Methods and devices for slope disease identification based on UAV and machine vision technologies

Abstract

A method and apparatus for slope disease identification based on UAV and machine vision technology includes a UAV equipped with a high-resolution camera that acquires multi-angle images of a slope area along a preset flight path; preprocessing the images, extracting features, and enhancing the images; identifying and classifying disease features using an improved YOLOv8-seg instance segmentation algorithm, covering slope diseases such as cracks, landslides, and collapses; quantifying the risk level of slope diseases based on the identification results using rule models or multi-factor analysis methods; and using a Long Short-Term Memory (LSTM) network to dynamically predict the changes in the acquired time-series images of the diseases, predicting the development trend and potential instability time. Based on the image data acquired by the UAV, real-time monitoring and dynamic analysis of slope diseases are performed, enabling timely detection of potential hazard areas and accurate identification of disease features, avoiding subjective errors that may be caused by human judgment in traditional methods.

Description

Technical Field

[0001] This invention relates to the fields of drones, computer vision, and other technologies, and in particular to a method and apparatus for identifying slope defects based on drone and machine vision technologies. Background Technology

[0002] Currently, slope damage monitoring methods mainly rely on manual inspection, fixed monitoring equipment (such as inclinometers and crack gauges), and satellite remote sensing technology. However, manual inspection is limited by human resources, inefficient, and poses safety hazards in dangerous areas; fixed monitoring equipment is complex to install and has limited coverage, making it difficult to cope with large-area or dynamically changing slope areas; satellite remote sensing, while having a wide coverage, has low resolution, making it difficult to capture subtle damage features, and is also limited by weather and imaging cycles, making real-time monitoring difficult. These traditional methods are insufficient to meet the needs of efficient, accurate, and dynamic monitoring in complex and variable slope environments. With the rapid development of UAV technology, its high flexibility, high-resolution image acquisition capabilities, and low cost have shown great potential in geological disaster monitoring. Combined with machine vision technology, image processing, and deep learning algorithms can automatically identify slope damage features such as cracks, landslides, and collapses, achieving high-precision detection and classification of damage. Simultaneously, based on the analysis of time-series image data, the evolution trend of slope damage can be dynamically tracked, providing an efficient and intelligent solution for risk level assessment and disaster early warning. This combination of drones and machine vision technology is expected to overcome the limitations of traditional monitoring methods and significantly improve the accuracy and efficiency of slope disease monitoring. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention discloses a method for slope disease identification based on unmanned aerial vehicle (UAV) and machine vision technology, the technical solution of which is as follows:

[0004] S1, a drone equipped with a high-resolution camera, collects multi-angle images of the slope area along a preset flight path;

[0005] S2 employs various methods such as denoising, correction, and stitching to preprocess the image, and performs feature extraction and image enhancement.

[0006] S3 combines the improved YOLOv8-seg instance segmentation algorithm to identify and classify disease features, covering common slope diseases such as cracks, landslides, and collapses;

[0007] S4. Based on the identification results, use rule models or multi-factor analysis methods (such as characteristics of disease scale, location, morphology, etc.) to quantify the risk level of slope disease;

[0008] S5 uses a Long Short-Term Memory (LSTM) network to dynamically predict changes in the acquired disease time-series images, predicting the disease's development trend and potential instability time.

[0009] This invention also discloses a device for slope disease identification based on UAV and machine vision technology, comprising:

[0010] The drone platform module is used to carry GPS modules, inertial navigation systems (INS), attitude sensors and other devices to support autonomous flight path planning and stable flight. It also carries high-resolution cameras for image acquisition and wireless communication technologies (such as 5G, Wi-Fi or 4G) to transmit image data.

[0011] The image processing module is used to perform image preprocessing on high-resolution images of the slope area captured by the drone camera, such as noise reduction, image enhancement, image correction, and image stitching, and uses a high-capacity storage device to retain the original and processed data.

[0012] The disease identification module extracts key features of diseases such as cracks, landslides, and collapses. It uses an improved YOLOv8-seg instance segmentation algorithm to classify disease types and locate disease areas. Simultaneously, it assesses the risk level of diseases based on disease features and slope topography data, providing a visual assessment report.

[0013] The time-series analysis module generates point cloud data through multi-view image matching, constructs a high-precision digital elevation model of the slope, marks the location and morphology of disease, and is used for dynamic monitoring. Based on the time-series data of disease, it analyzes the dynamic changes of slope characteristics and predicts the evolution trend of disease using an LSTM time-series model.

[0014] The present invention also discloses a non-volatile storage medium, characterized in that the non-volatile storage medium includes a stored program, wherein the program, when running, controls the device where the non-volatile storage medium is located to execute the above-mentioned method for slope disease identification based on UAV and machine vision technology.

[0015] The present invention also discloses an electronic device for slope disease identification based on UAV and machine vision technology, characterized in that it includes a processor and a memory; the memory stores computer-readable instructions, and the processor is used to run the computer-readable instructions, wherein the computer-readable instructions execute the above-described method for slope disease identification based on UAV and machine vision technology.

[0016] Beneficial effects

[0017] (1) Based on the improved YOLOv8-seg model, the UAV images of slopes can be segmented to realize the automated detection and classification of slope diseases;

[0018] (2) By using time-series image data obtained by UAVs and combining it with the LSTM time-series analysis algorithm, the changing trend of slope diseases can be dynamically monitored and the potential development direction and severity of diseases can be predicted. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the slope disease identification process of the present invention;

[0020] Figure 2 This is a diagram of the SPPF structure in YOLOv8-seg of the present invention;

[0021] Figure 3 This is a diagram of the SKAPPF structure in the improved YOLOv8-seg of this invention;

[0022] Figure 4 This is a structural diagram of the LSTM timing analysis algorithm model in this invention;

[0023] Figure 5 This is a schematic diagram of the slope disease identification device based on UAV and machine vision technology in this invention. Detailed Implementation

[0024] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. These embodiments are provided so that the present invention can be more thoroughly and completely conveyed to those skilled in the art.

[0025] The first aspect of this invention provides a method for slope disease identification based on unmanned aerial vehicle (UAV) and machine vision technology, as illustrated in the schematic diagram of an embodiment of the method. Figure 1 As shown, it includes the following steps:

[0026] S1, a drone equipped with a high-resolution camera, collects multi-angle images of the slope area along a preset flight path;

[0027] Specifically, based on the slope's terrain features and obstacle distribution, a reasonable flight path is designed, employing autonomous flight or manual control modes, combined with waypoint navigation, fixed-altitude cruise, and multi-angle shooting strategies to ensure coverage of the entire target area. Simultaneously, camera parameters such as resolution, shutter speed, and ISO sensitivity are adjusted according to the lighting conditions and environmental characteristics of the slope to ensure clear image capture.

[0028] Specifically, by adjusting the drone's attitude, such as tilt angle, heading, and camera pitch angle, the slope area can be photographed from multiple perspectives (directly above, side, slope, etc.) to obtain comprehensive image information, including areas with concentrated potential defects and areas with weak structures.

[0029] Furthermore, equip the drone with a high-speed memory card of sufficient capacity to record high-resolution image data. Use a wireless communication module (such as 5G or Wi-Fi) to transmit the collected data to a ground station or cloud server in real time for subsequent processing.

[0030] S2 employs various methods such as denoising, correction, and stitching to preprocess the image, and performs feature extraction and image enhancement.

[0031] Specifically, Gaussian filtering or median filtering is used to remove noise caused by wind interference in UAV imagery, and grainy noise is eliminated in low-light areas to highlight slope texture details and improve image quality. Geometric correction methods are used to correct barrel or pincushion distortion caused by the UAV lens, and orthographic projection correction is used to present the slope area in true proportion. At the same time, slope images taken from multiple angles are stitched into a panoramic image through feature point matching to make the slope area complete. Brightness equalization and seam smoothing techniques are used to solve the problems of uneven lighting or color difference in the stitched image.

[0032] Specifically, the formula used in the Gaussian filtering method is as follows:

[0033]

[0034] Where G(x,y) is the filter weight, x,y are the pixel offsets, and σ is the standard deviation of the Gaussian kernel, used to control the smoothness.

[0035] The formula used for image correction is:

[0036]

[0037] Where x, y are the x and y coordinates of the original image, x', y' are the x and y coordinates of the transformed image, H is a 3×3 perspective transformation matrix, and s is a scaling factor.

[0038] Furthermore, histogram equalization is used to enhance image contrast, making slope details more prominent, especially cracks and signs of landslides. For low-light environments, brightness and Gamma values ​​can be adjusted to ensure clear detail in all areas of the image.

[0039] Furthermore, the processed images are uniformly cropped and normalized to ensure consistency and stability in subsequent machine vision algorithm processing. The processed images are then stored in a standard format (such as GeoTIFF) to preserve geographic information.

[0040] S3 combines the improved YOLOv8-seg instance segmentation algorithm to identify and classify disease features, covering common slope diseases such as cracks, landslides, and collapses;

[0041] Specifically, based on the improved YOLOv8-seg instance segmentation algorithm, the algorithm accurately identifies and classifies slope defects using processed UAV multi-angle imagery. By optimizing the network structure, the detection accuracy of defects such as cracks, landslides, and collapses is improved, and an attention mechanism is incorporated to enhance the ability to capture defects in small areas. The algorithm can output bounding boxes, category labels, and masks for defect areas, automatically extract defect features such as crack length and landslide area, and visualize them on the imagery.

[0042] Specifically, the slope images collected by drones were compiled into a dataset, which was then divided into training, testing, and validation sets. The labelMe software was used to label areas with cracks, landslides, and collapses in the images. The labels for this training set represent the category and coordinates of each target in the image.

[0043] Specifically, the YOLOv8-seg model consists of three main parts: spine, neck, and head. The spine part, as a feature extraction network, integrates concepts from CSPDarkNet. The YOLOv8 backbone network replaces the C3 module of YOLOv5 with the C2f module, enhancing gradient flow through additional branches. The SPPF (Spatial Pyramid Pooling-Fast) module, such as... Figure 2 As shown in the diagram, MaxPool2d represents the pooling layer, and CBS stands for Cross-Stage Partial Network, whose main function is to introduce cross-stage partial connections into the YOLOv8 network structure. The SPPF module combines three consecutive pooling layers, combining the outputs of each layer to ensure multi-scale fusion, dynamically converting feature maps of different sizes to a fixed size, thus facilitating the simultaneous fusion of local and global feature information.

[0044] Specifically, the SKAPPF (Separable Kernel Attention Pyramid) module replaces the SPPF module in the original YOLOv8-seg model, addressing the shortcomings of the SPPF module in detail feature extraction and improving the original model. SKAPPF integrates various aspects of SPPF and LSKA, such as... Figure 3 As shown, k is the maximum receptive field, and d is the dilatation rate. The LSKA block represents the Hadamard product, where DW-D-Conv represents the residual network and Conv represents the convolutional block. As shown in the figure, the LSKA block employs a residual design with four deep convolutional blocks, enhanced by pointwise convolutional blocks equipped with a single-size kernel designed to linearly combine the output channels from the deep convolution. LSKA decomposes the 2D convolutional kernels of deep convolutional layers into cascaded horizontal and vertical 1D kernels. This decomposition allows direct use of the large kernels of deep convolutional layers within the attention module without requiring any additional blocks, reducing computational complexity and memory footprint while improving the model's adaptability to different kernel sizes. Compared to the original SPPF module, the SKAPPF module enhances feature fusion capabilities by fully utilizing and reusing multi-scale features, thereby improving the processing of fine-grained feature information.

[0045] Specifically, the SKAPPF module utilizes feature maps at different scales, drawn from various layers of the YOLOv8 network. By introducing a self-awareness mechanism, it evaluates the importance of each scale's feature map, enabling the model to dynamically adjust the contribution of feature maps at different scales. This preserves more useful information, particularly in learning object edges and texture details. Next, a progressive fusion strategy is employed, fusing low-resolution and high-resolution features with specific weights to fully leverage the strengths of each layer and reduce the possibility of feature information loss. Through this progressively enhanced information fusion, the model can generate more accurate instance segmentation results, demonstrating significant advantages, especially in handling complex scenes.

[0046] Furthermore, by using transfer learning techniques to train the dataset with the improved YOLOv8-seg model, a weight file is obtained to identify and classify slope disease areas, while extracting disease features such as crack length and landslide area.

[0047] S4. Based on the identification results, use multi-factor analysis methods (such as disease scale, location, morphology and other characteristics) to quantify the risk level of slope diseases;

[0048] Specifically, the identification results extract information such as the area of ​​the affected region, crack width and length, landslide volume and shape, and the location of the affected area on the slope (top, toe, or middle). Based on predefined risk assessment criteria, the characteristics of the affected area are matched with risk levels. For cracks, cracks wider than 3cm and longer than 10m are defined as having a higher risk level; for landslides, landslides with an area greater than 100m² are defined as having a higher risk level. 2 And the volume is greater than 1000m³ 3 Areas that are considered high-risk are defined as such; as for landslides, areas where the accumulation of rubble exceeds 20% of the total slope area are defined as high-risk areas.

[0049] The formula for calculating the area when extracting disease features is:

[0050]

[0051] Where A is the area of ​​the diseased region, which is in pixels, and I(i,j) is the pixel value, where 1 represents the diseased region and 0 represents the background.

[0052] The formula for calculating crack length is:

[0053]

[0054] Where L represents the crack length, (x k ,y k ), (x k+1 ,y k+1 Let x and y be the coordinates of the two endpoints on the crack profile, respectively. k x k+1 Let x and y be the x-coordinates of the two endpoints, respectively. k y k+1 , where k is the ordinate of the two endpoints, k is the crack number, and n is the number of crack outline points.

[0055] Furthermore, the impact of disease characteristics on risk was quantified using a multi-factor comprehensive weighting analysis method. An AHP (Analog-Hybrid Hierarchical Analysis) approach was employed for comprehensive evaluation, with assessment indicators including disease scale, morphology, and location. Weights were assigned based on the degree of influence of each indicator on slope stability, and a risk score was calculated for each area, categorized into low-risk, medium-risk, and high-risk levels.

[0056] Specifically, the Analytic Hierarchy Process (AHP) is used to comprehensively evaluate slope disease risk. First, a hierarchical structure of objectives, criteria, and indicators needs to be constructed, clearly defining specific elements such as disease scale, location, morphology, and environmental characteristics. A judgment matrix is ​​constructed through pairwise comparisons, and the weight vectors of each factor are calculated based on the importance scaling method. Consistency checks are then performed to ensure the matrix's rationality. Finally, a comprehensive score is calculated by combining the weights and actual data for each indicator, classifying slope diseases into low, medium, and high risk levels.

[0057] Specifically, a low-risk level indicates a small area of ​​slope damage with no significant changes in morphology, and a limited impact on overall stability. A medium-risk level indicates a moderate area of ​​damage, with characteristics such as deepening cracks or landslide expansion potentially threatening localized areas. A high-risk level indicates a large area of ​​damage in a critical location, posing a risk of collapse or large-scale landslides.

[0058] Specifically, the formula for calculating the weight vector in the Risk Weight Vector Calculation Method (AHP method) is as follows:

[0059]

[0060] Where, ω i represents the weight of the i-th indicator, where i and j represent the number of rows and columns of the matrix, and a ij Let be the element in the i-th row and j-th column of the matrix, and n be the order of the matrix.

[0061] The formula for calculating the overall risk score is:

[0062]

[0063] Where R is the overall risk score, ω i Let x be the weight of the i-th indicator. i This is the normalized value of the i-th feature (such as crack width or landslide area).

[0064] Furthermore, the predicted slope damage levels are marked on the slope image with different colors: green for low risk, yellow for medium risk, and red for high risk. Based on these results, a visual assessment report is automatically generated, including damage characteristics and risk levels.

[0065] S5 uses a Long Short-Term Memory (LSTM) network to dynamically predict changes in the acquired disease time-series images, predicting the disease's development trend and potential instability time.

[0066] Specifically, drones are used to conduct regular inspections of the slope area multiple times, collecting time-series image features containing disease characteristics, such as crack expansion and changes in landslide extent. The diseased areas in the images are segmented and feature extracted (crack length, width, landslide area, morphological changes) to generate corresponding time-series data. The disease features are then standardized to the [0,1] range to eliminate the influence of units. For missing data points, interpolation methods are used to complete the time series, and continuous time-series data are divided into training, validation, and test sets to ensure temporal consistency.

[0067] Furthermore, multidimensional temporal features, such as crack width variation over time and landslide area growth rate, are used as input layers to the LSTM model. The regression values—predicted future values ​​of the disease characteristics (crack length or landslide extent at the next moment)—are used as the model output. The masking layer contains multiple LSTM units to capture the long-term dependencies and short-term fluctuation patterns of the disease characteristics.

[0068] Furthermore, the forget gate determines whether to discard the cell state information from the previous time step at the current time step according to the following formula:

[0069] f t =σ(W f ·[h t-1 ,x t ]+b f (7)

[0070] Where f t The output of the forget gate takes values ​​in the range [0,1], representing the proportion of information retained from the previous time step, h. t-1 x is the hidden state from the previous time step. t For the current input, W f Let b be the weight matrix. f σ is the bias term used to adjust the calculation results. σ is the Sigmoid activation function with an output range of [0,1]. Here, t represents the current time step, t-1 represents the previous time step, and f is the subscript of the forget gate calculation parameter.

[0071] The input gate formula is:

[0072] i t =σ(W i ·[h t-1 ,x t ]+b i (8)

[0073] Where i t The input gate output has a value between [0,1], representing the new information written to the cell state C at the current moment. t The proportion, h t-1 x is the hidden state from the previous time step. t For the current input, W i Let b be the weight matrix. i σ is the bias term used to adjust the calculation results. σ is the Sigmoid activation function with an output range of [0,1]. Here, t represents the current time, t-1 represents the previous time, and i is the index of the input gate calculation parameter.

[0074] The formula for candidate memory values ​​is:

[0075]

[0076] in h represents candidate cell state values, indicating how new input updates long-term memory. t-1 x is the hidden state from the previous time step. t For the current input, W c Let b be the weight matrix. c This is a bias term used to adjust the calculation results. tanh is the hyperbolic tangent function with an output range of [-1, 1], used to represent positive and negative effects, where t represents the current time, t-1 represents the previous time, and c is the index of the candidate memory value calculation parameter.

[0077] The formula for updating cell state is:

[0078]

[0079] Where C t The current cell state represents long-term memory information, C. t-1 This represents the cell state at the previous time step, where 'e' indicates element-wise multiplication, used to fuse information from the forget gate, input gate, and candidate memories. f represents the candidate cell state value. t For the output of the forget gate, i t The input gate output is t, where t represents the current time and t-1 represents the previous time.

[0080] The output gate formula is:

[0081] o t =σ(W o ·[h t-1 ,x t ]+b o (11)

[0082] Among them o t The output of the gate is a value between [0,1], representing the current cell state C. t Which information will affect the hidden state h? t h t-1 x is the hidden state from the previous time step. t For the current input, W o Let b be the weight matrix. o σ is the bias term used to adjust the calculation results. σ is the Sigmoid activation function with an output range of [0,1]. Here, t represents the current time, t-1 represents the previous time, and o is the subscript of the output gate calculation parameter.

[0083] The formula for hidden state is:

[0084] h t =o t etanh(C t (12)

[0085] Where h t The hidden state at the current moment contains current input and long-term memory information, and is an important source of information for output prediction. t For output gate output, C t The cell state at the current moment is represented by 'e', ​​where 'e' represents element-wise multiplication, and 'tanh' is the hyperbolic tangent function with an output range of [-1, 1].

[0086] The formula for predicting the output is:

[0087]

[0088] in W represents predicted values ​​for the next moment, such as future crack width, landslide area, and other disease characteristics.y Let b be the weight matrix of the output layer. y h is the bias term for the output layer. t This represents the current hidden state.

[0089] Specifically, mean squared error (MSE) is used as the loss function to calculate the difference between the predicted and true values. The Adam optimizer is used to update model parameters and set hyperparameters such as the learning rate to optimize the convergence speed. The past ten observations are set as a fixed-length time window to predict future changes in disease characteristics. Simultaneously, training is stopped when the validation set performance no longer improves during the model training phase to prevent overfitting.

[0090] The mean squared error is used as the loss function, and the calculation formula is as follows:

[0091]

[0092] Where MSE is the mean squared error, y i The true value represents the actual characteristics of the disease. is the predicted value, and n is the number of samples.

[0093] Specifically, the prediction results are recursively used as input for the next prediction to obtain the future trend of the slope instability. A slope instability prediction curve (the increasing trend of crack width and landslide area over time) is plotted, and the timing of instability is estimated by combining this with a preset instability threshold. The prediction results are then combined with a risk level assessment to provide early warning of slope instability.

[0094] The formula for predicting the instability time of disease is:

[0095]

[0096] Where T failure T is the predicted instability time point. current At the current observation time, R rate This represents the rate of change of disease characteristics.

[0097] A second aspect of the present invention provides a device for slope disease identification based on unmanned aerial vehicle (UAV) and machine vision technology, such as... Figure 5 As shown, it includes:

[0098] The drone platform module is used to carry GPS modules, inertial navigation systems (INS), attitude sensors and other devices to support autonomous flight path planning and stable flight. It also carries high-resolution cameras for image acquisition and wireless communication technologies (such as 5G, Wi-Fi or 4G) to transmit image data.

[0099] The image processing module is used to perform image preprocessing on high-resolution images of the slope area captured by the drone camera, such as noise reduction, image enhancement, image correction, and image stitching, and uses a high-capacity storage device to retain the original and processed data.

[0100] The disease identification module extracts key features of diseases such as cracks, landslides, and collapses. It uses an improved YOLOv8-seg instance segmentation algorithm to classify disease types and locate disease areas. Simultaneously, it assesses the risk level of diseases based on disease features and slope topography data, providing a visual assessment report.

[0101] The time-series analysis module is used to generate point cloud data through multi-view image matching, construct a high-precision digital elevation model of the slope, mark the location and morphology of the disease, and use it for dynamic monitoring. Based on the time-series data of the disease, it analyzes the dynamic changes of slope characteristics and predicts the evolution trend of the disease using an LSTM time-series model. This invention also discloses a non-volatile storage medium, characterized in that the non-volatile storage medium includes a stored program, wherein, when the program runs, it controls the device where the non-volatile storage medium is located to execute the above-described method for detecting the compaction degree of recycled aggregate from construction solid waste based on image processing.

[0102] In summary, this invention provides a method and apparatus for slope disease identification based on UAV and machine vision technology, comprising: a UAV equipped with a high-resolution camera acquiring multi-angle images of a slope area along a preset flight path; preprocessing the images using various methods such as denoising, correction, and stitching, followed by feature extraction and image enhancement; identifying and classifying disease features using an improved YOLOv8-seg instance segmentation algorithm, covering common slope diseases such as cracks, landslides, and collapses; quantifying the risk level of slope diseases based on the identification results using rule models or multi-factor analysis methods (such as disease scale, location, and morphology); and using a Long Short-Term Memory (LSTM) network to dynamically predict the changes in the acquired time-series images of the diseases, predicting the development trend and potential instability time. Based on the image data acquired by the UAV, slope diseases can be monitored and dynamically analyzed in real time, potential hazard areas can be identified in a timely manner, providing a scientific basis for early warning and management of geological disasters, and reducing casualties and property losses. At the same time, it can accurately identify the characteristics of diseases such as cracks, landslides, and collapses, and has the advantages of automation and intelligence, avoiding the subjective errors that may be caused by human judgment in traditional methods.

[0103] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for slope disease identification based on UAV and machine vision technology, characterized by: S1, a drone equipped with a high-resolution camera, collects multi-angle images of the slope area along a preset flight path; S2, preprocesses the acquired images and performs feature extraction and image enhancement; S3, combined with the improved YOLOv8-seg instance segmentation algorithm, identifies and classifies slope diseases based on their characteristics; S4. Based on the identification results, use rule models or multi-factor analysis methods to quantify the risk level of slope diseases; Extract the area of ​​the affected area, crack width and length, landslide volume and shape, and the location of the affected area on the slope from the identification results; match the affected characteristics with the risk level based on predefined risk assessment criteria; The formula for calculating the area when extracting disease features is: （3） Where A is the area of ​​the diseased region, which is in pixels, and I(i,j) is the pixel value, where 1 represents the diseased region and 0 represents the background. The formula for calculating crack length is: （4） Where L represents the crack length, (x k ,y k ), (x k+1 ,y k+1 ) are the coordinates of the two endpoints on the crack profile, and n is the number of crack profile points; To comprehensively evaluate the risk of slope diseases using the AHP (Analog-Hybrid Analysis) method, it is first necessary to construct a hierarchical structure of objectives, criteria, and indicators, and clarify the specific elements of the evaluation indicators, such as disease scale, location, morphology, and environmental characteristics. A judgment matrix is ​​constructed by pairwise comparisons. The weight vector of each factor is calculated based on the importance scaling method, and a consistency check is performed to ensure the rationality of the matrix. Finally, the comprehensive score is calculated by combining the weights and the actual data of each indicator, and the slope disease is divided into low, medium and high risk levels. The formula for calculating the weight vector in the AHP (Advanced Risk Hierarchy Process) method is as follows: （5） Where, ω i represents the weight of the i-th indicator, where i and j represent the number of rows and columns of the matrix, and a ij Let be the element in the i-th row and j-th column of the matrix, and n be the order of the matrix; The formula for calculating the overall risk score is: （6） Where R is the overall risk score, ω i Let x be the weight of the i-th indicator. i is the normalized value of the i-th feature; The predicted slope disease levels are marked on the slope image with different colors, with low risk in green, medium risk in yellow, and high risk in red; based on the above results, a visual assessment report is automatically generated, which includes disease characteristics and risk levels. S5 uses a long short-term memory network to dynamically predict changes in the collected disease time-series images, predicting the development trend and potential instability time of the disease.

2. The method for slope disease identification based on UAV and machine vision technology according to claim 1, characterized in that: S2 further includes the following: using Gaussian filtering or median filtering to remove noise caused by wind interference in the UAV imagery, and eliminating grainy noise in low-light areas to highlight the details of the slope texture; using geometric correction methods to correct barrel or pincushion distortion caused by the UAV lens, and using orthophoto correction to make the slope area present a true proportion; at the same time, the slope images taken from multiple angles are stitched into a panoramic image by feature point matching to make the slope area complete, and brightness equalization and seam smoothing technology is used to solve uneven lighting or color difference in the stitched image.

3. The method for slope disease identification based on UAV and machine vision technology according to claim 1, characterized in that: S3 further includes the following: using processed UAV multi-angle images to accurately identify and classify slope diseases; improving the detection accuracy of cracks, landslides, and collapses by optimizing the network structure, and enhancing the ability to capture small-area disease features by combining an attention mechanism. The algorithm can output the bounding box, category label, and mask of the disease area, automatically extract disease features such as crack length and landslide area, and visualize them on the images.

4. The method for slope disease identification based on UAV and machine vision technology according to claim 3, characterized in that: S3 further includes the following: The separable kernel attention pyramid SKAPPF module is used to replace the SPPF module in the original YOLOv8-seg model. SKAPPF integrates various aspects of SPPF and LSKA, and enhances the feature fusion capability by utilizing and reusing multi-scale features, thereby enhancing the processing of fine-grained feature information. The improved YOLOv8-seg model was trained on the dataset using transfer learning techniques to obtain a weight file, which enables the identification and classification of slope disease areas, and extracts disease features such as crack length and landslide area information.

5. A device for slope disease identification based on UAV and machine vision technology, the device performing the method of claim 1, characterized in that: include: The drone platform module is used to carry a GPS module, an inertial navigation system, and an attitude sensor device to support autonomous flight path planning and stable flight. It also carries a high-resolution camera for image acquisition and wireless communication technology to transmit image data. The image processing module is used to perform image preprocessing steps such as noise reduction, image enhancement, image correction, and image stitching on high-resolution images of the slope area captured by the drone camera, and uses a high-capacity storage device to retain the original and processed data; The disease identification module is used to extract key features of cracks, landslides, and collapses. It uses an improved YOLOv8-seg instance segmentation algorithm to classify disease types and locate disease areas. At the same time, it assesses the risk level of diseases based on disease features and slope topography data and provides a visual assessment report. The time series analysis module is used to generate point cloud data through multi-view image matching, construct a high-precision digital elevation model of the slope, mark the location and morphology of the disease for dynamic monitoring, and analyze the dynamic changes of slope characteristics based on the disease time series data, and predict the disease evolution trend by combining the LSTM time series model.

6. A non-volatile storage medium, characterized in that, The non-volatile storage medium includes a stored program, wherein the program, when running, controls the device where the non-volatile storage medium is located to execute the slope disease identification method based on UAV and machine vision technology as described in any one of claims 1-4.

7. An electronic device for slope disease identification based on UAV and machine vision technology, characterized in that, It includes a processor and a memory; the memory stores computer-readable instructions, and the processor is used to execute the computer-readable instructions, wherein the computer-readable instructions, when executed, perform the slope disease identification method based on UAV and machine vision technology as described in any one of claims 1-4.

Images