Loess slope stability intelligent monitoring method and system

Abstract

The application discloses a loess slope stability intelligent monitoring method and system, the method comprises the following steps: regionally differentiating modeling of multi-source data of a target area, outputting slope change data; according to the periodical displacement time series data, the slope deformation trend is predicted based on the LSTM algorithm; the slope change data, the slope deformation trend and the meteorological rainfall data are input into the slope stability evaluation model to predict the safety factor, if the safety factor is lower than the preset threshold, the warning information is issued. The application provides a loess slope stability intelligent monitoring method and system, by accurately extracting slope change, deformation trend and other data, with the help of cross-modal attention, multi-source data and meteorological rainfall data are fused, then the random forest model is used to learn the multi-factor coupling law, finally, the safety factor is predicted by comprehensively analyzing multi-factor data to issue a warning, solving the problem that it is difficult to fuse slope, deformation and meteorological rainfall multi-factor to accurately evaluate the slope stability, resulting in low warning reliability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of slope monitoring, in particular to a loess slope stability intelligent monitoring method and system. BACKGROUND

[0002] Loess slopes are prone to instability and landslides under the influence of rainfall and engineering activities. Conducting stability monitoring can identify and warn of risks in advance, reduce damage to transportation, water conservancy and other facilities, and effectively protect the safety of residents' lives and property. At the same time, monitoring data can support the safe operation of infrastructure, assist in regional ecological protection and scientific planning of land use, and help build a safe living environment and promote sustainable development in the loess area. Traditional slope monitoring is mainly achieved through manual inspection combined with single-point instrument monitoring. Workers need to go to the site regularly to observe whether cracks appear on the slope and whether the slope surface deforms, and make detailed records. At the same time, devices such as inclinometers, settlement gauges and osmometers are arranged at key positions of the slope to monitor data such as deep soil displacement, ground settlement and pore water pressure. The data is manually sorted or stored by simple devices, and when the monitoring index exceeds the preset safety value, the early warning process is started.

[0003] However, the traditional slope monitoring technology has obvious shortcomings. Manual inspection requires regular on-site work, which is low in efficiency and limited in coverage, making it difficult to capture the dynamic of the slope in real time. Single-point instrument monitoring data is only limited to the local area and cannot reflect the overall condition of the slope. The data needs to be manually sorted or simply stored, which may cause lag and lead to delayed response to early warning. Not only is the labor input large and the cost high, but also errors may be introduced due to subjective judgment by humans, and it is difficult to integrate slope, deformation and meteorological rainfall factors to achieve accurate stability evaluation, which cannot effectively support reliable early warning. SUMMARY

[0004] The present application provides a loess slope stability intelligent monitoring method and system to solve the problem of low reliability of early warning due to the difficulty in integrating slope, deformation and meteorological rainfall factors to accurately evaluate the stability of the slope.

[0005] To achieve the above-mentioned purpose, the present application provides a loess slope stability intelligent monitoring method, comprising:

[0006] obtaining multi-source data and periodic displacement time series data of a target area;

[0007] The multi-source data is subjected to regional differential modeling, and slope change data is output, specifically: the multi-spectral image and the original point cloud in the multi-source data are corrected and aligned to obtain precise registration data; the precise registration data is input into an improved PSPNet network to segment the slope crack and gully development area, and a binary segmentation mask is obtained; the precise registration data is divided into different regional data according to the binary segmentation mask and spliced according to the spatial coordinates to generate a DEM model covering the entire slope; and the slope change data is output according to the DEM model.

[0008] According to the periodical displacement time series data, a slope deformation trend is predicted based on an LSTM algorithm.

[0009] The slope change data, the slope deformation trend and meteorological rainfall data are input into a slope stability evaluation model to predict a safety coefficient, and if the safety coefficient is lower than a preset threshold, an early warning information is issued; wherein the slope stability evaluation model is established based on a random forest algorithm by learning the mapping rule between a fusion feature vector and a stability level; and the fusion feature vector and the stability level are obtained by cross-modal attention fusion calculation and vector classification on historical multi-source data and meteorological rainfall data.

[0010] The improved PSPNet is used to segment the crack and gully area, the key instability area of the slope can be accurately identified, the DEM model is generated by regional differential modeling, the slope change data can reflect the characteristics of different regions, and the slope change data can be accurately corresponded with the deformation trend and the rainfall influence; and by generating the DEM model covering the whole area, the slope change data is output, which can avoid the model establishment deviation caused by the slope data problem. The LSTM algorithm is used to predict the slope deformation trend based on the periodical displacement time series data, which can accurately quantify the time evolution characteristics of the deformation and provide reliable data for the deformation factors. The random forest algorithm is used to learn the mapping rule between the fusion feature vector and the stability level, and the slope stability evaluation model can quantify the non-linear correlation of the slope, deformation and meteorological rainfall, and the safety coefficient is predicted by inputting the multi-factor data into the model, the prediction result is more close to the real state, the multi-factor fusion problem is solved, the artificial inspection cost is greatly reduced, and the early warning reliability is improved.

[0011] Compared with the prior art, the slope change, deformation trend and other single-factor data are accurately extracted, the multi-source data and meteorological rainfall data are fused by cross-modal attention, the random forest model is used to learn the multi-factor coupling rule, and finally the safety coefficient is predicted by comprehensively considering the slope, deformation and meteorological rainfall, so that the problem that it is difficult to accurately evaluate the slope stability by fusing the slope, deformation and meteorological rainfall is solved, and the early warning reliability is improved.

[0012] As a preferred solution, the multispectral images and the original point cloud in the multi-source data are corrected and aligned to obtain fine registration data, specifically:

[0013] The multispectral images in the multi-source data are orthorectified and radiometrically corrected to obtain corrected multispectral images;

[0014] For the original point cloud in the multi-source data, outliers with a distance mean greater than a preset multiple standard deviation are removed by statistical filtering to obtain a filtered point cloud;

[0015] Taking the target coordinates as a reference, the corrected multispectral images and the filtered point cloud are aligned according to the ICP algorithm to obtain coarse registration parameters in the form of a rotation matrix and a translation vector;

[0016] A preliminary mask is generated according to the corrected multispectral images and the filtered point cloud, and a crack light-dark boundary line and a gully terrain mutation line are extracted from the preliminary mask to obtain image edge points and point cloud edge points;

[0017] Taking the image edge points and the point cloud edge points as constraints, the coarse registration parameters are optimized according to the ICP algorithm, so that the registration error of the corrected multispectral images and the filtered point cloud is less than a preset threshold, to obtain the fine registration data.

[0018] In this preferred solution, for multispectral images and three-dimensional laser point clouds, the data quality is first improved through orthorectification, radiometric correction and filtering processing, then coarse registration is realized in combination with target coordinates and the ICP algorithm, and finally the edge points are used as constraints to optimize the parameters to ensure that the registration error is less than a preset threshold; this process not only removes noise interference, but also improves the spatial consistency of multi-source data through a double registration strategy, providing high-precision basic data for subsequent slope feature extraction and modeling, which can effectively avoid analysis errors caused by data misplacement.

[0019] As a preferred solution, a preliminary mask is generated according to the corrected multispectral images and the filtered point cloud, specifically:

[0020] Based on the corrected multispectral images and the filtered point cloud, a spatial correlation between image pixel coordinates and point cloud three-dimensional coordinates is established according to the coarse registration parameters;

[0021] A contrast enhancement process is performed on the green band in the corrected multispectral images, and a target continuous dark edge is extracted to generate an image dark edge mask;

[0022] Based on the pixel coordinates of the image dark edge mask, a corresponding three-dimensional coordinate region in the point cloud is obtained through the spatial correlation, and the Z coordinate gradient of the three-dimensional coordinate region is calculated, and a point cloud region with a Z coordinate gradient greater than a preset value is screened out, and is converted into a point cloud gradient anomaly mask;

[0023] The image dark edge mask and the point cloud gradient anomaly mask are subjected to spatial intersection operation, and the coincident region is retained, to obtain the preliminary mask about the cracks and gullies.

[0024] The preferred scheme establishes the spatial correlation of the image and the point cloud based on the coarse registration parameters, extracts the image dark edge through the green light band enhancement, screens the abnormal region in combination with the Z coordinate gradient of the point cloud, and finally generates the preliminary mask through the spatial intersection operation, which can fully utilize the spectral characteristics of the image and the three-dimensional geometric information of the point cloud, accurately lock the crack and gully region, reduce the interference of irrelevant regions, provide reliable constraints for subsequent fine registration parameter optimization, and improve the identification accuracy of the key feature region of the slope.

[0025] As a preferred scheme, after the fine registration data is divided into different region data according to the binary segmentation mask, the spatial coordinates are spliced to generate a DEM model covering the entire slope, specifically:

[0026] According to the binary segmentation mask, the fine registration data is divided into crack region point cloud, gully region point cloud and complete region point cloud;

[0027] The crack region point cloud is subjected to encryption resampling processing, the gully region point cloud is subjected to smoothing filtering and the target side wall steep edge is retained, and the complete region point cloud is subjected to downsampling processing, to obtain a processed comprehensive semantic enhanced point cloud;

[0028] Based on the crack region encryption point cloud in the comprehensive semantic enhanced point cloud, the depth discontinuity of the crack region is retained through Poisson reconstruction and edge constraint, to generate a crack region DEM;

[0029] Based on the gully region denoising point cloud in the comprehensive semantic enhanced point cloud, a gully region DEM is generated through inverse distance weighted interpolation and fluid mechanics correction;

[0030] Based on the complete region simplified point cloud in the comprehensive semantic enhanced point cloud, a complete region DEM is generated through Kriging interpolation;

[0031] The crack region DEM, the gully region DEM and the complete region DEM are spliced according to the spatial coordinates, to generate the DEM model covering the entire slope.

[0032] The preferred scheme differentiates the point clouds in different regions according to a binary segmentation mask, generates DEMs in different regions and splices them; this regional modeling strategy not only ensures the fine expression of key regions such as cracks and gullies, but also improves the overall modeling efficiency through simplification, and the generated DEM model can comprehensively reflect the complex topographic features of the slope, providing a high-precision basis for slope analysis.

[0033] As a preferred scheme, the slope change data is output according to the DEM model, specifically:

[0034] For each grid in the DEM model, the initial slope is calculated by the Horn algorithm to obtain an initial slope map;

[0035] For the initial slope map, the crack area range is locked and the crack direction is extracted according to the segmentation mask, the slope difference values on both sides of the cracks missed due to smoothing in the previous model construction are calculated along the crack direction, the calculated data is supplemented to the initial slope data of the corresponding grid, and the corrected slope data of the crack area is obtained;

[0036] For the initial slope map, the gully area range is locked according to the segmentation mask, and the initial slope value of the gully side wall in the gully area range is adjusted in combination with the cumulative rainfall of the target area within the target days to obtain the corrected slope data of the gully area; wherein the cumulative rainfall and the initial slope value of the gully side wall are positively correlated;

[0037] Integrate the corrected slope data of the crack area, the corrected slope data of the gully area and the initial slope data of the complete area in the initial slope map to obtain a corrected slope map; wherein the complete area refers to the area in the initial slope map, excluding the crack area range and the gully area range;

[0038] Based on the corrected slope map, the average slope of the complete area, and the average slope of the crack area and the gully area, the slope change data is calculated.

[0039] The preferred scheme calculates the initial slope by the Horn algorithm, corrects the slope difference of the crack area in combination with the crack direction, adjusts the slope of the gully area according to the rainfall, and finally integrates to obtain a corrected slope map; this method considers the information loss caused by smoothing in the crack area and the influence of rainfall on the gully, so that the slope data is more consistent with the actual slope condition, and the slope change data obtained by calculating the average slope in different regions can more accurately reflect the key indicators of slope stability.

[0040] As a preferred scheme, the fusion feature vector and the stability level are obtained by cross-modal attention fusion calculation and vector classification on historical multi-source data and meteorological rainfall data, specifically:

[0041] obtaining the historical multi-source data including historical slope change data and historical slope deformation trend, and the historical meteorological rainfall data;

[0042] transforming the historical slope change data, the historical slope deformation trend and the historical meteorological rainfall data into slope feature vectors, deformation feature vectors and rainfall feature vectors respectively;

[0043] calculating the cosine similarity of the slope feature vectors and the deformation feature vectors, normalizing the cosine similarity into 0-1 space attention weight, and performing pixel-level weighting on the slope feature vectors and the deformation feature vectors respectively according to the space attention weight to obtain weighted slope feature vectors and weighted deformation feature vectors;

[0044] splicing the weighted slope feature vectors, the weighted deformation feature vectors and the rainfall feature vectors, extracting the importance of each channel through global average pooling, and assigning several times weight to the channels with rainfall influence index greater than the preset rainfall, to obtain the fusion feature vector;

[0045] For several grid fusion feature vectors in the fusion feature vector, reading the specific values of slope change value, deformation acceleration and rainfall influence index, determining the grade category of the read values according to the preset stability grade determination rule, to obtain the stability grade of several grids.

[0046] The preferred scheme converts historical data into feature vectors, calculates space attention weight through cosine similarity, generates fusion feature vectors by combining rainfall influence index and assigning channel weight, and marks stability grade according to rules; this process effectively integrates multi-source data features through cross-modal attention fusion, highlights key influencing factors, and generates fusion feature vectors and stability grade, which provides high-quality training data for slope stability evaluation model and improves model learning effect.

[0047] As a preferred scheme, the historical slope change data, the historical slope deformation trend and the historical meteorological rainfall data are respectively transformed into slope feature vectors, deformation feature vectors and rainfall feature vectors, specifically:

[0048] In the historical slope change data, the area where the slope change of continuous several grids exceeds the preset degree is marked as a slip zone, and the area marked as the slip zone is processed by global average pooling to obtain the slope feature vector;

[0049] In the historical slope deformation trend, the area where the displacement acceleration exceeds the preset speed is marked as an acceleration risk zone, and the area marked as the acceleration risk zone is processed by convolution compression to obtain the deformation feature vector;

[0050] The historical meteorological rainfall data is converted into a multi-dimensional rainfall influence index table, and the rainfall influence index table is mapped to a feature space according to a preset full connection layer, so as to obtain the rainfall feature vector.

[0051] The present preferred scheme maps the rainfall data into a feature vector by marking the slip zone and the acceleration risk zone, and pooling and convolving the historical slope change data and the deformation trend, respectively. This feature conversion method accurately extracts the key regional features of the slope stability, converts the unstructured data into a structured feature vector, retains important information, reduces the data dimension, and lays a good foundation for cross-modal fusion calculation.

[0052] As a preferred scheme, the improved PSPNet network is obtained in the following manner:

[0053] An initial PSPNet network is established according to the data processing module, the double-branch cross-modal feature fusion module and the dynamic pyramid pooling module.

[0054] The model parameters of the initial PSPNet network are adjusted through back propagation according to the difference between the binary segmentation mask and the corresponding real mask, so as to obtain the improved PSPNet network.

[0055] The data processing module is used to project the point cloud after precise registration to the pixel plane of the image after precise registration based on the precise registration data, add geometric features including depth value and normal vector to each pixel, and generate a depth-assisted multi-spectral image.

[0056] The improved PSPNet network constructed in the present preferred scheme generates a depth-assisted multi-spectral image by adding geometric features to the pixels, and optimizes the parameters based on the difference between the segmentation masks, which can effectively fuse the spectral and geometric features, improve the segmentation accuracy of the slope crack and gully region, provide accurate binary segmentation mask for subsequent regional modeling, and enhance the recognition ability of the model for complex slope features.

[0057] As a preferred scheme, the double-branch cross-modal feature fusion module is used to convert the depth-assisted multi-spectral image into a cross-modal fusion feature map, and the dynamic pyramid pooling module is used to generate the binary segmentation mask according to the cross-modal fusion feature map, specifically as follows:

[0058] The double-branch cross-modal feature fusion module is used to extract the texture features of the depth-assisted multi-spectral image to generate a spectral feature map, extract the geometric features of the point cloud after precise registration, and convert them into a geometric feature map through data projection; the spectral feature map and the geometric feature map are sequentially subjected to spatial attention weighting calculation and channel attention weight distribution, and the obtained results are spliced to obtain the cross-modal fusion feature map.

[0059] The dynamic pyramid pooling module is used for four-scale pooling processing on the cross-modal fusion feature map, and dynamically allocates pooling weights according to the crack density to obtain a multi-scale feature map; the multi-scale feature map is subjected to convolution compression and up-sampling processing to obtain a super-resolution feature map; the super-resolution feature map is subjected to binary classification according to pixel values to obtain the binary segmentation mask.

[0060] In the preferred scheme, the double-branch cross-modal feature fusion module fuses spectral and geometric feature maps through an attention mechanism, and the dynamic pyramid pooling module allocates weights according to crack density and generates a binary segmentation mask; this design not only makes full use of the advantages of multi-modal data, but also highlights key areas such as cracks through dynamic weight allocation, improves the pertinence and accuracy of segmentation, and generates a binary segmentation mask that can accurately divide different regions of the slope, providing a reliable basis for subsequent differential processing.

[0061] The application also provides a loess slope stability intelligent monitoring system, comprising a data module, a slope module, a deformation module and a warning module.

[0062] The data module is configured to acquire multi-source data and periodic displacement time series data of a target region.

[0063] The slope module is configured to perform regional differential modeling on the multi-source data to output slope change data, specifically: correcting and aligning multi-spectral images and original point clouds in the multi-source data to obtain precise registration data; inputting the precise registration data into an improved PSPNet network to segment slope crack and gully development regions to obtain a binary segmentation mask; dividing the precise registration data into different regional data according to the binary segmentation mask and then splicing according to spatial coordinates to generate a DEM model covering the entire slope; and outputting the slope change data according to the DEM model.

[0064] The deformation module is configured to predict a slope deformation trend based on an LSTM algorithm according to the periodic displacement time series data.

[0065] The warning module is configured to input the slope change data, the slope deformation trend and meteorological rainfall data into a slope stability evaluation model to predict a safety factor, and if the safety factor is lower than a preset threshold, an early warning information is issued; the slope stability evaluation model is established based on a random forest algorithm by learning a mapping rule between a fusion feature vector and a stability level; the fusion feature vector and the stability level are obtained by cross-modal attention fusion calculation and vector classification on historical multi-source data and meteorological rainfall data. BRIEF DESCRIPTION OF DRAWINGS

[0066] Figure 1is a flowchart of a loess slope stability intelligent monitoring method provided by an embodiment of the present application.

[0067] Figure 2 is a structural diagram of a loess slope stability intelligent monitoring system provided by an embodiment of the present application. DETAILED DESCRIPTION

[0068] The technical solutions in the embodiments of the present application will be clearly and completely described in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0069] In the description of the present application, it should be understood that the terms "first" and "second" are used only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first" and "second" can explicitly or implicitly include one or more of the features. In the description of the present application, unless otherwise specified, the meaning of "several" is two or more.

[0070] Embodiment one:

[0071] Please refer to Figure 1 The embodiment of the present application provides a loess slope stability intelligent monitoring method, which comprises S1-S4, and the specific implementation steps are as follows:

[0072] S1, obtaining multi-source data and periodical displacement time series data of a target area.

[0073] The step S1 of the embodiment of the present application is specifically:

[0074] In a preset time window, a multi-spectral camera and a three-dimensional laser scanner are used for cooperative operation to obtain multi-source data of the target area. The multi-spectral camera is carried by a UAV and covers red, green, blue, near-infrared and red edge bands, synchronously records GPS (Global Positioning System) and IMU (Inertial Measurement Unit) data, and outputs TIFF format 5-band original multi-spectral images; the three-dimensional laser scanner scans the slope with a point spacing less than or equal to 2mm, outputs LAS format original point cloud containing XYZ coordinates and reflectivity, and synchronously arranges at least 5 spherical reflective targets and records their three-dimensional coordinates, thereby forming multi-source data covering the target area.

[0075] And, through the time series InSAR (Interferometric Synthetic Aperture Radar) technology, the target area is monitored to carry out the slope surface monitoring, and the time series displacement data is obtained, which contains the core information such as the three-dimensional coordinates of the monitoring point, the observation time, and the surface displacement amount; and the weather station is synchronously acquired to obtain the weather rainfall data which is completely matched with the displacement monitoring period, and the key parameters such as the daily rainfall, the cumulative rainfall in the past 15 days, and the daily maximum rainfall are specifically covered.

[0076] Among them, the target area can be a typical loess slope scene such as a highway slope and a pure slope.

[0077] S2, the multi-source data is differentially modeled in different regions, and the slope change data is output, specifically: the multi-spectral images and the original point cloud in the multi-source data are corrected and aligned to obtain the fine registration data; the fine registration data is input into the improved PSPNet network to segment the slope crack and the gully development area, and the binary segmentation mask is obtained; the fine registration data is divided into different regional data according to the binary segmentation mask, and then spliced according to the spatial coordinates to generate a DEM model covering the entire slope; and the slope change data is output according to the DEM model.

[0078] The step S2 of the embodiment of the application includes S2.1-S2.5, specifically:

[0079] S2.1, according to the GPS and IMU data obtained in the multi-source data acquisition stage, the original multi-spectral images in the multi-source data are subjected to double correction processing: one is orthographic correction, which eliminates the geometric distortion of the image caused by the terrain undulation through spatial coordinate matching, and ensures that the image and the actual space position are accurately corresponding; the second is radiation correction, which mainly removes external interference factors such as atmospheric scattering and uneven illumination, and restores the true spectral information of the image to obtain the corrected multi-spectral image. The image completely retains the red, green, blue, near-infrared, and red edge five core spectral bands, among which the green band can support the subsequent dark edge extraction of the slope crack and gully area.

[0080] For the original point cloud in the multi-source data, the statistical filtering is used to remove the outliers with a distance mean greater than 3 times the standard deviation, and the effective terrain points are retained to obtain the filtered point cloud carrying XYZ coordinates;

[0081] Taking the target coordinates as the reference, the corrected multi-spectral images and the filtered point cloud are aligned according to the ICP algorithm (Iterative Closest Point Algorithm), and the coarse registration parameters in the form of rotation matrix and translation vector are obtained.

[0082] S2.2, based on the corrected multi-spectral image and the filtered point cloud, establish the spatial correlation between the image pixel coordinates and the point cloud three-dimensional coordinates according to the coarse registration parameters;

[0083] Perform contrast enhancement processing on the green band of the corrected multi-spectral image to highlight the difference between the dark tone of the cracks and gullies and the background, and extract the continuous dark edges of the target to generate an image dark edge mask, which marks the image area that may be cracks and gullies;

[0084] Based on the pixel coordinates of the image dark edge mask, obtain the corresponding three-dimensional coordinate region in the point cloud through spatial correlation, and calculate the Z coordinate gradient of the three-dimensional coordinate region. Screen out the point cloud region with Z coordinate gradient greater than 0.5mm / pixel, and convert it into a point cloud gradient anomaly mask;

[0085] Perform spatial intersection operation on the image dark edge mask and the point cloud gradient anomaly mask, retain the overlapping region, and eliminate the misjudgment region such as shadow which only has dark edge in image but no gradient anomaly in point cloud. Finally, a preliminary mask about cracks and gullies is obtained.

[0086] Extract the crack light-dark boundary line and gully terrain mutation line from the preliminary mask, and correspondingly obtain the image edge point and point cloud edge point;

[0087] Constrain the image edge point and point cloud edge point, optimize the coarse registration parameters according to the ICP algorithm, so that the registration error of the corrected multi-spectral image and the filtered point cloud is less than ±1mm, and obtain the fine registration data including the fine registration image and the fine registration point cloud.

[0088] In this embodiment S2.2, the spatial correlation between the image and the point cloud is established based on the coarse registration parameters. The image dark edge is extracted by green band enhancement, and the abnormal region is screened by combining the point cloud Z coordinate gradient. Finally, the preliminary mask is generated by spatial intersection operation. It can make full use of the spectral characteristics of the image and the three-dimensional geometric information of the point cloud, accurately lock the crack and gully region, reduce the interference of irrelevant regions, and provide reliable constraints for subsequent fine registration parameter optimization, thereby improving the identification accuracy of the key feature region of the slope.

[0089] In this embodiment S2.1~S2.2, for multi-spectral image and three-dimensional laser point cloud, first, the data quality is improved by orthorectification, radiometric correction and filtering processing, then coarse registration is realized by combining target coordinates and ICP algorithm, and finally the parameters are optimized by constraining the edge points to ensure that the registration error is less than the preset threshold; This process not only removes noise interference, but also improves the spatial consistency of multi-source data through double registration strategy, providing high-precision basic data for subsequent slope feature extraction and modeling, which can effectively avoid analysis errors caused by data misplacement.

[0090] S2.3, establishing an initial PSPNet network according to the data processing module, the double-branch cross-modal feature fusion module and the dynamic pyramid pooling module; inputting preset fine registration data into the initial PSPNet network for calculation to obtain a binary segmentation mask for training, i.e., a prediction mask; wherein the "preset fine registration data" is calculated in the same manner as in the above step S2.2 and is historical data specially used for model training;

[0091] According to the difference between the prediction mask and the corresponding real mask, the model parameters of the initial PSPNet network are adjusted through back propagation, and when the edge positioning error is less than or equal to 0.8 mm, an improved PSPNet network is obtained.

[0092] The fine registration data calculated according to step S2.2 is input into the improved PSPNet network to segment the slope crack and gully development area, and a binary segmentation mask is obtained. The binary segmentation mask is a pixel-level binary label map consistent in size with the fine registration multi-spectral image, "0" corresponds to the intact slope area, "1" corresponds to the crack and gully area, and is generated by the improved PSPNet network processing fine registration data; it is used to guide data processing in different regions, support the generation of DEM in different regions, and also assist in slope correction, laying the foundation for accurately obtaining slope change data.

[0093] The data processing module, the double-branch cross-modal feature fusion module and the dynamic pyramid pooling module are described in detail as follows:

[0094] (1) Purpose of the data processing module:

[0095] Based on the fine registration data, the fine registration point cloud is projected onto the pixel plane of the fine registration image, and geometric features including depth values and normal vectors are added to each pixel, wherein the depth value is the Z coordinate of the point cloud, which is used to reflect the altitude; the normal vector is calculated by PCA analysis of 3x3 neighborhood point cloud, which is used to reflect the terrain slope direction. Subsequently, the "red, green, blue (RGB) + near-infrared + red edge" five spectral channels (i.e., the core spectral bands retained after the correction of the early multi-spectral image) and the depth value and normal vector added by the projection of the fine registration point cloud onto the image pixel plane are integrated to generate a depth-assisted multi-spectral image.

[0096] (2) Purpose of the double-branch cross-modal feature fusion module:

[0097] Taking ResNeXt-50 as the backbone, color or texture features such as crack dark tone and gully spectral mutation are extracted from the first 5 spectral channels of the depth-assisted multi-spectral image to generate a spectral feature map;

[0098] With PointNet++ as the backbone, the geometric features of the precisely registered point cloud, such as the crack edge curvature and the gully depth gradient, are extracted and converted into a geometric feature map through data projection.

[0099] The spatial attention weighting calculation and the channel attention weight distribution are sequentially performed on the spectral feature map and the geometric feature map, and the obtained results are spliced to obtain a cross-modal fusion feature map, specifically as follows:

[0100] With the spectral feature map and the geometric feature map as inputs, feature vectors are extracted for corresponding pixels of the two feature maps, the cosine similarity of the pixel-level feature vectors is calculated, the cosine similarity is normalized as a spatial attention weight, and the spectral feature map and the geometric feature map are respectively subjected to pixel-level weighting according to the weight to obtain a spatially weighted spectral feature map and a spatially weighted geometric feature map.

[0101] For the spatially weighted spectral feature map, the feature importance of 256 channels thereof is extracted through global average pooling, and higher channel weights are assigned to near-infrared channels (which are sensitive to dry crack response); for the spatially weighted geometric feature map, the feature importance of 128 channels thereof is extracted through global average pooling, and higher channel weights are assigned to depth gradient channels (which are sensitive to gully depth change response), so as to respectively complete channel-level reinforcement of the two types of feature maps to obtain a channel-weighted spectral feature map and a channel-weighted geometric feature map.

[0102] The channel-weighted spectral feature map and the channel-weighted geometric feature map are spliced in the channel dimension to obtain a cross-modal fusion feature map.

[0103] (3) Purpose of the dynamic pyramid pooling module:

[0104] The cross-modal fusion feature map is subjected to four-scale pooling processing, and the pooling weights are dynamically assigned according to the crack density, that is, when the crack density is greater than 0.6, the small-scale pooling weight accounts for 70%, and the fine topographic details of cracks and gullies are mainly reserved; when the crack density is less than 0.3, the large-scale pooling weight accounts for 60%, and the global topographic correlation of the slope is mainly captured, and finally a multi-scale feature map is obtained.

[0105] Subsequently, the multi-scale fusion feature map is processed: first, the channel number is compressed to 128 through 1x1 convolution to reduce the calculation amount, and then ESRGAN super-resolution technology is used to perform 4 times up-sampling to obtain a super-resolution feature map; finally, the super-resolution feature map is subjected to binary classification through a Softmax function, and a binary segmentation mask is output; wherein, the pixel value of 1 represents a crack (width ≥ 1 mm) or a gully (depth ≥ 5 mm), and the pixel value of 0 represents an intact slope.

[0106] The improved PSPNet network constructed in this embodiment S2.3 generates a depth auxiliary multi-spectral image by adding geometric features to the pixels, and optimizes the parameters based on the difference of the segmentation mask, can effectively fuse the spectral and geometric features, improve the segmentation accuracy of the slope crack and gully area, provide accurate binary segmentation mask for subsequent regional modeling, and enhance the recognition ability of the model to complex slope features;

[0107] Moreover, the dual-branch cross-modal feature fusion module fuses the spectral and geometric feature maps through an attention mechanism, and the dynamic pyramid pooling module allocates weights according to the crack density and generates a binary segmentation mask; this design not only makes full use of the advantages of multi-modal data, but also highlights key areas such as cracks through dynamic weight allocation, improves the relevance and accuracy of segmentation, and generates a binary segmentation mask that can accurately divide different areas of the slope, providing a reliable basis for subsequent differential processing.

[0108] S2.4, according to the binary segmentation mask, the fine registration data is divided into crack area point cloud, gully area point cloud and complete area point cloud;

[0109] The crack area point cloud is subjected to encryption resampling processing to retain the subtle depth changes of the cracks; the gully area point cloud is subjected to smoothing filtering and the target side wall steep edge is retained; and the complete area point cloud is subjected to downsampling processing to retain 1 / 5 of the original points; after the processing, the comprehensive semantic enhanced point cloud including the crack area encryption point cloud, the gully area denoising point cloud and the complete area simplified point cloud is obtained;

[0110] Based on the crack area encryption point cloud in the comprehensive semantic enhanced point cloud, the depth discontinuity of the crack area is retained through Poisson reconstruction and edge constraint to generate a crack area DEM;

[0111] Based on the gully area denoising point cloud in the comprehensive semantic enhanced point cloud, a gully area DEM is generated through inverse distance weighted interpolation and fluid mechanics correction;

[0112] Based on the complete area simplified point cloud in the comprehensive semantic enhanced point cloud, a complete area DEM is generated through Kriging interpolation;

[0113] The crack area DEM, the gully area DEM and the complete area DEM are spliced according to the spatial coordinates, the splicing seam is eliminated by using an edge transition algorithm, and a DEM model covering the entire slope is generated.

[0114] This embodiment S2.4 differentially processes the point clouds of different areas according to the binary segmentation mask, and generates DEMs in different areas and splices them; this regional modeling strategy not only ensures the fine expression of key areas such as cracks and gullies, but also improves the overall modeling efficiency through simplification processing, and the generated DEM model can fully reflect the complex topographic features of the slope and provide a high-precision basis for slope analysis.

[0115] S2.5, for each grid in the DEM model, calculate the initial slope by the Horn algorithm to obtain an initial slope map;

[0116] For the initial slope map, lock the crack area range and extract the crack direction with the segmentation mask as a guide, calculate the slope difference value of the two sides of the crack missed due to the smoothing processing in the previous model construction along the crack direction, and supplement the calculated data to the initial slope data of the corresponding grid to make the slope of the crack area truly reflect its local steepness and the terrain undulation on both sides, to obtain the corrected slope data of the crack area.

[0117] Based on the meteorological rainfall data, for the initial slope map, lock the gully area range with the segmentation mask as a guide, and adjust the initial slope value of the gully side wall in the gully area range combined with the cumulative rainfall of the target area within 15 days to obtain the corrected slope data of the gully area; wherein the cumulative rainfall and the initial slope value of the gully side wall are positively correlated with the initial slope value of the gully side wall, such as an increase of 100 mm in rainfall, an increase of 2% in the initial slope of the gully side wall, to compensate for the steepness deviation caused by rainwater erosion, so that the slope of the gully area is more in line with the actual terrain state.

[0118] Directly retain the initial slope data of the complete area as the basis for subsequent calculation; wherein the complete area refers to the area in the initial slope map, excluding the crack area range and the gully area range;

[0119] Integrate the corrected slope data of the crack area, the corrected slope data of the gully area, and the initial slope data of the complete area in the initial slope map, and eliminate invalid data such as blank areas outside the slope in the initial slope map to obtain a corrected slope map, which only retains "precise correction slope of crack area, compensated slope of gully area, and reference slope of complete area".

[0120] Based on the corrected slope map, the average slope S0 of the complete area, and the average slope S1 of the crack area and the gully area, the slope change data ΔS is calculated, i.e. ΔS=S1-S0.

[0121] The embodiment S2.5 calculates the initial slope by the Horn algorithm, corrects the slope difference of the crack area combined with the crack direction, adjusts the slope of the gully area according to the rainfall, and finally integrates to obtain the corrected slope map; this method considers the information loss caused by the smoothing processing of the crack area and the influence of rainfall on the gully, so that the slope data is more in line with the actual slope condition, and the slope change data obtained by calculating the average slope can more accurately reflect the key indicators of slope stability.

[0122] S3, according to the time series data of the displacement of each period, the slope deformation trend is predicted based on the LSTM algorithm.

[0123] The step S3 of the embodiment of the present application is specifically:

[0124] The time series data of displacement and meteorological rainfall data obtained in the multi-source data collection stage are integrated and processed; the integrated data set is standardized, and a time series window is constructed according to the rule of 'predicting the next 2 periods with the past 5 periods', and then the training set and the validation set are divided.

[0125] A 2-layer stacked LSTM structure is constructed, an Adam optimizer and an RMSE loss function are selected to train the model, and the training is iterated until the relative error of the validation set prediction is less than 7% and the model accuracy is stable, and finally the trained LSTM prediction model is obtained.

[0126] The latest 5-period time series data of displacement and meteorological rainfall data are input into the trained LSTM prediction model to predict the future 1-2 month slope displacement, and finally the slope deformation trend is output.

[0127] S4, input the slope change data, slope deformation trend and meteorological rainfall data into the slope stability evaluation model to predict the safety factor, if the safety factor is lower than the preset threshold, the warning information is issued; wherein the slope stability evaluation model is based on a random forest algorithm, and the mapping rule between the feature vector and the stability level is learned and fused; the feature vector and the stability level are obtained by cross-modal attention fusion calculation and vector classification on historical multi-source data and meteorological rainfall data.

[0128] The step S4 of the embodiment of the application includes S4.1-S4.2, which are specifically:

[0129] S4.1, based on the DEM model and the LSTM prediction model, obtain historical multi-source data including historical slope change data and historical slope deformation trend, and obtain contemporaneous historical meteorological rainfall data;

[0130] In the historical slope change data, the area where the slope change of the continuous grid is more than 1° is marked as a sliding crack zone, and the area marked as the sliding crack zone is subjected to global average pooling processing to compress the dimension to obtain a slope feature vector;

[0131] In the historical slope deformation trend, the area where the displacement acceleration is more than 0.5mm / month 2 is marked as an acceleration risk zone, and the area marked as the acceleration risk zone is subjected to convolution compression processing to obtain a deformation feature vector;

[0132] The historical meteorological rainfall data is converted into a rainfall influence index table including multiple dimensions such as a cumulative rainfall index and a heavy rainfall index, and the rainfall influence index table is mapped to a feature space according to a preset two-layer full connection layer; subsequently, a correction term is added in combination with the slope type, the initial feature vector of each dimension is multiplied by 1.2 for a soil slope with higher rainfall sensitivity, and the initial feature vector of each dimension is multiplied by 0.8 for a rock slope with lower rainfall sensitivity, so as to adapt to the differences in rainfall sensitivity of different slopes, and finally a rainfall feature vector is obtained. The first layer of the two-layer full connection layer is set to 64 hidden units, and the second layer is set to 32 hidden units, and both layers use a ReLU activation function.

[0133] The embodiment S4.1 maps the rainfall data to a feature vector by marking the slip zone and the acceleration risk zone and respectively performing pooling and convolution processing on the historical slope change data and the deformation trend; this feature conversion mode accurately extracts the key regional features of slope stability, converts the unstructured data into a structured feature vector, retains important information and reduces the data dimension, and lays a good foundation for cross-modal fusion calculation.

[0134] S4.2, calculate the cosine similarity of the slope feature vector and the deformation feature vector, normalize the cosine similarity to a 0-1 space attention weight, and respectively perform pixel-level weighting on the slope feature vector and the deformation feature vector according to the space attention weight, to obtain a weighted slope feature vector and a weighted deformation feature vector; wherein, for the space attention weight, the weight 1 corresponds to “the slope change is greater than 2°, and the acceleration is greater than 1 mm / month 2 ”;

[0135] The weighted slope feature vector, the weighted deformation feature vector and the rainfall feature vector are spliced, the importance of each channel is extracted through global average pooling, and the channels with a rainfall influence index greater than a preset rainfall amount are assigned a weight of 2, to calculate a fusion feature vector;

[0136] For some grid fusion feature vectors in the fusion feature vector, the specific values of the slope change value, the deformation acceleration and the rainfall influence index are read, the grade category of the read values is determined according to a preset stability grade determination rule, and the stability grade of the grid is obtained.

[0137] The “fusion feature vector and stability grade” is input into a random forest model as a data set, wherein the fusion feature vector is the core input of the model, each vector corresponds to the comprehensive risk features of a single slope grid, and covers key information such as slope mutation, deformation acceleration and heavy rainfall coupling; relying on the multiple decision trees of the random forest algorithm, the mapping rule between the fusion feature vector and the stability grade is learned, for example, the corresponding relationship between the feature combination of “high slope mutation + high deformation acceleration + heavy rainfall” and the “instability risk” label is automatically identified, and finally a trained slope stability evaluation model is obtained.

[0138] wherein the "stability level determination rule" is divided into three categories of "stable, sub-stable, and unstable risk" according to the slope change value (A, unit °), deformation acceleration (B, unit mm / month 2 ), and rainfall influence index (C, unitless) as the core determination index, and the specific rules are as follows:

[0139] (1) Stable: simultaneously meet "A≤0.5, B≤0.3, C≤20", or only a single index slightly exceeds the threshold value (such as A=0.8, B and C meet the threshold);

[0140] (2) Sub-stable: simultaneously meet "0.5<A≤2, 0.3<B≤1", or "20<C≤50, and either A / B slightly exceeds the threshold";

[0141] (3) Unstable risk: simultaneously meet "A>2, B>1", or "C>50, A>1, or B>0.8".

[0142] The embodiments S4.1-S4.2 convert historical data into feature vectors, calculate spatial attention weights through cosine similarity, combine channel weights according to the rainfall influence index to generate a fusion feature vector, and label the stability level according to the rules. This process effectively integrates multi-source data features through cross-modal attention fusion, highlights key influencing factors, and generates a fusion feature vector and a stability level, which provides high-quality training data for the slope stability evaluation model and improves the model learning effect.

[0143] S4.3, the slope change data calculated by the high-precision DEM model, the slope deformation trend output by the LSTM prediction model, and the meteorological rainfall data monitored by the meteorological station at the same period are jointly input into the trained slope stability evaluation model; and it is necessary to clarify that the three types of data are all for the same time period data in the target area for 1-3 consecutive monitoring periods;

[0144] The model first strengthens the coupling risk features of "slope mutation, deformation acceleration, and heavy rainfall" through the cross-modal attention mechanism, and then relies on the multi-tree decision tree operation of the random forest algorithm to output the safety factor of each monitoring grid of the slope and the corresponding stability level. If the safety factor of any monitoring grid is lower than the preset loess slope engineering safety threshold, the early warning process is automatically triggered, and early warning information including the three-dimensional coordinates of the risk area, the risk level, and the core risk factors is published in real time, providing timely decision basis for slope safety control;

[0145] wherein the stability level is divided into "stable, sub-stable, and unstable risk", which can support graded control, positioning of risk factors, and serve as an intuitive carrier of "risk level" in the early warning information.

[0146] Overall, the embodiment has the following beneficial effects:

[0147] The application uses improved PSPNet to segment the crack and gully area, can accurately identify the key instability area of the slope, and then generates a DEM model through regional differential modeling, so that the slope change data can reflect the characteristics of different regions, instead of general overall data, which can be accurately corresponded with the deformation trend and the influence of rainfall. Through the generation of DEM model covering the whole area, the output of slope change data can avoid the deviation of model establishment caused by slope data problems. The LSTM algorithm predicts the deformation trend of the slope based on the time series data of the displacement, which can accurately quantify the time evolution characteristics of the deformation and provide reliable data for the deformation factors. Through the random forest algorithm, the mapping rule between the feature vector and the stability level is learned, and the established slope stability evaluation model can quantify the non-linear correlation of slope, deformation and meteorological rainfall. The safety factor is predicted by inputting the multi-factor data into the model. Due to the comprehensive consideration of the interaction of different factors, the prediction result is more in line with the real state, which can solve the problem of multi-factor fusion, greatly reduce the cost of artificial inspection, and improve the reliability of early warning;

[0148] In summary, the application aims to solve the core demand of high cost of artificial inspection and lag of early warning response in loess slope stability monitoring. The unmanned aerial vehicle multispectral camera and the three-dimensional laser scanner are used for collaborative operation to collect multi-source data. Through multi-period data comparison, the key deformation characteristics such as slope crack and gully can be accurately identified and analyzed, and the automatic collection and intelligent analysis of slope monitoring data can be realized, which effectively reduces the labor and time cost of artificial inspection, and finally significantly improves the accuracy of early warning in practical monitoring scenes such as highway slope.

[0149] Embodiment two:

[0150] Please refer to Figure 2 The embodiment of the application provides a loess slope stability intelligent monitoring system, which comprises a data module 10, a slope module 20, a deformation module 30 and an early warning module 40.

[0151] The data module 10 is used for acquiring multi-source data and time series data of displacement of a target area.

[0152] The slope module 20 is used for regional differential modeling of the multi-source data and outputs slope change data, specifically: correcting and aligning the multispectral image and the original point cloud in the multi-source data to obtain accurate registration data; inputting the accurate registration data into the improved PSPNet network to segment the crack and gully development area of the slope, and obtaining a binary segmentation mask; dividing the accurate registration data into different regional data according to the binary segmentation mask, and then splicing according to the spatial coordinates to generate a DEM model covering the entire slope; and outputting the slope change data according to the DEM model.

[0153] a deformation module 30 configured to predict a slope deformation trend based on an LSTM algorithm according to the time-series displacement data;

[0154] an early warning module 40 configured to input the slope change data, the slope deformation trend and the meteorological rainfall data into a slope stability evaluation model to predict a safety factor, and if the safety factor is lower than a preset threshold, issue an early warning information; wherein the slope stability evaluation model is established based on a random forest algorithm by learning a mapping rule between a fusion feature vector and a stability level; and the fusion feature vector and the stability level are obtained by performing cross-modal attention fusion calculation and vector classification on historical multi-source data and meteorological rainfall data.

[0155] In one embodiment, the data module 10 specifically comprises:

[0156] Within a preset time window, the multi-source data of the target area is obtained by using a multi-spectral camera and a three-dimensional laser scanner of a UAV to work cooperatively. The UAV is equipped with a five-lens multi-spectral camera covering red, green, blue, near-infrared and red edge bands, and simultaneously records GPS (Global Positioning System) and IMU (Inertial Measurement Unit) data, and outputs TIFF format 5-band original multi-spectral images. The three-dimensional laser scanner scans the slope with a point spacing less than or equal to 2 mm, and outputs LAS format original point cloud containing XYZ coordinates and reflectivity, and at least five spherical reflective targets are arranged synchronously and their three-dimensional coordinates are recorded, thereby forming multi-source data covering the target area.

[0157] Further, time-series InSAR (Interferometric Synthetic Aperture Radar) technology is used to carry out slope surface monitoring on the target area to obtain time-series displacement data, which contains core information such as three-dimensional coordinates of monitoring points, observation time and surface displacement amount; meteorological rainfall data completely matched with the displacement monitoring period is also obtained synchronously, which specifically includes key parameters such as daily rainfall, cumulative rainfall in the past 15 days and daily maximum rainfall.

[0158] The target area can be a typical loess slope scene such as a highway slope or a pure slope.

[0159] In one embodiment, the slope module 20 comprises a coarse registration unit, a fine registration unit, a mask unit, a model unit and a slope unit.

[0160] The coarse registration unit is configured to perform double correction processing on the original multi-spectral image in the multi-source data according to the GPS and IMU data obtained in the multi-source data acquisition stage. The first correction is orthographic correction, which eliminates the geometric distortion of the image caused by the terrain through spatial coordinate matching, and ensures that the image and the actual spatial position correspond accurately. The second correction is radiation correction, which focuses on removing external interference factors such as atmospheric scattering and uneven illumination, and restores the true spectral information of the image to obtain the corrected multi-spectral image. The image retains the red, green, blue, near-infrared, and red edge five core spectral bands, among which the green band can support the subsequent dark edge extraction of the slope crack and gully area.

[0161] The coarse registration unit is further configured to remove outliers with a distance mean greater than 3 times the standard deviation from the original point cloud in the multi-source data through statistical filtering, and retain effective terrain points to obtain a filtered point cloud carrying XYZ coordinates.

[0162] The coarse registration unit is further configured to align the corrected multi-spectral image and the filtered point cloud according to the ICP algorithm (Iterative Closest Point Algorithm) based on the target coordinates to obtain coarse registration parameters in the form of a rotation matrix and a translation vector.

[0163] The fine registration unit is configured to establish a spatial correlation between the image pixel coordinates and the point cloud three-dimensional coordinates based on the corrected multi-spectral image and the filtered point cloud according to the coarse registration parameters.

[0164] The fine registration unit is further configured to perform contrast enhancement processing on the green band in the corrected multi-spectral image to highlight the difference between the dark tone of the crack and gully and the background, and extract the continuous dark edge of the target to generate an image dark edge mask, which marks the image area that may be a crack and gully.

[0165] The fine registration unit is further configured to obtain the corresponding three-dimensional coordinate region in the point cloud based on the pixel coordinates of the image dark edge mask through the spatial correlation, and calculate the Z coordinate gradient of the three-dimensional coordinate region to screen out the point cloud region with a Z coordinate gradient greater than 0.5 mm / pixel, and convert it into a point cloud gradient anomaly mask.

[0166] The fine registration unit is further configured to perform spatial intersection operation on the image dark edge mask and the point cloud gradient anomaly mask, retain the overlapping region, and eliminate the misjudgment region such as shadow that only has a dark edge in the image but has no gradient anomaly in the point cloud, and finally obtain a preliminary mask about the crack and gully.

[0167] The fine registration unit is further configured to extract the crack light-dark boundary line and the gully terrain mutation line from the preliminary mask, and correspondingly obtain the image edge point and the point cloud edge point.

[0168] The fine registration unit is also configured to optimize the coarse registration parameters according to the ICP algorithm with the image edge points and the point cloud edge points as constraints, so that the registration error of the corrected multi-spectral image and the filtered point cloud is less than ±1 mm, and fine registration data including the fine-registered image and the fine-registered point cloud are obtained.

[0169] The fine registration unit of the embodiment establishes the spatial correlation of the image and the point cloud based on the coarse registration parameters, extracts the dark edges of the image through green band enhancement, filters the abnormal areas in combination with the Z coordinate gradient of the point cloud, and finally generates a preliminary mask through spatial intersection operation. The preliminary mask can fully utilize the spectral characteristics of the image and the three-dimensional geometric information of the point cloud, accurately lock the crack and gully areas, reduce the interference of irrelevant areas, provide reliable constraints for subsequent fine registration parameter optimization, and improve the identification accuracy of the key feature areas of the slope.

[0170] In the coarse registration unit and the fine registration unit of the embodiment, for the multi-spectral image and the three-dimensional laser point cloud, the data quality is first improved through orthorectification, radiometric correction and filtering processing, then the coarse registration is realized in combination with the target coordinates and the ICP algorithm, and finally the parameters are optimized with the edge points as constraints to ensure that the registration error is less than a preset threshold. This process not only removes noise interference, but also improves the spatial consistency of multi-source data through a double registration strategy, providing high-precision basic data for subsequent slope feature extraction and modeling, and effectively avoiding analysis errors caused by data misplacement.

[0171] The mask unit is configured to establish an initial PSPNet network according to the data processing module, the double-branch cross-modal feature fusion module and the dynamic pyramid pooling module; input the preset fine registration data into the initial PSPNet network for calculation to obtain a binary segmentation mask for training, i.e., a predicted mask; wherein the "preset fine registration data" is calculated in the same way as the fine registration unit and is historical data specially used for model training;

[0172] The mask unit is also configured to adjust the model parameters of the initial PSPNet network through back propagation according to the difference between the predicted mask and the corresponding real mask, and obtain an improved PSPNet network when the edge positioning error is less than or equal to 0.8 mm.

[0173] The mask unit is also configured to input the fine registration data calculated according to the fine registration unit into the improved PSPNet network to segment the slope crack and gully development area, and obtain a binary segmentation mask. The binary segmentation mask is a pixel-level binary label map consistent in size with the fine-registered multi-spectral image, "0" corresponds to the complete area of the slope, and "1" corresponds to the crack and gully area. The binary segmentation mask is generated by the improved PSPNet network processing the fine registration data. It is used to guide the data processing in different regions, support the generation of DEM in different regions, and also assist in slope correction, laying a foundation for accurately obtaining slope change data.

[0174] The data processing module, the double-branch cross-modal feature fusion module, and the dynamic pyramid pooling module are described below:

[0175] (1) Purpose of the data processing module:

[0176] Based on the fine registration data, the fine registration point cloud is projected onto the pixel plane of the fine registration image, and geometric features including depth values and normal vectors are added to each pixel. The depth value is the Z coordinate of the point cloud, which reflects the altitude. The normal vector is calculated by PCA analysis of a 3x3 neighborhood of point clouds, which reflects the terrain slope direction. Subsequently, the "red, green, blue (RGB) + near-infrared + red edge" five spectral channels (i.e., the core spectral bands retained after the pre-multiple spectral image correction) and the depth values and normal vectors added by the fine registration point cloud projected onto the image pixel plane are integrated to generate a depth-assisted multi-spectral image.

[0177] (2) Purpose of the double-branch cross-modal feature fusion module:

[0178] Taking ResNeXt-50 as the backbone, color or texture features such as crack dark tone and gully spectral mutation are extracted from the first five spectral channels of the depth-assisted multi-spectral image to generate a spectral feature map.

[0179] Taking PointNet++ as the backbone, geometric features such as crack edge curvature and gully depth gradient are extracted from the fine registration point cloud and converted into a geometric feature map through data projection.

[0180] The spectral feature map and the geometric feature map are sequentially subjected to spatial attention weighting calculation and channel attention weight distribution, and the results are spliced to obtain a cross-modal fusion feature map, which is:

[0181] Taking the spectral feature map and the geometric feature map as input, feature vectors are extracted for the corresponding pixels of the two feature maps. The cosine similarity of the pixel-level feature vectors is calculated, and the cosine similarity is normalized to spatial attention weights. According to the weights, pixel-level weighting is performed on the spectral feature map and the geometric feature map respectively to obtain the spatially weighted spectral feature map and the spatially weighted geometric feature map.

[0182] For the spatially weighted spectral feature map, the feature importance of its 256 channels is extracted through global average pooling, and higher channel weights are assigned to the near-infrared channel (which is sensitive to dry cracks). For the spatially weighted geometric feature map, the feature importance of its 128 channels is also extracted through global average pooling, and higher channel weights are assigned to the depth gradient channel (which is sensitive to gully depth changes). In this way, channel-level reinforcement is completed for the two types of feature maps respectively to obtain the channel-weighted spectral feature map and the channel-weighted geometric feature map.

[0183] The channel-weighted spectral feature map and the channel-weighted geometric feature map are spliced in the channel dimension to obtain a cross-modal fusion feature map.

[0184] (3) Purpose of the dynamic pyramid pooling module:

[0185] The cross-modal fusion feature map is subjected to four-scale pooling processing, and the pooling weights are dynamically allocated according to the crack density, that is, when the crack density is greater than 0.6, the small-scale pooling weight accounts for 70%, and the fine topographic details of cracks and gullies are mainly retained; when the crack density is less than 0.3, the large-scale pooling weight accounts for 60%, and the global topographic correlation of the slope is mainly captured, and finally a multi-scale feature map is obtained.

[0186] Subsequently, the multi-scale fusion feature map is processed: first, the channel number is compressed to 128 through 1x1 convolution to reduce the calculation amount, and then ESRGAN super-resolution technology is used to perform 4 times up-sampling to obtain a super-resolution feature map; finally, the super-resolution feature map is subjected to binary classification through a Softmax function, and a binary segmentation mask is output; wherein, the pixel value of 1 represents a crack (width ≥ 1mm) or a gully (depth ≥ 5mm), and the pixel value of 0 represents a complete slope body.

[0187] The improved PSPNet network constructed by the mask unit in this embodiment generates a depth auxiliary multi-spectral image by adding geometric features to the pixels, and optimizes the parameters based on the difference in the segmentation mask, which can effectively fuse the spectral and geometric features, improve the segmentation accuracy of the crack and gully regions of the slope, provide accurate binary segmentation mask for subsequent regional modeling, and enhance the recognition ability of the model for complex slope features;

[0188] Moreover, the dual-branch cross-modal feature fusion module fuses the spectral and geometric feature maps through an attention mechanism, and the dynamic pyramid pooling module allocates weights according to the crack density and generates a binary segmentation mask; this design not only fully utilizes the advantages of multi-modal data, but also highlights the key areas such as cracks through dynamic weight allocation, improves the pertinence and accuracy of segmentation, and generates a binary segmentation mask that can accurately divide different regions of the slope, providing a reliable basis for subsequent differential processing.

[0189] The model unit is configured to divide the fine registration data into crack region point cloud, gully region point cloud and complete region point cloud according to the binary segmentation mask.

[0190] The model unit is further configured to perform encryption resampling processing on the crack region point cloud to retain the fine depth changes of the cracks, perform smoothing filtering on the gully region point cloud and retain the steep edges of the target side wall, and perform down-sampling processing on the complete region point cloud to retain 1 / 5 of the original points, so as to obtain comprehensive semantic enhanced point cloud including crack region encrypted point cloud, gully region denoised point cloud and complete region simplified point cloud after the processing.

[0191] The model unit is further configured to enhance the point cloud in the crack region based on the comprehensive semantics, encrypt the point cloud, retain the depth discontinuity of the crack region through Poisson reconstruction and edge constraint, and generate a crack region DEM;

[0192] The model unit is further configured to enhance the point cloud in the gully region based on the comprehensive semantics, denoise the point cloud, generate a gully region DEM through inverse distance weighted interpolation and fluid mechanics correction;

[0193] The model unit is further configured to simplify the point cloud in the complete region based on the comprehensive semantics, and generate a complete region DEM through Kriging interpolation;

[0194] The model unit is further configured to splice the crack region DEM, the gully region DEM and the complete region DEM according to spatial coordinates, eliminate the splicing joint through an edge transition algorithm, and generate a DEM model covering the entire slope.

[0195] The model unit of the embodiment differentiates the point clouds in different regions according to the binary segmentation mask, generates DEMs in different regions, and splices them; this regional modeling strategy not only ensures the fine expression of key regions such as cracks and gullies, but also improves the overall modeling efficiency through simplification processing, and the generated DEM model can fully reflect the complex topographic features of the slope and provide a high-precision basis for slope analysis.

[0196] The slope unit is configured to calculate an initial slope for each grid in the DEM model through a Horn algorithm to obtain an initial slope map;

[0197] The slope unit is further configured to lock the crack region range and extract the crack direction of the initial slope map guided by the segmentation mask, calculate the slope difference values of the two sides of the crack that are missed due to the smoothing processing in the previous model construction along the crack direction, supplement the calculated data to the initial slope data of the corresponding grid, make the slope of the crack region truly reflect the local steepness and the terrain undulation on both sides, and obtain the corrected slope data of the crack region.

[0198] The slope unit is further configured to lock the gully region range of the initial slope map guided by the segmentation mask based on meteorological rainfall data, adjust the initial slope value of the gully side wall in the gully region range in combination with the cumulative rainfall of the target region within 15 days, obtain the corrected slope data of the gully region, and wherein the cumulative rainfall and the initial slope value of the gully side wall are positively correlated with the up-regulation amount, for example, if the rainfall increases by 100 mm, the initial slope of the gully side wall is up-regulated by 2%, so as to compensate for the steepness deviation caused by rainwater erosion, and make the slope of the gully region more consistent with the actual terrain state;

[0199] The slope unit is further configured to directly retain the initial slope data of the complete region as a reference for subsequent calculation; wherein the complete region refers to the region in the initial slope map other than the crack region range and the gully region range.

[0200] The slope unit is further configured to integrate the corrected slope data of the crack area, the corrected slope data of the gully area, and the initial slope data of the complete area in the initial slope map, and remove invalid data such as blank areas outside the slope in the initial slope map, to obtain a corrected slope map, in which only the “precise corrected slope of the crack area, the compensated slope of the gully area, and the reference slope of the complete area” are reserved.

[0201] The slope unit is further configured to calculate slope change data ΔS based on the corrected slope map, according to the average slope S0 of the complete area, and the average slope S1 of the crack area and the gully area, that is, ΔS = S1-S0.

[0202] The slope unit of the embodiment calculates the initial slope by the Horn algorithm, corrects the slope difference in the crack area in combination with the crack direction, adjusts the slope in the gully area according to the rainfall, and finally integrates to obtain the corrected slope map; this method considers the information loss caused by the smoothing processing of the crack area and the influence of rainfall on the gully, so that the slope data is more in line with the actual slope condition, and the slope change data obtained by calculating the average slope in different areas can more accurately reflect the key indicators of slope stability.

[0203] In one embodiment, the deformation module 30 specifically comprises:

[0204] The integrated data set is subjected to standardization processing, and a time sequence window is constructed according to the rule of “predicting the next 2 periods with the past 5 periods of data”, and then the training set and the validation set are divided.

[0205] A 2-layer stacked LSTM structure is constructed, the Adam optimizer and the RMSE loss function are selected to train the model, and the training is iterated until the relative error of the validation set prediction is less than 7% and the model accuracy is stable, and finally the trained LSTM prediction model is obtained.

[0206] The latest 5-period time sequence data of the displacement and the meteorological rainfall data are input into the trained LSTM prediction model to predict the future 1-2 month slope displacement, and finally the slope deformation trend is output.

[0207] In one embodiment, the warning module 40 comprises a vector unit, a calculation unit, and a warning unit.

[0208] The vector unit is configured to obtain historical multi-source data including historical slope change data and historical slope deformation trend based on the DEM model and the LSTM prediction model, and obtain contemporaneous historical meteorological rainfall data.

[0209] The vector unit is further configured to mark, in the historical slope gradient change data, an area where a plurality of consecutive grid slope gradients exceed 1° as a sliding fracture zone, and perform global average pooling processing on the area marked as the sliding fracture zone to compress dimensions, so as to obtain a slope gradient feature vector;

[0210] The vector unit is further configured to mark, in the historical slope deformation trend, an area where displacement acceleration exceeds 0.5 mm / month 2 as an acceleration risk zone, and perform convolution compression processing on the area marked as the acceleration risk zone, so as to obtain a deformation feature vector;

[0211] The vector unit is further configured to convert historical meteorological rainfall data into a rainfall influence index table including a plurality of dimensions such as a cumulative rainfall index and a heavy rainfall index, and map the rainfall influence index table to a feature space according to a preset two-layer fully connected layer. Then, a correction term is added in combination with a slope type. The preliminary feature vector of each dimension of a soil slope is multiplied by 1.2, considering that the soil slope has a higher rainfall sensitivity. The preliminary feature vector of each dimension of a rock slope is multiplied by 0.8, considering that the rock slope has a lower rainfall sensitivity. In this way, the rainfall sensitivity difference of different slopes is adapted, and finally a rainfall feature vector is obtained. The first layer of the two-layer fully connected layer is set to 64 hidden units, and the second layer is set to 32 hidden units. Both layers use a ReLU activation function.

[0212] The vector unit of this embodiment marks the sliding fracture zone and the acceleration risk zone, respectively performs pooling and convolution processing on the historical slope gradient change data and the deformation trend, and maps the rainfall data to a feature vector. This feature conversion method accurately extracts the key area features of slope stability, converts unstructured data into structured feature vectors, retains important information, reduces the data dimension, and lays a good foundation for cross-modal fusion calculation.

[0213] The calculation unit is configured to calculate a cosine similarity of the slope gradient feature vector and the deformation feature vector, normalize the cosine similarity into a 0-1 space attention weight, and perform pixel-level weighting on the slope gradient feature vector and the deformation feature vector according to the space attention weight, so as to obtain a weighted slope gradient feature vector and a weighted deformation feature vector. For the space attention weight, a weight of 1 corresponds to “a slope gradient change greater than 2° and an acceleration greater than 1 mm / month 2 ”;

[0214] The calculation unit is further configured to splice the weighted slope gradient feature vector, the weighted deformation feature vector, and the rainfall feature vector, extract the importance of each channel through global average pooling, and assign a double weight to a channel with a rainfall influence index greater than a preset rainfall, so as to obtain a fusion feature vector;

[0215] The computing unit is further configured to read specific values of the slope change value, the deformation acceleration and the rainfall influence index for the fusion feature vector of each grid, determine the grade category of the read values according to a preset stability grade determination rule, and obtain the stability grade of each grid.

[0216] The computing unit is further configured to input the fusion feature vector and the stability grade as a data set into a random forest model, wherein the fusion feature vector is taken as a core input of the model, each vector corresponds to the comprehensive risk feature of a single slope grid, and covers key information such as slope mutation, deformation acceleration and strong rainfall coupling; relying on the multiple decision trees of the random forest algorithm, the mapping rule between the fusion feature vector and the stability grade is learned, for example, the corresponding relationship between the feature combination of “high slope mutation + high deformation acceleration + strong rainfall” and the “instability risk” label is automatically identified, and finally a trained slope stability evaluation model is obtained.

[0217] The “stability grade determination rule” is divided into three categories of “stable, sub-stable and instability risk” according to the slope change value (A, unit °), the deformation acceleration (B, unit mm / month 2 ), and the rainfall influence index (C, unitless) as the core determination indexes, and the specific rules are as follows:

[0218] (1) stable: simultaneously satisfying “A≤0.5, B≤0.3, C≤20”, or only a single index slightly exceeding the threshold value (such as A=0.8, B and C meet the standard);

[0219] (2) sub-stable: simultaneously satisfying “0.5<A≤2, 0.3<B≤1”, or “20<C≤50, and A / B either slightly exceeds the threshold value”;

[0220] (3) instability risk: simultaneously satisfying “A>2, B>1”, or “C>50, A>1, or B>0.8”.

[0221] The vector unit and the computing unit of the embodiment convert historical data into feature vectors, calculate spatial attention weights through cosine similarity, assign channel weights in combination with rainfall influence indexes to generate fusion feature vectors, and mark the stability grade according to the rules; this process effectively integrates multi-source data features through cross-modal attention fusion, highlights key influencing factors, and generates fusion feature vectors and stability grades, which provide high-quality training data for the slope stability evaluation model and improve the learning effect of the model.

[0222] The early warning unit is used for inputting the slope change data calculated by the high-precision DEM model, the slope deformation trend output by the LSTM prediction model and the meteorological rainfall data monitored by the meteorological station at the same period into the slope stability evaluation model which has been trained, and it is necessary to clarify that the three types of data are all the same period data in the target region in the last 1-3 continuous monitoring periods;

[0223] The early warning unit is also used for strengthening the coupling risk characteristics of the slope mutation, deformation acceleration and heavy rainfall by the cross-modal attention mechanism first, and then outputting the safety factor of each monitoring grid of the slope and the corresponding stability level by the multi-tree decision tree operation of the random forest algorithm; if the safety factor of any monitoring grid is lower than the preset loess slope engineering safety threshold, the early warning process is automatically triggered, and early warning information including the three-dimensional coordinates of the risk area, the risk level and the core risk factors is published in real time, thereby providing a timely decision basis for the slope safety control;

[0224] The stability level is divided into stable, less stable and unstable risk, which can support the hierarchical control, locate the risk factors and serve as an intuitive carrier of the risk level in the early warning information.

[0225] Overall, the embodiment has the following beneficial effects:

[0226] The improved PSPNet is used for segmenting the crack and gully area, the key unstable area of the slope can be accurately identified, the DEM model is generated by the differential modeling of different regions, the slope change data can reflect the characteristics of different regions, instead of general overall data, and can be accurately corresponded with the deformation trend and the rainfall influence; and by generating the DEM model covering the whole region and outputting the slope change data, the deviation of the model establishment caused by the slope data problem can be avoided. The LSTM algorithm is used for predicting the slope deformation trend based on the time series displacement data, the time evolution characteristics of the deformation can be accurately quantified, and reliable data with prediction for the deformation factors are provided. The random forest algorithm is used for learning the mapping rule between the feature vector and the stability level, the slope stability evaluation model can quantize the non-linear correlation of the slope, deformation and meteorological rainfall, the safety factor is predicted by inputting the multi-factor data into the model, the prediction result is more close to the real state due to the comprehensive consideration of the interaction of different factors, the multi-factor fusion problem can be solved, the artificial inspection cost is greatly reduced, and the early warning reliability is improved;

[0227] In summary, the present application aims at the core demand of high cost of artificial inspection and lagging response of early warning in loess slope stability monitoring, adopts cooperative operation of unmanned aerial vehicle multispectral camera and three-dimensional laser scanner, collects multi-source data, accurately identifies and analyzes key deformation features such as slope cracks and gullies through multi-period data comparison, can realize automatic collection and intelligent analysis of slope monitoring data, effectively reduces the labor and time cost of artificial inspection, and finally significantly improves the early warning accuracy in practical monitoring scenes such as highway slopes.

[0228] Embodiment three:

[0229] The embodiment of the present application provides a computer readable storage medium, the computer readable storage medium comprises a stored computer program, wherein the computer program controls the device where the computer readable storage medium is located to execute the loess slope stability intelligent monitoring method when the computer program runs.

[0230] The loess slope stability intelligent monitoring method can be stored in a computer readable storage medium if it is realized in the form of a software function unit and used as an independent product. Based on this understanding, all or part of the processes in the above-mentioned embodiment methods can also be completed by a computer program instructing related hardware, and the computer program can be stored in a computer readable storage medium. The computer program can implement the steps of each method embodiment when executed by a processor. The computer program includes computer program code, which can be in the form of source code, object code, executable files or some intermediate forms, etc. The computer readable medium can include any entity or device that can carry the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal and software distribution medium, etc.

[0231] The above is the preferred embodiment of the present application, and it should be noted that for ordinary skilled persons in the technical field, several improvements and refinements can be made without departing from the principles of the present application, and these improvements and refinements are also considered to be within the protection scope of the present application.

Claims

1. A method for intelligent monitoring of loess slope stability, characterized in that, include: Acquire multi-source data and periodic displacement time-series data of the target area; The multi-source data is used to perform regional differential modeling to output slope change data. Specifically, the multispectral images and original point clouds in the multi-source data are corrected and aligned to obtain finely registered data; the finely registered data is input into an improved PSPNet network to segment the slope crack and gully development areas to obtain a binary segmentation mask; the finely registered data is divided into different regional data according to the binary segmentation mask and then stitched together according to spatial coordinates to generate a DEM model covering the entire slope; the slope change data is output based on the DEM model. Based on the time-series displacement data, the slope deformation trend is predicted using the LSTM algorithm. The slope change data, slope deformation trend, and meteorological rainfall data are input into the slope stability evaluation model to predict the safety factor. If the safety factor is lower than a preset threshold, an early warning information is issued. The slope stability evaluation model is based on the random forest algorithm and is established by learning the mapping law between the fused feature vector and the stability level. The fused feature vector and the stability level are obtained by cross-modal attention fusion calculation and vector classification of historical multi-source data and meteorological rainfall data.

2. The intelligent monitoring method for loess slope stability as described in claim 1, characterized in that, The multispectral images and original point clouds in the multi-source data are corrected and aligned to obtain finely registered data, specifically as follows: Orthorectification and radiometric correction are performed on the multispectral images in the multi-source data to obtain corrected multispectral images; For the original point cloud in the multi-source data, outliers with a distance mean greater than a preset multiple of the standard deviation are removed by statistical filtering to obtain a filtered point cloud; Using the target coordinates as a reference, the corrected multispectral image and the filtered point cloud are aligned according to the ICP algorithm to obtain coarse registration parameters including rotation matrix and translation vector. A preliminary mask is generated based on the corrected multispectral image and the filtered point cloud. Crack light-dark boundary lines and gully terrain abrupt change lines are extracted from the preliminary mask, and corresponding image edge points and point cloud edge points are obtained. Using the image edge points and the point cloud edge points as constraints, the coarse registration parameters are optimized according to the ICP algorithm to make the registration error between the corrected multispectral image and the filtered point cloud less than a preset threshold, thus obtaining the fine registration data.

3. The intelligent monitoring method for loess slope stability as described in claim 2, characterized in that, A preliminary mask is generated based on the corrected multispectral image and the filtered point cloud, specifically as follows: Based on the corrected multispectral image and the filtered point cloud, a spatial relationship between the image pixel coordinates and the point cloud three-dimensional coordinates is established according to the coarse registration parameters. Contrast enhancement processing is performed on the green band in the corrected multispectral image, and continuous dark edges of the target are extracted to generate an image dark edge mask; Based on the pixel coordinates of the dark edge mask of the image, the corresponding three-dimensional coordinate region in the point cloud is obtained through the spatial association, and the Z coordinate gradient of the three-dimensional coordinate region is calculated. Point cloud regions with Z coordinate gradients greater than a preset value are filtered out and transformed into point cloud gradient anomaly masks. Perform a spatial intersection operation on the image dark edge mask and the point cloud gradient anomaly mask, retain the overlapping areas, and obtain the preliminary mask for cracks and gullies.

4. The intelligent monitoring method for loess slope stability as described in claim 1, characterized in that, After dividing the finely registered data into different regions using the binary segmentation mask, the data is stitched together according to spatial coordinates to generate a DEM model covering the entire slope, specifically: Based on the binary segmentation mask, the fine registration data is divided into point clouds of crack areas, point clouds of gully areas, and point clouds of intact areas; The point cloud in the crack area is densified and resampled, the point cloud in the gully area is smoothed and filtered while retaining the steep edges of the target sidewalls, and the point cloud in the intact area is downsampled to obtain the processed comprehensive semantically enhanced point cloud. Based on the crack region densified point cloud in the comprehensive semantically enhanced point cloud, the depth discontinuity of the crack region is preserved through Poisson reconstruction and edge constraints to generate a crack region DEM; Based on the denoised point cloud of the gully area in the comprehensive semantically enhanced point cloud, a DEM of the gully area is generated through inverse distance weighted interpolation and hydrodynamic correction. Based on the complete region simplified point cloud in the comprehensive semantically enhanced point cloud, a complete region DEM is generated by kriging interpolation; The DEMs of the cracked area, the gully area, and the complete area are stitched together according to spatial coordinates to generate the DEM model covering the entire slope.

5. The intelligent monitoring method for loess slope stability as described in claim 1, characterized in that, The slope change data is output based on the DEM model, specifically as follows: For each grid cell in the DEM model, the initial slope is calculated using the Horn algorithm to obtain an initial slope map; For the initial slope map, the crack area is locked and the crack direction is extracted by using the segmentation mask as a guide. The slope difference value on both sides of the crack that was missed due to the smoothing process in the previous model construction is calculated along the crack direction. The calculated data is added to the initial slope data of the corresponding grid to obtain the crack area corrected slope data. For the initial slope map, the gully area is locked using a segmentation mask as a guide. The initial slope value of the gully sidewalls within the gully area is adjusted based on the cumulative rainfall in the target area within the target number of days to obtain the corrected slope data of the gully area. The cumulative rainfall and the upward adjustment of the initial slope value of the gully sidewalls are positively correlated. By integrating the corrected slope data of the crack area, the corrected slope data of the gully area, and the initial slope data of the complete area in the initial slope map, a corrected slope map is obtained; wherein, the complete area refers to the area in the initial slope map excluding the range of the crack area and the range of the gully area; Based on the corrected slope map, the slope variation data is calculated according to the average slope of the intact area, as well as the average slope of the cracked area and the gully area.

6. The intelligent monitoring method for loess slope stability as described in claim 1, characterized in that, The fused feature vector and the stability level are obtained by cross-modal attention fusion calculation and vector classification of historical multi-source data and meteorological precipitation data, specifically as follows: Acquire the historical multi-source data, including historical slope change data and historical slope deformation trends, as well as historical meteorological and rainfall data; The historical slope change data, the historical slope deformation trend, and the historical meteorological rainfall data are respectively converted into slope feature vectors, deformation feature vectors, and rainfall feature vectors; Calculate the cosine similarity between the slope feature vector and the deformation feature vector, normalize the cosine similarity to a spatial attention weight of 0-1, and perform pixel-level weighting on the slope feature vector and the deformation feature vector according to the spatial attention weight to obtain the weighted slope feature vector and the weighted deformation feature vector. The weighted slope feature vector, the weighted deformation feature vector, and the rainfall feature vector are concatenated. The importance of each channel is extracted by global average pooling, and channels with a rainfall impact index greater than the preset rainfall amount are assigned a weight by several times to calculate the fused feature vector. For the fused feature vector of several grids, the specific values ​​of slope change, deformation acceleration and rainfall impact index are read, and the level category of the read values ​​is determined according to the preset stability level determination rules to obtain the stability level of several grids.

7. The intelligent monitoring method for loess slope stability as described in claim 6, characterized in that, The historical slope change data, the historical slope deformation trend, and the historical meteorological rainfall data are respectively converted into slope feature vectors, deformation feature vectors, and rainfall feature vectors, as follows: In the historical slope change data, areas where the slope change of several consecutive grids exceeds a preset degree are marked as slip zones, and global average pooling is performed on the areas marked as slip zones to obtain the slope feature vector. In the historical slope deformation trend, the area where the displacement acceleration exceeds the preset speed is marked as the acceleration risk zone. The area marked as the acceleration risk zone is subjected to convolution compression processing to obtain the deformation feature vector. The historical meteorological rainfall data is transformed into a multi-dimensional rainfall impact index table, and the rainfall impact index table is mapped to the feature space according to a preset fully connected layer to obtain the rainfall feature vector.

8. A method for intelligent monitoring of loess slope stability as described in any one of claims 1 to 7, characterized in that, The improved method for obtaining the PSPNet network is as follows: An initial PSPNet network is established based on the data processing module, the dual-branch cross-modal feature fusion module, and the dynamic pyramid pooling module. Based on the difference between the binary segmentation mask and the corresponding real mask, the model parameters of the initial PSPNet network are adjusted through backpropagation to obtain the improved PSPNet network. The data processing module is used to project the finely registered point cloud onto the pixel plane of the finely registered image based on the finely registered data, and to add geometric features including depth value and normal vector to each pixel to generate a depth-assisted multispectral image.

9. The intelligent monitoring method for loess slope stability as described in claim 8, characterized in that, The dual-branch cross-modal feature fusion module is used to convert the depth-assisted multispectral image into a cross-modal fusion feature map, and the dynamic pyramid pooling module is used to generate the binary segmentation mask based on the cross-modal fusion feature map, specifically: The dual-branch cross-modal feature fusion module is used to extract the texture features of the depth-assisted multispectral image and generate a spectral feature map; and to extract the geometric features of the finely registered point cloud and convert them into a geometric feature map through data projection. The spectral feature map and the geometric feature map are sequentially subjected to spatial attention weighting calculation and channel attention weight allocation, and the results are spliced ​​together to obtain the cross-modal fusion feature map. The dynamic pyramid pooling module is used to perform four-scale pooling on the cross-modal fusion feature map and dynamically allocate pooling weights according to the crack density to obtain a multi-scale feature map; the multi-scale feature map is then subjected to convolutional compression and upsampling to obtain a super-resolution feature map; the super-resolution feature map is then classified into two categories according to pixel values ​​to obtain the binary segmentation mask.

10. An intelligent monitoring system for loess slope stability, characterized in that, It includes a data module, a slope module, a deformation module, and an early warning module; The data module is used to acquire multi-source data and periodic displacement time series data of the target area. The slope module is used to perform regional differential modeling on the multi-source data and output slope change data. Specifically, it involves: correcting and aligning the multispectral images and original point clouds in the multi-source data to obtain finely registered data; inputting the finely registered data into an improved PSPNet network to segment the slope crack and gully development areas to obtain a binary segmentation mask; dividing the finely registered data into different regional data according to the binary segmentation mask and then stitching them together according to spatial coordinates to generate a DEM model covering the entire slope; and outputting the slope change data based on the DEM model. The deformation module is used to predict the slope deformation trend based on the LSTM algorithm according to the phase-by-phase displacement time series data. The early warning module is used to input the slope change data, the slope deformation trend, and meteorological rainfall data into the slope stability evaluation model to predict the safety factor. If the safety factor is lower than a preset threshold, an early warning information is issued. The slope stability evaluation model is based on the random forest algorithm and is established by learning the mapping law between the fused feature vector and the stability level. The fused feature vector and the stability level are obtained by cross-modal attention fusion calculation and vector classification of historical multi-source data and meteorological rainfall data.

Images