Ice landslide monitoring and early warning method and system based on AI image recognition

Abstract

The invention provides an ice landslide monitoring and early warning method and system based on AI image recognition, and the method comprises the steps: firstly collecting a dynamic image sequence of a target glacier region, extracting the three-dimensional spatial-temporal features of the surface texture, crack distribution pattern and displacement rate of an ice surface, generating an ice surface state feature set, inputting the feature set into a pre-trained ice landslide monitoring model, and carrying out the recognition of the ice landslide; performing spatial correlation analysis through a multi-scale feature fusion layer, outputting an ice surface stability evaluation parameter, generating an ice surface risk grade label according to the ice surface stability evaluation parameter, matching an early warning trigger threshold value in historical disaster data, and if the risk grade label meets the threshold value, generating an ice landslide risk early warning signal; meanwhile, the model parameter weight is dynamically adjusted to adapt to the environment change, and finally an early warning signal and ice surface texture visual data are sent to the monitoring terminal, so that effective monitoring and early warning of the ice landslide are realized, and potential disasters can be dealt with in time.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an ice landslide monitoring and early warning method and system based on AI image recognition. Background Art

[0002] In the field of glacier disaster monitoring and early warning technology, traditional methods mainly rely on manual inspections, fixed-point sensor monitoring or numerical simulations based on physical models, which have significant technical limitations. Manual inspections are limited by the harsh natural environment and high-risk working conditions, making it difficult to achieve high-frequency, full-coverage dynamic monitoring of glaciers; although fixed-point sensors can provide physical parameters in local areas, they cannot capture the microscopic deformation characteristics of the glacier surface and the global evolution of crack expansion; numerical simulation methods based on physical models are highly dependent on precise initial conditions and boundary parameters, and it is difficult to ensure the reliability of prediction results in complex and changeable glacier environments.

[0003] In the existing technology, traditional early warning systems mostly use static threshold judgment mechanisms, which cannot adapt to the dynamic characteristics of the glacier environment with seasonal and climatic changes, and are prone to false alarms or missed alarms. In terms of risk visualization, the existing technology is mostly limited to the simple presentation of numerical parameters, lacking an intuitive glacier surface state visualization interface, making it difficult to support decision makers to quickly understand the spatial and temporal distribution characteristics of risks. Summary of the invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides an ice landslide monitoring and early warning method based on AI image recognition, the method comprising: Collect dynamic image sequences of the target glacier area, extract the three-dimensional spatiotemporal characteristics of ice surface texture, crack distribution morphology and displacement rate in the dynamic image sequences, and generate an ice surface state feature set; Inputting the ice surface state feature set into a pre-trained ice landslide monitoring model, performing spatial correlation analysis on the ice surface state feature set through a multi-scale feature fusion layer in the ice landslide monitoring model, and outputting ice surface stability assessment parameters; Generate an ice surface risk level label according to the ice surface stability assessment parameter, and match the warning trigger threshold in the historical disaster data based on the ice surface risk level label; When the ice surface risk level label meets the warning trigger threshold, an ice landslide risk warning signal is generated, and the parameter weights of the ice landslide monitoring model are dynamically adjusted to adapt to environmental changes in the target glacier area; The ice landslide risk warning signal and the corresponding ice surface texture visualization data are sent to the monitoring terminal.

[0005] On the other hand, an embodiment of the present invention also provides an ice landslide monitoring and early warning system based on AI image recognition, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0006] Based on the above aspects, the embodiment of the present application realizes the multi-scale stereoscopic characterization of the dynamic evolution of the glacier by collecting the dynamic image sequence of the target glacier area and extracting the three-dimensional spatiotemporal characteristics of the ice surface texture, crack distribution morphology and displacement rate. On this basis, the pre-trained ice landslide monitoring model uses a multi-scale feature fusion layer to perform a deep spatial correlation analysis on the ice surface state feature set. The output ice surface stability assessment parameters not only quantify the mechanical stability of the glacier structure, but also establish a risk warning paradigm with spatiotemporal adaptive capabilities through a dynamic threshold matching mechanism between risk level labels and historical disaster data. When the ice surface risk level is monitored to reach the warning trigger threshold, a visual warning signal can be generated synchronously and the online optimization of the model parameter weights can be started, which not only ensures the real-time nature of the warning response, but also realizes the continuous adaptation to the environmental changes in the target glacier area through the dynamic adjustment of the model parameters, significantly improving the robustness and generalization ability of the warning system. The final output of the ice landslide risk warning signal and the ice surface texture visualization data not only provides a multi-dimensional decision-making basis for disaster prevention decisions, but also reduces the professional threshold through an intuitive visualization interface, so that the monitoring terminal can accurately intervene based on the global risk situation. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 It is a schematic diagram of the execution flow of the ice landslide monitoring and early warning method based on AI image recognition provided in an embodiment of the present invention.

[0008] Figure 2 It is a schematic diagram of exemplary hardware and software components of an ice landslide monitoring and early warning system based on AI image recognition provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0009] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 The figure is a flow chart of an ice landslide monitoring and early warning method based on AI image recognition provided by an embodiment of the present invention. The ice landslide monitoring and early warning method based on AI image recognition is introduced in detail below.

[0010] Step S110, collecting a dynamic image sequence of the target glacier area, extracting the three-dimensional spatiotemporal characteristics of the ice surface texture, crack distribution morphology and displacement rate in the dynamic image sequence, and generating an ice surface state feature set.

[0011] For example, for a certain glacier area, multispectral imagers, thermal infrared sensors and synthetic aperture radars can be used to perform data collection. The visible light image sequence acquired by the multispectral imager can reflect the basic appearance characteristics of the ice surface, the thermal infrared image sequence acquired by the thermal infrared sensor can capture information related to the temperature of the ice body, and the radar interference image sequence collected by the synthetic aperture radar helps to accurately measure the displacement of the ice surface. Then, the crack extension direction vector, the thermal anomaly zone diffusion path vector and the ice surface displacement rate field in the dynamic image sequence can be extracted, and then mapped to the same three-dimensional spatial coordinate system to construct a multidimensional spatiotemporal feature matrix containing crack length, temperature gradient and displacement acceleration. The spatiotemporal pyramid pooling layer is used to perform hierarchical aggregation, and the fused feature vectors of ice surface crack growth rate, thermal diffusion intensity and displacement acceleration at different time scales are extracted. The fused feature vectors are measured for similarity with the standard feature template in the historical glacier deformation database, and the abnormal feature components caused by instantaneous environmental noise (such as short-term sunlight reflection, etc.) are removed to generate a three-dimensional spatiotemporal feature set containing ice surface structural stability indicators, that is, an ice surface state feature set.

[0012] Step S120, inputting the ice surface state feature set into a pre-trained ice landslide monitoring model, performing spatial correlation analysis on the ice surface state feature set through a multi-scale feature fusion layer in the ice landslide monitoring model, and outputting ice surface stability assessment parameters.

[0013] In detail, the ice landslide monitoring model is trained with a large amount of historical glacier data. The multi-scale feature fusion layer analyzes the ice surface state feature set and can output ice surface stability assessment parameters. The ice surface stability assessment parameters represent the stability of the ice surface in different areas in numerical form, such as between 0 and 1, where 0 represents extremely unstable and 1 represents very stable.

[0014] Step S130, generating an ice surface risk level label according to the ice surface stability assessment parameter, and matching an early warning trigger threshold in historical disaster data based on the ice surface risk level label.

[0015] In this embodiment, the ice surface stability assessment parameters are used as core inputs and are converted into intuitive and easy-to-understand ice surface risk level labels through analysis and processing. The risk level labels not only reflect the current structural stability status of the ice surface, but also provide a direct basis for subsequent early warning decisions.

[0016] Step S140, when the ice surface risk level label meets the warning trigger threshold, an ice landslide risk warning signal is generated, and the parameter weights of the ice landslide monitoring model are dynamically adjusted to adapt to the environmental changes in the target glacier area.

[0017] In this embodiment, when the ice surface risk level label reaches the warning trigger threshold, for example, the ice surface risk level label of a certain area is displayed as high risk and meets the pre-set threshold, an ice landslide risk warning signal can be generated. The ice landslide risk warning signal contains detailed information of the risk area, such as coordinates, risk level, etc.

[0018] At the same time, in order to adapt to environmental changes in the glacier area, the parameter weights of the ice landslide monitoring model are dynamically adjusted to ensure the accuracy of the ice landslide monitoring model in monitoring the glacier status.

[0019] Step S150, sending the ice landslide risk warning signal and the corresponding ice surface texture visualization data to the monitoring terminal.

[0020] In this embodiment, after the ice landslide risk warning signal and ice surface texture visualization data are generated, they can be sent to the monitoring terminal so that relevant personnel can obtain accurate ice landslide risk information and take measures.

[0021] Based on the above steps, the embodiment of the present application realizes the multi-scale stereoscopic characterization of the dynamic evolution of the glacier by collecting the dynamic image sequence of the target glacier area and extracting the three-dimensional spatiotemporal characteristics of the ice surface texture, crack distribution morphology and displacement rate. On this basis, the pre-trained ice landslide monitoring model uses a multi-scale feature fusion layer to perform a deep spatial correlation analysis on the ice surface state feature set. The output ice surface stability assessment parameters not only quantify the mechanical stability of the glacier structure, but also establish a risk warning paradigm with spatiotemporal adaptive capabilities through a dynamic threshold matching mechanism between risk level labels and historical disaster data. When the ice surface risk level is monitored to reach the warning trigger threshold, a visual warning signal can be generated synchronously and the online optimization of the model parameter weights can be started, which not only ensures the real-time nature of the warning response, but also realizes the continuous adaptation to the environmental changes in the target glacier area through the dynamic adjustment of the model parameters, significantly improving the robustness and generalization ability of the warning system. The final output of the ice landslide risk warning signal and the ice surface texture visualization data not only provides a multi-dimensional decision-making basis for disaster prevention decisions, but also reduces the professional threshold through an intuitive visualization interface, so that the monitoring terminal can accurately intervene based on the global risk situation.

[0022] In a possible implementation, step S110 includes: Step S111, synchronously collect visible light image sequence, thermal infrared image sequence and radar interferometric image sequence of the target glacier area in a continuous time window through a multispectral imager, a thermal infrared sensor and a synthetic aperture radar, and perform spatiotemporal alignment according to a unified timestamp.

[0023] Taking the large glacier area mentioned earlier as an example, the continuous time window is set to one week, and data is collected every 3 hours during this week. The visible light image sequence collected by the multispectral imager can intuitively present the appearance characteristics of the ice surface, the thermal infrared image sequence obtained by the thermal infrared sensor helps to understand the temperature distribution of the ice body, and the radar interferometric image sequence collected by the synthetic aperture radar can measure the displacement of the ice surface and other information. During the acquisition process, through time synchronization technology, a unified timestamp is given to each set of collected visible light images, thermal infrared images, and radar interferometric images to ensure that these different types of images are accurately registered in time and space. For example, the visible light images, thermal infrared images, and radar interferometric images collected at a certain moment all reflect the state of the glacier at the same location and at the same time.

[0024] Step S112, performing adaptive histogram equalization processing on the visible light image sequence to enhance the contrast between the ice cracks and the background, using a multi-scale edge detection algorithm to extract crack edge pixel clusters of the ice surface texture, and performing optical flow tracking on the crack edge pixel clusters of adjacent frames to generate crack expansion direction vectors.

[0025] After collecting the visible light image sequence, the contrast between the ice cracks and the background in the original image may be low, which is not conducive to accurate analysis of the cracks, so adaptive histogram equalization processing is required. Taking a frame of visible light image as an example, the color of the ice crack part in the original image is similar to the surrounding background color. After adaptive histogram equalization processing, the color of the crack part becomes darker, forming a sharp contrast with the background, making the crack easier to identify.

[0026] Step S113, performing region growing segmentation on the thermal infrared image sequence, identifying the boundary coordinates of the pixel region where the temperature difference of the ice body exceeds a preset threshold, and generating a diffusion path vector of the thermal anomaly zone on the ice surface based on the direction of temperature gradient change.

[0027] In this embodiment, it is assumed that the temperature difference of the ice body is set to be more than 3 degrees Celsius as the preset threshold. When the temperature difference between the ice body in a certain pixel area and the surrounding ice body exceeds 3 degrees Celsius, the pixel area is identified. For example, in a certain area in the thermal infrared image, the temperature of the ice body in the central part is 3.5 degrees Celsius higher than the temperature of the surrounding ice body. Through the regional growth segmentation algorithm, starting from this central pixel with a higher temperature, it gradually expands outward until it encounters a pixel with a temperature difference of less than 3 degrees Celsius, thereby determining the boundary coordinates of the pixel area where the temperature difference of the ice body exceeds the preset threshold. Then, the diffusion path vector of the thermal anomaly zone on the ice surface is generated based on the direction of temperature gradient change. Since heat always diffuses from the high temperature area to the low temperature area, the direction of temperature gradient change is determined according to the rate of change of temperature in different directions. For example, in this high temperature area, the temperature gradually decreases around, and the temperature decreases fastest in the north direction. Then the direction of the diffusion path vector of the thermal anomaly zone is northward, and the diffusion path vector can intuitively show the direction of heat diffusion from the high temperature area to the low temperature area.

[0028] Step S114, performing phase unwrapping processing on the radar interference image sequence, calculating the displacement vector of each pixel point on the ice surface in the three-dimensional space coordinate system, and generating the ice surface displacement rate field through time differential interferometry.

[0029] Take a pixel point at the edge of the glacier as an example. Within a certain period of time, by analyzing phase changes and known radar wavelength and other parameters, the displacement vector of the pixel point in the three-dimensional space coordinate system is calculated. Assuming that within a month, the displacement of the pixel point in the x direction is 0.5 meters, the displacement in the y direction is 0.3 meters, and the displacement in the z direction is 0.1 meters (here it is assumed that the z direction is the vertical direction, and the x and y directions are the horizontal directions), then the displacement vector of the pixel point is (0.5, 0.3, 0.1) meters. Then, the ice surface displacement rate field is generated by time differential interferometry. The calculation method is to subtract the displacement of each pixel point at different time points and then divide it by the time interval. For example, in the first month, the displacement vector of the pixel point is (0.5, 0.3, 0.1) meters, and in the second month, the displacement vector is (0.8, 0.4, 0.2) meters. The time interval is 1 month. Then the displacement rate in the x-direction is (0.8-0.5) / 1=0.3 meters / month, the displacement rate in the y-direction is (0.4-0.3) / 1=0.1 meters / month, and the displacement rate in the z-direction is (0.2-0.1) / 1=0.1 meters / month. The displacement rate of the pixel point is thus obtained, and the displacement rates of each pixel point on the entire ice surface constitute the ice surface displacement rate field.

[0030] Step S115, mapping the crack extension direction vector, the thermal anomaly zone diffusion path vector and the ice surface displacement rate field to the same three-dimensional space coordinate system, and constructing a multi-dimensional spatiotemporal feature matrix including crack length, temperature gradient and displacement acceleration.

[0031] Take a pixel point at the edge of the glacier as an example. By analyzing phase changes and known radar wavelength and other parameters within a certain period of time, the displacement vector of the pixel point in the three-dimensional space coordinate system can be calculated. Assuming that within a month, the displacement of the pixel point in the x direction is 0.5 meters, the displacement in the y direction is 0.3 meters, and the displacement in the z direction is 0.1 meters (here it is assumed that the z direction is the vertical direction, and the x and y directions are the horizontal directions), then the displacement vector of the pixel point is (0.5, 0.3, 0.1) meters. Then, the ice surface displacement rate field is generated by time differential interferometry. The calculation method is to subtract the displacement of each pixel point at different time points and then divide it by the time interval. For example, in the first month, the displacement vector of the pixel point is (0.5, 0.3, 0.1) meters, and in the second month, the displacement vector is (0.8, 0.4, 0.2) meters. The time interval is 1 month. Then the displacement rate in the x-direction is (0.8-0.5) / 1=0.3 meters / month, the displacement rate in the y-direction is (0.4-0.3) / 1=0.1 meters / month, and the displacement rate in the z-direction is (0.2-0.1) / 1=0.1 meters / month. The displacement rate of the pixel point is thus obtained, and the displacement rates of each pixel point on the entire ice surface constitute the ice surface displacement rate field.

[0032] Step S116, using a spatiotemporal pyramid pooling layer to perform hierarchical aggregation on the multi-dimensional spatiotemporal feature matrix, and extracting fused feature vectors of ice surface crack growth rate, thermal diffusion intensity and displacement acceleration at different time scales.

[0033] In detail, the spatiotemporal pyramid pooling layer can process the data in the matrix according to different time scales. For example, for data with a shorter time scale, more attention may be paid to the changes in the local ice surface crack growth rate, thermal diffusion intensity and displacement acceleration. For data with a longer time scale, the changing trends of these features in a larger range and longer time period will be comprehensively considered. Through this hierarchical aggregation method, the fused feature vectors of ice surface crack growth rate, thermal diffusion intensity and displacement acceleration are extracted from different time scales. For example, at a shorter time scale, the ice surface crack growth rate in a certain area is faster, the thermal diffusion intensity is smaller, and the displacement acceleration is larger. These features are fused according to the set weights (the weights are determined according to the importance and relevance of the data) to obtain partial components of a fused feature vector. At a longer time scale, the changes in these features of the entire glacier area are integrated and then fused into the feature vector, thereby obtaining a complete fused feature vector containing information at different time scales.

[0034] Step S117, performing similarity measurement on the fused feature vector and the standard feature template in the historical glacier deformation database, eliminating abnormal feature components caused by instantaneous environmental noise, and generating a three-dimensional spatiotemporal feature set containing ice surface structure stability indicators as the ice surface state feature set.

[0035] For example, the distance between the fused feature vector and each standard feature template can be calculated (the distance here can be a suitable measurement method such as Euclidean distance). The smaller the distance, the higher the similarity. Assuming that the distance between the fused feature vector and a standard feature template is 0.1, and the distance between the fused feature vector and another standard feature template is 0.5, then it is more similar to the standard feature template with a distance of 0.1. In this process, due to the transient environmental noise (such as short-term sunlight reflection, local meteorological fluctuations, etc.), some abnormal feature components may be caused. These abnormal feature components caused by the transient environmental noise need to be removed. For example, if the value of a certain feature component obviously deviates from the value range of other normal conditions, it is determined by analysis that it is caused by the transient environmental noise, and the feature component is removed. After such similarity measurement and removal of abnormal feature components, a three-dimensional spatiotemporal feature set containing ice surface structure stability indicators is generated. The three-dimensional spatiotemporal feature set is used as the ice surface state feature set, which contains various information that can reflect the stability of the ice surface structure, such as the development of cracks, changes in thermal distribution, displacement rate and acceleration.

[0036] In a possible implementation, step S112 includes: Step S1121, input the preprocessed visible light image sequence into a two-branch convolutional network, the first branch convolutional network uses a large-scale convolution kernel to extract the trend line of the macro cracks on the ice surface, and the second branch convolutional network uses a small-scale convolution kernel to detect the bifurcation endpoints of the micro cracks.

[0037] Taking the previously collected visible light image sequence of the large glacier area as an example, after preprocessing, the visible light image sequence is input into the two-branch convolutional network. The first branch convolutional network uses a large-scale convolution kernel, for example, the convolution kernel size is 5×5. In a certain area of ​​the image, the large-scale convolution kernel detects the trend line of the macroscopic cracks on the ice surface through convolution operation. The trend line indicates the main extension direction of the cracks on the ice surface. At the same time, the second branch convolutional network uses a small-scale convolution kernel, assuming a 3×3 convolution kernel, to detect the bifurcation endpoints of microscopic cracks. Based on the macroscopic cracks, these small-scale convolution kernels can find the bifurcation endpoints of microscopic cracks in more detail.

[0038] Step S1122, performing pixel-level weighted fusion on the edge feature map output by the dual-branch convolutional network to generate a multi-resolution crack edge probability map, and after removing pseudo-edge noise through a non-maximum suppression algorithm, binarizing the crack edge probability map to extract crack pixel clusters whose connected domain area is greater than a preset threshold, and recording the centroid coordinates and principal axis direction angles of each pixel cluster.

[0039] For the edge feature map output by the dual-branch convolution network, different weights are assigned according to the importance of the features output by different convolution kernels for pixel-level weighted fusion. For example, the weight of the macro crack trend line feature output by the large-scale convolution kernel is 0.6, and the weight of the micro crack bifurcation endpoint feature output by the small-scale convolution kernel is 0.4 (the weights here are determined based on experience or experiments, and may be adjusted according to specific circumstances in practical applications). Then these two features are weighted and calculated according to the weights to obtain a multi-resolution crack edge probability map. The multi-resolution crack edge probability map represents the probability that each pixel belongs to the crack edge. Next, the pseudo-edge noise is removed by the non-maximum suppression algorithm. In the multi-resolution crack edge probability map, there may be some local peaks that are not real edges. The non-maximum suppression algorithm compares the probability values ​​of each pixel with the surrounding pixels, and reduces the probability values ​​of those pixels that are not local maxima (i.e., pseudo-edge noise), thereby removing these interferences. After that, the crack edge probability map is binarized, and a threshold value (for example, 0.5) is set. If the probability value of a pixel point is greater than 0.5, it is considered to be a crack edge pixel, and if it is less than 0.5, it is considered not to be. Then extract the crack pixel clusters whose connected domain area is greater than the preset threshold (assuming it is 50 square pixels). For example, in a connected crack pixel area in the image, after calculating its area to be 60 square pixels, it meets the condition of being greater than 50 square pixels, so the pixel cluster is extracted, and the centroid coordinates and main axis direction angle of the pixel cluster are recorded. Assuming that the centroid coordinates of the pixel cluster are calculated to be (150, 250), and the main axis direction angle is 45 degrees, the centroid coordinates represent the center position of the pixel cluster in the image, and the main axis direction angle represents the main extension direction of the crack in the pixel cluster.

[0040] Step S1123, between the visible light image sequences of adjacent time stamps, a sparse feature point correspondence relationship of the crack pixel cluster is established based on the centroid coordinates, and a pyramid Lucas-Kanade optical flow algorithm is used to calculate the motion vector of the feature point.

[0041] Between the visible light image sequences of adjacent time stamps, the sparse feature point correspondence of the fissure pixel clusters is established based on the centroid coordinates of the previously recorded fissure pixel clusters. For example, in the image at a certain moment, the centroid coordinates of a fissure pixel cluster are (150, 250). In the image at the adjacent moment, the corresponding feature points are determined based on the pixel points near the centroid coordinates. Then the pyramid Lucas-Kanade optical flow algorithm is used to calculate the motion vectors of these feature points. Taking a certain feature point as an example, the optical flow of the feature point between two adjacent frames is calculated at different layers of the pyramid (for example, starting from the bottom layer and gradually calculating to the upper layers). Assuming that at the bottom layer, the optical flow calculation result of the feature point in the horizontal direction is 3 pixels and the vertical direction is 2 pixels, then the result is adjusted according to the hierarchical relationship of the pyramid and some parameters in the algorithm (such as the scale factor, etc.), and finally the motion vector of the feature point moving 3 pixels in the horizontal direction and 2 pixels in the vertical direction is obtained.

[0042] Step S1124, generating the velocity components of each crack pixel cluster in three dimensions of horizontal displacement, vertical settlement and lateral expansion according to the projection components of the motion vector in the three-dimensional space coordinate system.

[0043] The calculated motion vector needs to be converted into the velocity components of horizontal displacement, vertical settlement and lateral expansion in the three-dimensional space coordinate system. Take the motion vector of a feature point that has moved 3 pixels horizontally and 2 pixels vertically as an example. Assume that the horizontal direction of the image corresponds to the x direction in the three-dimensional space coordinate system, the vertical direction corresponds to the z direction (here, the z direction is assumed to be the vertical direction), and the lateral expansion direction is the y direction (perpendicular to the x and z directions). First, determine the projection relationship. Assume that the ratio of the horizontal motion vector projected onto the horizontal displacement dimension is 0.8, the ratio of the vertical motion vector projected onto the vertical settlement dimension is 0.6, and the horizontal expansion dimension is calculated based on the relationship between the horizontal and vertical vectors (calculated by geometric relationship, here it is assumed that the ratio is 0.5 calculated according to the Pythagorean theorem). Then the horizontal displacement rate component is 3×0.8=2.4 pixels / unit time, the vertical settlement rate component is 2×0.6=1.2 pixels / unit time, and the horizontal expansion rate component is sqrt(3²+2²)×0.5=1.8 pixels / unit time (here sqrt means square root, and the text description is the square root of the sum of the square of 3 plus the square of 2). For multiple feature points in the same crack pixel cluster, their respective rate components are calculated in this way.

[0044] Step S1125, median filtering is performed on the velocity components of multiple feature points of the same crack pixel cluster, and after eliminating local motion outliers, the average expansion rate and acceleration in the main axis direction of each crack are calculated.

[0045] Take a crack pixel cluster as an example, assuming that there are 5 feature points in the pixel cluster, and their horizontal displacement rate components are 2.2, 2.5, 2.3, 2.6, and 2.4 pixels / unit time respectively. First, these rate components are median filtered, and the 5 values ​​are sorted from small to large as 2.2, 2.3, 2.4, 2.5, and 2.6. The middle value 2.4 is the result after median filtering. The vertical settlement rate component and the lateral expansion rate component are also median filtered in the same way.

[0046] When calculating the average expansion rate in the main axis direction of the crack, the main axis direction angle of the crack needs to be considered. Assume that the main axis direction angle of the crack pixel cluster recorded previously is 45 degrees. For the horizontal displacement rate component and the lateral expansion rate component, the projection component in the main axis direction is calculated according to the trigonometric function relationship. The projection of the horizontal displacement rate component in the main axis direction is the filtered horizontal displacement rate component multiplied by cos(45 degrees), and the projection of the lateral expansion rate component in the main axis direction is the filtered lateral expansion rate component multiplied by sin(45 degrees), and then these two projection components are added to obtain the rate component in the main axis direction. Assume that after calculation, the rate component in the main axis direction at a certain moment is 3.0 pixels / unit time, and after the next unit time, the rate component becomes 3.3 pixels / unit time.

[0047] When calculating acceleration, according to the definition of acceleration, that is, the change in velocity divided by the time interval. Here, the time interval is 1 unit time, so the acceleration is (3.3-3.0) / 1=0.3 pixels / unit time².

[0048] Step S1126, output a vector set including the spatial position, expansion direction and rate of the crack as the crack expansion direction vector, and perform spatial superposition verification with the thermal anomaly zone diffusion path vector.

[0049] In this embodiment, the calculated vector set including the spatial position of the crack (expressed as the centroid coordinates of the crack pixel cluster, such as (150, 250)), the expansion direction (expressed as the main axis direction angle of 45 degrees) and the rate (such as the average expansion rate and acceleration information) can be used as the crack expansion direction vector. Then, it is spatially superimposed and verified with the diffusion path vector of the thermal anomaly area. For example, in a three-dimensional spatial coordinate system, the crack expansion direction vector and the thermal anomaly area diffusion path vector are drawn. If the area pointed by the crack expansion direction vector overlaps with the area pointed by the thermal anomaly area diffusion path vector, or there is a set association in space (such as the crack expansion direction is in the direction of the thermal anomaly area), it means that there is a set spatial correlation between the two, which is helpful to further analyze the state of the glacier and determine whether the glacier has a higher risk of ice landslides in these areas.

[0050] In a possible implementation, the training method of the ice landslide monitoring model includes: Step S210, obtaining an image training set of the target glacier area before the historical ice landslide event occurs, wherein the image training set includes a visible light image sequence of the ice surface, a thermal infrared image sequence, and ice displacement rate measurement data at corresponding timestamps.

[0051] For the large glacier area mentioned above, image data before previous ice landslide events can be collected to construct an image training set. The visible light image sequence in the image training set can reflect the appearance texture of the ice surface, and the thermal infrared image sequence can reflect the temperature distribution of the ice body. The ice displacement rate measurement data under the corresponding timestamp accurately records the movement speed of the ice body at different times. For example, this embodiment selects data from a set time period (such as one month before the event) before multiple ice landslide events occurred from the historical database of glacier monitoring. During this time period, there are corresponding visible light images, thermal infrared images, and ice displacement rate measurement values ​​every hour. The visible light image clearly shows the different forms of the ice surface, the thermal infrared image shows subtle differences in the temperature of the ice body, and the ice displacement rate measurement data accurately records the speed of the ice body moving in different directions at various times.

[0052] Step S220, grayscale processing is performed on the visible light image sequence to extract the pixel coordinate set of the edge of the ice surface crack, and temperature threshold segmentation is performed on the thermal infrared image sequence to identify the boundary coordinates of the area where the temperature difference of the ice body changes suddenly.

[0053] Taking a frame of visible light image as an example, the original image may contain multiple color information. After grayscale processing, it is converted into a single grayscale value, which is more convenient for the subsequent analysis of the edge of the ice crack. Through the existing algorithm, the grayscale image is processed to identify the pixels at the edge of the ice crack, so as to extract the pixel coordinate set of the edge of the ice crack. For example, in a certain area in the upper left corner of the image, a series of pixels are detected to constitute the edge of the ice crack. The coordinates of these pixels are recorded to form the pixel coordinate set of the edge of the ice crack. At the same time, for the thermal infrared image sequence, temperature threshold segmentation is performed. A temperature threshold is set, such as 3 degrees Celsius. When the temperature change of the ice body exceeds the threshold, it is identified as a temperature difference mutation area. By traversing each pixel in the thermal infrared image and comparing its temperature difference with the surrounding pixels, once the temperature difference exceeds 3 degrees Celsius, the pixel is marked, and the boundary coordinates of these temperature difference mutation areas are gradually determined. For example, in the central area of ​​the thermal infrared image, a circular area is found, in which the temperature of the pixels inside is more than 3 degrees Celsius different from that of the surrounding area. After algorithm processing, the boundary coordinates of the temperature difference mutation area can be obtained.

[0054] Step S230, spatially superimposing the pixel coordinate set of the edge of the ice surface crack with the boundary coordinates of the area where the temperature difference of the ice body suddenly changes, to generate polygonal mask data of the fragile area of ​​the ice surface structure.

[0055] In detail, in a three-dimensional spatial coordinate system, the pixel coordinates of the edge of the ice crack are mapped to the corresponding spatial position, and the boundary coordinates of the area where the temperature difference of the ice body suddenly changes are also located in the same spatial coordinate system. For example, a certain section of pixel coordinates at the edge of the ice crack is located between the coordinates (x1, y1, z1) and (x2, y2, z2), and the corresponding boundary coordinates of the area where the temperature difference of the ice body suddenly changes overlap or are adjacent to these coordinates. Through the above spatial superposition, the spatial information of the two is integrated to generate polygonal mask data of the fragile area of ​​the ice surface structure. The polygonal mask data can accurately represent the fragile area in the ice surface structure where there are both cracks and temperature anomalies. For example, the generated polygonal mask data may present an irregular polygonal shape, and the area covered by the polygonal shape is the fragile area of ​​the ice surface structure.

[0056] Step S240: calculating an acceleration change curve of the displacement rate in the area covered by the polygonal mask data according to the timestamp of the ice body displacement rate measurement data, and aligning the acceleration change curve with the occurrence time of the ice body rupture event.

[0057] For example, taking the measurement data of ice displacement rate in a certain time period as an example, it is assumed that in the area covered by the polygonal mask data, the displacement rate of the ice at a certain time t1 is v1, and after a period of time Δt, the displacement rate of the ice at time t2 is v2. According to the definition of acceleration, acceleration a=(v2-v1) / Δt. Through such calculation, the acceleration values ​​at different times are obtained. These acceleration values ​​are connected in time sequence to form an acceleration change curve of the displacement rate in the area covered by the polygonal mask data. Then, the acceleration change curve is aligned with the time of occurrence of the ice rupture event. For example, it is known that a certain ice rupture event occurred at time T. The time axis in the acceleration change curve is adjusted so that each data point in the acceleration change curve has an accurate time correspondence with the time T of the ice rupture event. This makes it easier to analyze the relationship between the change trend of acceleration before the ice ruptures and the ice rupture event.

[0058] Step S250, inputting the visible light image sequence, thermal infrared image sequence and corresponding polygonal mask data into a three-dimensional convolutional neural network to extract the deformation characteristics of the ice surface texture over time and the thermal diffusion characteristics of the fragile area.

[0059] For example, in a three-dimensional convolutional neural network, for a visible light image sequence, the convolution kernel performs convolution operations on the spatial dimension (x, y direction) and time dimension (t direction) of the image. Taking a frame of visible light image as an example, the convolution kernel starts from the upper left corner of the image and slides in sequence according to the set step size. The pixels in each sliding window are convoluted to extract features related to the deformation of the ice surface texture over time. For example, through the convolution operation, the deformation of the ice surface texture at different times, such as distortion and stretching, can be detected. For a thermal infrared image sequence, similarly, the convolution kernel performs convolution operations in three-dimensional space (including the time dimension) to extract the thermal diffusion characteristics of the fragile area. For example, in a thermal infrared image sequence, the thermal diffusion characteristics such as the direction, speed, and diffusion range of heat diffusion from the high temperature area to the low temperature area can be identified. At the same time, the polygonal mask data, as an auxiliary information, guides the network to pay more attention to the feature extraction of the fragile area of ​​the ice surface structure and enhances the feature capture capability of the key areas of the ice surface.

[0060] Step S260, splicing the deformation feature and the thermal diffusion feature in time series to generate an ice surface state evolution tensor, and aligning the ice surface state evolution tensor with the acceleration change curve in time step.

[0061] For example, suppose that at a certain time t, the deformation feature of the ice surface texture is F1, and the thermal diffusion feature of the fragile area is F2. F1 and F2 are combined in chronological order to form a composite feature containing ice surface texture and thermal diffusion information. Multiple such composite features are arranged in chronological order to generate the ice surface state evolution tensor. Then, the ice surface state evolution tensor is aligned with the previously obtained acceleration change curve in time step. Make sure that each time step in the ice surface state evolution tensor corresponds to the corresponding time step in the acceleration change curve, so that the relationship between the ice surface state evolution and the acceleration of the ice body displacement rate can be analyzed in the same time frame.

[0062] Step S270, using a bidirectional long short-term memory network to perform time series modeling on the aligned ice surface state evolution tensor, and output the predicted value of ice surface stability at each time step.

[0063] Taking a certain time step in the ice surface state evolution tensor as an example, in the forward propagation layer of the bidirectional long short-term memory network, starting from the initial moment, the information before the time step is processed in chronological order, and a hidden state is calculated. At the same time, in the backward propagation layer of the bidirectional long short-term memory network, starting from the last moment, the information after the time step is processed in reverse chronological order, and a hidden state is also calculated. The two hidden states are feature spliced ​​to obtain a comprehensive hidden state. For example, the hidden state calculated by the forward propagation layer is H1, and the hidden state calculated by the backward propagation layer is H2. H1 and H2 are spliced ​​together to obtain a new hidden state H. The hidden state H is input into the fully connected layer, and after the linear transformation of the fully connected layer, it is mapped to the initial probability distribution of the ice surface stability prediction value. Assume that the initial probability distribution is P=[p1, p2, p3], where p1, p2, and p3 represent the initial probabilities of the ice surface in the three states of stable, relatively stable, and unstable, respectively. The initial probability distribution is subjected to sliding average filtering in the time dimension to eliminate instantaneous noise interference. For example, for the initial probability distribution of several consecutive time steps, the corresponding average value is calculated to obtain a smoothed ice surface stability prediction value sequence, and each value in the ice surface stability prediction value sequence is used as the ice surface stability prediction value of the corresponding time step.

[0064] Step S280, performing loss calculation on the predicted value of ice surface stability and the actual predicted value of ice surface stability, and back-propagating to adjust the weight parameters of the three-dimensional convolutional neural network and the bidirectional long short-term memory network.

[0065] In this embodiment, the output ice surface stability prediction value is compared with the actual ice surface stability prediction value. For example, at a certain set time step, the ice surface stability prediction value predicted by the model is predicted to be in a relatively stable state with a probability of 0.6, while the actual ice surface stability measurement value indicates that the ice surface is in an unstable state. By calculating the difference between the two, such as using a cross entropy loss function to calculate the loss value. According to the loss value, using the back propagation algorithm, starting from the output layer, the weight parameters of the three-dimensional convolutional neural network and the bidirectional long short-term memory network are adjusted in sequence along the structure of the network toward the input layer. For example, in the bidirectional long short-term memory network, if a certain weight parameter contributes more to the loss value, the weight parameter is adjusted according to the set learning rate to reduce the loss value and improve the prediction accuracy of the model.

[0066] Step S290, during the training process, randomly mask the pixel blocks in the ice surface crack area in the visible light image sequence, and use the generative adversarial network to reconstruct the texture continuity of the masked area, and add the reconstructed image to the image training set to enhance the robustness of the model to local feature loss.

[0067] During the training process, for the visible light image sequence, randomly select pixel blocks in the ice crack area for masking. For example, in a certain frame of visible light image, select a pixel block in a rectangular area of ​​the ice crack area, and set the values ​​of these pixels to fixed values ​​(such as 0) to indicate masking. Then, the generative adversarial network is used to reconstruct the texture continuity of the masked area. The generator in the generative adversarial network tries to generate a texture similar to the original image to fill the masked area, while the discriminator judges the similarity between the generated image and the original image. After multiple iterations, the generator gradually learns the texture features of the ice crack area, and the generated image can better restore the texture continuity of the masked area. This reconstructed image is added to the image training set. In this way, in the subsequent training process, the adaptability to local feature loss can be enhanced in the case of such local feature loss, and the robustness of the model can be improved. For example, when the ice crack area is partially missing or blocked again, it is still possible to accurately extract features such as ice surface texture to predict ice surface stability.

[0068] For example, in a possible implementation, step S270 includes: Step S271, dividing the aligned ice surface state evolution tensor into continuous tensor subsequences according to the time step, each tensor subsequence contains the deformation characteristic sequence of the ice surface texture over time and the thermal diffusion characteristics of the fragile area at the current time step.

[0069] Taking the aligned ice surface state evolution tensor obtained by the previous processing as an example, it is assumed that the ice surface state evolution tensor covers the ice surface state information from the initial time t0 to the final time tn. According to the preset time step, such as one time step per hour, the ice surface state evolution tensor is divided. For a certain set time step t1, the corresponding tensor subsequence contains the deformation feature sequence of the ice surface texture over time and the thermal diffusion characteristics of the fragile zone under the time step. The deformation feature sequence of the ice surface texture over time records the changes in the ice surface texture from the previous several time steps to the current time step t1, such as how the ice surface texture is twisted, stretched or compressed. The thermal diffusion characteristics of the fragile zone reflect the propagation and diffusion of heat in the fragile zone within the same time range, including information such as the change in heat intensity and diffusion direction. Each tensor subsequence is such a set containing the ice surface texture and thermal diffusion information under the set time step. These tensor subsequences are arranged in chronological order and fully describe the evolution of the ice surface state over time.

[0070] Step S272, input the tensor subsequence into the forward propagation layer of the bidirectional long short-term memory network, calculate the hidden state of each time step in chronological order, and generate a forward hidden state sequence.

[0071] In this embodiment, starting from the initial tensor quantum sequence, the processing is carried out one by one in chronological order. Taking the first time step in the tensor quantum sequence as an example, in the forward propagation layer of the bidirectional long short-term memory network, the input ice surface texture deformation feature sequence over time and the thermal diffusion feature of the fragile zone are first processed, and the neurons in the network calculate the hidden state of the time step according to the pre-set weights and activation functions. For example, for a certain feature component in the deformation feature sequence of the ice surface texture over time, assuming that its value is x1, an intermediate result is obtained after multiplying it with the corresponding weight w1, and then the bias term b1 is added, and after the activation function f, a partial hidden state related to the feature component is obtained. Such calculations are performed on all feature components in the deformation feature sequence of the ice surface texture over time and the thermal diffusion feature of the fragile zone, and these partial hidden states are combined according to the set rules to obtain the hidden state h1 of the time step. Then, the subsequent time steps are processed in the same way to obtain a series of hidden states h1, h2, h3, etc. These hidden states are arranged in chronological order to form a forward hidden state sequence. This forward hidden state sequence reflects an internal representation of the ice surface state in a forward time sequence starting from the initial moment and over time, integrating the influence of the ice surface texture and thermal diffusion information at previous time steps.

[0072] Step S273, synchronously input the tensor subsequence into the backward propagation layer of the bidirectional long short-term memory network in reverse order, calculate the hidden state of each time step in reverse time order, and generate a backward hidden state sequence.

[0073] At the same time, the same tensor sequence is input in reverse order to the backward propagation layer of the bidirectional long short-term memory network. Starting from the last time step, the processing is carried out in reverse chronological order. For example, for the last time step tn in the tensor sequence, in the backward propagation layer of the bidirectional long short-term memory network, the deformation characteristic sequence of the ice surface texture over time and the thermal diffusion characteristics of the fragile area are also calculated. First, the characteristic components are operated similarly to the forward propagation layer, such as multiplying the value x2 of a certain characteristic component by the corresponding weight w2, adding the bias term b2, and after the action of the activation function g, a partial hidden state is obtained, and then all the partial hidden states are combined to obtain the hidden state hn of the time step. Then the previous time steps, such as tn-1, tn-2, etc., are processed in sequence to obtain a series of hidden states hn, hn-1, hn-2, etc. These hidden states are arranged in reverse chronological order to form a backward hidden state sequence. This backward hidden state sequence captures information about the state of the ice surface from the opposite time direction, reflecting an internal representation of the state of the ice surface in reverse time order when tracing back from the final moment, and includes information about the impact of subsequent time steps on the current time step.

[0074] Step S274, aligning the forward hidden state sequence and the backward hidden state sequence on a time axis, performing feature concatenation on the forward hidden state and the backward hidden state at the same time step, and generating a bidirectional fused hidden state sequence.

[0075] Since the forward hidden state sequence and the backward hidden state sequence are generated in different time orders, they need to be aligned on the time axis. Make sure that the first time step in the forward hidden state sequence corresponds to the last time step in the backward hidden state sequence, the second time step corresponds to the second-to-last time step, and so on. For example, the forward hidden state sequence is h1, h2, h3, and the backward hidden state sequence is hn, hn-1, hn-2. H1 corresponds to hn-2, h2 corresponds to hn-1, and h3 corresponds to hn. Then, for each pair of corresponding forward hidden states and backward hidden states, concatenate their features. Taking h1 and hn-2 as an example, assuming that h1 is a vector containing n1 elements and hn-2 is a vector containing n2 elements, concatenate these two vectors in order to obtain a new vector containing n1+n2 elements. Perform this operation on all corresponding hidden states to generate a bidirectional fused hidden state sequence. This bidirectional fusion hidden state sequence integrates the ice surface state information in both the forward and backward directions, and can more comprehensively describe the ice surface state at each time step.

[0076] Step S275, inputting the bidirectional fused hidden state sequence into a fully connected layer, mapping the fused hidden state of each time step to an initial probability distribution of the ice surface stability prediction value.

[0077] In this embodiment, the generated bidirectional fusion hidden state sequence can be input into the fully connected layer. The neurons in the fully connected layer have a connection relationship with each element in the bidirectional fusion hidden state sequence. For the fused hidden state of each time step in the bidirectional fusion hidden state sequence, the neurons in the fully connected layer process it according to their own weights and activation functions. For example, for the fused hidden state vector of a certain time step, each element thereof is multiplied by the weight of the neuron in the fully connected layer, and then the bias term is added. After the action of the activation function (such as the softmax function), the fused hidden state is mapped to the initial probability distribution of the ice surface stability prediction value. Assume that the initial probability distribution is a vector [p1, p2, p3] containing three elements, where p1 represents the initial probability that the ice surface is in a stable state, p2 represents the initial probability that the ice surface is in a relatively stable state, and p3 represents the initial probability that the ice surface is in an unstable state. The initial probability distribution reflects the preliminary prediction result of the ice surface stability based on the bidirectional fusion hidden state at the current time step.

[0078] Step S276, performing sliding average filtering on the time dimension on the initial probability distribution, eliminating instantaneous noise interference, and generating a smoothed ice surface stability prediction value sequence.

[0079] The initial probability distribution of the ice surface stability prediction value obtained at each time step is subjected to a sliding average filter in the time dimension. Taking three consecutive time steps as an example, it is assumed that the initial probability distribution at time step t1 is [p11, p21, p31], the initial probability distribution at time step t2 is [p12, p22, p32], and the initial probability distribution at time step t3 is [p13, p23, p33]. For the probability component p1 of the stable state, the result after the sliding average filter is (p11+p12+p13) / 3. Similarly, the probability component p2 of the relatively stable state and the probability component p3 of the unstable state are also calculated in this way to obtain the probability distribution after the sliding average filter. All time steps are subjected to sliding average filtering in this way to eliminate the noise interference caused by measurement errors or other instantaneous factors, thereby generating a smoothed ice surface stability prediction value sequence, which can more accurately reflect the changing trend of ice surface stability and avoid the influence of instantaneous noise on the prediction results.

[0080] Step S277, point-to-point matching of the smoothed ice surface stability prediction value sequence and the time step of the acceleration change curve is performed, and the error gradient between the prediction value of each time step and the actual ice surface stability measurement value is calculated.

[0081] Match the smoothed ice surface stability prediction value sequence with the previously obtained acceleration change curve point-to-point according to the time step. For example, the first time step in the smoothed ice surface stability prediction value sequence corresponds to the first time step in the acceleration change curve, the second time step corresponds to the second time step, and so on. For each time step, calculate the error gradient between the prediction value and the actual ice surface stability measurement value. Assume that at a certain time step t, the smoothed ice surface stability prediction value predicts that the ice surface is in a relatively stable state with a probability of 0.6, while the actual ice surface stability measurement value indicates that the ice surface is in an unstable state. Use cross entropy as the loss function to calculate the error gradient. First, construct a target probability distribution based on the actual ice surface stability measurement value, such as [0, 0, 1] (indicating an unstable state), and then calculate the error between the predicted probability distribution [0.6, 0.3, 0.1] and the target probability distribution [0, 0, 1] according to the cross entropy formula. This error is the error between the predicted value and the actual measurement value at this time step. By taking the partial derivative of this error with respect to the weight of the bidirectional long short-term memory network, the error gradient can be obtained. This calculation is performed for each time step to obtain the error gradient of the entire sequence, which reflects the degree of difference between the predicted value and the actual value and the changing trend of the error as the weight changes.

[0082] Step S278, reversely adjusting the forget gate weight and the input gate weight of the bidirectional long short-term memory network according to the error gradient, so that when the hidden state of the subsequent time step is updated, the deformation characteristics and thermal diffusion characteristics that are strongly correlated with the ice breakage event are preferentially retained.

[0083] According to the calculated error gradient, the back propagation algorithm is used to adjust the forget gate weight and input gate weight of the bidirectional long short-term memory network. The forget gate weight and input gate weight play a key role in updating the hidden state in the bidirectional long short-term memory network. Taking the forget gate weight as an example, assume that at a certain moment, the error gradient indicates that the current prediction result has a large deviation from the actual result, and the analysis finds that it is because some deformation features or thermal diffusion features that are strongly related to the ice breakup event are not effectively retained in the hidden state. According to the direction and size of the error gradient, the forget gate weight is appropriately increased or decreased. If the error gradient of a feature at the current time step indicates that the feature has an important impact on the ice breakup event, then the forget gate weight is adjusted so that when the hidden state is updated in the subsequent time step, the probability of retaining the feature increases. A similar adjustment is also made to the input gate weight, so that when updating the hidden state, it is more inclined to input deformation features and thermal diffusion features that are strongly related to the ice breakup event, thereby improving the accuracy of the model's prediction of ice surface stability.

[0084] Step S279, during the time series modeling process, when it is detected that the difference between the predicted value of the ice surface stability at the current time step and the predicted value at the previous time step exceeds a preset mutation threshold, the cell state of the backward propagation layer is forcibly reset to avoid interference of historically irrelevant features with the prediction of the current time step.

[0085] During the time series modeling process, the predicted value of ice surface stability at each time step is continuously monitored. Assuming that the preset mutation threshold is 0.3, when it is found that the difference between the predicted value of ice surface stability at the current time step t and the predicted value at the previous time step t-1 exceeds 0.3, for example, the ice surface is predicted to be in a relatively stable state at time step t-1 with a probability of 0.6, while the ice surface is predicted to be in an unstable state at time step t with a probability of 0.8. In this case, the cell state of the backward propagation layer is forced to be reset. The cell state of the backward propagation layer contains the information of the previous time step. When such a mutation occurs, the previous information may interfere with the prediction of the current time step. By resetting the cell state, these historically irrelevant features that may interfere with the current prediction can be cleared, so that the backward propagation layer can start again from the current time step and accurately calculate the hidden state in reverse chronological order, thereby improving the accuracy of the prediction of the current time step.

[0086] Step S2710, output the predicted value of ice surface stability at each time step, and spatially correlate and map the time series change rate of the predicted value with the thermal distribution data of the ice surface displacement rate to verify the spatial consistency of the predicted value.

[0087] Finally, the predicted values ​​of ice surface stability at each time step are output, and these predicted values ​​of ice surface stability reflect the stability of the ice surface at different time steps. Then, the time series change rate of these predicted values ​​of ice surface stability is spatially correlated with the thermal distribution data of the ice surface displacement rate. The thermal distribution data of the ice surface displacement rate reflects the displacement rate and thermal distribution of the ice surface at different positions. For example, the displacement rate of the ice surface in a certain area is faster, and the thermal distribution of the area also has a set pattern. The time series change rate of the predicted value of ice surface stability is correlated with the ice surface displacement rate and thermal distribution data at these spatial positions. If the time series change rate of the predicted value shows a reasonable correlation with the thermal distribution data of the ice surface displacement rate in space, for example, in areas with fast ice surface displacement rate and abnormal thermal distribution, the time series change rate of the predicted value of ice surface stability also shows an unstable trend, then the spatial consistency of the predicted value is verified. This verification of spatial consistency can further confirm the rationality of the model prediction results and ensure that the model can accurately reflect the stability state of the ice surface in space and time.

[0088] In a possible implementation, step S250 includes: Step S251 , performing pixel normalization processing on the visible light image sequence frame by frame to generate a visible light normalization tensor, and performing temperature calibration processing on the thermal infrared image sequence to generate a thermal infrared temperature tensor.

[0089] For the visible light image sequence acquired from the target glacier area, pixel normalization is performed frame by frame. Taking one frame of visible light image as an example, each pixel in the image has a set numerical range, which may vary due to factors such as lighting conditions and sensor characteristics. In order to make the pixel values ​​between different frames comparable, pixel normalization is performed. Assuming that the original numerical range of the pixels in the image is from 0 to 255, it is converted to a new numerical range, such as between 0 and 1, through the existing normalization algorithm. This operation is performed for each frame of visible light image, and all normalized frames constitute a visible light normalized tensor. For the thermal infrared image sequence, temperature calibration is performed. Since thermal infrared images reflect the temperature information of the ice surface, the numerical representation of temperature may be different under different sensors or acquisition conditions. Through temperature calibration, the temperature value in the thermal infrared image is converted into a unified numerical representation with clear physical meaning. For example, the temperature value in the original thermal infrared image is represented by a sensor-specific code. After temperature calibration, it is converted into a temperature value in degrees Celsius. In this way, all thermal infrared images constitute a thermal infrared temperature tensor.

[0090] Step S252, aligning the visible light normalized tensor and the thermal infrared temperature tensor according to the timestamp, correcting the sensor viewing angle offset by the optical flow method, and generating a visible light tensor and thermal infrared tensor that are aligned in time and space.

[0091] Both the visible light normalized tensor and the thermal infrared temperature tensor contain data collected in chronological order. Since the glacier area is collected at the same time, the corresponding timestamps are hidden. Taking the collected data at a certain moment as an example, the normalized tensor of the visible light image and the temperature tensor of the thermal infrared image corresponding to the moment should reflect the different characteristics of the glacier at the same moment. However, in the actual collection process, there may be a sensor perspective offset due to the installation position of the sensor or slight shaking during the collection process. The optical flow method is used to correct this perspective offset. The optical flow method estimates the change in perspective by analyzing the movement of pixels in the image. For example, in two adjacent frames of visible light images, the position of some feature points (such as set marks on the ice surface or stable geomorphic features) in the image has changed. According to the displacement of these feature points, the offset of the perspective can be calculated, and then the offset can be applied to the thermal infrared image, so that the visible light normalized tensor and the thermal infrared temperature tensor are also aligned in space, and finally a visible light tensor and thermal infrared tensor aligned in time and space are generated. The two tensors not only correspond in time, but also accurately reflect the glacier characteristics of the same area in space.

[0092] Step S253, converting the polygonal mask data into a binary weight matrix, wherein pixels in the fragile area of ​​the ice surface covered by the mask are assigned a first weight value, and pixels in the non-fragile area are assigned a second weight value.

[0093] In this embodiment, the polygonal mask data can be converted into a binary weight matrix. For example, for pixels in the fragile area of ​​the ice surface, a first weight value is assigned, assuming that the first weight value is 1, indicating that these pixels have a higher importance in subsequent calculations. For pixels in the non-fragile area, a second weight value is assigned, assuming that the second weight value is 0, indicating a relatively low importance. In this way, the entire polygonal mask data is converted into a binary weight matrix, each element in the binary weight matrix corresponds to a pixel in the ice surface image, and the fragile area and the non-fragile area of ​​the ice surface are distinguished by different weight values.

[0094] Step S254, multiplying the spatiotemporally aligned visible light tensor by the binarized weight matrix pixel by pixel to generate a visible light weighted tensor, and multiplying the thermal infrared tensor by the binarized weight matrix to generate a thermal infrared weighted tensor.

[0095] For the spatiotemporally aligned visible light tensor, it is multiplied pixel by pixel with the binarized weight matrix. Taking a frame of image in the visible light tensor as an example, each pixel in this frame is multiplied by the corresponding element in the binarized weight matrix. If a pixel is located in the fragile zone of the ice surface, its corresponding weight value is 1, then the result of the multiplication is the original value of the pixel; if it is located in the non-fragile zone, the weight value is 0, and the result of the multiplication is 0. Performing this operation on each pixel of the entire frame of the image will result in a new frame of image, and all frame images processed in this way constitute the visible light weighted tensor. Similarly, for the thermal infrared tensor, it is also multiplied pixel by pixel with the binarized weight matrix to obtain the thermal infrared weighted tensor.

[0096] Step S255, inputting the visible light weighted tensor into the first branch of the three-dimensional convolutional neural network, extracting the displacement gradient of the ice surface cracks between adjacent frames through the three-dimensional convolutional layer, and generating an ice surface texture deformation feature map.

[0097] In this embodiment, in the first branch of the three-dimensional convolutional neural network, the three-dimensional convolutional layer begins to process the visible light weighted tensor. Taking two adjacent frames of visible light weighted images as an example, the convolution kernel of the three-dimensional convolutional layer performs convolution operations in the spatial dimension (x, y direction) and the time dimension (t direction). The convolution kernel starts from the upper left corner of the image and slides in sequence according to the set step size to perform convolution calculations on the pixels in each sliding window. In this process, by analyzing the changes in pixel values ​​between adjacent frames, the displacement information of the ice cracks can be extracted. For example, in a certain area, if the edge pixels of the ice cracks in the previous frame image are displaced from the corresponding pixels in the next frame image, the convolutional layer can detect this displacement and calculate the size and direction of the displacement, thereby obtaining the displacement gradient of the ice cracks between adjacent frames. This operation is performed on all adjacent frames in the entire visible light weighted tensor, and finally an ice surface texture deformation feature map is generated, which intuitively shows the deformation of the ice surface texture at different time and space positions, especially the displacement change of the ice surface cracks.

[0098] Step S256, performing time axis sliding window segmentation on the ice surface texture deformation feature map, extracting the mean and variance of the crack displacement gradient in each time window, and generating a deformation feature sequence of the ice surface texture over time.

[0099] For the generated ice surface texture deformation feature map, the time axis sliding window is segmented. Assume that the size of the time window is set to 3 time steps. Starting from the starting position of the time axis, the first time window contains the ice surface texture deformation feature map parts corresponding to the 1st, 2nd, and 3rd time steps. In this time window, the displacement gradients of all ice surface cracks are statistically analyzed. The mean of the displacement gradient of each ice surface crack is calculated, that is, the values ​​of the displacement gradients of all ice surface cracks in the time window are added, and then divided by the number of cracks. For example, there are 5 ice surface cracks in the time window, and the corresponding displacement gradients are 0.1, 0.2, 0.15, 0.18, and 0.22, respectively. These values ​​are added to get 1.05, and then divided by 5 to get the mean of 0.21. At the same time, the variance of the displacement gradient is calculated. The variance is calculated by first finding the square of the difference between each displacement gradient and the mean. For example, for a crack with a displacement gradient of 0.1, its difference with the mean of 0.21 is -0.11, and the square of the difference is 0.0121. This calculation is performed for all cracks, and then the squares of these differences are added and divided by the number of cracks to obtain the variance. Through this calculation, the mean and variance of the crack displacement gradient in the time window are obtained. Then, in chronological order, the sliding window is moved back one time step in turn, and the same mean and variance calculation is performed for each new time window, and finally a deformation feature sequence of the ice surface texture over time is generated. This deformation feature sequence contains the deformation features of the ice surface texture over time in different time windows, and reflects the changes in the displacement gradient of the ice surface cracks in the form of mean and variance.

[0100] Step S257, inputting the thermal infrared weighted tensor into the second branch of the three-dimensional convolutional neural network, extracting the diffusion direction and fluctuation intensity of the temperature field in the fragile area by expanding the three-dimensional convolutional layer, and generating a thermal diffusion characteristic map.

[0101] The thermal infrared weighted tensor is input into the dilated three-dimensional convolution layer in the second branch of the three-dimensional convolutional neural network. Taking the thermal infrared weighted tensor at a certain moment as an example, the convolution kernel of the dilated three-dimensional convolution layer performs convolution operation in three-dimensional space (including time dimension). By analyzing the changes of temperature values ​​in the thermal infrared weighted tensor at different spatial positions and time steps, the information of the temperature field in the fragile zone can be extracted. For example, in the fragile zone, the convolution layer can detect the propagation direction of temperature from the high temperature area to the low temperature area, that is, the diffusion direction of the temperature field. At the same time, by analyzing the fluctuation of temperature at different time steps, such as the amplitude change of temperature rising first and then falling in a certain area, the fluctuation intensity of the temperature field is calculated. The entire thermal infrared weighted tensor is processed in this way to generate a thermal diffusion feature map, which shows the diffusion direction of the temperature field in the fragile zone and the distribution of the fluctuation intensity at different spatial positions and time steps.

[0102] Step S258, performing a spatial pooling operation on the thermal diffusion feature map in the time dimension, extracting the temperature change extreme value of each spatial position in the continuous time step, and generating the thermal diffusion feature of the fragile area.

[0103] Perform spatial pooling operation on the time dimension for the generated thermal diffusion feature map. Take a certain spatial position as an example. In continuous time steps, the temperature value of the spatial position will change. Through spatial pooling operation, statistical analysis is performed on the temperature value of the spatial position in the time dimension. Find the highest and lowest temperature values ​​of the spatial position in all time steps. These two temperature values ​​are the extreme values ​​of temperature change at the spatial position. For example, at this spatial position, after 5 consecutive time steps of observation, the temperature values ​​are -2℃, -1.5℃, -1℃, -2.5℃, and -3℃, respectively. Then the highest temperature value is -1℃ and the lowest temperature value is -3℃. Perform such operation on each spatial position in the thermal diffusion feature map, and finally generate the thermal diffusion feature of the fragile zone. The thermal diffusion feature of the fragile zone contains the extreme value information of temperature change at each spatial position in the fragile zone in continuous time steps, which can reflect the characteristics of thermal diffusion in the fragile zone, such as the range of temperature fluctuation.

[0104] Step S259, outputting the deformation feature sequence of the ice surface texture over time and the thermal diffusion characteristics of the fragile area as feature extraction results of the three-dimensional convolutional neural network.

[0105] Finally, the previously calculated deformation feature sequence of ice surface texture over time and the thermal diffusion characteristics of the fragile zone are output as the feature extraction results of the three-dimensional convolutional neural network. These feature results contain important information related to the ice surface state extracted from the visible light image sequence and thermal infrared image sequence. The deformation feature sequence of ice surface texture over time reflects the deformation of ice surface texture, especially ice cracks, in time and space, while the thermal diffusion characteristics of the fragile zone reflect the diffusion and fluctuation characteristics of the temperature field in the fragile zone. These features will be used for subsequent operations such as ice surface stability analysis.

[0106] In a possible implementation, step S130 includes: Step S131, dividing the ice surface stability assessment parameters into subsets of continuous time series according to a preset time window, normalizing the parameters in each subset and splicing them into a multi-dimensional feature vector.

[0107] Taking the ice surface stability assessment parameters previously obtained from monitoring in large glacier areas as an example, assume that the preset time window is one day. The entire ice surface stability assessment parameters are divided according to the time window of one day, thereby obtaining multiple subsets of continuous time series. For the ice surface stability assessment parameters in each subset, since these parameters may have different numerical ranges and dimensions, normalization is required to facilitate subsequent unified processing and analysis. For example, a subset contains stability values ​​of different areas of the ice surface, which may be between 0 and 100. Through the existing normalization algorithm, it is converted to the range of 0 to 1. After all parameters in each subset are normalized, these normalized parameters are spliced ​​together in the set order to form a multidimensional feature vector, which integrates the ice surface stability related information in each time window.

[0108] Step S132, performing time series decomposition on the multi-dimensional feature vector, extracting the acceleration extreme value of the ice surface displacement rate and the mutation point position of the crack expansion rate, and generating time domain abnormal fluctuation characteristics.

[0109] For the spliced ​​multidimensional feature vectors, a time series decomposition operation is performed. Taking the ice surface displacement rate as an example, by analyzing the part of the multidimensional feature vector related to the ice surface displacement rate, its changes at different time points are determined. In this process, the acceleration extreme value of the ice surface displacement rate is found. For example, in the time period corresponding to a certain subset, the acceleration of the ice surface displacement rate will change greatly at certain moments. By comparing the acceleration values ​​at different time points, the maximum and minimum values ​​are found, which are the acceleration extreme values ​​of the ice surface displacement rate. At the same time, for the crack expansion rate, its changes in the time series are analyzed to determine the location of the mutation point of the crack expansion rate. For example, at continuous time points, the crack expansion rate has remained relatively stable, but suddenly increased or decreased at a certain moment, and this moment is the mutation point of the crack expansion rate. The information such as the acceleration extreme value of the ice surface displacement rate and the location of the mutation point of the crack expansion rate are combined to generate the time domain abnormal fluctuation characteristics. These time domain abnormal fluctuation characteristics reflect the unstable factors of the ice surface in the time dimension and are an important basis for judging the ice surface risk area.

[0110] Step S133, based on the spatial distribution density of the time domain abnormal fluctuation characteristics, a density clustering algorithm is used to identify the geometric center coordinates and radiation radius of the ice surface risk area.

[0111] According to the generated time-domain abnormal fluctuation characteristics, the spatial distribution density of the characteristics in the glacier area is considered. For example, in some areas of the glacier, the time-domain abnormal fluctuation characteristics are relatively concentrated, while in other areas they are relatively sparse. The density clustering algorithm is used to perform cluster analysis on these time-domain abnormal fluctuation characteristics. Taking a certain area as an example, when the number of time-domain abnormal fluctuation characteristics in the area reaches the set density, the algorithm identifies it as a potential ice surface risk area. In this process, the geometric center coordinates of the ice surface risk area are determined. The geometric center coordinates are determined by calculating the average coordinates of all time-domain abnormal fluctuation feature points in the ice surface risk area. For example, there are several time-domain abnormal fluctuation feature points in the area, and their coordinates are (x1, y1, z1), (x2, y2, z2), etc. The x values ​​of these coordinates are added and divided by the number of feature points to obtain the x coordinate of the geometric center, and the y coordinate and z coordinate are obtained in the same way. At the same time, the radiation radius is determined according to the distribution range of the time-domain abnormal fluctuation characteristics in the ice surface risk area. For example, the distance from the time-domain abnormal fluctuation feature point farthest from the geometric center in the area to the geometric center is calculated, and this distance is the radiation radius.

[0112] Step S134, mapping the geometric center coordinates to the spatial position set of ice breakage events in the historical disaster database, and calculating the spatial overlap and displacement trend similarity between the current risk area and the historical breakage area.

[0113] Map the geometric center coordinates of the identified ice surface risk area to the spatial position set of ice breakage events in the historical disaster database. Take an ice breakage event in the historical disaster database as an example, which records the spatial position information of the ice breakage area. Calculate the spatial overlap between the current ice surface risk area and the historical breakage area. For example, the current ice surface risk area is regarded as a geometric shape in three-dimensional space (such as a sphere, determined by the geometric center coordinates and the radiation radius), and the historical breakage area is also regarded as a similar geometric shape. Calculate the ratio of the volume of the intersection of the two geometric shapes to the volume of the current ice surface risk area, and this ratio is the spatial overlap. At the same time, analyze the similarity of the displacement trend between the current ice surface risk area and the historical breakage area. The displacement trend similarity is determined by comparing the displacement rate change trends of the ice in different directions in the two areas. For example, in a certain direction, the displacement rate of the ice in the current ice surface risk area is gradually increasing, and the ice in the historical breakage area in the same direction also has a similar displacement rate increase trend before breaking. The displacement trend similarity is calculated based on the similarity of this trend.

[0114] Step S135, adjusting the dynamic threshold interval of the preset risk level according to the spatial overlap and the displacement trend similarity, and performing weighted scoring on the time domain abnormal fluctuation characteristics based on the dynamic threshold interval.

[0115] According to the calculated spatial overlap and displacement trend similarity, the dynamic threshold interval of the preset risk level is adjusted. For example, the threshold interval of the ice surface stability assessment parameter corresponding to the preset low risk level is originally 0.6 to 1. If the spatial overlap between the current ice surface risk area and the historical rupture area is high and the displacement trend similarity is also high, the threshold interval is adjusted to 0.7 to 1. Then, the time domain abnormal fluctuation characteristics are weighted and scored based on the adjusted dynamic threshold interval. For example, for the acceleration extreme value of the ice surface displacement rate in the time domain abnormal fluctuation characteristics, if the acceleration extreme value is within the adjusted threshold interval, the corresponding weight score is given according to its specific position in the interval. Assume that the closer the acceleration extreme value is to the lower limit of the threshold interval, the lower the weight score is given, and the closer it is to the upper limit, the higher the weight score is given. Other time domain abnormal fluctuation characteristics such as the position of the mutation point of the crack expansion rate are also weighted and scored in a similar way, and these weighted scores are combined to obtain a weighted scoring result for the time domain abnormal fluctuation characteristics.

[0116] Step S136, inputting the weighted scoring result into a preset fuzzy logic classifier, matching the membership function corresponding to the ice surface risk level label, and generating an independent risk level for each risk area and a global ice surface comprehensive risk level.

[0117] The weighted scoring results are input into a preset fuzzy logic classifier, in which the membership function corresponding to the ice surface risk level label is pre-set. For example, for the low risk level, there is a membership function, which specifies the degree of belonging to the low risk level corresponding to different weighted scoring results. When the weighted scoring results are input, the degree of belonging to the low risk level is calculated according to the membership function. Similarly, there are corresponding membership functions for the medium risk level and the high risk level. In this way, the degree to which each risk area belongs to different risk levels is calculated respectively, so as to determine the independent risk level of each risk area. At the same time, the situation of all risk areas is comprehensively considered to generate a global comprehensive risk level of the ice surface. For example, if most risk areas are judged to be at a higher risk level, then the global comprehensive risk level of the ice surface may also be at a higher risk level.

[0118] Step S137, performing a confidence check on the independent risk level and the global ice surface comprehensive risk level, eliminating the false detection areas with confidence levels lower than a preset threshold, and updating the spatial distribution matrix of the ice surface risk level labels.

[0119] Perform confidence checks on the independent risk levels and global ice surface comprehensive risk levels obtained. For example, the preset confidence threshold is 0.6. For the independent risk level of a risk area, if its confidence is lower than 0.6, it means that there may be errors in the judgment of the risk area, and it is regarded as a false detection area and eliminated. A similar check is also performed on the global ice surface comprehensive risk level. After the false detection areas are eliminated, the spatial distribution matrix of the ice surface risk level labels is updated. The spatial distribution matrix records the risk level label information of each ice surface area.

[0120] The matching process between the spatial distribution matrix and the warning trigger threshold adopts a sliding window traversal strategy, and the window size is dynamically adjusted according to the real-time monitoring value of the ice body displacement rate.

[0121] In the process of matching the spatial distribution matrix with the warning trigger threshold, a sliding window traversal strategy is adopted. The window size is dynamically adjusted based on the real-time monitoring value of the ice displacement rate. For example, when the ice displacement rate is fast, it means that the ice surface state changes more dramatically. At this time, the size of the sliding window is appropriately increased to more comprehensively consider the risk of the ice surface. Conversely, when the ice displacement rate is slow, the size of the sliding window is appropriately reduced to improve the matching accuracy. Through this sliding window traversal strategy of dynamically adjusting the window size, the spatial distribution matrix and the warning trigger threshold can be matched more accurately, so as to better judge whether the ice surface meets the warning conditions.

[0122] In a possible implementation, step S132 includes: Step S1321 : Slidingly divide the multidimensional feature vector into a plurality of subsequence segments on the time axis at a fixed step length, and perform decomposition processing of trend terms and residual terms on each subsequence segment.

[0123] Taking the multidimensional feature vector of a large glacier area obtained previously as an example, assume that a fixed step size of one hour is set on the time axis. Starting from the starting time of the multidimensional feature vector, the entire multidimensional feature vector is slid and divided into multiple subsequence segments according to the fixed step size of each hour. For each subsequence segment, the trend term and the residual term are decomposed. Taking a certain subsequence segment as an example, the data in the subsequence segment is decomposed into a trend term and a residual term through the existing time series analysis method. The trend term represents the main change trend of the data in the subsequence segment, for example, the ice surface displacement rate tends to gradually increase or decrease as a whole within this hour. The residual term is the difference between the actual data and the trend term, which contains the fluctuation information in the data.

[0124] Step S1322, detecting the zero-crossing point of the second-order derivative of the ice surface displacement rate in the residual term, marking the moment of acceleration direction reversal and calculating the acceleration integral area between adjacent reversal moments.

[0125] Among the residual terms that have been obtained, focus on the part related to the displacement rate of the ice surface. The second-order derivative of the displacement rate of the ice surface reflects the rate of change of acceleration. By calculating the second-order derivative of the displacement rate of the ice surface in the residual term, find the moment when its second-order derivative crosses the zero point. For example, at a certain moment t1, the second-order derivative of the displacement rate of the ice surface changes from a positive value to a negative value. This moment is the moment when the acceleration direction is reversed. After marking these acceleration direction reversal moments, calculate the acceleration integral area between adjacent reversal moments. Taking t1 and the next acceleration direction reversal moment t2 as an example, calculate the integral area of ​​the acceleration in the time period from t1 to t2. The calculation of this integral area is obtained by accumulating the values ​​of the acceleration in this time period. Assuming that there are several time points between t1 and t2, and the acceleration values ​​of each time point are a1, a2, a3, etc., these acceleration values ​​are accumulated according to the time interval, and the result is the acceleration integral area, which reflects the cumulative effect of the acceleration of the displacement rate of the ice surface in this time period.

[0126] Step S1323, threshold filtering is performed on the acceleration integrated area, subsequence segments exceeding a preset acceleration threshold are retained, a timestamp set of acceleration extreme values ​​is generated, and a wavelet transform is synchronously performed on the residual term of the crack propagation rate to extract the energy peak point of the high-frequency component and mark it as a crack propagation mutation point.

[0127] In this embodiment, the acceleration threshold is a specific value, which can be determined based on historical data and the actual characteristics of the glacier. The acceleration integral area calculated for each subsequence segment is subjected to threshold filtering. If the acceleration integral area of ​​a subsequence segment exceeds the preset acceleration threshold, the subsequence segment is retained. For example, among multiple subsequence segments, the acceleration integral area of ​​subsequence segment 1 exceeds the threshold, while the acceleration integral area of ​​subsequence segment 2 does not exceed the threshold, then subsequence segment 1 is retained. The time corresponding to all retained subsequence segments is marked, and these time marks constitute a timestamp set of acceleration extremes. At the same time, a wavelet transform is performed on the residual term of the crack expansion rate. The wavelet transform can decompose the residual term of the crack expansion rate into components of different frequencies. Among these components, the high-frequency component contains the mutation information of the crack expansion rate. By analyzing the energy distribution of the high-frequency component, the energy peak point is found. For example, in the energy distribution curve of the high-frequency component, the energy value corresponding to a certain moment t3 is the maximum value in the entire curve. The moment t3 is the energy peak point of the high-frequency component, and the energy peak point is marked as the crack expansion mutation point.

[0128] Step S1324, aligning and matching the timestamp set of the acceleration extreme value with the timestamp of the crack extension mutation point, and calculating the standard deviation and co-occurrence probability of the time interval between the two.

[0129] Align and match the timestamp set of the acceleration extreme value and the timestamp of the crack extension mutation point. For example, the timestamp set of the acceleration extreme value contains timestamps t1, t2, etc., and the timestamp set of the crack extension mutation point contains timestamps t3, t4, etc. These timestamps are aligned in chronological order. Then calculate the standard deviation of the time intervals between the two. Taking t1 and t3 as an example, calculate the time interval d1 between them, calculate such time intervals for all corresponding timestamp pairs, and then calculate the standard deviation of these time intervals according to the standard deviation calculation formula. At the same time, calculate the co-occurrence probability. The co-occurrence probability refers to the probability that the timestamp of the acceleration extreme value and the timestamp of the crack extension mutation point appear at the same time or appear successively in a short period of time. For example, count the proportion of the number of timestamp pairs with a time interval less than a set value (such as 10 minutes) in the total number of timestamp pairs among all timestamp pairs, and this proportion is the co-occurrence probability.

[0130] Step S1325: Probability weighting is performed on the time domain abnormal fluctuation intensity in each subsequence segment according to the standard deviation and the co-occurrence probability to generate a time domain abnormal fluctuation feature vector with a timestamp. The time resolution of the time domain abnormal fluctuation feature vector is consistent with the update frequency of the three-dimensional spatiotemporal features of the ice surface state feature set.

[0131] According to the calculated standard deviation and co-occurrence probability, the time domain abnormal fluctuation intensity in each subsequence segment is probability weighted. For example, the time domain abnormal fluctuation intensity of a subsequence segment is I. According to the standard deviation and co-occurrence probability, a weighting coefficient w is calculated through the existing weighting algorithm. The time domain abnormal fluctuation intensity I is multiplied by the weighting coefficient w to obtain the weighted time domain abnormal fluctuation intensity Iw. This operation is performed on each subsequence segment, and then these weighted time domain abnormal fluctuation intensities are combined with the corresponding timestamps to generate a time domain abnormal fluctuation feature vector marked with a timestamp. The time resolution of the time domain abnormal fluctuation feature vector is the same as the update frequency of the three-dimensional spatiotemporal characteristics of the ice surface state feature set, for example, they are updated once every hour. This ensures that in the subsequent analysis, the two can be accurately matched in time so as to comprehensively consider various state information of the ice surface.

[0132] In a possible implementation, step S140 includes: Step S141, collecting environmental noise data of the target glacier area in real time, wherein the environmental noise data includes glacier vibration signals, wind speed fluctuation frequency, and precipitation pulse intensity.

[0133] In large glacier areas, special sensors are set up to collect environmental noise data in real time. For glacier vibration signals, they are obtained through vibration sensors installed at different locations on the glacier. For example, vibration sensors are installed at key locations such as the edge and center of the glacier. These sensors can accurately detect the tiny vibrations of the glacier and record the intensity and frequency of the vibrations. At the same time, wind speed sensors are installed to obtain the wind speed fluctuation frequency. The wind speed sensor can measure the wind speed in the glacier area in real time and record the changes in wind speed over time to obtain the wind speed fluctuation frequency. For precipitation pulse intensity, it is collected through rain gauges and other equipment. When precipitation occurs, the rain gauge can measure the amount of precipitation per unit time and determine the precipitation pulse intensity based on the changes in precipitation. These environmental noise data (glacier vibration signals, wind speed fluctuation frequency, precipitation pulse intensity) can reflect the changes in the external environment of the glacier area, and these changes may affect the accuracy of the ice landslide monitoring model, so the model parameter weights need to be dynamically adjusted.

[0134] Step S142, converting the environmental noise data into a frequency spectrum feature vector, and inputting the frequency spectrum feature vector into the noise suppression layer in the ice landslide monitoring model for feature dimension reduction, and then weightedly fusing the frequency spectrum feature vector after dimension reduction with the current model parameters through an online learning algorithm to generate a parameter adjustment gradient.

[0135] First, the collected environmental noise data (glacier vibration signal, wind speed fluctuation frequency, precipitation pulse intensity) are converted into spectral feature vectors. Taking the glacier vibration signal as an example, the vibration signal is converted from a time domain signal to a frequency domain signal by Fourier transform, and the spectral characteristics of the vibration signal are obtained. Similarly, the wind speed fluctuation frequency and precipitation pulse intensity are also converted accordingly, and the spectral characteristics of these three types of environmental noise data are combined to form a spectral feature vector. Then the spectral feature vector is input into the noise suppression layer in the ice landslide monitoring model. In the noise suppression layer, the spectral feature vector is reduced in dimension by the existing dimensionality reduction algorithm. For example, the principal component analysis (PCA) algorithm is used to reduce multiple feature dimensions in the spectral feature vector to a few main dimensions, which can retain most of the information of the original spectral feature vector. After feature dimensionality reduction, the reduced spectral feature vector is weightedly fused with the current model parameters through an online learning algorithm. The online learning algorithm assigns different weights to the reduced spectral feature vector and the current model parameters according to pre-set rules, and then performs weighted summation and other operations on them according to the weights to generate parameter adjustment gradients, which reflect the direction and magnitude of adjustment of the model parameters according to the environmental noise data.

[0136] Step S143, updating the convolution kernel size and channel attention weight of the multi-scale feature fusion layer according to the parameter adjustment gradient, wherein the adjustment range of the convolution kernel size is positively correlated with the rising rate of the ice surface risk level label.

[0137] The convolution kernel size and channel attention weight of the multi-scale feature fusion layer are updated according to the generated parameter adjustment gradient. Taking the convolution kernel size as an example, if the parameter adjustment gradient indicates that the convolution kernel size needs to be increased, then the size parameters such as the side length of the convolution kernel are increased according to the set rules. At the same time, the channel attention weight is also updated accordingly. Among them, the adjustment range of the convolution kernel size is positively correlated with the rising rate of the ice surface risk level label. For example, when the rising rate of the ice surface risk level label is fast, it means that the ice surface state changes greatly, and the model needs to adapt to this change more quickly. At this time, the adjustment range of the convolution kernel size is large. If the ice surface risk level label rises rapidly from a low risk level to a high risk level, the convolution kernel size of the multi-scale feature fusion layer may increase significantly, and the channel attention weight will also be adjusted accordingly, so that the model can better focus on the key features related to ice landslides and improve the model's ability to monitor changes in ice surface states.

[0138] In a possible implementation, step S142 includes: Step S1421, constructing a mapping relationship table between the environmental noise data and changes in model parameters in historical disaster events, wherein the mapping relationship table includes model learning rate adjustment coefficients corresponding to different noise intensity intervals.

[0139] Taking the glacier vibration signal as an example, according to the data from historical disaster events, when the intensity of the glacier vibration signal is within a certain range, the adjustment of the model parameters at that time is recorded, especially the adjustment coefficient of the model learning rate. Similarly, a similar analysis is performed on the wind speed fluctuation frequency and precipitation pulse intensity, and a mapping relationship table containing different noise intensity intervals (glacier vibration signal intensity interval, wind speed fluctuation frequency interval, precipitation pulse intensity interval) and the corresponding model learning rate adjustment coefficient is constructed. For example, when the glacier vibration signal intensity is in the low intensity interval, the model learning rate adjustment coefficient is a1; when the vibration signal intensity is in the medium intensity interval, the model learning rate adjustment coefficient is a2; when the vibration signal intensity is in the high intensity interval, the model learning rate adjustment coefficient is a3. There are also similar model learning rate adjustment coefficients corresponding to different intervals for wind speed fluctuation frequency and precipitation pulse intensity.

[0140] Step S1422, matching the optimal learning rate adjustment coefficient from the mapping relationship table according to the amplitude distribution of the frequency spectrum feature vector after dimension reduction.

[0141] Take the spectral component corresponding to the glacier vibration signal in the spectral feature vector as an example, and observe the distribution of its amplitude. If the amplitude is distributed in the low-intensity range, according to the previously constructed mapping relationship table, the corresponding model learning rate adjustment coefficient is matched to a1. Similarly, a similar analysis is performed on the spectral components corresponding to the wind speed fluctuation frequency and precipitation pulse intensity in the spectral feature vector. After comprehensive consideration, an optimal learning rate adjustment coefficient is determined. This optimal learning rate adjustment coefficient is most suitable for the current environmental noise situation, which can make the model more reasonable and accurate when adjusting parameters according to environmental noise.

[0142] Step S1423, calculating the iteration step size and direction vector of the parameter adjustment gradient based on the optimal learning rate adjustment coefficient, wherein the direction vector is used to constrain the parameter update path of the online learning algorithm to prevent the model weight from deviating from the core feature dimension of disaster warning.

[0143] For example, assuming that the optimal learning rate adjustment coefficient is α, the iterative step size of the parameter adjustment gradient is calculated according to the formula of the online learning algorithm. The iterative step size determines the magnitude of each adjustment of the model parameters. At the same time, the direction vector is calculated. The calculation of the direction vector is based on the relationship between the environmental noise data and the model parameters and the optimal learning rate adjustment coefficient. The role of the direction vector is to constrain the parameter update path of the online learning algorithm. For example, during the model parameter update process, if there is no constraint of the direction vector, the model weights may deviate from the core feature dimensions related to disaster warning, such as the model weights corresponding to key features such as ice surface displacement rate and crack expansion. Through the constraint of the direction vector, the model parameters are always adjusted in a direction that is conducive to accurate disaster warning during the update process, ensuring that the model can better adapt to environmental changes in the glacier area and accurately monitor the state of the ice surface.

[0144] For example, in a possible implementation, the method for generating ice surface texture visualization data includes: Step S151, selecting the ice surface visible light image and the corresponding ice surface displacement rate data of the current time period from the dynamic image sequence.

[0145] In this embodiment, for a large glacier area that has been monitored before, its dynamic image sequence contains glacier-related data for multiple time periods. This embodiment selects the visible light image of the ice surface and the corresponding ice surface displacement rate data in the current time period (for example, within the current 24 hours). The visible light image of the ice surface can intuitively show the appearance texture, cracks, etc. of the ice surface. These images are collected by a multi-spectral imager previously set in the glacier area. In the current time period, a frame of visible light image is collected every set time (such as every hour), and these images constitute the visible light image sequence of the ice surface in the current time period. At the same time, the corresponding ice surface displacement rate data is also measured by synthetic aperture radar and other equipment in the same time period. The ice surface displacement rate data records in detail the displacement speed of each position on the ice surface at each measurement time relative to the previous time. These data are key information for accurately quantifying the motion state of the ice surface, corresponding to the visible light image of the ice surface, and provide a basis for the subsequent generation of visualization data.

[0146] Step S152, performing crack edge enhancement processing on the ice surface visible light image to obtain an enhanced visible light image.

[0147] Taking the ice surface visible light image selected in the current time period as an example, since the crack edge of the ice surface in the original image may not be clear enough, it is not conducive to intuitively observing and analyzing the condition of the ice surface, and the crack edge enhancement processing is required. First, the ice surface visible light image is processed by the existing image edge detection algorithm. This algorithm identifies the edge by analyzing the gray value change of pixels in the image. For the edge of the ice surface crack, the gray value of the surrounding pixels usually changes greatly. The algorithm will enhance the contrast of this gray value change, making the crack edge more obvious. For example, in the original image, the difference between the pixel gray value of the crack edge and the surrounding pixels may be small. After being processed by the edge detection algorithm, the gray value of the crack edge pixel is adjusted to be darker or brighter (depending on the specific settings of the algorithm), thereby forming a sharper contrast with the surrounding pixels. After all the crack edges in the entire ice surface visible light image are processed in this way, an enhanced visible light image is obtained. In the enhanced visible light image, the outline of the ice surface crack is clearer, and the distribution and direction of the crack can be more easily observed.

[0148] Step S153, converting the ice surface displacement rate data into a color gradient layer, wherein the color depth of the color gradient layer indicates the speed of displacement, and superimposing the color gradient layer on the enhanced visible light image to obtain a superimposed image.

[0149] According to the obtained ice surface displacement rate data, it is converted into a color gradient layer. Assume that a color mapping rule is set. For example, the color corresponding to the ice surface area with the slowest displacement rate is light blue. As the displacement rate increases, the color gradually darkens, and the color corresponding to the area with the fastest displacement rate is dark blue. Take a certain area in the ice surface displacement rate data as an example. The ice surface displacement rate in this area is 0.1 meters per hour. According to the color mapping rule, this area is assigned light blue in the color gradient layer. The ice surface displacement rate in another area is 0.5 meters per hour. This area is assigned a darker blue in the color gradient layer. In this way, the entire ice surface displacement rate data is converted into a color gradient layer. Then the color gradient layer is superimposed on the previously obtained enhanced visible light image. During the superposition process, ensure that the color gradient layer and the enhanced visible light image are accurately corresponding in space, that is, each pixel position in the color gradient layer is consistent with the corresponding pixel position in the enhanced visible light image. By superimposing, a superimposed image is obtained. In the superimposed image, the texture and cracks of the ice surface can be clearly seen (from the enhanced visible light image), and the displacement rate of different areas of the ice surface can be intuitively understood (through the color gradient layer).

[0150] Step S154: marking the risk area corresponding to the ice surface risk level label in the superimposed image, and using dynamic arrows to indicate the moving direction and speed of the ice body in the risk area.

[0151] In the obtained superimposed image, the corresponding risk area is marked according to the previously generated ice surface risk level label. For example, a series of analyses have been conducted to determine the high-risk area, medium-risk area, and low-risk area of ​​the ice surface. For the high-risk area, the area is circled with a set mark (such as a red border) in the superimposed image. At the same time, in each risk area, a dynamic arrow is used to indicate the moving direction and speed of the ice body. Taking a certain point in the high-risk area as an example, by analyzing the ice surface displacement rate data, it is determined that the moving direction of the ice body at this point is northeast, and the moving speed is 0.3 meters per hour. According to the speed and direction, a dynamic arrow is drawn at this point in the image. The direction of the arrow points to the northeast, indicating the moving direction of the ice body. The length of the arrow is set according to the speed. The faster the speed, the longer the arrow. The dynamic arrows are drawn for other points in the risk area in the same way. In this way, in the superimposed image, by marking the risk area and drawing the dynamic arrow, the risk status of different areas of the ice surface and the movement of the ice body in these areas can be intuitively displayed.

[0152] Step S155, marking a color scale bar on the edge of the superimposed image according to the range of the warning trigger threshold, so as to visually compare the correlation between the current displacement rate and the historical disaster data, and obtain a marked image.

[0153] For example, the warning trigger threshold sets the displacement rate range of the ice surface corresponding to different risk levels. The displacement rate range corresponding to the lowest risk level is 0 to 0.1 meters per hour, the range corresponding to the medium risk level is 0.1 to 0.3 meters per hour, and the range corresponding to the high risk level is more than 0.3 meters per hour. According to the threshold range, a color scale bar is drawn at the edge of the superimposed image. The color division of the color scale bar corresponds to the color mapping rule for converting the ice surface displacement rate data into a color gradient layer. For example, the leftmost end of the scale bar corresponds to light blue, indicating a displacement rate of 0 meters per hour. As the scale moves to the right, the color gradually darkens, and the rightmost end corresponds to dark blue, indicating a displacement rate of more than 0.3 meters per hour. Through the color scale bar, the correlation between the displacement rate of the current ice surface and the warning trigger threshold set in the historical disaster data can be intuitively compared. If the color of a certain area of ​​the current ice surface is close to the color range corresponding to the high risk level on the color scale bar, then the area has a higher risk. After such marking, the marked image is obtained.

[0154] Step S156, binding the marked image to the geographic coordinates of the ice landslide risk warning signal to generate a scalable map-type warning image, wherein the boundary shape and arrow length of the risk area in the map-type warning image change in real time with the update of the dynamic image sequence and are displayed through the monitoring terminal.

[0155] For example, the central geographic coordinates of a high-risk area are (x1, y1, z1), and the geographic coordinates are bound to the corresponding risk area in the marked image. In this way, a scalable map-type warning image is generated. In the map-type warning image, the boundary shape and arrow length of the risk area will change in real time as the dynamic image sequence is updated. As time goes by, the state of the glacier will change, and the displacement of the ice surface, the expansion of cracks, etc. will affect the boundary shape of the risk area and the movement speed of the ice body. For example, as the ice surface moves, a risk area may expand or shrink, and its boundary shape will change accordingly. At the same time, the movement speed of the ice body in the risk area may also change, resulting in a change in the length of the arrow indicating the direction and speed of the ice body. The map-type warning image is displayed to relevant personnel through the monitoring terminal. Relevant personnel can zoom in and out on the map-type warning image on the monitoring terminal to view the situation of different areas of the ice surface in more detail, understand the risk status of the ice surface in a timely manner, and take corresponding measures.

[0156] Figure 2The structural diagram of the ice landslide monitoring and early warning system 10 based on AI image recognition provided by the embodiment of the present invention includes a processor 102, a memory 104, and a bus 106. Among them, the memory 104 is used to store execution instructions, including a memory and an external memory. The memory can also be understood as an internal memory, which is used to temporarily store the operation data in the processor 102 and the data exchanged with the external memory such as the hard disk. The processor 102 exchanges data with the external memory through the memory. When the ice landslide monitoring and early warning system 10 based on AI image recognition is running, the processor 102 communicates with the memory 104 through the bus 106, so that the processor 102 executes the ice landslide monitoring and early warning method based on AI image recognition according to the embodiment of the present invention.

[0157] For ease of explanation, only one processor is described in the ice landslide monitoring and early warning system 10 based on AI image recognition. However, it should be noted that the ice landslide monitoring and early warning system 10 based on AI image recognition in the present application may also include multiple processors, so the steps performed by one processor described in the present application may also be performed jointly or individually by multiple processors. For example, if the processor of the ice landslide monitoring and early warning system 10 based on AI image recognition executes step A and step B, it should be understood that step A and step B may also be executed jointly by two different processors or individually in one processor. For example, the first processor executes step A, the second processor executes step B, or the first processor and the second processor execute steps A and B together.

[0158] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer executable instructions are preset. When a processor executes the computer executable instructions, the above-mentioned ice landslide monitoring and early warning method based on AI image recognition is implemented.

[0159] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, various features are sometimes combined into one embodiment, drawing or description thereof.

Claims

1. An ice landslide monitoring and early warning method based on AI image recognition, characterized in that: The method comprises: Collect dynamic image sequences of the target glacier area, extract the three-dimensional spatiotemporal characteristics of ice surface texture, crack distribution morphology and displacement rate in the dynamic image sequences, and generate an ice surface state feature set; Inputting the ice surface state feature set into a pre-trained ice landslide monitoring model, performing spatial correlation analysis on the ice surface state feature set through a multi-scale feature fusion layer in the ice landslide monitoring model, and outputting ice surface stability assessment parameters; Generate an ice surface risk level label according to the ice surface stability assessment parameter, and match the warning trigger threshold in the historical disaster data based on the ice surface risk level label; When the ice surface risk level label meets the warning trigger threshold, an ice landslide risk warning signal is generated, and the parameter weights of the ice landslide monitoring model are dynamically adjusted to adapt to environmental changes in the target glacier area; The ice landslide risk warning signal and the corresponding ice surface texture visualization data are sent to the monitoring terminal.

2. The ice landslide monitoring and early warning method based on AI image recognition according to claim 1 is characterized in that: The method collects a dynamic image sequence of the target glacier area, extracts the three-dimensional spatiotemporal characteristics of ice surface texture, crack distribution morphology and displacement rate in the dynamic image sequence, and generates an ice surface state feature set, including: The visible light image sequence, thermal infrared image sequence and radar interferometric image sequence of the target glacier area in a continuous time window are collected synchronously by multispectral imager, thermal infrared sensor and synthetic aperture radar, and the time and space registration is performed according to the unified timestamp; Adaptive histogram equalization is performed on the visible light image sequence to enhance the contrast between the ice cracks and the background, a multi-scale edge detection algorithm is used to extract crack edge pixel clusters of the ice surface texture, and optical flow tracking is performed on the crack edge pixel clusters of adjacent frames to generate crack expansion direction vectors; Performing region growing segmentation on the thermal infrared image sequence, identifying the boundary coordinates of the pixel region where the temperature difference of the ice body exceeds a preset threshold, and generating a diffusion path vector of the thermal anomaly zone on the ice surface based on the direction of temperature gradient change; Performing phase unwrapping processing on the radar interference image sequence, calculating the displacement vector of each pixel point on the ice surface in the three-dimensional space coordinate system, and generating the ice surface displacement rate field through time differential interferometry; Mapping the crack extension direction vector, the thermal anomaly zone diffusion path vector and the ice surface displacement rate field to the same three-dimensional space coordinate system, and constructing a multi-dimensional spatiotemporal feature matrix including crack length, temperature gradient and displacement acceleration; The multi-dimensional spatiotemporal feature matrix is ​​hierarchically aggregated using a spatiotemporal pyramid pooling layer to extract fused feature vectors of ice surface crack growth rate, thermal diffusion intensity, and displacement acceleration at different time scales; The fused feature vector is measured for similarity with the standard feature template in the historical glacier deformation database, and after eliminating the abnormal feature components caused by instantaneous environmental noise, a three-dimensional spatiotemporal feature set containing ice surface structure stability indicators is generated as the ice surface state feature set.

3. The ice landslide monitoring and early warning method based on AI image recognition according to claim 2 is characterized in that: The method uses a multi-scale edge detection algorithm to extract crack edge pixel clusters of the ice surface texture, and performs optical flow tracking on the crack edge pixel clusters of adjacent frames to generate crack expansion direction vectors, including: The preprocessed visible light image sequence is input into a two-branch convolutional network. The first branch convolutional network uses a large-scale convolution kernel to extract the trend line of the macro cracks on the ice surface, and the second branch convolutional network uses a small-scale convolution kernel to detect the bifurcation endpoints of the micro cracks. The edge feature map output by the dual-branch convolutional network is subjected to pixel-level weighted fusion to generate a multi-resolution crack edge probability map, and after removing pseudo-edge noise through a non-maximum suppression algorithm, the crack edge probability map is binarized to extract crack pixel clusters whose connected domain area is greater than a preset threshold, and the centroid coordinates and principal axis direction angles of each pixel cluster are recorded; Between visible light image sequences of adjacent time stamps, a sparse feature point correspondence relationship of a crack pixel cluster is established based on the centroid coordinates, and a pyramid Lucas-Kanade optical flow algorithm is used to calculate a motion vector of the feature point; Generate the velocity components of each crack pixel cluster in three dimensions of horizontal displacement, vertical settlement and lateral expansion according to the projection components of the motion vector in the three-dimensional space coordinate system; The velocity components of multiple feature points of the same crack pixel cluster are subjected to median filtering to eliminate local motion outliers, and then the average expansion rate and acceleration in the main axis direction of each crack are calculated. The vector set including the spatial position, expansion direction and rate of the crack is output as the crack expansion direction vector, and is spatially superimposed and verified with the diffusion path vector of the thermal anomaly area.

4. The ice landslide monitoring and early warning method based on AI image recognition according to claim 1 is characterized in that: The training method of the ice landslide monitoring model includes: Obtain an image training set of a target glacier area before a historical ice landslide event occurs, wherein the image training set includes a visible light image sequence of the ice surface, a thermal infrared image sequence, and ice displacement rate measurement data at a corresponding time stamp; Gray-scale processing is performed on the visible light image sequence to extract the pixel coordinate set of the edge of the ice surface crack, and temperature threshold segmentation is performed on the thermal infrared image sequence to identify the boundary coordinates of the area where the temperature difference of the ice body changes suddenly; The pixel coordinate set of the edge of the ice surface crack is spatially superimposed with the boundary coordinates of the area where the temperature difference of the ice body suddenly changes, so as to generate polygonal mask data of the fragile area of ​​the ice surface structure; Calculating an acceleration change curve of the displacement rate within the area covered by the polygonal mask data according to the timestamp of the ice body displacement rate measurement data, and aligning the acceleration change curve with the occurrence time of the ice body rupture event; Inputting the visible light image sequence, thermal infrared image sequence and corresponding polygonal mask data into a three-dimensional convolutional neural network to extract the deformation characteristics of the ice surface texture over time and the thermal diffusion characteristics of the fragile area; splicing the deformation feature and the thermal diffusion feature in time series to generate an ice surface state evolution tensor, and aligning the ice surface state evolution tensor with the acceleration change curve in time step; A bidirectional long short-term memory network is used to perform time series modeling on the aligned ice surface state evolution tensor, and the predicted value of ice surface stability at each time step is output; Perform loss calculation on the predicted value of ice surface stability and the actual predicted value of ice surface stability, and adjust weight parameters of the three-dimensional convolutional neural network and the bidirectional long short-term memory network by back propagation; During the training process, pixel blocks in the ice surface crack area in the visible light image sequence are randomly masked, and the texture continuity of the masked area is reconstructed using a generative adversarial network, and the reconstructed image is added to the image training set.

5. The ice landslide monitoring and early warning method based on AI image recognition according to claim 4 is characterized in that: The visible light image sequence, thermal infrared image sequence and corresponding polygonal mask data are input into a three-dimensional convolutional neural network to extract the deformation characteristics of ice surface texture over time and the thermal diffusion characteristics of the fragile area, including: Performing pixel normalization processing on the visible light image sequence frame by frame to generate a visible light normalization tensor, and performing temperature calibration processing on the thermal infrared image sequence to generate a thermal infrared temperature tensor; Aligning the visible light normalized tensor and the thermal infrared temperature tensor according to timestamps, correcting the sensor viewing angle offset by optical flow method, and generating a visible light tensor and thermal infrared tensor aligned in time and space; Converting the polygonal mask data into a binary weight matrix, wherein pixels in the fragile area of ​​the ice surface covered by the mask are assigned a first weight value, and pixels in the non-fragile area are assigned a second weight value; Multiplying the spatiotemporally aligned visible light tensor by the binarized weight matrix pixel by pixel to generate a visible light weighted tensor, and multiplying the thermal infrared tensor by the binarized weight matrix to generate a thermal infrared weighted tensor; Inputting the visible light weighted tensor into the first branch of the three-dimensional convolutional neural network, extracting the displacement gradient of the ice surface cracks between adjacent frames through the three-dimensional convolutional layer, and generating an ice surface texture deformation feature map; The ice surface texture deformation feature map is segmented into a time axis sliding window, the mean and variance of the crack displacement gradient in each time window are extracted, and a deformation feature sequence of the ice surface texture over time is generated; The thermal infrared weighted tensor is input into the second branch of the three-dimensional convolutional neural network, and the diffusion direction and fluctuation intensity of the temperature field in the fragile area are extracted by expanding the three-dimensional convolutional layer to generate a thermal diffusion characteristic map; Performing a spatial pooling operation on the thermal diffusion feature map in the time dimension, extracting the temperature change extreme value of each spatial position in continuous time steps, and generating a thermal diffusion feature of the fragile area; The deformation feature sequence of the ice surface texture over time and the thermal diffusion characteristics of the fragile area are output as feature extraction results of the three-dimensional convolutional neural network.

6. The ice landslide monitoring and early warning method based on AI image recognition according to claim 1 is characterized in that: The step of generating an ice surface risk level label according to the ice surface stability assessment parameter, and matching an early warning trigger threshold in historical disaster data based on the ice surface risk level label includes: Dividing the ice surface stability assessment parameters into subsets of continuous time series according to a preset time window, normalizing the parameters in each subset and splicing them into a multidimensional feature vector; Decomposing the multidimensional feature vector in time series, extracting the acceleration extreme value of the ice surface displacement rate and the mutation point position of the crack expansion rate, and generating time domain abnormal fluctuation characteristics; Based on the spatial distribution density of the abnormal fluctuation characteristics in the time domain, a density clustering algorithm is used to identify the geometric center coordinates and radiation radius of the ice surface risk area; Mapping the geometric center coordinates to the spatial position set of ice breakup events in the historical disaster database, and calculating the spatial overlap and displacement trend similarity between the current risk area and the historical breakup area; According to the spatial overlap and displacement trend similarity, the dynamic threshold interval of the preset risk level is adjusted, and the time domain abnormal fluctuation characteristics are weightedly scored based on the dynamic threshold interval; Input the weighted scoring result into a preset fuzzy logic classifier, match the membership function corresponding to the ice surface risk level label, and generate an independent risk level for each risk area and a global ice surface comprehensive risk level; Performing a confidence check on the independent risk level and the global ice surface comprehensive risk level, eliminating the false detection areas with confidence levels lower than a preset threshold, and updating the spatial distribution matrix of the ice surface risk level labels; The matching process between the spatial distribution matrix and the warning trigger threshold adopts a sliding window traversal strategy, and the window size is dynamically adjusted according to the real-time monitoring value of the ice body displacement rate.

7. The ice landslide monitoring and early warning method based on AI image recognition according to claim 6 is characterized in that: The multi-dimensional feature vector is decomposed into time series, the acceleration extreme value of the ice surface displacement rate and the mutation point position of the crack expansion rate are extracted, and the time domain abnormal fluctuation characteristics are generated, including: Sliding the multidimensional feature vector into multiple subsequence segments at a fixed step size on the time axis, and performing a decomposition process of a trend term and a residual term on each subsequence segment; Detecting the zero-crossing point of the second-order derivative of the ice surface displacement rate in the residual term, marking the moment of acceleration direction reversal and calculating the acceleration integral area between adjacent reversal moments; Performing threshold filtering on the acceleration integrated area, retaining subsequence segments exceeding a preset acceleration threshold, generating a timestamp set of acceleration extreme values, and synchronously performing wavelet transform on the residual term of the crack expansion rate, extracting the energy peak point of the high-frequency component, and marking it as a crack expansion mutation point; Align and match the timestamp set of the acceleration extreme value with the timestamp of the crack extension mutation point, and calculate the standard deviation and co-occurrence probability of the time interval between the two; According to the standard deviation and the co-occurrence probability, the time domain abnormal fluctuation intensity in each subsequence segment is probability weighted to generate a time domain abnormal fluctuation feature vector with a timestamp; wherein the time resolution of the time domain abnormal fluctuation feature vector is consistent with the updating frequency of the three-dimensional spatiotemporal characteristics of the ice surface state feature set.

8. The ice landslide monitoring and early warning method based on AI image recognition according to claim 1 is characterized in that: The dynamically adjusting the parameter weights of the ice landslide monitoring model includes: Collecting environmental noise data of the target glacier area in real time, the environmental noise data including glacier vibration signals, wind speed fluctuation frequency and precipitation pulse intensity; The environmental noise data is converted into a frequency spectrum feature vector, and the frequency spectrum feature vector is input into the noise suppression layer in the ice landslide monitoring model for feature dimension reduction, and then the frequency spectrum feature vector after dimension reduction is weightedly fused with the current model parameters through an online learning algorithm to generate a parameter adjustment gradient; The convolution kernel size and the channel attention weight of the multi-scale feature fusion layer are updated according to the parameter adjustment gradient; wherein the adjustment amplitude of the convolution kernel size is positively correlated with the rising rate of the ice surface risk level label.

9. The ice landslide monitoring and early warning method based on AI image recognition according to claim 8 is characterized in that: The method of weighting and fusing the reduced-dimensional spectrum feature vector with the current model parameters through an online learning algorithm to generate a parameter adjustment gradient includes: Constructing a mapping relationship table between the environmental noise data and the model parameter changes in historical disaster events, wherein the mapping relationship table includes model learning rate adjustment coefficients corresponding to different noise intensity intervals; According to the amplitude distribution of the frequency spectrum feature vector after dimension reduction, matching the optimal learning rate adjustment coefficient from the mapping relationship table; The iterative step size and direction vector of the parameter adjustment gradient are calculated based on the optimal learning rate adjustment coefficient; wherein the direction vector is used to constrain the parameter update path of the online learning algorithm to prevent the model weight from deviating from the core feature dimension of disaster warning.

10. An ice landslide monitoring and early warning system based on AI image recognition, characterized in that: The ice landslide monitoring and early warning system based on AI image recognition includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the ice landslide monitoring and early warning method based on AI image recognition as described in any one of claims 1 to 9 above.

Images