Intelligent monitoring method and system for stability of loess slope

Abstract

The invention discloses a loess slope stability intelligent monitoring method and system, and the method comprises the steps: carrying out the regional differentiation modeling of multi-source data of a target region, and outputting slope change data; according to the period-by-period displacement time sequence data, a slope deformation trend is obtained through prediction based on an LSTM algorithm; and inputting the slope change data, the slope deformation trend and the meteorological rainfall data into a slope stability evaluation model to predict a safety coefficient, and if the safety coefficient is lower than a preset threshold value, issuing early warning information. According to the loess slope stability intelligent monitoring method and system provided by the invention, data such as gradient change and deformation trend are accurately extracted, multi-source data and meteorological rainfall data are fused by means of cross-modal attention, a random forest model is used for learning a multi-factor coupling rule, and finally multi-factor data prediction is integrated to obtain a safety coefficient for early warning. The problem of low early warning reliability caused by difficulty in accurately evaluating the slope stability by fusing multiple factors such as gradient, deformation and meteorological rainfall can be solved.

Description

Technical Field

[0001] This invention relates to the field of slope monitoring technology, and in particular to an intelligent monitoring method and system for the stability of loess slopes. Background Technology

[0002] Loess slopes are prone to instability and landslides due to rainfall and engineering activities. Stability monitoring can identify and warn of risks in advance, reducing damage to transportation, water conservancy, and other facilities, and effectively protecting the lives and property of residents. Simultaneously, monitoring data can provide strong support for the safe operation of infrastructure, assist in regional ecological protection and scientific land use planning, and help build a safe living environment and promote sustainable development in the Loess Plateau region. Traditional slope monitoring relies heavily on manual inspections combined with single-point instrument monitoring for early warning. Staff need to regularly visit the site to observe and record details such as cracks and surface deformation. Simultaneously, inclinometers, settlement gauges, and piezometers are deployed at key locations on the slope to monitor deep soil displacement, surface settlement, and pore water pressure. The data is manually processed or stored using simple equipment. When monitored indicators exceed preset safety values, an early warning process is initiated.

[0003] However, traditional slope monitoring technologies have significant shortcomings. Manual inspections require regular on-site operations, which are inefficient and have limited coverage, making it difficult to capture slope dynamics in real time. Single-point instrument monitoring data is limited to a local area and cannot reflect the overall condition of the slope. Data needs to be manually processed or simply stored, which can lead to delays and untimely early warning responses. In addition to high manpower and costs, errors may be introduced due to subjective human judgment. Furthermore, it is difficult to integrate multiple factors such as slope, deformation, and weather and rainfall to achieve accurate stability evaluation, and thus cannot effectively support reliable early warning. Summary of the Invention

[0004] This invention provides an intelligent monitoring method and system for loess slope stability to solve the problem of low early warning reliability caused by the difficulty in accurately assessing slope stability by integrating multiple factors such as slope, deformation and meteorological rainfall.

[0005] To achieve the above objectives, this application provides an intelligent monitoring method for loess slope stability, comprising: 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.

[0006] This invention utilizes an improved PSPNet to segment crack and gully areas, accurately identifying key instability zones on slopes. A differentiated regional modeling process generates a Deformation Model (DEM), allowing slope variation data to specifically reflect the characteristics of different regions, rather than providing general data. This ensures accurate correlation with deformation trends and rainfall impacts. Furthermore, by generating a DEM covering the entire region and outputting slope variation data, it avoids model building biases caused by slope data issues. The LSTM algorithm predicts slope deformation trends based on periodic displacement time-series data, accurately quantifying the temporal evolution characteristics of deformation and providing reliable predictive data for deformation factors. By learning the mapping relationship between fusion feature vectors and stability levels through the Random Forest algorithm, the established slope stability evaluation model quantifies the nonlinear correlation between slope, deformation, and meteorological rainfall. Inputting multi-factor data into the model to predict the safety factor, and considering the interaction of different factors, the prediction results more closely resemble the actual state, solving the multi-factor fusion problem, significantly reducing manual inspection costs, and improving early warning reliability.

[0007] Compared with existing technologies, this invention accurately extracts single-factor data such as slope changes and deformation trends, integrates multi-source data and meteorological and rainfall data with cross-modal attention, and then uses a random forest model to learn the coupling law of multiple factors. Finally, it integrates slope, deformation, and meteorological and rainfall predictions to obtain a safety factor for early warning. Therefore, it can solve the problem of low reliability of early warning caused by the difficulty in accurately assessing slope stability by integrating multiple factors such as slope, deformation, and meteorological and rainfall.

[0008] As a preferred embodiment, 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.

[0009] In this preferred scheme, for multispectral images and 3D laser point clouds, the data quality is first improved by orthorectification, radiometric correction and filtering. Then, coarse registration is achieved by combining target coordinates and ICP algorithm. Finally, the parameters are optimized by using edge points as constraints 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 the dual registration strategy, providing high-precision basic data for subsequent slope feature extraction and modeling, and can effectively avoid analysis errors caused by data misalignment.

[0010] As a preferred embodiment, 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.

[0011] This preferred scheme establishes the spatial relationship between the image and the point cloud based on coarse registration parameters, extracts the dark edges of the image by enhancing the green light band, and filters out abnormal areas by combining the Z-coordinate gradient of the point cloud. Finally, a preliminary mask is generated through spatial intersection operation. This scheme can make full use of the spectral characteristics of the image and the three-dimensional geometric information of the point cloud to accurately locate 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 key feature areas of the slope.

[0012] As a preferred embodiment, the finely registered data is divided into different regions based on the binary segmentation mask, and then 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.

[0013] This preferred scheme differentiates the point clouds of different regions based on a binary segmentation mask, and then generates DEMs by region and stitches them together. This regional modeling strategy not only ensures the detailed representation of key areas such as cracks and gullies, but also improves the overall modeling efficiency through simplified processing. The generated DEM model can fully reflect the complex terrain features of the slope and provide a high-precision basis for slope analysis.

[0014] As a preferred embodiment, 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.

[0015] This preferred scheme calculates the initial slope using the Horn algorithm, corrects the slope difference in the crack area by combining 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 takes into account the information loss caused by the smoothing of the crack area and the impact of rainfall on the gully, making the slope data more consistent with the actual slope conditions. The slope change data obtained by calculating the average slope by partition can more accurately reflect the key indicators of slope stability.

[0016] As a preferred embodiment, the fused 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 precipitation data, specifically: Acquire the historical multi-source data, including historical slope change data and historical slope deformation trends, as well as the 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.

[0017] This preferred scheme transforms historical data into feature vectors, calculates spatial attention weights using cosine similarity, and generates fused feature vectors by assigning channel weights in conjunction with the rainfall impact index. Stability levels are then labeled according to rules. This process effectively integrates multi-source data features through cross-modal attention fusion, highlighting key influencing factors. The generated fused feature vectors and stability levels provide high-quality training data for the slope stability evaluation model, improving the model's learning performance.

[0018] As a preferred embodiment, 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, specifically 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.

[0019] This preferred scheme marks the slip zone and the accelerated risk zone, and performs pooling and convolution processing on historical slope change data and deformation trends respectively, mapping rainfall data into feature vectors. This feature transformation method accurately extracts the key features of slope stability, transforms unstructured data into structured feature vectors, retains important information and reduces data dimensionality, laying a good foundation for cross-modal fusion computing.

[0020] As a preferred embodiment, the improved PSPNet network is obtained 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.

[0021] The improved PSPNet network constructed in this preferred scheme generates depth-assisted multispectral images by adding geometric features to pixels and optimizing parameters based on segmentation mask differences. It can effectively fuse spectral and geometric features, improve the segmentation accuracy of slope cracks and gullies, provide an accurate binary segmentation mask for subsequent regional modeling, and enhance the model's ability to identify complex slope features.

[0022] As a preferred embodiment, 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; extract the geometric features of the finely registered point cloud and convert them into a geometric feature map through data projection; perform spatial attention weighting calculation and channel attention weight allocation on the spectral feature map and the geometric feature map in sequence, and stitch the results 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.

[0023] In this preferred scheme, the dual-branch cross-modal feature fusion module fuses spectral and geometric feature maps through an attention mechanism, and the dynamic pyramid pooling module assigns weights according to crack density and generates a binary segmentation mask. This design not only makes full use of the advantages of multimodal data, but also highlights key areas such as cracks through dynamic weight allocation, thereby improving the targeting and accuracy of segmentation. The generated binary segmentation mask can accurately divide different areas of the slope, providing a reliable basis for subsequent differentiated processing.

[0024] This application also provides an intelligent monitoring system for loess slope stability, including 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. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating an intelligent monitoring method for loess slope stability provided in an embodiment of this application. Figure 2 This is a schematic diagram of the structure of an intelligent monitoring system for loess slope stability provided in an embodiment of this application. Detailed Implementation

[0026] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0027] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" and "second" may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "several" means two or more.

[0028] Example 1: Please see Figure 1The embodiments of this application provide an intelligent monitoring method for the stability of loess slopes, including S1~S4, and the specific implementation steps are as follows: S1. Obtain multi-source data and periodic displacement time series data of the target area.

[0029] Step S1 in this embodiment of the application is specifically as follows: Within a preset time window, a multispectral camera from a drone and a 3D laser scanner work together to acquire multi-source data of the target area. The drone is equipped with a five-lens multispectral camera covering red, green, blue, near-infrared, and red-edge bands, simultaneously recording GPS (Global Positioning System) and IMU (Inertial Measurement Unit) data, and outputting raw TIFF format 5-band multispectral images. The 3D laser scanner scans the slope with a point spacing of less than or equal to 2mm, outputting raw point clouds in LAS format containing XYZ coordinates and reflection intensity. At least five spherical reflective targets are simultaneously deployed and their 3D coordinates are recorded, thus forming multi-source data covering the target area.

[0030] Furthermore, the slope surface of the target area is monitored using time-series InSAR (Interferometric Synthetic Aperture Radar) technology to obtain periodic displacement time-series data, which includes core information such as the three-dimensional coordinates of the monitoring points, observation time, and surface displacement. Simultaneously, meteorological rainfall data that perfectly matches the displacement monitoring cycle is acquired from the meteorological station, specifically covering key parameters such as daily rainfall, cumulative rainfall in the past 15 days, and maximum daily rainfall.

[0031] The target area can be a typical loess slope scene such as a highway slope or a pure slope.

[0032] S2. Perform regional differential modeling on multi-source data and output slope change data. Specifically, correct and align the multispectral images and original point clouds in the multi-source data to obtain finely registered data; input the finely registered data into the improved PSPNet network to segment the slope crack and gully development areas to obtain a binary segmentation mask; divide the finely registered data into different regional data according to the binary segmentation mask and then stitch them together according to spatial coordinates to generate a DEM model covering the entire slope; output slope change data based on the DEM model.

[0033] Step S2 in this embodiment includes S2.1 to S2.5, specifically as follows: S2.1 Based on the GPS and IMU data acquired during the multi-source data acquisition phase, the original multispectral imagery from the multi-source data undergoes dual correction processing: firstly, orthorectification, which eliminates geometric distortion caused by terrain undulations through spatial coordinate matching, ensuring accurate correspondence between the imagery and the actual spatial location; secondly, radiometric correction, which focuses on removing external interference factors such as atmospheric scattering and uneven illumination, restoring the true spectral information of the imagery, resulting in the corrected multispectral imagery. This imagery fully retains the five core spectral bands: red, green, blue, near-infrared, and red edge. The green band can specifically support subsequent work on extracting the dark edges of slope cracks and gully areas.

[0034] For the original point cloud in multi-source data, outliers with a distance greater than 3 times the standard deviation from the mean are removed by statistical filtering, and effective points are retained to obtain a filtered point cloud carrying XYZ coordinates; Using the target coordinates as a reference, the corrected multispectral image and the filtered point cloud are aligned according to the ICP algorithm (Iterative Closest Point Algorithm) to obtain coarse registration parameters in the form of rotation matrix and translation vector.

[0035] S2.2 Based on the corrected multispectral image and filtered point cloud, establish the spatial relationship between image pixel coordinates and point cloud three-dimensional coordinates according to the coarse registration parameters; Contrast enhancement processing is performed on the green band of the corrected multispectral image to highlight the difference between the dark tones of cracks and gullies and the background. At the same time, continuous dark edges of the target are extracted to generate a dark edge mask of the image, which marks the image areas that may be cracks and gullies. Based on the pixel coordinates of the dark edge mask in the image, the corresponding three-dimensional coordinate region in the point cloud is obtained through spatial correlation, and the Z coordinate gradient of the three-dimensional coordinate region is calculated. Point cloud regions with Z coordinate gradient greater than 0.5mm / pixel are selected and transformed into point cloud gradient anomaly masks. Spatial intersection operation is performed on the image dark edge mask and the point cloud gradient anomaly mask to retain the overlapping area and remove misjudged areas such as shadows that only have dark edges in the image but no gradient anomalies in the point cloud. Finally, a preliminary mask for cracks and gullies is obtained.

[0036] Extract the light-dark boundary line of the crack and the abrupt change line of the gully terrain from the initial mask, and obtain the corresponding image edge points and point cloud edge points; Using image edge points and 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 ±1mm, thus obtaining fine registration data including finely registered image and finely registered point cloud.

[0037] In this embodiment, S2.2 establishes the spatial relationship between the image and the point cloud based on the coarse registration parameters. The dark edges of the image are extracted by enhancing the green light band, and abnormal areas are screened by combining the Z-coordinate gradient of the point cloud. Finally, a preliminary mask is generated through spatial intersection operation. This can make full use of the spectral characteristics of the image and the three-dimensional geometric information of the point cloud to accurately locate 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 key feature areas of the slope.

[0038] In S2.1 to S2.2 of this embodiment, for multispectral images and three-dimensional laser point clouds, the data quality is first improved by orthorectification, radiometric correction and filtering. Then, coarse registration is achieved by combining target coordinates and ICP algorithm. Finally, the parameters are optimized by using edge points as constraints 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 the dual registration strategy, providing high-precision basic data for subsequent slope feature extraction and modeling, and can effectively avoid analysis errors caused by data misalignment.

[0039] S2.3. Establish an initial PSPNet network based on the data processing module, the dual-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 prediction mask; wherein, the "preset fine registration data" is historical data specifically used for model training, calculated in the same way as in step S2.2 above. Based on the difference between the predicted mask and the corresponding real mask, the model parameters of the initial PSPNet network are adjusted through backpropagation. When the edge localization error is less than or equal to 0.8 mm, an improved PSPNet network is obtained.

[0040] The finely registered data calculated in step S2.2 is input into the improved PSPNet network to segment the slope crack and gully development areas, resulting in a binary segmentation mask. This binary segmentation mask is a pixel-level binary marker map with the same size as the finely registered multispectral image; "0" corresponds to the intact slope area, and "1" corresponds to the crack and gully area. It is generated by the improved PSPNet network processing the finely registered data. It guides the regional data processing, supports the generation of regional DEMs, and assists in slope correction, laying the foundation for accurately acquiring slope change data.

[0041] The following provides a detailed explanation of the data processing module, the dual-branch cross-modal feature fusion module, and the dynamic pyramid pooling module: (1) Purpose of the data processing module: Based on the finely registered data, the finely registered point cloud is projected onto the pixel plane of the finely registered image. Geometric features, including depth and normal vectors, are added to each pixel. The depth value is the Z-coordinate of the point cloud, reflecting the altitude; the normal vector is calculated through PCA analysis of the 3×3 neighborhood point cloud, reflecting the slope direction. Subsequently, the five spectral channels (red, green, blue (RGB), near-infrared, and red edge) are integrated, along with the depth value and normal vector added by projecting the finely registered point cloud onto the image pixel plane, to generate a depth-assisted multispectral image.

[0042] (2) Applications of the dual-branch cross-modal feature fusion module: Using ResNeXt-50 as the backbone, color or texture features, such as dark tones in cracks and spectral abrupt changes in gullies, are extracted from the first 5 spectral channels of depth-assisted multispectral images to generate spectral feature maps. Using PointNet++ as the backbone, geometric features of the finely registered point cloud, such as crack edge curvature and gully depth gradient, are extracted and transformed into geometric feature maps through data projection.

[0043] Spatial attention weighting and channel attention weighting are performed sequentially on the spectral feature map and the geometric feature map. The results are then concatenated to obtain the cross-modal fusion feature map, specifically: Using spectral feature maps and geometric feature maps as input, feature vectors are extracted from corresponding pixels of the two feature maps respectively. By calculating the cosine similarity of the pixel-level feature vectors, the cosine similarity is normalized into spatial attention weights. Based on these weights, pixel-level weighting is performed on the spectral feature maps and geometric feature maps respectively to obtain spatially weighted spectral feature maps and spatially weighted geometric feature maps. For the spatially weighted spectral feature map, the feature importance of its 256 channels is extracted by global average pooling, and a higher channel weight is assigned to the near-infrared channel (which is sensitive to dry crack response). For the spatially weighted geometric feature map, the feature importance of its 128 channels is extracted by global average pooling, and a higher channel weight is assigned to the depth gradient channel (which is sensitive to gully depth changes). In this way, channel-level enhancement of the two types of feature maps is completed, resulting in channel-weighted spectral feature maps and channel-weighted geometric feature maps. By concatenating the channel-weighted spectral feature map and the channel-weighted geometric feature map along the channel dimension, a cross-modal fusion feature map is obtained.

[0044] (3) The purpose of the dynamic pyramid pooling module: The cross-modal fusion feature map is subjected to four-scale pooling processing, and the pooling weight is dynamically allocated according to the crack density. That is, when the crack density is >0.6, the small-scale pooling weight accounts for 70%, which focuses on preserving the fine topographic details of cracks and gullies; when the crack density is <0.3, the large-scale pooling weight accounts for 60%, which focuses on capturing the global topographic correlation of the slope, and finally a multi-scale feature map is obtained. The multi-scale fusion feature map is then processed: first, the number of channels is compressed to 128 through 1×1 convolution to reduce the computational load, and then ESRGAN super-resolution technology is used to perform 4x upsampling to obtain the super-resolution feature map; finally, the super-resolution feature map is binary classified by the Softmax function to output a binary segmentation mask; where a pixel value of 1 represents a crack (width ≥ 1 mm) or a gully (depth ≥ 5 mm), and a pixel value of 0 represents a complete slope.

[0045] The improved PSPNet network constructed in S2.3 of this embodiment generates depth-assisted multispectral images by adding geometric features to pixels and optimizing parameters based on segmentation mask differences. It can effectively fuse spectral and geometric features, improve the segmentation accuracy of slope cracks and gullies, provide an accurate binary segmentation mask for subsequent regional modeling, and enhance the model's ability to identify complex slope features. Furthermore, the dual-branch cross-modal feature fusion module fuses spectral and geometric feature maps through an attention mechanism, while the dynamic pyramid pooling module assigns weights based on crack density and generates a binary segmentation mask. This design not only makes full use of the advantages of multimodal data but also highlights key areas such as cracks through dynamic weight allocation, improving the targeting and accuracy of segmentation. The generated binary segmentation mask can accurately divide different areas of the slope, providing a reliable basis for subsequent differentiated processing.

[0046] S2.4. Based on the binary segmentation mask, the fine registration data is divided into point clouds of crack area, point clouds of gully area, and point clouds of intact area; The point cloud in the crack area is densified and resampled to preserve the subtle depth changes of the cracks; the point cloud in the gully area is smoothed and filtered to preserve the steep edges of the target sidewalls; and the point cloud in the intact area is downsampled to retain 1 / 5 of the original points. After processing, a comprehensive semantically enhanced point cloud is obtained, which includes the densified point cloud in the crack area, the denoised point cloud in the gully area, and the simplified point cloud in the intact area. Based on the crack region densified point cloud in the comprehensive semantically enhanced point cloud, the depth discontinuity of the crack region is preserved by Poisson reconstruction and edge constraints to generate crack region DEM; Based on the denoised point cloud of the gully area in the comprehensive semantically enhanced point cloud, the DEM of the gully area is generated by inverse distance weighted interpolation and hydrodynamic correction. Based on the complete region simplification point cloud in the comprehensive semantic enhancement point cloud, the complete region DEM is generated by kriging interpolation; By splicing the DEMs of the cracked area, gully area, and intact area according to spatial coordinates, and using the edge transition algorithm to eliminate the splicing seams, a DEM model covering the entire slope is generated.

[0047] In this embodiment, S2.4, point clouds in different regions are differentiated based on a binary segmentation mask, and then DEMs are generated and stitched together by region. This regional modeling strategy not only ensures the detailed representation of key areas such as cracks and gullies, but also improves the overall modeling efficiency through simplified processing. The generated DEM model can fully reflect the complex terrain features of the slope, providing a high-precision foundation for slope analysis.

[0048] S2.5 For each grid cell in the DEM model, the initial slope is calculated using the Horn algorithm to obtain the 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 so that the slope of the crack area can truly reflect its local steepness and the topographic undulation on both sides, and the corrected slope data of the crack area is obtained. Based on meteorological rainfall data, the initial slope map is used as a guide to locate the gully area. The initial slope value of the gully sidewalls within the gully area is adjusted by combining the cumulative rainfall of the target area over 15 days, resulting in the corrected slope data of the gully area. The cumulative rainfall and the adjustment of the initial slope value of the gully sidewalls are positively correlated. For example, for every 100mm increase in rainfall, the initial slope of the gully sidewalls is increased by 2%, which compensates for the slope steepening deviation caused by rainwater erosion, making the slope of the gully area more consistent with the actual terrain. The initial slope data of the complete area is directly retained as the benchmark for subsequent calculations; the complete area refers to the area in the initial slope map excluding the crack area and the gully area. 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, and removing invalid data such as blank areas outside the slope in the initial slope map, a corrected slope map is obtained. This map only retains the "precisely corrected slope of the crack area, the compensated slope of the gully area, and the baseline slope of the complete area". Based on the corrected slope map, the slope variation data ΔS is calculated according to the average slope S0 of the intact area and the average slope S1 of the cracked area and gully area, i.e., ΔS=S1-S0.

[0049] In this embodiment, S2.5, the initial slope is calculated using the Horn algorithm, the slope difference in the crack area is corrected by combining the crack direction, the slope of the gully area is adjusted according to the rainfall, and finally the corrected slope map is obtained by integration. This method takes into account the information loss caused by the smoothing of the crack area and the impact of rainfall on the gully, so that the slope data is more in line with the actual slope condition. The slope change data obtained by calculating the average slope by partition can more accurately reflect the key indicators of slope stability.

[0050] S3. Based on the displacement time series data, the slope deformation trend is predicted using the LSTM algorithm.

[0051] Step S3 in this embodiment of the application is specifically as follows: The time-series displacement data and meteorological precipitation data acquired during the multi-source data acquisition phase are integrated and processed. The integrated dataset is then standardized and a time-series window is constructed according to the rule of "predicting the next two periods based on the past five periods of data" to divide the dataset into training and validation sets.

[0052] A two-layer stacked LSTM structure was constructed, and the Adam optimizer and RMSE loss function were selected to train the model. The training was iterated until the relative error of the prediction on the validation set was less than 7% and the model accuracy was stable. Finally, the trained LSTM prediction model was obtained.

[0053] Input the latest five periods of displacement time series data and meteorological and rainfall data into the trained LSTM prediction model to predict the slope displacement in the next 1-2 months, and finally output the slope deformation trend.

[0054] S4. Input 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, 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 stability level are obtained by cross-modal attention fusion calculation and vector classification of historical multi-source data and meteorological rainfall data.

[0055] Step S4 in this embodiment includes S4.1 to S4.2, specifically as follows: S4.1 Based on the above DEM model and LSTM prediction model, obtain historical multi-source data including historical slope change data and historical slope deformation trends, as well as historical meteorological rainfall data for the same period. In historical slope change data, areas with consecutive grid slope changes exceeding 1° are marked as slip zones, and global average pooling is performed on the areas marked as slip zones to compress dimensions, resulting in slope feature vectors. In the historical slope deformation trend, the displacement acceleration exceeds 0.5 mm / month.2 The regions marked as acceleration risk zones are then subjected to convolutional compression to obtain deformed feature vectors. Historical meteorological rainfall data is transformed into a multi-dimensional rainfall impact index table, including cumulative rainfall index and heavy rainfall index. This table is then mapped to the feature space using a pre-defined two-layer fully connected layer. Subsequently, correction terms are added based on slope type. Considering that soil slopes are more sensitive to rainfall, each dimension of the initial feature vector is multiplied by 1.2, while the dimensions of rock slopes, which are less sensitive, are multiplied by 0.8. This adapts to the differences in rainfall sensitivity among different slope types, ultimately yielding the rainfall feature vector. The first layer of the two fully connected layers has 64 hidden units, and the second layer has 32 hidden units, both using the ReLU activation function.

[0056] In this embodiment, S4.1, by marking the slip zone and the accelerated risk zone, the historical slope change data and deformation trend are processed by pooling and convolution, respectively, and the rainfall data is mapped into feature vectors. This feature transformation method accurately extracts the key features of slope stability, transforms unstructured data into structured feature vectors, retains important information and reduces data dimensionality, and lays a good foundation for cross-modal fusion computing.

[0057] S4.2 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 deformation feature vector according to the spatial attention weight to obtain the weighted slope feature vector and the weighted deformation feature vector; where, for the spatial attention weight, weight 1 corresponds to "slope change greater than 2° and acceleration greater than 1 mm / month". 2 ”; The weighted slope feature vector, weighted deformation feature vector, and 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 of 2 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 judgment rules to obtain the stability level of several grids.

[0058] Input the "fusion feature vector and stability level" as the dataset into the random forest model. Among them, the fusion feature vector is the core input of the model, and each vector corresponds to the comprehensive risk characteristics of a single slope grid, covering key information such as slope mutation, deformation acceleration, and strong rainfall coupling. Relying on multiple decision trees of the random forest algorithm, by learning the mapping rule between the fusion feature vector and the stability level, for example, automatically identifying the corresponding relationship between the feature combination of "high slope mutation + high deformation acceleration + strong rainfall" and the "instability risk" label, finally, a trained slope stability evaluation model is obtained.

[0059] Among them, the "stability level determination rule" is divided into three categories: "stable, sub - stable, instability risk", with the slope change value (A, unit: °), deformation acceleration (B, unit: mm / month 2 ) and rainfall impact index (C, unitless) as the core determination indicators. The specific rules are as follows: (1) Stable: Simultaneously satisfy "A ≤ 0.5, B ≤ 0.3, C ≤ 20", or only a single indicator slightly exceeds the threshold (such as A = 0.8, B and C meet the standards); (2) Sub - stable: Simultaneously satisfy "0.5 < A ≤ 2, 0.3 < B ≤ 1", or "20 < C ≤ 50, and either A or B slightly exceeds the threshold"; (3) Instability risk: Simultaneously satisfy "A > 2, B > 1", or "C > 50, A > 1, or B > 0.8".

[0060] In this embodiment, in S4.1 - S4.2, historical data is transformed into feature vectors, the spatial attention weights are calculated through cosine similarity, the channel weights are assigned in combination with the rainfall impact index to generate a fusion feature vector, and the stability level is marked according to the rules. This process effectively integrates the multi - source data features through cross - modal attention fusion, highlights the key influencing factors, and the generated fusion feature vector and stability level provide high - quality training data for the slope stability evaluation model, improving the model learning effect.

[0061] S4.3: Input the slope change data calculated through the high - precision DEM model in the early stage, the slope deformation trend output by the LSTM prediction model, and the meteorological rainfall data monitored by the meteorological station during the same period into the trained slope stability evaluation model; and it should be clear that all three types of data are for the target area and are data for the same time period within the recent 1 - 3 consecutive monitoring cycles; The model first strengthens the coupled risk characteristics of "slope change, deformation acceleration and heavy rainfall" through a cross-modal attention mechanism, and then relies on the multi-decision tree operation of the random forest algorithm to output the safety coefficient and corresponding stability level of each monitoring grid of the slope. If the safety coefficient of any monitoring grid is lower than the preset safety threshold for loess slope engineering, 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 released in real time, providing timely decision-making basis for slope safety management. Among them, the stability level is divided into "stable, under-stable, and unstable risk", which can support hierarchical management and control, locate risk factors, and serve as an intuitive carrier of "risk level" in early warning information.

[0062] Overall, this embodiment has the following beneficial effects: This invention utilizes an improved PSPNet to segment crack and gully areas, accurately identifying key slope instability zones. A regionally differentiated DEM model is then generated, allowing slope variation data to specifically reflect the characteristics of different areas, rather than providing general data. This ensures accurate correlation with deformation trends and rainfall impacts. Furthermore, by generating a DEM model covering the entire region and outputting slope variation data, it avoids model building biases caused by slope data issues. The LSTM algorithm predicts slope deformation trends based on periodic displacement time-series data, accurately quantifying the temporal evolution characteristics of deformation and providing reliable predictive data for deformation factors. By learning the mapping relationship between fusion feature vectors and stability levels through the random forest algorithm, the established slope stability evaluation model quantifies the nonlinear correlation between slope, deformation, and meteorological rainfall. Inputting multi-factor data into the model to predict the safety factor, and considering the interaction of different factors, the prediction results more closely match the actual state, solving the multi-factor fusion problem, significantly reducing manual inspection costs, and improving early warning reliability. In summary, this invention addresses the core needs of loess slope stability monitoring, such as high costs and delayed early warning responses caused by manual inspections. It employs a multispectral camera from a drone and a 3D laser scanner to collect multi-source data. By comparing data from multiple periods, it accurately identifies and analyzes key deformation features such as slope cracks and gullies. Simultaneously, it enables automated acquisition and intelligent analysis of slope monitoring data, effectively reducing the manpower and time costs of manual inspections. Ultimately, in actual monitoring scenarios such as highway slopes, it significantly improves the accuracy of early warnings.

[0063] Example 2: Please see Figure 2 The embodiments of this application provide an intelligent monitoring system for loess slope stability, including a data module 10, a slope module 20, a deformation module 30, and an early warning module 40; Among them, data module 10 is used to acquire multi-source data and periodic displacement time series data of the target area; The slope module 20 is used to perform regional differential modeling on multi-source data and output slope change data. Specifically, it corrects and aligns the multispectral images and original point clouds in the multi-source data to obtain finely registered data; inputs the finely registered data into an improved PSPNet network to segment the slope crack and gully development areas to obtain a binary segmentation mask; divides the finely registered data into different regional data according to the binary segmentation mask and then stitches them according to spatial coordinates to generate a DEM model covering the entire slope; and outputs slope change data based on the DEM model. Deformation module 30 is used to predict the slope deformation trend based on the LSTM algorithm according to the displacement time series data. The early warning module 40 is used to input 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, 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.

[0064] In one embodiment, data module 10 specifically comprises: Within a preset time window, a multispectral camera from a drone and a 3D laser scanner work together to acquire multi-source data of the target area. The drone is equipped with a five-lens multispectral camera covering red, green, blue, near-infrared, and red-edge bands, simultaneously recording GPS (Global Positioning System) and IMU (Inertial Measurement Unit) data, and outputting raw TIFF format 5-band multispectral images. The 3D laser scanner scans the slope with a point spacing of less than or equal to 2mm, outputting raw point clouds in LAS format containing XYZ coordinates and reflection intensity. At least five spherical reflective targets are simultaneously deployed and their 3D coordinates are recorded, thus forming multi-source data covering the target area.

[0065] Furthermore, the slope surface of the target area is monitored using time-series InSAR (Interferometric Synthetic Aperture Radar) technology to obtain periodic displacement time-series data, which includes core information such as the three-dimensional coordinates of the monitoring points, observation time, and surface displacement. Simultaneously, meteorological rainfall data that perfectly matches the displacement monitoring cycle is acquired from the meteorological station, specifically covering key parameters such as daily rainfall, cumulative rainfall in the past 15 days, and maximum daily rainfall.

[0066] The target area can be a typical loess slope scene such as a highway slope or a pure slope.

[0067] In one embodiment, the slope module 20 includes a coarse registration unit, a fine registration unit, a mask unit, a model unit, and a slope unit; The coarse registration unit performs dual correction processing on the raw multispectral imagery from the multi-source data, based on GPS and IMU data acquired during the multi-source data acquisition phase. The first correction is orthorectification, which eliminates geometric distortion caused by terrain undulations through spatial coordinate matching, ensuring accurate correspondence between the imagery and the actual spatial location. The second correction is radiometric correction, which focuses on removing external interference factors such as atmospheric scattering and uneven illumination to restore the true spectral information of the imagery, resulting in a corrected multispectral image. This imagery fully retains five core spectral bands: red, green, blue, near-infrared, and red edge. The green band can specifically support subsequent dark edge extraction work for slope cracks and gully areas.

[0068] The coarse registration unit is also used to remove outliers with a distance greater than 3 times the standard deviation from the mean in the original point cloud of multi-source data through statistical filtering, while retaining effective shape points to obtain a filtered point cloud carrying XYZ coordinates. The coarse registration unit is also used to align the corrected multispectral image and the filtered point cloud with the target coordinates as a reference, according to the ICP algorithm (Iterative Closest Point Algorithm), to obtain coarse registration parameters in the form of rotation matrix and translation vector.

[0069] The fine registration unit is used to establish the spatial relationship between image pixel coordinates and point cloud 3D coordinates based on the corrected multispectral image and filtered point cloud, according to the coarse registration parameters. The fine registration unit is also used to perform contrast enhancement processing on the green band in the corrected multispectral image to highlight the difference between the dark tones of cracks and gullies and the background. At the same time, it extracts the continuous dark edges of the target and generates a dark edge mask for the image, which marks the image areas that may be cracks and gullies. The fine registration unit is also used to obtain the corresponding three-dimensional coordinate region in the point cloud based on the pixel coordinates of the dark edge mask of the image, and calculate the Z coordinate gradient of the three-dimensional coordinate region. Point cloud regions with Z coordinate gradient greater than 0.5mm / pixel are selected and converted into point cloud gradient anomaly masks. The fine registration unit is also used to perform spatial intersection operations on the image dark edge mask and the point cloud gradient anomaly mask, retaining the overlapping area of ​​the two and removing misjudged areas such as shadows that only have dark edges in the image but no gradient anomalies in the point cloud, and finally obtaining the preliminary mask for cracks and gullies.

[0070] The fine registration unit is also used to extract the light-dark boundary line of cracks and the abrupt change line of gully terrain from the preliminary mask, and to obtain the image edge points and point cloud edge points accordingly. The fine registration unit is also used to optimize the coarse registration parameters according to the ICP algorithm, using image edge points and point cloud edge points as constraints, so that the registration error between the corrected multispectral image and the filtered point cloud is less than ±1mm, and obtain fine registration data including finely registered image and finely registered point cloud.

[0071] In this embodiment, the fine registration unit establishes the spatial relationship between the image and the point cloud based on the coarse registration parameters. It extracts the dark edges of the image by enhancing the green light band and filters out abnormal areas by combining the Z-coordinate gradient of the point cloud. Finally, a preliminary mask is generated through 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 to accurately locate 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 key feature areas of the slope.

[0072] In the coarse registration unit and fine registration unit of this embodiment, for multispectral images and 3D laser point clouds, the data quality is first improved by orthorectification, radiometric correction and filtering. Then, coarse registration is achieved by combining target coordinates and ICP algorithm. Finally, the parameters are optimized with edge points as constraints 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 the dual registration strategy, providing high-precision basic data for subsequent slope feature extraction and modeling, and can effectively avoid analysis errors caused by data misalignment.

[0073] The mask unit is used to build the initial PSPNet network based on the data processing module, the dual-branch cross-modal feature fusion module, and the dynamic pyramid pooling module. The preset fine registration data is input into the initial PSPNet network for calculation to obtain the binary segmentation mask used for training, i.e., the prediction mask. The "preset fine registration data" is historical data specifically used for model training, which is calculated in the same way as the fine registration unit. The mask unit is also used to adjust the model parameters of the initial PSPNet network through backpropagation based on the difference between the predicted mask and the corresponding real mask. When the edge localization error is less than or equal to 0.8 mm, an improved PSPNet network is obtained.

[0074] The mask unit is also used to input the finely registered data calculated by the finely registered unit into the improved PSPNet network to segment the slope crack and gully development areas, obtaining a binary segmentation mask. The binary segmentation mask is a pixel-level binary marker map with the same size as the finely registered multispectral image; "0" corresponds to the intact slope area, and "1" corresponds to the crack and gully area. Specifically, it is generated by the improved PSPNet network processing the finely registered data. It guides the regional data processing, supports the generation of regional DEMs, and assists in slope correction, laying the foundation for accurately acquiring slope change data.

[0075] The following provides a detailed explanation of the data processing module, the dual-branch cross-modal feature fusion module, and the dynamic pyramid pooling module: (1) Purpose of the data processing module: Based on the finely registered data, the finely registered point cloud is projected onto the pixel plane of the finely registered image. Geometric features, including depth and normal vectors, are added to each pixel. The depth value is the Z-coordinate of the point cloud, reflecting the altitude; the normal vector is calculated through PCA analysis of the 3×3 neighborhood point cloud, reflecting the slope direction. Subsequently, the five spectral channels (red, green, blue (RGB), near-infrared, and red edge) are integrated, along with the depth value and normal vector added by projecting the finely registered point cloud onto the image pixel plane, to generate a depth-assisted multispectral image.

[0076] (2) Applications of the dual-branch cross-modal feature fusion module: Using ResNeXt-50 as the backbone, color or texture features, such as dark tones in cracks and spectral abrupt changes in gullies, are extracted from the first 5 spectral channels of depth-assisted multispectral images to generate spectral feature maps. Using PointNet++ as the backbone, geometric features of the finely registered point cloud, such as crack edge curvature and gully depth gradient, are extracted and transformed into geometric feature maps through data projection.

[0077] Spatial attention weighting and channel attention weighting are performed sequentially on the spectral feature map and the geometric feature map. The results are then concatenated to obtain the cross-modal fusion feature map, specifically: Using spectral feature maps and geometric feature maps as input, feature vectors are extracted from corresponding pixels of the two feature maps respectively. By calculating the cosine similarity of the pixel-level feature vectors, the cosine similarity is normalized into spatial attention weights. Based on these weights, pixel-level weighting is performed on the spectral feature maps and geometric feature maps respectively to obtain spatially weighted spectral feature maps and spatially weighted geometric feature maps. For the spatially weighted spectral feature map, the feature importance of its 256 channels is extracted by global average pooling, and a higher channel weight is assigned to the near-infrared channel (which is sensitive to dry crack response). For the spatially weighted geometric feature map, the feature importance of its 128 channels is extracted by global average pooling, and a higher channel weight is assigned to the depth gradient channel (which is sensitive to gully depth changes). In this way, channel-level enhancement of the two types of feature maps is completed, resulting in channel-weighted spectral feature maps and channel-weighted geometric feature maps. By concatenating the channel-weighted spectral feature map and the channel-weighted geometric feature map along the channel dimension, a cross-modal fusion feature map is obtained.

[0078] (3) The purpose of the dynamic pyramid pooling module: The cross-modal fusion feature map is subjected to four-scale pooling processing, and the pooling weight is dynamically allocated according to the crack density. That is, when the crack density is >0.6, the small-scale pooling weight accounts for 70%, which focuses on preserving the fine topographic details of cracks and gullies; when the crack density is <0.3, the large-scale pooling weight accounts for 60%, which focuses on capturing the global topographic correlation of the slope, and finally a multi-scale feature map is obtained. The multi-scale fusion feature map is then processed: first, the number of channels is compressed to 128 through 1×1 convolution to reduce the computational load, and then ESRGAN super-resolution technology is used to perform 4x upsampling to obtain the super-resolution feature map; finally, the super-resolution feature map is binary classified by the Softmax function to output a binary segmentation mask; where a pixel value of 1 represents a crack (width ≥ 1 mm) or a gully (depth ≥ 5 mm), and a pixel value of 0 represents a complete slope.

[0079] The improved PSPNet network constructed by the mask unit in this embodiment generates depth-assisted multispectral images by adding geometric features to pixels and optimizing parameters based on segmentation mask differences. It can effectively fuse spectral and geometric features, improve the segmentation accuracy of slope cracks and gullies, provide accurate binary segmentation masks for subsequent regional modeling, and enhance the model's ability to identify complex slope features. Furthermore, the dual-branch cross-modal feature fusion module fuses spectral and geometric feature maps through an attention mechanism, while the dynamic pyramid pooling module assigns weights based on crack density and generates a binary segmentation mask. This design not only makes full use of the advantages of multimodal data but also highlights key areas such as cracks through dynamic weight allocation, improving the targeting and accuracy of segmentation. The generated binary segmentation mask can accurately divide different areas of the slope, providing a reliable basis for subsequent differentiated processing.

[0080] The model unit is used to divide the finely registered data into point clouds of crack areas, gully areas, and intact areas based on the binary segmentation mask; The model unit is also used to perform densification and resampling processing on the point cloud of the crack area, preserving the subtle depth changes of the crack; to perform smoothing filtering on the point cloud of the gully area and preserve the steep edges of the target sidewall; and to perform downsampling processing on the point cloud of the intact area, retaining 1 / 5 of the original points. After processing, a comprehensive semantically enhanced point cloud is obtained, including the densified point cloud of the crack area, the denoised point cloud of the gully area, and the simplified point cloud of the intact area. The model unit is also used to encrypt the point cloud based on the crack region in the comprehensive semantically enhanced point cloud, and to generate the crack region DEM by preserving the depth discontinuity of the crack region through Poisson reconstruction and edge constraints. The model unit is also used to denoise the gully area point cloud based on the comprehensive semantically enhanced point cloud, and generate the gully area DEM through inverse distance weighted interpolation and hydrodynamic correction; The model unit is also used to simplify the point cloud based on the complete region in the comprehensive semantically enhanced point cloud, and generate the complete region DEM through kriging interpolation; The model unit is also used to splice the DEM of the cracked area, the DEM of the gully area and the DEM of the intact area according to spatial coordinates, and to eliminate the splicing seams using the edge transition algorithm to generate a DEM model covering the entire slope.

[0081] In this embodiment, the model unit performs differentiated processing on point clouds in different regions based on a binary segmentation mask, and then generates DEMs by region and stitches them together. This regional modeling strategy not only ensures the detailed representation of key areas such as cracks and gullies, but also improves the overall modeling efficiency through simplified processing. The generated DEM model can fully reflect the complex terrain features of the slope and provide a high-precision foundation for slope analysis.

[0082] Slope cells are used to calculate the initial slope for each grid cell in the DEM model using the Horn algorithm, thus obtaining the initial slope map. The slope unit is also used to lock the range of the crack area and extract the crack direction for the initial slope map by using the segmentation mask as a guide. It calculates the slope difference value on both sides of the crack that was missed due to the smoothing process in the previous model construction along the crack direction, and supplements the calculated data into the initial slope data of the corresponding grid so that the slope of the crack area can truly reflect its local steepness and the topographic undulation on both sides, and obtains the corrected slope data of the crack area. The slope unit is also used to lock the gully area range based on meteorological rainfall data. For the initial slope map, the segmentation mask is used as a guide. The initial slope value of the gully sidewall in the gully area range is adjusted by combining the cumulative rainfall of the target area in 15 days to obtain the corrected slope data of the gully area. Among them, the cumulative rainfall and the adjustment of the initial slope value of the gully sidewall are positively correlated. For example, for every 100mm increase in rainfall, the initial slope of the gully sidewall is increased by 2%, so as to compensate for the slope steepening deviation caused by rainwater erosion and make the slope of the gully area more consistent with the actual terrain. The slope unit is also used to directly retain the initial slope data of the complete area as a benchmark for subsequent calculations; where the complete area refers to the area in the initial slope map excluding the crack area and gully area. The slope unit is also used 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. It removes invalid data such as blank areas outside the slope in the initial slope map to obtain the corrected slope map, which only retains the "precise corrected slope of the crack area, the compensated slope of the gully area, and the baseline slope of the complete area". The slope unit is also used to calculate the slope change data ΔS based on the corrected slope map, according to the average slope S0 of the intact area and the average slope S1 of the cracked area and gully area, i.e., ΔS=S1-S0.

[0083] In this embodiment, the slope unit calculates the initial slope using the Horn algorithm, corrects the slope difference in the crack area by combining 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 takes into account the information loss caused by the smoothing of the crack area and the impact of rainfall on the gully, making the slope data more consistent with the actual slope conditions. The slope change data obtained by calculating the average slope by partition can more accurately reflect the key indicators of slope stability.

[0084] In one embodiment, the deformation module 30 specifically comprises: The time-series displacement data and meteorological precipitation data acquired during the multi-source data acquisition phase are integrated and processed. The integrated dataset is then standardized and a time-series window is constructed according to the rule of "predicting the next two periods based on the past five periods of data" to divide the dataset into training and validation sets.

[0085] A two-layer stacked LSTM structure was constructed, and the Adam optimizer and RMSE loss function were selected to train the model. The training was iterated until the relative error of the prediction on the validation set was less than 7% and the model accuracy was stable. Finally, the trained LSTM prediction model was obtained.

[0086] Input the latest five periods of displacement time series data and meteorological and rainfall data into the trained LSTM prediction model to predict the slope displacement in the next 1-2 months, and finally output the slope deformation trend.

[0087] In one embodiment, the early warning module 40 includes a vector unit, a calculation unit, and an early warning unit; Among them, the vector unit is used to obtain historical multi-source data, including historical slope change data and historical slope deformation trend, based on the above-mentioned DEM model and LSTM prediction model, as well as historical meteorological rainfall data for the same period. The vector unit is also used to mark areas with slope changes exceeding 1° in several consecutive grids as slip zones in historical slope change data, and to perform global average pooling on the areas marked as slip zones to compress the dimensions and obtain slope feature vectors. Vector elements are also used to analyze historical slope deformation trends, specifically displacement accelerations exceeding 0.5 mm / month. 2 The regions marked as acceleration risk zones are then subjected to convolutional compression to obtain deformed feature vectors. The vector unit is also used to transform historical meteorological rainfall data into a multi-dimensional rainfall impact index table, including cumulative rainfall index and heavy rainfall index. This table is then mapped to the feature space using a pre-defined two-layer fully connected layer. Subsequently, correction terms are added based on slope type. Considering that soil slopes are more sensitive to rainfall, each dimension of the initial feature vector is multiplied by 1.2, while the dimensions of rock slopes, which are less sensitive, are multiplied by 0.8. This adapts to the differences in rainfall sensitivity among different slopes, ultimately yielding the rainfall feature vector. The first layer of the two fully connected layers has 64 hidden units, and the second layer has 32 hidden units, both using the ReLU activation function.

[0088] In this embodiment, the vector unit marks the slip zone and the accelerated risk zone, and performs pooling and convolution processing on historical slope change data and deformation trends, respectively, mapping rainfall data into feature vectors. This feature transformation method accurately extracts the key features of slope stability, transforms unstructured data into structured feature vectors, retains important information while reducing data dimensionality, and lays a good foundation for cross-modal fusion computing.

[0089] The computational unit calculates the cosine similarity between the slope feature vector and the deformation feature vector, normalizes the cosine similarity to a spatial attention weight of 0-1, and performs pixel-level weighting on the slope feature vector and deformation feature vector according to the spatial attention weight, resulting in weighted slope feature vector and weighted deformation feature vector. Specifically, for the spatial attention weight, a weight of 1 corresponds to a slope change greater than 2° and an acceleration greater than 1 mm / month. 2 ”; The calculation unit is also used to splice the weighted slope feature vector, the weighted deformation feature vector, and the rainfall feature vector. It extracts the importance of each channel through global average pooling and assigns a weight of 2 times to channels whose rainfall impact index is greater than the preset rainfall amount, and calculates the fused feature vector. The calculation unit is also used to read the specific values ​​of slope change, deformation acceleration and rainfall impact index of the fused feature vector of several grids in the fused feature vector, and determine the level category of the read values ​​according to the preset stability level judgment rules to obtain the stability level of several grids.

[0090] The calculation unit is also used to input the "fusion feature vector and stability level" as a data set into the random forest model. Among them, the fusion feature vector is used as the core input of the model, and each vector corresponds to the comprehensive risk characteristics of a single slope grid, covering key information such as slope mutation, deformation acceleration, and strong rainfall coupling. Relying on multiple decision trees of the random forest algorithm, by learning the mapping rule between the fusion feature vector and the stability level, for example, automatically identifying the corresponding relationship between the feature combination of "high slope mutation + high deformation acceleration + strong rainfall" and the "instability risk" label, the trained slope stability evaluation model is finally obtained.

[0091] Among them, the "stability level determination rule" is divided into three categories: "stable, sub-stable, and instability risk", with the slope change value (A, unit: °), deformation acceleration (B, unit: mm / month 2 ) and rainfall impact index (C, unitless) as the core determination indicators. The specific rules are as follows: (1) Stable: simultaneously satisfying "A ≤ 0.5, B ≤ 0.3, C ≤ 20", or only a single index slightly exceeding the threshold (such as A = 0.8, B and C meeting the standards); (2) Sub-stable: simultaneously satisfying "0.5 < A ≤ 2, 0.3 < B ≤ 1", or "20 < C ≤ 50, and either A or B slightly exceeding the threshold"; (3) Instability risk: simultaneously satisfying "A > 2, B > 1", or "C > 50, A > 1, or B > 0.8".

[0092] In this embodiment, the vector unit and the calculation unit convert historical data into feature vectors, calculate the spatial attention weights through cosine similarity, allocate channel weights in combination with the rainfall impact index to generate fusion feature vectors, and mark the stability level according to the rules. This process effectively integrates the features of multi-source data through cross-modal attention fusion, highlights the key influencing factors, and the generated fusion feature vectors and stability levels provide high-quality training data for the slope stability evaluation model, improving the learning effect of the model.

[0093] The warning unit is used to input the slope change data calculated by the high-precision DEM model in the early stage, the slope deformation trend output by the LSTM prediction model, and the meteorological rainfall data monitored by the meteorological station during the same period into the trained slope stability evaluation model. And it should be clear that all three types of data are for the target area and are data for the same time period within the recent 1 - 3 consecutive monitoring cycles; The early warning unit is also used to first strengthen the coupled risk characteristics of "slope change, deformation acceleration and heavy rainfall" through cross-modal attention mechanism, and then rely on the multi-decision tree operation of the random forest algorithm to output the safety coefficient and corresponding stability level of each monitoring grid of the slope; if the safety coefficient of any monitoring grid is lower than the preset safety threshold of loess slope engineering, the early warning process is automatically triggered, and early warning information including the three-dimensional coordinates of the risk area, risk level and core risk factors is released in real time, providing timely decision-making basis for slope safety management; Among them, the stability level is divided into "stable, under-stable, and unstable risk", which can support hierarchical management and control, locate risk factors, and serve as an intuitive carrier of "risk level" in early warning information.

[0094] Overall, this embodiment has the following beneficial effects: This invention utilizes an improved PSPNet to segment crack and gully areas, accurately identifying key slope instability zones. A regionally differentiated DEM model is then generated, allowing slope variation data to specifically reflect the characteristics of different areas, rather than providing general data. This ensures accurate correlation with deformation trends and rainfall impacts. Furthermore, by generating a DEM model covering the entire region and outputting slope variation data, it avoids model building biases caused by slope data issues. The LSTM algorithm predicts slope deformation trends based on periodic displacement time-series data, accurately quantifying the temporal evolution characteristics of deformation and providing reliable predictive data for deformation factors. By learning the mapping relationship between fusion feature vectors and stability levels through the random forest algorithm, the established slope stability evaluation model quantifies the nonlinear correlation between slope, deformation, and meteorological rainfall. Inputting multi-factor data into the model to predict the safety factor, and considering the interaction of different factors, the prediction results more closely match the actual state, solving the multi-factor fusion problem, significantly reducing manual inspection costs, and improving early warning reliability. In summary, this invention addresses the core needs of loess slope stability monitoring, such as high costs and delayed early warning responses caused by manual inspections. It employs a multispectral camera from a drone and a 3D laser scanner to collect multi-source data. By comparing data from multiple periods, it accurately identifies and analyzes key deformation features such as slope cracks and gullies. Simultaneously, it enables automated acquisition and intelligent analysis of slope monitoring data, effectively reducing the manpower and time costs of manual inspections. Ultimately, in actual monitoring scenarios such as highway slopes, it significantly improves the accuracy of early warnings.

[0095] Example 3: This application provides a computer-readable storage medium, which includes a stored computer program, wherein the computer program controls the device where the computer-readable storage medium is located to execute the intelligent monitoring method for loess slope stability when it is running. The intelligent monitoring method for loess slope stability, when implemented as a software functional unit and used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0096] The above are preferred embodiments of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

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 the 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