution model.

[0153] When collecting historical deformation monitoring data, data can be obtained from relevant monitoring equipment or databases. This data may include information on slope displacement, tilt, and other deformation variables. Simultaneously, corresponding meteorological data, such as rainfall and temperature, should be collected. This data should be organized and analyzed, and slope deformation variables should be categorized according to different meteorological conditions. For example, slope deformation variables corresponding to different rainfall and temperature ranges can be statistically analyzed separately. Then, data analysis methods, such as regression analysis and machine learning algorithms, can be used to establish the historical meteorological-deformation relationship.

[0154] Step S465: Based on the historical weather-deformation relationship, adjust the parameters of the time evolution model to achieve the optimal consistency between the instability risk probability distribution output by the model and the historical deformation data. The adjustment process is achieved through iterative optimization.

[0155] In iterative optimization, the initial parameters of the time evolution model are first set. Then, historical meteorological data is input into the model to obtain the instability risk probability distribution output by the model. This distribution is compared with historical deformation data, and the errors between them, such as mean squared error (MSE) and mean absolute error (MAE), are calculated. Based on the magnitude and direction of the error, optimization algorithms (such as gradient descent, Newton's method, etc.) are used to adjust the model parameters. Taking gradient descent as an example, the gradient of the error function with respect to the model parameters is calculated, and then the parameters are updated along the opposite direction of the gradient. This process is repeated iteratively until the value of the error function is less than a preset threshold or the maximum number of iterations is reached. Through this iterative optimization process, the instability risk probability distribution output by the model is optimized to match the historical deformation data, thereby improving the model's prediction accuracy.

[0156] Step S466: Apply the adjusted time evolution model to the current probability distribution of instability risk, predict the slope deformation for each future time period, and generate an updated deformation prediction curve. The deformation prediction curve reflects the slope deformation trend after considering the influence of meteorology.

[0157] When applying the adjusted time evolution model, the current instability risk probability distribution and meteorological data are obtained and input into the model. Based on the input data and its own parameters, the model calculates the slope deformation for each future time period. For example, for the t-th future time period, the model outputs the corresponding slope deformation yt. The slope deformations y1, y2, ..., yn for all future time periods are arranged in chronological order, and the updated deformation prediction curve is plotted with time on the horizontal axis and slope deformation on the vertical axis.

[0158] Step S500: Based on the deformation prediction curve and the disaster threshold library in the historical disaster database, a matching analysis is performed to generate a slope maintenance decision table that includes displacement critical value, risk level and maintenance priority.

[0159] In one implementation, step S500 may specifically include the following steps S510-S560:

[0160] Step S510: Extract thresholds from the disaster threshold library in the historical disaster database, obtain the critical displacement index and deformation rate index corresponding to different instability types, determine the safe range of each index, and generate a multi-index threshold matrix. The rows of the multi-index threshold matrix represent instability types, and the columns represent different critical indices.

[0161] The disaster threshold database in the historical disaster database is a database that stores critical index information corresponding to various instability types. These critical indexes are the basis for judging whether a slope is unstable. The displacement critical index is the maximum displacement corresponding to the slope instability, and the deformation rate critical index is the upper limit of the deformation rate at which the slope is unstable. Determining the safe range of each index clarifies within what range the slope is in a relatively safe state, and beyond which there may be a risk of instability.

[0162] During threshold extraction, the required critical displacement and deformation rate indices are selected based on the storage structure and query interface of the disaster threshold database. For each index, its safe range is determined; for example, the safe range for the critical displacement index can be from 0 to a set maximum value, and the safe range for the critical deformation rate index can be a set interval. The critical indices corresponding to different instability types are then organized into a matrix.

[0163] Step S520: Calculate the characteristic index of the deformation prediction curve, extract the predicted displacement index value and deformation rate index value for each future time period, and generate a prediction index sequence. Each element of the prediction index sequence corresponds to the index prediction value for one time period.

[0164] When calculating characteristic indices, the first step is to determine the functional expression or discrete data points of the deformation prediction curve. If it is a functional expression, the deformation rate index value can be calculated using mathematical methods such as differentiation. For the displacement index value, the displacement value at the corresponding time can be directly obtained from the function. If it is discrete data points, the deformation rate can be calculated using numerical difference methods. For example, for two adjacent time points t... i and t i+1 The corresponding displacement y i and y i+1 Deformation rate v i =(y i+1 -y i ) / (t i+1 -t i For each time period, the calculated predicted displacement index value and deformation rate index value are combined into an index vector. The index vectors of all time periods are arranged in chronological order to generate a predicted index sequence.

[0165] Step S530: Match the predicted index sequence with the multi-index threshold matrix index by index, compare the relationship between each predicted index value and the corresponding safe range of the index, and when the index value exceeds the safe range, record the instability type and the degree of exceeding the threshold corresponding to the index, and generate an index exceeding record. Each record includes the index name, the degree of exceeding the threshold and the corresponding instability type.

[0166] During index-by-index matching, each element in the predicted index sequence is traversed. For each time period's index vector, the predicted displacement index value and deformation rate index value are compared with the critical displacement index and deformation rate index corresponding to each instability type in the multi-index threshold matrix. Assume the k-th element in the predicted index sequence is (y... k ,v k The multi-index threshold matrix is ​​T. For the i-th type of instability, the critical index for displacement is T. i1 The critical index for deformation rate is T. i2 If y k >T i1 Or v k >T i2 If the indicator exceeds the safe range, record the indicator name (e.g., displacement or deformation rate) and the degree of exceeding the threshold (e.g., y). k -T i1 or v k -T i2 The records of exceeding the threshold are compiled into a list of instability records. Each record can be represented by a dictionary, such as {"Indicator Name":"Displacement","Degree of Exceeding Threshold":5.2,"Corresponding Instability Type":"Landslide"}.

[0167] Step S540: Based on the records of indicators exceeding the standard, a comprehensive risk level is determined. The exceedance of multiple indicators is considered. When multiple indicators exceed the standard at the same time, the risk level is determined to be higher than that of a single indicator exceeding the standard. A risk level determination result is generated, which includes the comprehensive risk level for each time period.

[0168] In one implementation, step S540 may specifically include the following steps S541 to S546:

[0169] Step S541: Quantify the degree of exceeding the threshold for each indicator in the record of exceeding the threshold, convert the degree of exceeding the threshold into a numerical form, the larger the value, the more serious the exceedance, and generate a quantitative value of the degree of exceedance.

[0170] Quantifying the degree of exceeding the threshold for each indicator in the record of indicators exceeding the standard is to represent the degree of exceeding the threshold in a unified and comparable numerical form. During the quantification of the degree of exceeding the threshold, the degree of exceeding the threshold information is extracted from each record in the record of indicators exceeding the standard. If the degree of exceeding the threshold is itself a numerical value, such as the specific numerical value of the displacement exceeding the threshold, this value can be directly used as the quantification value. If the degree of exceeding the threshold is a relative description, such as "slight exceeding the standard," "moderate exceeding the standard," "severe exceeding the standard," etc., it can be converted into a numerical value according to pre-defined rules. For example, "slight exceeding the standard" corresponds to the numerical value 1, "moderate exceeding the standard" corresponds to the numerical value 2, and "severe exceeding the standard" corresponds to the numerical value 3. The quantified value of the degree of exceeding the threshold for each indicator is then compiled into a list, such as [2,3,1,...].

[0171] Step S542: Assign importance weights to each indicator. The importance weights are determined based on the degree of contribution of the indicator to the instability warning in historical disaster cases. The higher the degree of contribution, the greater the weight. Generate an indicator weight vector.

[0172] When determining the importance weights of indicators, historical disaster case data is first collected and analyzed to statistically determine the frequency of each indicator's occurrence in instability early warning and its impact on instability judgment. For example, analysis reveals that in most landslide disaster cases, the deformation rate indicator has a greater impact on instability early warning, while the displacement indicator has a relatively smaller impact. Therefore, a larger weight can be assigned to the deformation rate indicator, and a smaller weight to the displacement indicator. Assuming there are two indicators: displacement and deformation rate, and analysis determines that the importance weight of the displacement indicator is 0.3 and the importance weight of the deformation rate indicator is 0.7, then the indicator weight vector is [0.3, 0.7].

[0173] Step S543: Calculate the comprehensive exceedance index by weighted summation of the quantitative value of the exceedance degree and the indicator weight vector. The comprehensive exceedance index reflects the overall severity of the exceedance of multiple indicators. The larger the index value, the more serious the overall exceedance.

[0174] Step S544: Divide the risk level intervals according to the range of the comprehensive excess index. Different intervals correspond to different risk levels. Generate risk level classification standards. Each interval includes the range of the comprehensive excess index and the corresponding risk level.

[0175] When dividing risk level ranges, first determine the range of values ​​for the comprehensive exceedance index. Based on historical data and experience, set different thresholds to divide the ranges. For example, define a comprehensive exceedance index less than 1 as a low-risk range, corresponding to a low-risk level; a comprehensive exceedance index between 1 and 2 as a medium-risk range, corresponding to a medium-risk level; and a comprehensive exceedance index greater than 2 as a high-risk range, corresponding to a high-risk level. Compile each range and its corresponding risk level into a list, such as [{"Comprehensive Exceedance Index Range":"0-1","Corresponding Risk Level":"Low Risk"},{"Comprehensive Exceedance Index Range":"1-2","Corresponding Risk Level":"Medium Risk"},{"Comprehensive Exceedance Index Range":"2-+∞","Corresponding Risk Level":"High Risk"}].

[0176] Step S545: Compare the calculated comprehensive exceedance index with the risk level classification standard to determine its risk level range and generate a preliminary risk level.

[0177] During the comparison, each interval in the risk level classification standard is traversed to determine whether the calculated comprehensive exceedance index falls within that interval. If it does, the risk level corresponding to that interval is determined as the preliminary risk level. For example, if the calculated comprehensive exceedance index is 2.5, and the "2-+∞" interval in the risk level classification standard corresponds to a high risk level, then the preliminary risk level is high risk.

[0178] Step S546: Perform multi-indicator collaborative verification on the preliminary risk level. When multiple indicators exceed the standard at the same time and the direction of exceeding the standard is consistent, the preliminary risk level is appropriately increased. When the direction of exceeding the standard of the indicators is inconsistent, the preliminary risk level is maintained, and a risk level judgment result is generated.

[0179] When multiple indicators exceed the standard simultaneously and in the same direction, it indicates that the slope instability risk may be higher than initially assessed, thus the initial risk level should be appropriately increased. When the indicators exceed the standard in inconsistent directions, it indicates that the abnormalities in multiple indicators may mutually constrain each other, and their impact on the instability risk is relatively small, thus the initial risk level should be maintained. Generating the risk level determination result involves organizing and recording the risk levels after multi-indicator collaborative verification to obtain the final risk level determination result.

[0180] When conducting multi-indicator collaborative verification, the first step is to analyze whether multiple indicators simultaneously exceed the limits and in the same direction. Consistent exceedance direction means that the exceedances of multiple indicators all point in an increased risk of slope instability; for example, both displacement and deformation rate exceed the limits significantly at the same time. If this situation exists, the initial risk level is appropriately increased according to pre-set rules, such as raising low risk to medium risk, and medium risk to high risk. If the exceedance directions are inconsistent, such as displacement exceeding the limit but deformation rate not exceeding it, or the exceedance trends of the two being contradictory, the initial risk level is maintained. The verified risk levels are then compiled into a risk level determination result list to provide accurate risk information for subsequent maintenance decisions.

[0181] Step S550: Based on the risk level determination result, query the preset maintenance measure plan library to obtain information such as maintenance methods, required resources, and implementation cycle for the corresponding risk level, and generate maintenance measure suggestions that match the risk level.

[0182] The pre-set maintenance measure plan library is a database storing information on maintenance measures corresponding to different risk levels. This information includes maintenance methods, required resources, and implementation cycles. Searching the pre-set maintenance measure plan library based on the risk level assessment result is to find suitable maintenance measures according to the actual risk level of the slope. Obtaining information such as maintenance methods, required resources, and implementation cycles for the corresponding risk level provides specific operational guidance and resource demand planning for slope maintenance work. Generating maintenance measure suggestions involves organizing and recording the retrieved maintenance measure information, matching the maintenance measure suggestions with the risk level to ensure the effectiveness and relevance of the maintenance measures.

[0183] When performing a query, for each risk level in the risk assessment results, the corresponding maintenance measures information is searched in the maintenance measure solution database. For example, if the risk level is low, the corresponding maintenance method might be regular inspections and small-scale reinforcement; if the risk level is high, the corresponding maintenance method might be large-scale reinforcement projects or emergency rescue measures. For each maintenance method, the required resources, such as manpower, material resources, and financial resources, as well as the implementation period, i.e., the time required to complete the maintenance work, are obtained. This information is compiled into a maintenance measure suggestion list, and each suggestion can be represented by a dictionary, such as {"Risk Level":"Low Risk","Maintenance Method":"Regular Inspections and Small-Scale Reinforcement","Required Resources":"2 Inspection Personnel, Small Amount of Reinforcement Materials","Implementation Period":"Weekly Inspections, Monthly Reinforcement"}.

[0184] Step S560: Based on the maintenance measures recommendations and the spatial location and surrounding environmental importance of the in-service slope, calculate the maintenance priority. The priority calculation comprehensively considers the risk level, maintenance urgency, and implementation difficulty to generate a maintenance priority ranking. Integrate the displacement critical value, risk level judgment result, maintenance measures recommendations, and maintenance priority ranking into a slope maintenance decision table.

[0185] When calculating maintenance priorities, appropriate weights are first assigned to factors such as risk level, maintenance urgency, and implementation difficulty. For example, the weight for risk level is 0.5, the weight for maintenance urgency is 0.3, and the weight for implementation difficulty is 0.2. For each in-service slope, a quantitative score is calculated based on its risk level, maintenance urgency, and implementation difficulty. The risk level can be quantified based on the previous risk level assessment results, such as low risk corresponding to 1 point, medium risk corresponding to 2 points, and high risk corresponding to 3 points. The maintenance urgency can be assessed based on the actual condition of the slope, such as slopes that are about to become unstable having high maintenance urgency, corresponding to 3 points; slopes with some risk but not likely to become unstable in the short term having medium maintenance urgency, corresponding to 2 points; and slopes with low risk and no significant changes in the short term having low maintenance urgency, corresponding to 1 point. The difficulty of implementation can be assessed based on factors such as the slope's spatial location and surrounding environment. For example, slopes with open spaces and no surrounding obstacles have low implementation difficulty, corresponding to 1 point; slopes with narrow spaces and some surrounding obstacles have medium implementation difficulty, corresponding to 2 points; and slopes with complex spaces and important surrounding facilities have high implementation difficulty, corresponding to 3 points. Then, a maintenance priority score is calculated for each slope according to the assigned weights, and all slopes are sorted from highest to lowest maintenance priority score to generate a maintenance priority ranking. Finally, the displacement threshold, risk level assessment results, maintenance measure recommendations, and maintenance priority ranking are integrated into a single table to form a slope maintenance decision table. This table clearly displays relevant information and maintenance decision recommendations for each in-service slope.

[0186] This invention provides a computer system including a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the program, it implements some or all of the steps in the above-described method.

[0187] Figure 2 A hardware entity diagram of a computer system provided as an embodiment of the present invention, such as... Figure 2 As shown, the hardware entity of the computer system 1000 includes a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can run on the processor 1001, and the processor 1001 executes the program to implement the steps in the method of any of the above embodiments.

[0188] The memory 1002 stores computer programs that can run on the processor. The memory 1002 is configured to store instructions and applications that can be executed by the processor 1001. It can also cache data to be processed or already processed (e.g., image data, audio data, voice communication data, and video communication data) of the processor 1001 and various modules in the computer system 1000. It can be implemented by flash memory or random access memory (RAM).

[0189] When the processor 1001 executes the program, it implements the steps of any of the above-mentioned machine vision-based in-service slope inspection methods. The processor 1001 typically controls the overall operation of the computer system 1000.

Claims

1. A machine vision-based method for inspecting in-service slopes, characterized in that, The method includes: Continuously acquired visible light image data and synchronously acquired infrared thermal imaging data of in-service slopes, as well as three-dimensional point cloud data with the same acquisition timestamp as the visible light image data and infrared thermal imaging data, are used to construct a composite dataset with spatiotemporal alignment. Multi-dimensional information extraction processing is performed on the composite dataset to obtain the texture change sequence of the in-service slope surface and the temperature gradient distribution corresponding to the infrared thermal imaging data; The texture change sequence and the temperature gradient distribution are input into a pre-trained transfer learning model. The model is then optimized using slope instability case data from a historical disaster database to output the probability distribution of instability risk for the in-service slope. Specifically, this includes: enhancing the texture change sequence across time scales by decomposing the texture change trends at different time periods into a multi-resolution set, generating a multi-scale texture feature set whose feature dimensions match the input layer dimension of the pre-trained transfer learning model; encoding the temperature gradient distribution using spatial features by converting the three-dimensional temperature gradient data into a multi-channel feature map, where each channel corresponds to a temperature gradient component in a different direction; enhancing gradient direction information through inter-channel correlation analysis to generate a temperature gradient feature map, the size of which is consistent with the spatial dimension of the multi-scale texture feature set; and fusing the multi-scale texture feature set and the temperature gradient feature map across modalities by constructing a cross-modal attention mechanism, calculating the mutual information between texture and temperature features, and allocating feature fusion based on the mutual information magnitude. Weighing and generating a fused feature tensor, which contains collaborative correlation information between texture and temperature; extracting slope instability case data from the historical disaster database, standardizing the texture changes and temperature gradient features in the case data to adapt their feature distribution to the distribution of the fused feature tensor, and establishing a case feature library containing feature patterns corresponding to different instability types; associating and matching the fused feature tensor with the case features in the case feature library, adding case similarity labels to the fused feature tensor through feature pattern similarity analysis, inputting the fused feature tensor with similarity labels into a pre-trained transfer learning model, adjusting the parameters of the model's feature extraction layer and output layer to enable the model to learn the feature association between the current slope and historical cases; after multiple rounds of iterative optimization, inputting the fused feature tensor into the optimized pre-trained transfer learning model, outputting the instability probability of each area of ​​the in-service slope in the future time period, and correcting the outlier values ​​of the instability probability of isolated areas by combining the slope spatial topology, generating an instability risk probability distribution covering the entire in-service slope; The probability distribution of instability risk is spatiotemporally matched with real-time meteorological data. The influence weights of rainfall and temperature changes on the stability of the in-service slope are quantified through a meteorological-geological coupling model, and the deformation prediction curve of the in-service slope is updated. Based on the deformation prediction curve and the disaster threshold database in the historical disaster database, a slope maintenance decision table is generated that includes displacement critical value, risk level and maintenance priority.

2. The method according to claim 1, characterized in that, The process of extracting multi-dimensional information from the composite dataset to obtain the texture change sequence of the in-service slope surface and the temperature gradient distribution corresponding to the infrared thermal imaging data includes: Spatiotemporal indexing is constructed on the continuously acquired visible light image data, synchronously acquired infrared thermal imaging data, and three-dimensional point cloud data contained in the composite dataset. A unified timestamp and absolute coordinate label are added to each data set to generate a three-dimensional data volume with spatiotemporal identifiers. Each data unit in the three-dimensional data volume contains the visible light pixel value, infrared temperature value, and point cloud three-dimensional coordinates at the corresponding time coordinate. Multimodal spatial registration is performed on continuously acquired visible light image data and synchronously acquired infrared thermal imaging data in the three-dimensional data volume. The coordinate mapping relationship between visible light pixels and infrared pixels is established by feature point matching. The registration deviation is corrected by combining the geometric constraints of the three-dimensional point cloud data to generate a registration data pair with spatial consistency. The spatial resolution of the registration data pair is adapted to the sampling density of the three-dimensional point cloud data. Dynamic window texture extraction is performed on the continuously acquired visible light image data in the registration data pair. The image is traversed with a preset spatial window, and the change amplitude and direction of gray value within the window are calculated. The texture feature evolution of the same window in the continuous acquisition period is tracked in combination with the timestamp information to generate an initial dynamic texture feature set. The initial dynamic texture feature set includes the texture change direction and change amplitude of each window. The synchronously acquired infrared thermal imaging data in the registration data pair are used to calculate the neighborhood temperature difference. Taking each pixel as the center, the temperature difference between it and the surrounding neighboring pixels is calculated. The elevation information of the three-dimensional point cloud data is combined to correct the influence of slope aspect on temperature distribution and generate a preliminary temperature gradient field. The spatial range of the preliminary temperature gradient field covers the entire surface of the in-service slope. The initial texture dynamic feature set is spatially topologically correlated with the three-dimensional point cloud data in the three-dimensional data volume. Based on the topological structure of the point cloud data, the slope surface is divided into multiple connected regions. The texture changes of all dynamic windows in each connected region are tracked. Through the synergistic analysis of texture changes in connected regions, a texture change sequence with spatial correlation is generated. The texture change sequence contains the overall texture evolution trend of each connected region in a continuous time period. The initial temperature gradient field is subjected to spatial structural surface constraint processing. Combined with the slope structural surface information extracted from the three-dimensional point cloud data, the abnormal values ​​of the temperature gradient near the structural surface are corrected. The angle relationship between the gradient direction and the orientation of the structural surface is calculated, and the planar temperature gradient field is converted into a three-dimensional temperature gradient distribution.

3. The method according to claim 2, characterized in that, The process involves spatially and topologically associating the initial dynamic texture feature set with the 3D point cloud data in the 3D data volume. Based on the topological structure of the point cloud data, the slope surface is divided into multiple connected regions. Texture changes of all dynamic windows within each connected region are tracked. Through synergistic analysis of texture changes within connected regions, a spatially correlated texture change sequence is generated, including: A topological structure is constructed on the three-dimensional point cloud data in the three-dimensional data volume, the adjacency relationship between each point in the point cloud is calculated, and a three-dimensional topological map of the slope surface is generated. The nodes of the three-dimensional topological map are the points in the point cloud data, and the edges represent the spatial proximity relationship between the points. Based on the three-dimensional topology map, the region is divided by growth. The region growth seed point is determined based on the curvature change. Points that are spatially adjacent and have a curvature difference of less than a preset threshold are aggregated into the same connected region to generate the connected region division result of the slope surface. Each connected region contains a set of point cloud data points with similar geometric features. Extract the spatial coordinates of each dynamic window from the initial texture dynamic feature set, assign the dynamic window to the corresponding connected region according to the connected region partitioning result, establish a region-window mapping relationship, and each connected region contains at least one dynamic window; The texture changes of all dynamic windows within each connected region are aligned over time, and the average and standard deviation of the texture change amplitude of all windows within the region are calculated to generate regional texture change statistics. The evolution trend of the regional texture change statistics over time is analyzed. By calculating the rate of change of the statistics within adjacent time periods, the acceleration or deceleration characteristics of regional texture change are determined. Combined with regional geometric features, it is determined whether the changes are synergistic. Arrange the collaborative texture change trends of each connected region in chronological order, and combine them with regional spatial coordinate information to generate a spatially correlated texture change sequence covering the entire surface of the in-service slope.

4. The method according to claim 1, characterized in that, The step of fusing the multi-scale texture feature set with the temperature gradient feature map across modalities, constructing a cross-modal attention mechanism, calculating the mutual information between texture features and temperature features, assigning feature fusion weights based on the mutual information magnitude, and generating a fused feature tensor includes: The multi-scale texture feature set is expanded in channel dimension, and texture features at different time scales are mapped to independent channels to generate a multi-channel texture feature map. The number of channels in the multi-channel texture feature map is the same as the number of time scales. The temperature gradient feature map is spatially enhanced by calculating the spatial distribution histogram of the gradient direction, encoding the gradient direction information into spatial features, and generating an enhanced temperature feature map. The spatial resolution of the enhanced temperature feature map is consistent with that of the multi-channel texture feature map. Calculate the mutual information between the multi-channel texture feature map and the enhanced temperature feature map at each spatial location. The mutual information value represents the degree of correlation between the two features at that spatial location, and generate a mutual information weight map. Based on the mutual information weight map, the multi-channel texture feature map and the enhanced temperature feature map are weighted and summed to generate a preliminary fused feature map. The number of channels in the preliminary fused feature map is the sum of the number of channels in the two feature maps. Feature interaction modeling is performed on the preliminary fused feature map. By calculating the feature correlation between different channels, channel attention weights are constructed to generate channel-enhanced fused features. The channel-enhanced fusion features are integrated with spatial and channel dimensions and reshaped into a three-dimensional tensor form, generating a fusion feature tensor containing texture-temperature co-processing information.

5. The method according to claim 1, characterized in that, The process of associating and matching the fused feature tensor with case features in the case feature library, adding case similarity labels to the fused feature tensor through feature pattern similarity analysis, inputting the fused feature tensor with similarity labels into a pre-trained transfer learning model, and adjusting the parameters of the model's feature extraction layer and output layer to enable the model to learn the feature association between the current slope and historical cases includes: The case features in the case feature library are clustered by feature pattern, and cases with similar feature distributions are grouped into one class. The feature center of each class of cases is extracted to generate a case feature pattern library, and each feature pattern represents a type of instability feature. Calculate the similarity between the fused feature tensor and each feature pattern in the case feature pattern library, and generate a similarity distribution vector, where each element of the similarity distribution vector corresponds to a similarity value of a feature pattern; Based on the feature pattern corresponding to the maximum similarity value in the similarity distribution vector, add case similarity labels to the fused feature tensor. The label content includes the instability type of the case and the degree of similarity. Input the fused feature tensor with similarity labels into the pre-trained transfer learning model, freeze the parameters of the model's bottom feature extraction layer, and only adjust the parameters of the top output layer to allow the model to initially adapt to the new feature distribution. After the model has initially adapted, some parameters of the bottom feature extraction layer are gradually unfrozen, and the unfreezing is performed layer by layer from the top to the bottom. After each layer is unfrozen, the parameters are fine-tuned so that the bottom feature extraction can capture the features of the current slope. Through multiple rounds of layer-by-layer unfreezing and parameter fine-tuning, the feature extraction layer of the pre-trained transfer learning model can simultaneously learn the general features of historical cases and the features of the current slope, thereby achieving feature association learning between the current slope and historical cases.

6. The method according to claim 2, characterized in that, After multiple rounds of iterative optimization, the fused feature tensor is input into the optimized pre-trained transfer learning model, which outputs the instability probability of each region of the in-service slope in the future time period. The model then combines the slope's spatial topology to correct outliers in isolated instability probability areas, generating an instability risk probability distribution covering the entire in-service slope, including: The fused feature tensor is input into the optimized pre-trained transfer learning model, and the model outputs the instability probability of each spatial location on the surface of the in-service slope in multiple future time periods, generating a preliminary instability probability map. The spatial resolution of the preliminary instability probability map is consistent with the spatial dimension of the fused feature tensor. Spatial topology analysis is performed on the preliminary instability probability map. Based on the connected region division results, the average value and distribution range of instability probability in each connected region are calculated. The existence of isolated high instability probability points in the connected regions is identified. The high instability probability points are points whose instability risk probability exceeds a preset multiple of the region average. For isolated high instability probability points identified, a neighborhood consistency check is performed. The difference between the point and the average instability probability of its connected region is compared. When the difference exceeds a preset range, it is judged as an outlier and corrected based on the neighborhood instability probability. After correcting outliers, the rate of change of the instability probability of each connected region over time is calculated, the accelerating trend of the rate of change is analyzed, the rate of change is compared with the rate of change before instability in historical cases, and the temporal distribution of the instability probability of connected regions is adjusted. Based on the spatial relationship of each connected region, the correlation of the instability probability of adjacent connected regions is calculated. Through the analysis of the instability risk transmission between connected regions, the instability probability value of the edge region is adjusted. The corrected instability probabilities of all connected regions are spliced ​​together according to the spatial topology to generate an instability risk probability distribution covering the entire surface of the in-service slope. The instability risk probability distribution includes the instability probability value of each spatial location in future time periods.

7. The method according to claim 2, characterized in that, The step of performing spatiotemporal matching of the instability risk probability distribution with real-time acquired meteorological data, quantifying the impact weights of rainfall and temperature changes on the stability of the in-service slope through a meteorological-geological coupling model, and updating the deformation prediction curve of the in-service slope includes: Time series are constructed from real-time meteorological data, and rainfall and temperature changes are organized in chronological order. A timestamp is added to each meteorological element to generate a meteorological time series. The time dimension of the instability risk probability distribution is extracted to obtain all the time period information contained therein, the start and end times of each time period are determined, and a risk time axis is generated. The meteorological time series is aligned with the risk time axis so that each meteorological element value corresponds to the corresponding time period of the instability risk probability distribution. The time granularity difference between meteorological data and risk data is compensated by time interpolation processing to generate spatiotemporally aligned data pairs. Each spatiotemporally aligned data pair contains meteorological elements of one time period and the instability risk probability of the corresponding period. Geological features are extracted from the three-dimensional point cloud data in the three-dimensional data volume, the geological attributes of the slope are analyzed, and a geological vulnerability feature set is generated. Each feature in the geological vulnerability feature set corresponds to a quantitative description of a geological attribute of the slope. The spatiotemporal aligned data pairs and the geological vulnerability feature set are input into the meteorological-geological coupling model to analyze the interaction between meteorological elements and geological vulnerability features, calculate the degree of infiltration impact of rainfall on different lithological regions, the weakening effect of temperature changes on the strength of structural surfaces, and generate an influence relationship matrix. The rows of the influence relationship matrix represent meteorological elements, the columns represent geological vulnerability features, and the elements represent the intensity of the interaction between the two. Based on the aforementioned influence relationship matrix, the influence weights of rainfall and temperature changes on the stability of in-service slopes are quantified. The weight values ​​reflect the degree of contribution of meteorological elements to the slope instability risk. The influence weights are introduced into the time evolution model of the probability distribution of instability risk, and the model parameters are adjusted in combination with historical deformation monitoring data to update the deformation prediction curve of in-service slopes.

8. The method according to claim 7, characterized in that, The step of aligning the meteorological time series with the risk time axis, so that each meteorological element value corresponds to the corresponding time period of the instability risk probability distribution, and using time interpolation to compensate for the time granularity difference between meteorological data and risk data to generate spatiotemporally aligned data pairs, includes: The metadata of the meteorological time series is parsed to obtain the start time, end time and collection interval of the meteorological data, and to determine the time base of the meteorological time axis; The time period division rules of the risk time axis are analyzed to determine the start and end times of each time period and generate a time period division table. Each table entry contains a period number and the corresponding time range. Based on the time period division table, the meteorological time series is periodically segmented, meteorological element data within each time period is extracted, the cumulative value or average value of meteorological elements within the period is calculated, and a periodic meteorological element set is generated. Each periodic meteorological element set contains the meteorological element statistical values ​​of one time period. When the collection interval of the meteorological time series is inconsistent with the time period of the risk time axis, the meteorological element data is resampled so that the time node of the resampled meteorological data is consistent with the time period node of the risk time axis. The time integrity of the interpolated periodic meteorological element set is checked to see if there are any missing meteorological data in the time period. If there are any missing data, trend filling is performed based on the meteorological data of the adjacent periods to ensure that each time period has a corresponding meteorological element value. The periodic meteorological element set for each time period is combined with the instability risk probability value for that period to generate spatiotemporally aligned data pairs. Each spatiotemporally aligned data pair includes a time period identifier, meteorological element statistics, and the instability risk probability for the corresponding period.

9. A computer system comprising a memory and a processor, the memory storing a computer program executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 8.

Images