A rock-soil body displacement identification method based on multi-temporal unmanned aerial vehicle images

Abstract

The present application relates to the field of multi-temporal unmanned aerial vehicle image, and particularly relates to a rock-soil mass displacement identification method based on multi-temporal unmanned aerial vehicle image. The content comprises: collecting multi-temporal image data of a target area and preprocessing, performing multi-scale filtering processing on the preprocessed multi-temporal image data, extracting multi-type image features, and constructing a final image feature vector of each phase; based on the final image feature vector, calculating an initial displacement amount; based on the initial displacement amount, constructing a space-time weighted model, and dynamically adjusting the space-time correlation weight; introducing a multi-level feedback mechanism, based on the space-time correlation weight, optimizing the initial displacement amount to obtain the final displacement identification result. The problems of single feature extraction dimension in the traditional rock-soil mass displacement identification method, fixed weight mode in space-time analysis, poor adaptability, simple displacement optimization mechanism, and insufficient monitoring accuracy and stability are solved.

Description

Technical Field

[0001] This invention relates to the field of multi-temporal UAV imagery, and more particularly to a method for identifying rock and soil displacement based on multi-temporal UAV imagery. Background Technology

[0002] Soil and rock displacement monitoring is a core component for ensuring construction and operational safety in fields such as mine slopes, highway subgrades, and foundation pit engineering, directly impacting project stability assessment and risk warning decisions. With the development of UAV (Unmanned Aerial Vehicle) technology and multispectral imaging, UAV multispectral imagery, due to its advantages of wide coverage, high timeliness, and non-contact monitoring, has gradually replaced traditional manual monitoring and total station monitoring methods, becoming the mainstream technology for soil and rock displacement monitoring and widely applied in dynamic monitoring across various engineering scenarios. Current monitoring solutions based on UAV imagery primarily involve acquiring multi-temporal images, combining image preprocessing and feature analysis to achieve displacement identification, providing data support for engineering safety management.

[0003] However, the above-mentioned methods for identifying soil and rock displacement still have significant shortcomings: First, the feature extraction dimension is too limited to fully capture the complex changes in soil and rock; second, the spatiotemporal analysis uses a fixed weight mode, which has poor adaptability; and third, the displacement optimization mechanism is too simple to cope with complex environmental interference. These problems result in insufficient monitoring accuracy and stability, making it difficult to fully meet the needs of engineering-level monitoring. Summary of the Invention

[0004] This invention provides a method for identifying soil and rock displacement based on multi-temporal UAV imagery, which solves the technical problems of traditional soil and rock displacement identification methods, such as single feature extraction dimension, fixed weight mode for spatiotemporal analysis with poor adaptability, and simple displacement optimization mechanism that is difficult to cope with complex environmental interference, resulting in insufficient monitoring accuracy and stability.

[0005] The present invention provides a method for identifying soil and rock displacement based on multi-temporal UAV imagery, comprising the following steps: S1. Multi-temporal image data of the target area is collected by a UAV equipped with a multispectral camera and preprocessed to obtain preprocessed multi-temporal image data; multi-scale Gaussian filtering is applied to the preprocessed multi-temporal image data to obtain Gaussian smoothed multi-temporal image data at different scales; based on the Gaussian smoothed multi-temporal image data at different scales, multiple types of image features are extracted, including texture features, spectral variation features, edge features, and morphological features, and the final image feature vector for each temporal phase is constructed; based on the final image feature vector, the change in the final image features of pixels in adjacent temporal phases is calculated to obtain the initial displacement. S2. Based on the initial displacement, quantify the overall correlation between displacements in different time phases, construct a time-weighted term, and combine it with spatial weighting to construct a spatiotemporal weighted model, dynamically adjusting the spatiotemporal correlation weights; introduce a multi-level feedback mechanism, optimize the initial displacement based on the spatiotemporal correlation weights, and obtain the final displacement identification result.

[0006] Preferably, S1 specifically includes: Image denoising, georegistration, geometric correction, and normalization are performed on multi-temporal image data to obtain preprocessed multi-temporal image data.

[0007] Preferably, S1 specifically includes: Multi-scale image features of the same type within each time phase are weighted and fused using an equal-weight method to obtain the final image features of the same type within each time phase. The same operation is performed on all feature types to generate the final image features of different types within the same time phase. The image feature values ​​of all feature types within the current time phase are concatenated to generate the final image feature vector for each time phase.

[0008] Preferably, S2 specifically includes: Based on the initial displacement, combined with the initial displacement of historical time phases, and introducing the time dynamic weight of historical time phases, the overall correlation of displacement between different time phases is quantified, and a time weighting term is constructed; the time dynamic weight of historical time phases is calculated by an exponential decay model based on time difference.

[0009] Preferably, S2 specifically includes: A spatial weighting term is constructed by introducing local image feature influence factors and combining them with the spatial distance between pixels.

[0010] Preferably, S2 specifically includes: A local window is introduced, and the variance of each type of final image feature within the local window is calculated and normalized to obtain the normalized local variance. Combined with the sensitivity score of the feature, the normalized local variance is weighted and summed to obtain the total variance. Based on the total variance, a local image feature influence factor is constructed.

[0011] Preferably, S2 specifically includes: Based on time-weighted and space-weighted terms, a spatiotemporal weighted model is constructed, with the specific formula as follows: ; in, Indicates phase pixel The spatiotemporal correlation weights; This represents the total number of temporal phases of the preprocessed multi-temporal image data involved in displacement identification calculations; It is an index of historical phases; , These are the current time phases. and historical phases The average displacement at time; For displacement-related terms, representing the time phase and historical phases The overall correlation of displacements between them; For the first The dynamic weight of each historical phase; It is a correlation index; Represents pixels In time phase The initial displacement; Represents pixels In historical phases The initial displacement; Time-weighted terms; Represents pixels The spatial distance between the center point of the preprocessed multi-temporal image data used in displacement identification calculation; For pixels In time phase Local image feature influence factors; The maximum spatial distance threshold; For spatial weighting terms.

[0012] Preferably, S2 specifically includes: The multi-level feedback mechanism, combining multi-level iteration and nonlinear threshold screening, optimizes the initial displacement.

[0013] Preferably, S2 specifically includes: In the implementation of the multi-level feedback mechanism, the initial displacement is iteratively corrected in multiple levels based on the spatiotemporal correlation weight: at each level, the offset degree of the displacement output by the previous level is calculated, and the single-level correction amount of the current level is generated by combining the level weight and the response steepness coefficient. The single-level correction amount of the current level is then superimposed on the displacement output by the previous level to obtain the displacement output of the current level. The displacement output of the final level is used as the final displacement recognition result.

[0014] The beneficial effects of the technical solution of the present invention are: 1. By combining multi-scale processing with multi-type image feature fusion, the limitations of traditional single features are overcome, effectively capturing multi-dimensional change information of soil and rock, and significantly reducing interference from false features.

[0015] 2. By dynamically allocating the weights of the final image features, the contribution of displacement-sensitive features is highlighted, irrelevant interference is reduced, and a high-precision foundation is laid for subsequent analysis.

[0016] 3. By constructing a spatiotemporal weighted model, we can dynamically adapt to spatiotemporal relationships, accurately identify key change areas, and avoid the omission problem of traditional fixed weight model.

[0017] 4. By introducing a multi-level feedback mechanism, displacement information is iteratively optimized, effectively suppressing false and abnormal displacements, and maintaining a stable and reliable recognition effect even in complex environments.

[0018] 5. The entire process is highly automated, adaptable to different monitoring scenarios and accuracy requirements, significantly improving monitoring efficiency, and the output results can directly meet the quantitative requirements of engineering decision-making. Attached Figure Description

[0019] Figure 1 This is a flowchart of a method for identifying rock and soil displacement based on multi-temporal UAV imagery, as described in this invention. Detailed Implementation

[0020] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0022] The following description, in conjunction with the accompanying drawings, details a specific scheme for a method for identifying rock and soil displacement based on multi-temporal UAV imagery provided by the present invention.

[0023] See attached document Figure 1 The diagram illustrates a flowchart of a method for identifying soil and rock displacement based on multi-temporal UAV imagery, provided by an embodiment of the present invention. The method includes the following steps: S1. Multi-temporal image data of the target area is collected by a UAV equipped with a multispectral camera and preprocessed to obtain preprocessed multi-temporal image data; multi-scale Gaussian filtering is applied to the preprocessed multi-temporal image data to obtain Gaussian smoothed multi-temporal image data at different scales; based on the Gaussian smoothed multi-temporal image data at different scales, multiple types of image features are extracted, including texture features, spectral variation features, edge features, and morphological features, and the final image feature vector for each temporal phase is constructed; based on the final image feature vector, the change in the final image features of pixels in adjacent temporal phases is calculated to obtain the initial displacement. First, using a drone equipped with a multispectral camera adapted to different precision requirements, multi-temporal image data of the target area is collected. Then, four standardized preprocessing operations—image denoising, georegistration, geometric correction, and normalization—are performed on the multi-temporal image data. The specific implementation process is as follows: Based on the multi-temporal image data, bilateral filtering is used for image denoising to obtain denoised multi-temporal image data. Considering the need to preserve weakly changing features such as edges and textures for soil and rock displacement identification, the ground resolution GSD (Ground Surface Depth, the actual ground length corresponding to each pixel) of the multi-temporal image data is used as a benchmark, and the filter window size is determined to be 5×5 to balance noise suppression and detail preservation. The spatial domain standard deviation is set to 1 / 2 to 3 / 4 of the filter window radius. When the wavelet window size is 5×5, its window radius is 2 pixels, and the corresponding spatial domain standard deviation is between 1.0 and 1.5. The grayscale value domain standard deviation is determined based on the statistical results of the standard deviation of the absolute values ​​of the grayscale differences of all adjacent pixels in the multi-temporal image data, and controlled between 15 and 25 to achieve random noise suppression while preserving edge abrupt change information. Subsequently, based on the denoised multi-temporal image data, georegistration is performed using an affine transformation method. The denoised multi-temporal image data of the first temporal phase is used as the reference image. Key feature points are extracted from the denoised multi-temporal image data of each temporal phase using the SIFT algorithm, and then the FLANN matching algorithm is used to quickly match the key feature points extracted from different temporal phases. Among them, Lowe's ratio The threshold for the Lochte ratio test was set to 0.7. After matching, the RANSAC algorithm was used to verify the geometric consistency of the matched key feature point pairs: the distance between the theoretical and actual positions of the matched points after affine transformation was calculated, and mismatched key feature points with a deviation exceeding 3 pixels were removed, finally obtaining the matched point pairs. Based on the matched point pairs, the 6-parameter affine transformation matrix of each denoised multi-temporal image data relative to the reference image was solved using the least squares method. Then, the 6-parameter affine transformation matrix was applied to the denoised multi-temporal image data, and bilinear interpolation was used for resampling to denoise other temporal phases. The multi-temporal image data is then mapped onto the reference image coordinate system to obtain registered and aligned multi-temporal image data. Next, based on the registered and aligned multi-temporal image data, a template matching iterative residual correction method based on normalized cross-correlation is adopted. Grid control points (spaced 20 pixels apart) are uniformly selected on the reference image. Within an 11×11 pixel search window centered on the corresponding position of each non-reference image in each temporal phase, the sub-pixel offset of each grid control point is calculated using the template matching method based on normalized cross-correlation. A local offset field is then constructed using bilinear interpolation to perform local correction on the image. This process is iterated and repeated until the residual RMSE ≤ 0.Three pixels or up to five iterations are used to eliminate local geometric distortions, resulting in geometrically corrected multi-temporal image data. Finally, based on the geometrically corrected multi-temporal image data, a max-min linear mapping method is used for grayscale normalization, mapping pixel values ​​uniformly to the [0,1] interval, resulting in preprocessed multi-temporal image data. The multi-temporal image data and the preprocessed multi-temporal image data are stored in the UAV image database. Furthermore, multi-scale Gaussian filtering is performed on the preprocessed multi-temporal image data to obtain Gaussian-smoothed multi-temporal image data at different scales. The specific processing steps are as follows: First, multi-scale Gaussian filtering is performed independently on the preprocessed image data for each temporal phase. Each scale corresponds to a Gaussian filtering window with a standard deviation, and the size of the Gaussian filtering window is based on the standard deviation. Determined, usually the window radius is taken. The Gaussian filter window size is Pixels, standard deviation Based on the ground resolution (GSD) of multi-temporal image data and the spatial scale of soil and rock displacement characteristics, values ​​of 0.5, 1, 2, and 4 pixels can be selected, corresponding to a smoothing effect from fine texture to large-scale structure; secondly, for each standard deviation... The corresponding Gaussian kernel weight matrix is ​​generated (calculated using a standard two-dimensional Gaussian function and normalized), and then convolved with the preprocessed image data of each time phase. Mirror padding or zero padding is applied to the boundary pixels of the preprocessed image data for each time phase to ensure that the Gaussian filter window completely covers each pixel. Finally, multiple images with different smoothing levels are output for each time phase, with each image corresponding to a... The values ​​were used to obtain multi-temporal image data after Gaussian smoothing at different scales.

[0024] Based on Gaussian-smoothed multi-temporal image data at different scales, image features are extracted sequentially through Gray-Level Co-occurrence Matrix (GLCM), spectral reflectance difference analysis, Canny edge detection algorithm, and morphological operations such as erosion, dilation, opening, closing, and morphological gradient. These features include four types: texture features, spectral variation features, edge features, and morphological features. Specifically, the gray values ​​of the Gaussian-smoothed multi-temporal image data at different scales are quantized to 16 levels. Then, a 15×15 pixel sliding window (step size 1 pixel) is moved pixel by pixel. Within each sliding window, the Gray-Level Co-occurrence Matrix (GLCM) is calculated, considering four directions: 0°, 45°, 90°, and 135° (pixel spacing of 1). Contrast, energy, correlation, and entropy are extracted. The texture features are analyzed, and the final texture feature of the center pixel of the window is obtained by averaging the four directions. This is used to reflect the texture changes on the surface of the soil and rock. A spectral reflectance difference analysis method is used to calculate the difference in gray values ​​of the same pixel position in multi-temporal image data after Gaussian smoothing at different time phases and scales. By statistically analyzing the change amplitude of gray values ​​of corresponding pixels in adjacent time phases, spectral change features are obtained to reflect the state changes of the soil and rock surface. The Canny edge detection algorithm is used to extract edges from the Gaussian smoothed image data at different time phases and scales, resulting in binary edge images. Edge features such as edge pixel density, edge length, and edge direction distribution are statistically analyzed based on these binary edge images. The parameters of the Canny edge detection algorithm are set as follows: gradient calculation uses 3×3... The Sobel operator uses the mean of the grayscale gradient magnitude of the entire image for low thresholding and twice the mean for high thresholding. Morphological processing, including erosion, dilation, opening, closing, and morphological gradient operations, is applied to the binary edge image to remove small-area noise, connect broken edges, and optimize the regional contour structure, resulting in a morphologically processed edge image. Based on this morphologically processed edge image, morphological features such as region connectivity, contour change rate, and boundary gradient are statistically analyzed. Finally, multi-scale image features of the same type within each time phase are weighted and fused using an equal-weight method to obtain the final image features of the same type within each time phase. The same operation is performed on all feature types to generate different types of final image features for the same time phase. For each pixel... and each phase The feature values ​​of all feature types within the current time phase are concatenated to generate the final image feature vector for each pixel and each time phase. ,in, For the temporal index during multi-temporal image data acquisition; and These are the row and column indices of pixels in the preprocessed multi-temporal image data, representing the two-dimensional spatial coordinates of the preprocessed multi-temporal image data. Indicates the row where the pixel is located. Indicates the column where the pixel is located.

[0025] Furthermore, the change in the final image features of each pixel in adjacent time phases is calculated and used as the initial displacement. The specific formula is as follows: ; in, Represents pixels In time phase The initial displacement; The dimension of the final image feature vector; It is the index of the dimension of the final image feature vector; Indicates phase Pixels in preprocessed multi-temporal image data The A final image feature, Indicates phase Pixels in preprocessed multi-temporal image data The A final image feature, and All are derived from the final image feature vector ; Indicates the first The normalized weights of the final image features reflect the... The contribution of each final image feature to the displacement is determined by calculating the variance of all final image features and then applying the variance of the first final image feature to the second final image feature. The variance of each final image feature is obtained by max-min normalization, with a value range of [0,1].

[0026] S2. Based on the initial displacement, quantify the overall correlation between displacements in different time phases, construct a time-weighted term, and combine it with spatial weighting to construct a spatiotemporal weighted model, dynamically adjusting the spatiotemporal correlation weights; introduce a multi-level feedback mechanism, optimize the initial displacement based on the spatiotemporal correlation weights, and obtain the final displacement identification result.

[0027] Based on the initial displacement By combining spatiotemporal context constraint theory and dynamic weighted fusion principle, a spatiotemporal weighted model is constructed. Based on the characteristics that recent phase data is more valuable temporally and that adjacent pixel displacements have stronger correlations spatially, the spatiotemporal correlation weights are dynamically adjusted to ensure more accurate displacement recognition. Especially when processing multi-temporal data, the spatiotemporal weighted model can adaptively increase or decrease the spatiotemporal correlation weights of data from different time periods to highlight key areas of change. The formula for the spatiotemporal weighted model is as follows: ; in, Indicates phase pixel The spatiotemporal correlation weights; This represents the total number of time phases of the preprocessed multi-temporal image data involved in displacement identification calculation. It is determined based on the theoretical design requirements of the monitoring cycle and monitoring frequency in the geotechnical engineering monitoring specifications, the specific monitoring plan of the project, and the frequency of UAV image acquisition. The value range is a positive real number. It is an index of historical phases; , These are the current time phases. and historical phases The average displacement over time is obtained by averaging the initial displacement of all pixels in the preprocessed multi-temporal image data that are involved in displacement recognition calculation. For displacement-related terms, representing the time phase and historical phases The overall correlation of displacements between them is calculated by the current time phase. and historical phases The Pearson correlation coefficient of the initial displacement is obtained, with a value range of [0,1]. For the first The time dynamic weights of each historical phase are used to represent the previous... The temporal dynamic importance of each historical phase in the displacement identification process is calculated using an exponential decay model based on time difference, as shown in the formula: , The decay factor is determined using the least squares method: an exponential decay function is used as the fitting model, with time difference sequences from different time phases as the basis. Using the corresponding displacement-related terms as the dependent variable, the attenuation factor is determined, with a value range of [0.01, 0.5]. For the current time phase The corresponding collection time, For the first Each historical time corresponds to a specific data collection time, which is recorded using a date and timestamp (with hours, minutes, and seconds) and converted into a numerical time difference in days during calculation. The range of values ​​for is (0,1]; The correlation index is used to control the importance of displacement correlation terms in spatiotemporal weighting, reflecting the degree of influence of displacement correlation on the final displacement identification result during the weighting process. Based on different actual scenarios (such as mining areas) and preprocessed historical multi-temporal image data, a random forest feature importance analysis model is used to analyze texture features, spectral variation features, edge features, and morphological features. The specific implementation process is as follows: the final image features of each pixel (including texture features, spectral variation features, edge features, and morphological features) are used as input, and the measured displacement value is used as the supervision label to construct training samples. Then, a random forest feature importance analysis model is constructed, with key parameters set as follows. The decision trees are set to 100, the maximum tree depth to 10, the minimum number of sample splits to 2, and the minimum number of samples per leaf node to 1. A random forest feature importance analysis model is trained, and the mean reduction in impurity (MDI) value of each final image feature is calculated as a sensitivity score for each feature to the displacement recognition result. The sensitivity scores of each feature are then summed with equal weights to obtain the initial correlation index. A five-fold cross-validation method is then used to divide the preprocessed historical multi-temporal image data into five groups. Four groups are selected as the training set and one group as the validation set each time. Within the pre-defined candidate interval [0,5] of the correlation index, which is based on engineering experience in soil and rock displacement monitoring and the requirement for displacement continuity, a grid search is used to traverse the data. (Step size 0.1), with the goal of minimizing the average verification error, determine the final correlation index; Represents pixels In historical phases The initial displacement; The time-weighted term enhances the dependence on time periods with strong correlations in displacement by weighting the displacement-related terms at different time phases. Represents pixels The spatial distance between the pixel and the center point of the preprocessed multi-temporal image data used in the displacement recognition calculation is used to measure the relative distance in spatial location. The farther the pixel is from the center point, the smaller its contribution to the final displacement recognition result. , where pixels It is the geometric center pixel of the preprocessed multi-temporal image data that participates in displacement recognition calculation. For the ground resolution of the image; For pixels In time phase The local image feature influence factor reflects the impact of local image features such as texture features and spectral variation features on displacement recognition. The calculation method is as follows: ,in, It is the total variance, which is calculated in pixels. Centered on the image, a 5×5 neighborhood is taken as the local window. The variance of all feature values ​​of each type of final image feature (texture feature, spectral variation feature, edge feature, morphological feature) within the local window is calculated to obtain the local variance. Then, the local variance is normalized by max-min to obtain the normalized local variance. The sensitivity score of each feature is used as the weight of each type of feature. The normalized local variance is then summed by weight to obtain the total variance. The maximum spatial distance threshold represents the maximum influence range of spatial distance in the spatiotemporal weighted model calculation. It prevents the influence of excessively far pixels by limiting the maximum distance between pixels. The value is determined based on the image size and the actual scene. For example, if a single aerial image of a region actually covers a rectangular area with a side length of 800m and the ground resolution of the image is 0.2m / pixel, the maximum spatial distance threshold is determined to be 120m, taking into account the actual deformation influence range of the mining area slope. The spatial weighting term adjusts the contribution of displacement data at different spatial distances to the recognition results by considering the spatial distance between pixels and local image features.

[0028] Next, based on the theory of multi-level iterative optimization and the principle of nonlinear threshold screening, and based on the initial displacement and spatiotemporal correlation weights, a multi-level feedback mechanism is introduced. This mechanism adjusts and optimizes the displacement information of each pixel multiple times, strengthening effective displacement information and suppressing interference to improve the accuracy of displacement recognition. The specific formula for the multi-level feedback mechanism is as follows: ; ; ; in, Indicates phase pixel The final displacement identification result after optimization by a multi-level feedback mechanism; This indicates the number of layers in the multi-level feedback mechanism. The feedback adjustment of each layer is based on the output of the previous layer to gradually optimize the displacement identification results. The initial number of candidate layers is preset according to the monitoring accuracy requirements and the complexity of the soil and rock displacement, and then substituted into the multi-level feedback mechanism. The layer with the smallest error is selected by comparing with the actual value on the ground. A hierarchical index for a multi-level feedback mechanism; It is the first Layer weights, i.e., hierarchical weights. It is the first The steepness coefficient of the layer's response, Initial values ​​are set based on different monitoring needs and the characteristics of soil and rock displacement, such as the first layer. Then, the gradient descent method is used for optimization. The gradient descent method uses the mean squared error (MSE) as the loss function. The training data consists of the final image features extracted from the preprocessed historical multi-temporal images and the corresponding measured displacement values. The learning rate is set to 0.01, and the number of iterations is 500. Finally, the final value is obtained to control the feedback intensity of the multi-level feedback mechanism in order to optimize the effect of the multi-level feedback mechanism. The value range is [0,1]. This is the displacement threshold, determined according to the engineering safety standards and monitoring accuracy requirements for soil and rock mass monitoring. The value range is [0.5, 2], used to screen areas of significant change. As a specific example, the value here can be 1.0. For the time phase pixel In the Displacement output of layer The difference between the value and the displacement threshold is used to reflect the degree of deviation of the displacement of the pixel. Here, is the denominator term of the Sigmoid function, where It can reflect the adjustment range of the Sigmoid function to the degree of displacement deviation, and can more accurately distinguish between displacements slightly exceeding the threshold and disturbances slightly below the threshold. For the first The single-level correction amount is fed back by a multi-level feedback mechanism.

[0029] Through a multi-level feedback mechanism, for each pixel The displacement recognition results were further optimized to reflect the actual displacement of each pixel in the soil and rock mass, providing quantitative data support for engineering safety assessment.

[0030] In summary, a method for identifying rock and soil displacement based on multi-temporal UAV imagery has been developed.

[0031] The order of the embodiments is for illustrative purposes only and does not represent the superiority or inferiority of the embodiments. The processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0032] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0033] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for identifying soil and rock displacement based on multi-temporal UAV imagery, characterized in that, Includes the following steps: S1. Collect multi-temporal image data of the target area using a drone equipped with a multispectral camera, and perform preprocessing to obtain preprocessed multi-temporal image data; Multi-scale Gaussian filtering is applied to the preprocessed multi-temporal image data to obtain Gaussian-smoothed multi-temporal image data at different scales. Based on the Gaussian-smoothed multi-temporal image data at different scales, multiple types of image features are extracted, including texture features, spectral variation features, edge features, and morphological features, and the final image feature vector for each temporal phase is constructed. Based on the final image feature vector, the change in the final image features of a pixel in adjacent temporal phases is calculated, and the weighted sum of the changes in the final image features of a pixel in all dimensions of adjacent temporal phases is obtained to obtain the initial displacement. S2. Based on the initial displacement, quantify the overall correlation of displacement between different time phases, construct a time-weighted term, and combine it with spatial weighting to construct a spatiotemporal weighted model, dynamically adjusting the spatiotemporal correlation weights. The specific formula is as follows: ； in, Indicates phase pixel The spatiotemporal correlation weights; This represents the total number of temporal phases of the preprocessed multi-temporal image data involved in displacement identification calculations; It is an index of historical phases; , These are the current time phases. and historical phases The average displacement at time; For displacement-related terms, representing the time phase and historical phases The overall correlation of displacements between them; For the first The dynamic weight of each historical phase; It is a correlation index; Represents pixels In time phase The initial displacement; Represents pixels In historical phases The initial displacement; Time-weighted terms; Represents pixels The spatial distance between the center point of the preprocessed multi-temporal image data used in displacement identification calculation; For pixels In time phase Local image feature influence factors; The maximum spatial distance threshold; Spatial weighting terms; A multi-level feedback mechanism is introduced, based on spatiotemporal correlation weights, combined with multi-level iteration and nonlinear threshold screening, to correct the initial displacement: at each level, the offset degree of the displacement output by the previous level is calculated, and combined with the level weight and response steepness coefficient, a single-level correction amount for the current level is generated, and the single-level correction amount of the current level is superimposed on the displacement output by the previous level to obtain the displacement output of the current level; the displacement output of the final level is used as the final displacement recognition result.

2. The method for identifying rock and soil displacement based on multi-temporal UAV imagery according to claim 1, characterized in that, S1 specifically includes: Image denoising, georegistration, geometric correction, and normalization are performed on multi-temporal image data to obtain preprocessed multi-temporal image data.

3. The method for identifying rock and soil displacement based on multi-temporal UAV imagery according to claim 1, characterized in that, S1 specifically includes: Multi-scale image features of the same type within each time phase are weighted and fused using an equal-weight method to obtain the final image features of the same type within each time phase. The same operation is performed on all feature types to generate the final image features of different types within the same time phase. The image feature values ​​of all feature types within the current time phase are concatenated to generate the final image feature vector for each time phase.

4. The method for identifying soil and rock displacement based on multi-temporal UAV imagery according to claim 1, characterized in that, S2 specifically includes: Based on the initial displacement, combined with the initial displacement of historical time phases, and introducing the time dynamic weight of historical time phases, the overall correlation of displacement between different time phases is quantified, and a time weighting term is constructed; the time dynamic weight of historical time phases is calculated by an exponential decay model based on time difference.

5. The method for identifying soil and rock displacement based on multi-temporal UAV imagery according to claim 1, characterized in that, S2 specifically includes: A local window is introduced, and the variance of each type of final image feature within the local window is calculated and normalized to obtain the normalized local variance. Combined with the sensitivity score of the feature, the normalized local variance is weighted and summed to obtain the total variance. Based on the total variance, a local image feature influence factor is constructed.

Images