AI Image Recognition-Based Ice Landslide Monitoring and Early Warning Method and System

Abstract

The present invention provides a method and system for ice landslide monitoring and early warning based on AI image recognition. First, a dynamic image sequence of the target glacier area is collected, and three-dimensional spatio-temporal features of the ice surface texture, fracture distribution pattern, and displacement rate are extracted to generate an ice surface state feature set. The feature set is input into a pre-trained ice landslide monitoring model, and spatial correlation analysis is performed through a multi-scale feature fusion layer to output ice surface stability evaluation parameters. An ice surface risk level label is generated based on the ice surface stability evaluation parameters, and the early warning trigger threshold in historical disaster data is matched. If the risk level label meets the threshold, an ice landslide risk early warning signal is generated. At the same time, the model parameter weights are dynamically adjusted to adapt to environmental changes. Finally, an early warning signal and ice surface texture visualization data are sent to the monitoring terminal to achieve effective monitoring and early warning of ice landslides and timely respond to potential disasters.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence, 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, and these methods have significant technical limitations. Manual inspections are restricted by harsh natural environments and high-risk operating conditions, making it difficult to achieve high-frequency and full-coverage dynamic monitoring of glaciers; fixed-point sensors can provide physical parameters of local areas, but cannot capture the microscopic deformation characteristics of the glacier surface and the global evolution process of crack expansion; numerical simulation methods based on physical models highly depend on accurate 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 prior art, traditional early warning systems mostly adopt static threshold determination mechanisms, which cannot adapt to the dynamic characteristics of the glacier environment changing with seasons and climate, and are prone to false alarms or missed alarms. In terms of risk visualization, the prior art is mostly limited to the simple presentation of numerical parameters, lacking an intuitive visualization interface for the glacier surface state, and it is difficult to support decision-makers in quickly understanding the spatio-temporal distribution characteristics of risks. Summary of the Invention

[0004] In view of the problems mentioned above, in combination with the first aspect of the present invention, embodiments of the present invention provide an ice landslide monitoring and early warning method based on AI image recognition, and the method includes:

[0005] Collect a dynamic image sequence of a target glacier area, extract three-dimensional spatio-temporal features of ice surface texture, crack distribution pattern and displacement rate in the dynamic image sequence, and generate an ice surface state feature set;

[0006] Input the ice surface state feature set into a pre-trained ice landslide monitoring model, perform spatial correlation analysis on the ice surface state feature set through a multi-scale feature fusion layer in the ice landslide monitoring model, and output an ice surface stability evaluation parameter;

[0007] According to the ice surface stability evaluation parameter, generate an ice surface risk level label, and match the early warning trigger threshold in historical disaster data based on the ice surface risk level label;

[0008] When the ice surface risk level label meets the early warning trigger threshold, generate an ice landslide risk early warning signal, and dynamically adjust the parameter weights of the ice landslide monitoring model to adapt to the environmental changes in the target glacier area;

[0009] Send the ice landslide risk warning signal and the corresponding visualized data of the ice surface texture to the monitoring terminal.

[0010] In another aspect, an embodiment of the present invention further provides an ice landslide monitoring and warning system based on AI image recognition, including a processor and a machine-readable storage medium. 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.

[0011] Based on the above aspects, the embodiments of the present application realize the multi-scale three-dimensional characterization of the dynamic evolution of glaciers by collecting the dynamic image sequences of the target glacier area and extracting the three-dimensional spatio-temporal features of the ice surface texture, fracture distribution pattern and displacement rate. On this basis, the pre-trained ice landslide monitoring model uses the multi-scale feature fusion layer to perform in-depth spatial correlation analysis on the ice surface state feature set. The output ice surface stability evaluation parameters not only quantify the mechanical stability of the glacier structure, but also establish a risk warning paradigm with spatio-temporal adaptability through the dynamic threshold matching mechanism of the risk level label and historical disaster data. When the ice surface risk level is monitored to reach the warning trigger threshold, a visualized 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 finally output ice landslide risk warning signal and the visualized data of the ice surface texture not only provide multi-dimensional decision-making basis for disaster prevention decision-making, but also reduce the professional threshold through an intuitive visualized interface, enabling the monitoring terminal to perform precise intervention based on the global risk situation. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 is a schematic execution flow diagram of an ice landslide monitoring and warning method based on AI image recognition provided by an embodiment of the present invention.

[0013] Figure 2 is a schematic diagram of an exemplary hardware and software component of an ice landslide monitoring and warning system based on AI image recognition provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0014] The present invention will be specifically described below in conjunction with the drawings of the specification. Figure 1 is a schematic flow diagram of an ice landslide monitoring and warning method based on AI image recognition provided by an embodiment of the present invention. The ice landslide monitoring and warning method based on AI image recognition will be introduced in detail below.

[0015] Step S110, collect the dynamic image sequence of the target glacier area, extract the three-dimensional spatio-temporal features of the ice surface texture, fracture distribution pattern and displacement rate in the dynamic image sequence, and generate an ice surface state feature set.

[0016] For example, for a certain glacier area, a multi-spectral imager, a thermal infrared sensor and a synthetic aperture radar can be used to perform data collection. The visible light image sequence obtained by the multi-spectral imager can reflect the basic appearance features of the ice surface. The thermal infrared image sequence collected by the thermal infrared sensor can capture information related to the ice body temperature. The radar interferometric image sequence collected by the synthetic aperture radar helps to accurately measure the displacement of the ice surface, etc. Then, the fracture expansion direction vector, the thermal anomaly area 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 space coordinate system to construct a multi-dimensional spatio-temporal feature matrix including fracture length, temperature difference gradient and displacement acceleration. The spatio-temporal pyramid pooling layer is used to perform hierarchical aggregation on it, extract the fusion feature vectors of the ice surface fracture growth rate, thermal diffusion intensity and displacement acceleration at different time scales, and perform similarity measurement on the fusion feature vectors and the standard feature templates in the historical glacier deformation database. After removing the abnormal feature components caused by instantaneous environmental noise (such as short-term sunlight reflection, etc.), a three-dimensional spatio-temporal feature set including the ice surface structure stability index, that is, the ice surface state feature set, is generated.

[0017] Step S120, input the ice surface state feature set into a pre-trained ice landslide monitoring model, perform spatial correlation analysis on the ice surface state feature set through the multi-scale feature fusion layer in the ice landslide monitoring model, and output ice surface stability evaluation parameters.

[0018] Specifically, the ice landslide monitoring model is trained with a large amount of historical glacier data. The multi-scale feature fusion layer in it analyzes the ice surface state feature set and can output ice surface stability evaluation parameters. The ice surface stability evaluation parameters represent the stability degree of the ice surface in different regions in numerical form, such as between 0 and 1, where 0 represents extremely unstable and 1 represents very stable.

[0019] Step S130, generate an ice surface risk level label according to the ice surface stability evaluation parameter, and match the warning trigger threshold in the historical disaster data based on the ice surface risk level label.

[0020] In this embodiment, the ice surface stability evaluation parameter is used as the core input, and through analysis and processing, it is converted into an intuitive and easy-to-understand ice surface risk level label. The risk level label not only reflects the current structural stability state of the ice surface, but also provides a direct basis for subsequent warning decisions.

[0021] Step S140: When the ice surface risk level label meets the warning trigger threshold, generate an ice avalanche risk warning signal and dynamically adjust the parameter weights of the ice avalanche monitoring model to adapt to the environmental changes in the target glacier area.

[0022] In this embodiment, when the ice surface risk level label reaches the warning trigger threshold, for example, when the ice surface risk level label in a certain area shows high risk and meets the pre-set threshold, an ice avalanche risk warning signal can be generated. This ice avalanche risk warning signal contains detailed information about the risk area, such as coordinates, risk level, etc.

[0023] At the same time, to adapt to the environmental changes in the glacier area, dynamically adjust the parameter weights of the ice avalanche monitoring model to ensure the accuracy of the ice avalanche monitoring model in monitoring the glacier state.

[0024] Step S150: Send the ice avalanche risk warning signal and the corresponding visualized ice surface texture data to the monitoring terminal.

[0025] In this embodiment, after generating the ice avalanche risk warning signal and the visualized ice surface texture data, they can be sent to the monitoring terminal so that relevant personnel can obtain accurate ice avalanche risk information and take measures.

[0026] Based on the above steps, the embodiment of this application realizes the multi-scale three-dimensional characterization of the dynamic evolution of the glacier by collecting the dynamic image sequence of the target glacier area and extracting the three-dimensional spatio-temporal features of the ice surface texture, fracture distribution pattern and displacement rate. On this basis, the pre-trained ice avalanche monitoring model uses the multi-scale feature fusion layer to conduct in-depth spatial correlation analysis on the ice surface state feature set. The output ice surface stability evaluation parameters not only quantify the mechanical stability of the glacier structure, but also establish a risk warning paradigm with spatio-temporal adaptive ability through the dynamic threshold matching mechanism of the risk level label and historical disaster data. When it is detected that the ice surface risk level reaches the warning trigger threshold, a visualized 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 finally output ice avalanche risk warning signal and the visualized ice surface texture data not only provide multi-dimensional decision-making basis for disaster prevention decisions, but also reduce the professional threshold through an intuitive visualized interface, enabling the monitoring terminal to perform precise intervention based on the global risk situation.

[0027] In a possible implementation manner, step S110 includes:

[0028] Step S111: Synchronously collect visible light image sequences, thermal infrared image sequences, and radar interferometry image sequences of the target glacier area within a continuous time window through a multispectral imager, a thermal infrared sensor, and a synthetic aperture radar, and perform spatio-temporal registration according to a unified timestamp.

[0029] Taking the large glacier area mentioned above as an example, set the continuous time window to one week. Within this week, data is collected every 3 hours. The visible light image sequences collected by the multispectral imager can intuitively present the appearance characteristics of the ice surface. The thermal infrared image sequences obtained by the thermal infrared sensor help to understand the temperature distribution of the ice body. The radar interferometry image sequences collected by the synthetic aperture radar can measure information such as the displacement of the ice surface. During the collection process, through time synchronization technology, a unified timestamp is assigned to each set of collected visible light images, thermal infrared images, and radar interferometry images to ensure that these different types of images are accurately registered in space and time. For example, the visible light image, thermal infrared image, and radar interferometry image collected at a certain moment all reflect the state of the same position of the glacier at the same moment.

[0030] Step S112: Perform adaptive histogram equalization processing on the visible light image sequences to enhance the contrast between ice surface fissures and the background, use a multi-scale edge detection algorithm to extract the fissure edge pixel clusters of the ice surface surface texture, and perform optical flow tracking on adjacent frame fissure edge pixel clusters to generate fissure expansion direction vectors.

[0031] After collecting the visible light image sequences, since the contrast between the ice surface fissures and the background in the original images may be low, which is not conducive to accurately analyzing the fissure situation, adaptive histogram equalization processing is required. Taking a certain frame of visible light image as an example, the color of the ice surface fissure part in the original image is similar to the surrounding background color. After adaptive histogram equalization processing, the color of the fissure part becomes darker, forming a sharp contrast with the background, making the fissures easier to identify.

[0032] Step S113: Perform region growing segmentation on the thermal infrared image sequences, identify the boundary coordinates of pixel regions where the ice temperature difference exceeds a preset threshold, and generate diffusion path vectors of the ice surface thermal anomaly area based on the temperature gradient change direction.

[0033] In this embodiment, it is assumed that the preset threshold is set such that the temperature difference of the ice exceeds 3 degrees Celsius. When the temperature difference between the ice in a certain pixel region and the surrounding ice exceeds 3 degrees Celsius, this pixel region is identified. For example, in a certain region of the thermal infrared image, the temperature of the ice in the central part is 3.5 degrees Celsius higher than the temperature of the surrounding ice. Starting from this central pixel with a higher temperature, the region growing segmentation algorithm is used to gradually expand outward until pixels with a temperature difference less than 3 degrees Celsius are encountered, thereby determining the boundary coordinates of the pixel region where the ice temperature difference exceeds the preset threshold. Then, a diffusion path vector of the thermal anomaly region on the ice surface is generated based on the direction of the temperature gradient change. Since heat always diffuses from a high-temperature region to a low-temperature region, the direction of the temperature gradient change is determined according to the rate of temperature change in different directions. For example, in this high-temperature region, the temperature gradually decreases in all directions, and the rate of temperature decrease is the fastest in the north direction. Then, the direction of the diffusion path vector of the thermal anomaly region is north, and this diffusion path vector can intuitively show the direction of heat diffusion from the high-temperature region to the low-temperature region.

[0034] Step S114: Perform phase unwrapping processing on the radar interferogram sequence, calculate the displacement vector of each pixel point on the ice surface in the three-dimensional space coordinate system, and generate an ice surface displacement rate field through time-differential interferometric measurement.

[0035] Taking a certain pixel point on the glacier edge as an example, within a certain time period, by analyzing parameters such as phase change and the known radar wavelength, the displacement vector of this pixel point in the three-dimensional space coordinate system is calculated. Suppose that within one month, the displacement of this 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 meter (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 this pixel point is (0.5, 0.3, 0.1) meters. Then, an ice surface displacement rate field is generated through time-differential interferometric measurement. The calculation method is to subtract the displacement amounts of each pixel point at different time points and then divide by the time interval. For example, the displacement vector of this pixel point in the first month is (0.5, 0.3, 0.1) meters, and the displacement vector in the second month is (0.8, 0.4, 0.2) meters, and 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. Thus, the displacement rate of this pixel point is obtained, and the displacement rates of all pixel points on the entire ice surface constitute the ice surface displacement rate field.

[0036] Step S115: Map the crack expansion direction vector, the thermal anomaly area diffusion path vector, and the ice surface displacement rate field to the same three-dimensional space coordinate system, and construct a multi-dimensional spatio-temporal feature matrix including crack length, temperature difference gradient, and displacement acceleration.

[0037] Taking a certain pixel point on the glacier edge as an example, within a certain time period, by analyzing parameters such as phase change and known radar wavelength, the displacement vector of this pixel point in the three-dimensional space coordinate system can be calculated. Suppose within one month, the displacement of this pixel point in the x direction is 0.5 meters, in the y direction is 0.3 meters, and in the z direction is 0.1 meter (assuming the z direction is the vertical direction and the x and y directions are horizontal directions here), then the displacement vector of this pixel point is (0.5, 0.3, 0.1) meters. Then, the ice surface displacement rate field is generated through time-differential interferometric measurement. The calculation method is to subtract the displacement amounts of each pixel point at different time points and then divide by the time interval. For example, the displacement vector of this pixel point in the first month is (0.5, 0.3, 0.1) meters, and in the second month is (0.8, 0.4, 0.2) meters, and 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. Thus, the displacement rate of this pixel point is obtained, and the displacement rates of all pixel points on the entire ice surface constitute the ice surface displacement rate field.

[0038] Step S116: Use a spatio-temporal pyramid pooling layer to perform hierarchical aggregation on the multi-dimensional spatio-temporal feature matrix, and extract the fusion feature vectors of the ice surface crack growth rate, thermal diffusion intensity, and displacement acceleration at different time scales.

[0039] Specifically, the spatio-temporal 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 local ice surface crack growth rate, thermal diffusion intensity, and displacement acceleration changes. For data with a longer time scale, the change trends of these features over a larger range and longer time period will be comprehensively considered. Through this hierarchical aggregation method, fused feature vectors of the 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 relatively fast, the thermal diffusion intensity is relatively small, and the displacement acceleration is relatively large. 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 change situations of these features in the entire glacier area are comprehensively considered and then fused into the feature vector, thereby obtaining a complete fused feature vector containing information from different time scales.

[0040] Step S117: Perform a similarity measurement on the fused feature vector and the standard feature templates in the historical glacier deformation database. After removing the abnormal feature components caused by instantaneous environmental noise, generate a three-dimensional spatio-temporal feature set containing ice surface structure stability indicators as the ice surface state feature set.

[0041] For example, the distance between the fused feature vector and each standard feature template can be calculated (the distance here can be an appropriate measurement method such as Euclidean distance). The smaller the distance, the higher the similarity. Suppose the distance between the fused feature vector and a certain 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. During this process, due to instantaneous environmental noise (such as short-term sunlight reflection, local meteorological fluctuations, etc.), some abnormal feature components may occur, and these abnormal feature components caused by instantaneous environmental noise need to be removed. For example, if the value of a certain feature component significantly deviates from the value range in other normal situations and is determined to be caused by instantaneous environmental noise after analysis, then this feature component is removed. After such similarity measurement and removal of abnormal feature components, a three-dimensional spatio-temporal feature set containing ice surface structure stability indicators is generated, and this three-dimensional spatio-temporal feature set is used as the ice surface state feature set, which contains various information that can reflect the ice surface structure stability, such as the development of cracks, changes in thermal distribution, and displacement rate and acceleration, etc.

[0042] In a possible implementation manner, step S112 includes:

[0043] Step S1121: Input the pre - processed visible - light image sequence into a dual - branch convolutional network. The first - branch convolutional network uses a large - scale convolutional kernel to extract the trend line of the macroscopic cracks on the ice surface, and the second - branch convolutional network uses a small - scale convolutional kernel to detect the bifurcation endpoints of the microscopic cracks.

[0044] Taking the visible - light image sequence of that large - scale glacier area collected previously as an example again, after pre - processing it, input the visible - light image sequence into the dual - branch convolutional network. The first - branch convolutional network uses a large - scale convolutional kernel, for example, the convolutional kernel size is 5×5. In a certain area of the image, this large - scale convolutional kernel detects the trend line of the macroscopic cracks on the ice surface through convolution operation. This trend line represents the main extension direction of the cracks on the ice surface macroscopically. At the same time, the second - branch convolutional network uses a small - scale convolutional kernel, assumed to be a 3×3 convolutional kernel, to detect the bifurcation endpoints of the microscopic cracks. Based on the macroscopic cracks, these small - scale convolutional kernels can more precisely find the bifurcation endpoints of the microscopic cracks.

[0045] Step S1122: Perform pixel - level weighted fusion on the edge feature maps output by the dual - branch convolutional network to generate a multi - resolution crack - edge probability map. After removing the pseudo - edge noise through the non - maximum suppression algorithm, perform binary processing on the crack - edge probability map, extract the crack pixel clusters with a connected - domain area greater than a preset threshold, and record the centroid coordinates and the principal axis direction angles of each pixel cluster.

[0046] For the edge feature maps output by the dual-branch convolutional network, pixel-level weighted fusion is performed by assigning different weights according to the importance of the features output by different convolutional kernels. For example, the weight of the macro crack trend line feature output by the large-scale convolutional kernel is 0.6, and the weight of the micro crack bifurcation endpoint feature output by the small-scale convolutional kernel is 0.4 (the weights here are determined based on experience or experiments and may be adjusted according to specific situations in actual applications). Then, these two features are weighted and calculated according to the weights to obtain a multi-resolution crack edge probability map. This multi-resolution crack edge probability map represents the probability that each pixel belongs to the crack edge. Next, the non-maximum suppression algorithm is used to remove the pseudo-edge noise. In the multi-resolution crack edge probability map, there may be some local peaks that are not real edges. The non-maximum suppression algorithm will compare the probability values of each pixel with those of its surrounding pixels and reduce the probability values of the pixels that are not local maxima (i.e., pseudo-edge noise), thereby removing these disturbances. After that, the crack edge probability map is binarized. A threshold (e.g., 0.5) is set. If the probability value of a pixel is greater than 0.5, it is considered a crack edge pixel; if it is less than 0.5, it is not. Then, crack pixel clusters with a connected domain area greater than a preset threshold (assumed to be 50 square pixels) are extracted. For example, in a connected crack pixel region in the image, after calculating its area to be 60 square pixels, which meets the condition of being greater than 50 square pixels, this pixel cluster is extracted, and the centroid coordinates and the principal axis direction angle of this pixel cluster are recorded. Suppose the calculated centroid coordinates of this pixel cluster are (150, 250) and the principal axis direction angle is 45 degrees. The centroid coordinates represent the central position of the pixel cluster in the image, and the principal axis direction angle represents the main extension direction of the crack within this pixel cluster.

[0047] Step S1123, between the visible light image sequences of adjacent timestamps, establish a sparse feature point correspondence of the crack pixel clusters based on the centroid coordinates, and use the pyramid Lucas-Kanade optical flow algorithm to calculate the motion vectors of the feature points.

[0048] Between visible light image sequences at adjacent timestamps, based on the centroid coordinates of previously recorded crack pixel clusters, establish sparse feature point correspondence relationships for the crack pixel clusters. For example, in an image at a certain moment, the centroid coordinates of a crack pixel cluster are (150, 250). In the image at an adjacent moment, determine the corresponding feature points based on the pixel points near the centroid coordinates. Then use the pyramid Lucas-Kanade optical flow algorithm to calculate the motion vectors of these feature points. Take a certain feature point as an example. First, calculate the optical flow of this feature point between two adjacent frames of images at different layers of the pyramid (for example, starting from the bottom layer and calculating layer by layer upwards). Suppose at the bottom layer, the calculated result of the optical flow of this feature point in the horizontal direction is 3 pixels, and in the vertical direction is 2 pixels. Then adjust this result according to the hierarchical relationship of the pyramid and some parameters in the algorithm (such as the scale factor, etc.). Finally, obtain the motion vector that this feature point has moved 3 pixels in the horizontal direction and 2 pixels in the vertical direction.

[0049] Step S1124, generate rate 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.

[0050] For the calculated motion vector, it needs to be converted into rate components in three dimensions of horizontal displacement, vertical settlement, and lateral expansion in the three-dimensional space coordinate system. Take the previously calculated motion vector of a certain feature point that has moved 3 pixels in the horizontal direction and 2 pixels in the vertical direction as an example. Suppose 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 it is assumed that the z direction is the vertical direction), and the lateral expansion direction is the y direction (perpendicular to the x and z directions). First, determine the projection relationship. Suppose the proportion of the horizontal motion vector projected onto the horizontal displacement dimension is 0.8, the proportion of the vertical motion vector projected onto the vertical settlement dimension is 0.6, and for the lateral expansion dimension, according to the relationship between the horizontal and vertical direction vectors (calculated through geometric relationships, here it is assumed to be calculated according to the Pythagorean theorem with a proportion of 0.5). Then the horizontal displacement rate component is 3×0.8 = 2.4 pixels per unit time, the vertical settlement rate component is 2×0.6 = 1.2 pixels per unit time, and the lateral expansion rate component is sqrt(3² + 2²)×0.5 = 1.8 pixels per unit time (here sqrt represents taking the square root, and the text description is taking the square root of the sum of the square of 3 and the square of 2). For multiple feature points in the same crack pixel cluster, calculate their respective rate components in this way.

[0051] Step S1125, perform median filtering on the rate components of multiple feature points of the same crack pixel cluster, eliminate local motion outliers, and then calculate the average expansion rate and acceleration in the main axis direction of each crack.

[0052] Taking a certain crack pixel cluster as an example, assume that there are 5 feature points in this pixel cluster, and their horizontal displacement rate components are 2.2, 2.5, 2.3, 2.6, and 2.4 pixels per unit time respectively. First, perform median filtering on these rate components. Sort these 5 values from smallest to largest as 2.2, 2.3, 2.4, 2.5, 2.6, and 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.

[0053] 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 this crack pixel cluster recorded previously is 45 degrees. For the horizontal displacement rate component and the lateral expansion rate component, calculate the projection components in the main axis direction 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). Then add these two projection components 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 per unit time, and after the next unit time, the rate component becomes 3.3 pixels per unit time.

[0054] When calculating the 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 per unit time².

[0055] 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 overlay verification with the diffusion path vector of the thermal anomaly area.

[0056] In this embodiment, the vector set calculated to include the spatial position of the crack (represented by the centroid coordinates of the crack pixel cluster, such as (150, 250)), the expansion direction (represented by the main axis direction angle of 45 degrees), and the rate (such as information on the average expansion rate and acceleration, etc.) can be used as the crack expansion direction vector. Then perform spatial overlay verification with the diffusion path vector of the thermal anomaly area. For example, in a three-dimensional space coordinate system, draw the crack expansion direction vector and the diffusion path vector of the thermal anomaly area. If the area pointed to by the crack expansion direction vector overlaps with the area pointed to by the diffusion path vector of the thermal anomaly area, or there is a set association in space (such as the crack expansion direction is towards the direction of the thermal anomaly area), it indicates that there is a set spatial correlation between the two, which helps to further analyze the state of the glacier and judge whether there is a higher risk of ice landslide in these areas of the glacier.

[0057] In a possible implementation manner, the training method of the ice landslide monitoring model includes:

[0058] Step S210: Obtain an image training set of a target glacier area before a historical ice landslide event occurs. The image training set includes a visible light image sequence, a thermal infrared image sequence on the ice surface layer, and ice body displacement rate measurement data under corresponding timestamps.

[0059] 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 this image training set can reflect the appearance texture of the ice surface, etc. The thermal infrared image sequence can reflect the temperature distribution of the ice body, and the ice body displacement rate measurement data under corresponding timestamps accurately records the moving speed of the ice body at different moments. For example, in this embodiment, data within a set time period (such as one month before the event) before multiple ice landslide events are selected from the historical database of glacier monitoring. Within this time period, there are corresponding visible light images, thermal infrared images, and ice body displacement rate measurement values every hour. The visible light images clearly show different forms of the ice surface, the thermal infrared images present subtle differences in the temperature of the ice body, and the ice body displacement rate measurement data accurately records the moving speed of the ice body along different directions at each moment.

[0060] Step S220: Perform grayscale processing on the visible light image sequence, extract the pixel coordinate set of the ice surface crack edge, and perform temperature threshold segmentation on the thermal infrared image sequence to identify the boundary coordinates of the ice body temperature difference mutation area.

[0061] Taking a certain frame of visible light image as an example, the original image may contain various color information. After grayscale processing, it is converted into a single grayscale value, which is more convenient for subsequent analysis of the ice surface crack edge. Through existing algorithms, the grayscale image is processed to identify the pixels of the ice surface crack edge, thereby extracting the pixel coordinate set of the ice surface crack edge. For example, in a certain area in the upper left corner of the image, a series of pixel points are detected to form the edge of the ice surface crack, and the coordinates of these pixel points are recorded to form the pixel coordinate set of the ice surface crack edge. At the same time, for the thermal infrared image sequence, temperature threshold segmentation is performed. A temperature threshold is set, for example, 3 degrees Celsius. When the temperature change of the ice body exceeds this threshold, it is identified as a temperature difference mutation area. By traversing each pixel point in the thermal infrared image and comparing its temperature difference with the surrounding pixel points, once the temperature difference exceeds 3 degrees Celsius, the pixel point 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, and the temperature of the internal pixel points is different from the surrounding by more than 3 degrees Celsius. Through algorithm processing, the boundary coordinates of this temperature difference mutation area can be obtained.

[0062] Step S230: Spatially superimpose the pixel coordinate set of the ice surface crack edge and the boundary coordinates of the ice temperature difference mutation region to generate polygon mask data for the ice surface structure vulnerable area.

[0063] Specifically, in a three-dimensional space coordinate system, map the pixel coordinates of the ice surface crack edge to the corresponding spatial positions, and at the same time locate the boundary coordinates of the ice temperature difference mutation region in the same spatial coordinate system. For example, a certain segment of pixel coordinates of the ice surface crack edge is between coordinates (x1, y1, z1) and (x2, y2, z2), and a part of the boundary coordinates of the corresponding ice temperature difference mutation region overlaps or is adjacent to these coordinates. Through the above spatial superposition, integrate the spatial information of the two to generate polygon mask data for the ice surface structure vulnerable area. This polygon mask data can accurately represent the vulnerable area in the ice surface structure where there are both cracks and temperature anomalies. For example, the generated polygon mask data may present an irregular polygon shape, and the area covered by this polygon shape is the ice surface structure vulnerable area.

[0064] Step S240: According to the timestamps of the ice body displacement rate measurement data, calculate the acceleration change curve of the displacement rate within the area covered by the polygon mask data, and align the acceleration change curve with the occurrence time of the ice body rupture event.

[0065] For example, taking the ice body displacement rate measurement data within a certain time period as an example, assume that within the area covered by the polygon mask data, the ice body displacement rate at a certain moment t1 is v1, and after a time interval Δt, the ice body displacement rate at moment t2 is v2. According to the definition of acceleration, acceleration a = (v2 - v1) / Δt. Through such calculations, obtain the acceleration values at different times. Connect these acceleration values in chronological order to form the acceleration change curve of the displacement rate within the area covered by the polygon mask data. Then, align this acceleration change curve with the occurrence time of the ice body rupture event. For example, it is known that a certain ice body rupture event occurs at moment T. Adjust the time axis in the acceleration change curve so that each data point in the acceleration change curve has an accurate time correspondence with the occurrence time T of the ice body rupture event, which is convenient for analyzing the relationship between the acceleration change trend before the ice body rupture and the ice body rupture event.

[0066] Step S250: Input the visible light image sequence, thermal infrared image sequence, and the corresponding polygon 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 vulnerable area.

[0067] For example, in a three-dimensional convolutional neural network, for a visible light image sequence, the convolutional kernel performs convolutional operations in the spatial dimensions (x and y directions) and the temporal dimension (t direction) of the image. Taking a certain frame of visible light image as an example, the convolutional kernel starts from the upper left corner of the image and slides sequentially according to the set stride, performing convolutional calculations on the pixels within each sliding window to extract features related to the deformation of the ice surface texture over time. For example, through convolutional operations, deformation situations such as distortion and stretching of the ice surface texture at different times can be detected. For the thermal infrared image sequence, similarly, the convolutional kernel performs convolutional operations in three-dimensional space (including the temporal dimension) to extract the thermal diffusion characteristics of the vulnerable area. For example, in the thermal infrared image sequence, the direction, speed, and diffusion range of the thermal diffusion, such as the diffusion of heat from high-temperature areas to low-temperature areas, can be identified. At the same time, the polygon mask data, as auxiliary information, guides the network to pay more attention to the feature extraction of the ice surface structure vulnerable area, enhancing the ability to capture features of the key areas of the ice surface.

[0068] Step S260: Concatenate the deformation features and the thermal diffusion features in a time series to generate an ice surface state evolution tensor, and align the time steps of the ice surface state evolution tensor with the acceleration change curve.

[0069] For example, assume that at a certain moment t, the deformation feature of the ice surface texture is F1, and the thermal diffusion feature of the vulnerable area is F2. Combine F1 and F2 in chronological order to form a composite feature containing ice surface texture and thermal diffusion information. Multiple such composite features arranged in chronological order generate the ice surface state evolution tensor. Then, align the time steps of this ice surface state evolution tensor with the previously obtained acceleration change curve. Ensure 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 within the same time frame.

[0070] Step S270: Use a bidirectional long short-term memory network to perform temporal modeling on the aligned ice surface state evolution tensor, and output the ice surface stability prediction value for each time step.

[0071] Taking one 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 this time step is processed sequentially in chronological order to calculate a hidden state. 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 this time step is processed in reverse chronological order to also calculate a hidden state. These two hidden states are concatenated in terms of features 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 concatenated together to obtain a new hidden state H. This hidden state H is input into the fully connected layer, and through the linear transformation of the fully connected layer, it is mapped to the initial probability distribution of the ice surface stability prediction value. Suppose this initial probability distribution is P = [p1, p2, p3], where p1, p2, and p3 respectively represent the initial probabilities of the ice surface in the stable, relatively stable, and unstable states. A moving average filter in the time dimension is performed on this initial probability distribution to eliminate instantaneous noise interference. For example, for the initial probability distributions of several consecutive time steps, the corresponding average values are calculated to obtain a smoothed sequence of ice surface stability prediction values. Each value in this ice surface stability prediction value sequence is used as the ice surface stability prediction value for the corresponding time step.

[0072] Step S280, calculate the loss between the ice surface stability prediction value and the actual ice surface stability prediction value, and backpropagate to adjust the weight parameters of the three-dimensional convolutional neural network and the bidirectional long short-term memory network.

[0073] 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 predicted ice surface stability prediction value of 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 the cross-entropy loss function to calculate the loss value. According to this loss value, using the backpropagation 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 sequentially along the network structure towards the input layer direction. For example, in the bidirectional long short-term memory network, if a certain weight parameter contributes greatly to the loss value, it is adjusted according to the set learning rate to reduce the loss value and improve the prediction accuracy of the model.

[0074] Step S290, during the training process, randomly cover the pixel blocks in the ice surface crack area of the visible light image sequence, and use the generative adversarial network to reconstruct the texture continuity of the covered area, and add the reconstructed image to the image training set to enhance the robustness of the model to local feature loss.

[0075] During the training process, for the visible light image sequence, pixel blocks in the ice crack area are randomly selected for masking. For example, in a certain frame of visible light image, a pixel block in a rectangular area of the ice crack area is selected, and the values of these pixels are set to a fixed value (such as 0) to represent masking. Then, a generative adversarial network is used to reconstruct the texture continuity of the masked area. The generator in the generative adversarial network attempts to generate textures 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, during the subsequent training process, in the case of such local feature loss, the adaptability to local feature loss can be enhanced, and the robustness of the model can be improved. For example, when encountering the situation where some pixels in the ice crack area are missing or blocked again, the features such as ice surface texture can still be accurately extracted to predict the stability of the ice surface.

[0076] For example, in a possible implementation manner, step S270 includes:

[0077] Step S271, splitting the aligned ice surface state evolution tensor into continuous tensor subsequences according to the time step, and each tensor subsequence contains the deformation feature sequence of the ice surface texture over time and the thermal diffusion feature of the vulnerable area at the current time step.

[0078] Taking the aligned ice surface state evolution tensor obtained from the previous processing as an example, assume that this 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, for example, one hour as a time step, this ice surface state evolution tensor is split. 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 feature of the vulnerable area at this time step. The deformation feature sequence of the ice surface texture over time records the change of the ice surface texture from several previous time steps to the current time step t1, such as how the ice surface texture is distorted, stretched or compressed. The thermal diffusion feature of the vulnerable area reflects the propagation and diffusion of heat in the vulnerable area within the same time range, including information such as the intensity change and diffusion direction of heat. Each tensor subsequence is such a set containing the ice surface texture and thermal diffusion information at the set time step, and these tensor subsequences are arranged in chronological order, completely describing the evolution process of the ice surface state over time.

[0079] Step S272, inputting the tensor subsequence into the forward propagation layer of the bidirectional long short-term memory network, and calculating the hidden state of each time step in chronological order to generate a forward hidden state sequence.

[0080] In this embodiment, starting from the initial tensor subsequence, processing is performed one by one in chronological order. Taking the first time step in the tensor subsequence as an example, in the forward propagation layer of the bidirectional long short-term memory network, first, the deformation feature sequence of the ice surface texture over time and the thermal diffusion feature of the vulnerable area are processed. Neurons in the network calculate the hidden state of this time step according to the preset weights and activation functions. For example, for a certain feature component in the deformation feature sequence of the ice surface texture over time, assuming its value is x1, after multiplying it by the corresponding weight w1, an intermediate result is obtained, and then adding the bias term b1, through the action of the activation function f, a partial hidden state related to this 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 vulnerable area, and these partial hidden states are combined according to the set rules to obtain the hidden state h1 of this time step. Then, subsequent time steps are processed in the same way in sequence, obtaining a series of hidden states h1, h2, h3, etc. These hidden states are arranged in chronological order, forming the forward hidden state sequence. This forward hidden state sequence reflects an internal representation of the ice surface state in the forward time order as time progresses from the initial moment, integrating the influence of the ice surface texture and thermal diffusion information of each previous time step.

[0081] Step S273, synchronously input the tensor subsequence in reverse order into the backward propagation layer of the bidirectional long short-term memory network, and calculate the hidden state of each time step in reverse chronological order to generate a backward hidden state sequence.

[0082] Meanwhile, input the same tensor subsequence in reverse order into the backward propagation layer of the bidirectional long short-term memory network. Starting from the last time step, processing is performed in reverse chronological order. For example, for the last time step tn in the tensor subsequence, in the backward propagation layer of the bidirectional long short-term memory network, the deformation feature sequence of the ice surface texture over time and the thermal diffusion feature of the vulnerable area are also calculated. First, similar operations to those in the forward propagation layer are performed on the feature components. For example, multiplying the value x2 of a certain feature component by the corresponding weight w2, adding the bias term b2, through the action of the activation function g, a partial hidden state is obtained, and then all partial hidden states are combined to obtain the hidden state hn of this time step. Then, the previous time steps are processed in sequence, such as tn - 1, tn - 2, etc., obtaining a series of hidden states hn, hn - 1, hn - 2, etc. These hidden states are arranged in reverse chronological order, forming the backward hidden state sequence. This backward hidden state sequence captures the information of the ice surface state from the opposite time direction, reflecting an internal representation of the ice surface state in the reverse time order when tracing back from the final moment, containing the influence information of subsequent time steps on the current time step.

[0083] Step S274, perform timeline alignment on the forward hidden state sequence and the backward hidden state sequence, and splice the features of the forward hidden state and the backward hidden state at the same time step to generate a bidirectional fusion hidden state sequence.

[0084] 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 timeline. Ensure 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, if the forward hidden state sequence is h1, h2, h3 and the backward hidden state sequence is hn, hn-1, hn-2, then h1 is corresponding to hn-2, h2 is corresponding to hn-1, and h3 is corresponding to hn. Then, for each pair of corresponding forward and backward hidden states, splice their features. Taking h1 and hn-2 as an example, assume h1 is a vector containing n1 elements and hn-2 is a vector containing n2 elements. Splice these two vectors in order to obtain a new vector containing n1 + n2 elements. Perform such operations on all corresponding hidden states to generate a bidirectional fusion hidden state sequence. This bidirectional fusion hidden state sequence synthesizes 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.

[0085] Step S275, input the bidirectional fusion hidden state sequence into a fully connected layer, and map the fusion hidden state at each time step to an initial probability distribution of the ice surface stability prediction value.

[0086] 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 fusion hidden state at 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 fusion hidden state vector at a certain time step, each element in it is multiplied by the weights of the neurons in the fully connected layer respectively, and then the bias term is added. After the action of the activation function (such as the softmax function), the fusion hidden state is mapped to an initial probability distribution of the ice surface stability prediction value. Assume that the initial probability distribution is a vector containing three elements [p1, p2, p3], 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. This 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.

[0087] Step S276: Perform a moving average filter on the initial probability distribution in the time dimension to eliminate instantaneous noise interference and generate a smoothed ice surface stability prediction value sequence.

[0088] Perform a moving average filter on the initial probability distribution of the ice surface stability prediction values obtained at each time step in the time dimension. Taking three consecutive time steps as an example, assume 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, calculate the result after the moving average filter as (p11 + p12 + p13) / 3. Similarly, perform such calculations on the probability component p2 of the relatively stable state and the probability component p3 of the unstable state to obtain the probability distribution after the moving average filter. Perform the moving average filter on all time steps in this way to eliminate 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 change trend of the ice surface stability and avoid the influence of instantaneous noise on the prediction result.

[0089] Step S277: Point-to-point match the smoothed ice surface stability prediction value sequence with the time step of the acceleration change curve, and calculate the error gradient between the predicted value and the actual ice surface stability measurement value at each time step.

[0090] Point-to-point match the smoothed ice surface stability prediction value sequence with the previously obtained acceleration change curve 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 predicted value and the actual ice surface stability measurement value. Assume that at a certain time step t, the smoothed ice surface stability prediction value indicates that the ice surface is in a relatively stable state with a probability of 0.6, while the actual ice surface stability measurement value shows 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 weights of the bidirectional long short-term memory network, the error gradient can be obtained. Perform such calculations 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 change trend of the error as the weights change.

[0091] Step S278, according to the error gradient, inversely adjust the forgetting gate weight and the input gate weight of the bidirectional long short-term memory network, so that when updating the hidden state at subsequent time steps, the deformation features and thermal diffusion features strongly related to the ice body rupture event are preferentially retained.

[0092] According to the calculated error gradient, use the backpropagation algorithm to adjust the forgetting gate weight and the input gate weight of the bidirectional long short-term memory network. The forgetting gate weight and the input gate weight play a key role in updating the hidden state in the bidirectional long short-term memory network. Taking the forgetting gate weight as an example, assume that at a certain moment, the error gradient indicates that there is a large deviation between the current prediction result and the actual result, and it is found through analysis that some deformation features or thermal diffusion features strongly related to the ice body rupture event are not effectively retained in the hidden state. According to the direction and magnitude of the error gradient, appropriately increase or decrease the forgetting gate weight. If the error gradient of a certain feature at the current time step indicates that the feature has an important impact on the ice body rupture event, then adjust the forgetting gate weight so that the probability of retaining the feature when updating the hidden state at subsequent time steps 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 the deformation features and thermal diffusion features strongly related to the ice body rupture event, thereby improving the accuracy of the model's prediction of ice surface stability.

[0093] Step S279, in the process of time series modeling, when it is detected that the difference between the predicted value of ice surface stability at the current time step and the predicted value at the previous time step exceeds the preset mutation threshold, forcibly reset the cell state of the backward propagation layer to avoid interference of historical irrelevant features on the prediction at the current time step.

[0094] In the process of time series modeling, continuously monitor the predicted value of ice surface stability at each time step. Assume 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, at time step t - 1, it is predicted that the ice surface is in a relatively stable state with a probability of 0.6, while at time step t, it is predicted that the ice surface is in an unstable state with a probability of 0.8. In this case, forcibly reset the cell state of the backward propagation layer. The cell state of the backward propagation layer contains information from previous time steps. When such a mutation occurs, the previous information may interfere with the prediction at the current time step. By resetting the cell state, these historical irrelevant features that may interfere with the current prediction can be cleared, enabling the backward propagation layer to accurately calculate the hidden state starting from the current time step in reverse chronological order, thereby improving the accuracy of the prediction at the current time step.

[0095] Step S2710: Output the predicted values of ice surface stability for each time step, and perform spatial correlation mapping on the temporal change rate of the predicted values and the thermal distribution data of the ice surface displacement rate to verify the spatial consistency of the predicted values.

[0096] Finally, output the predicted values of ice surface stability for each time step. These predicted values of ice surface stability reflect the stability status of the ice surface at different time steps. Then perform spatial correlation mapping on the temporal change rate of these predicted values of ice surface stability and 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, in a certain area, the ice surface displacement rate is relatively fast, and at the same time, the thermal distribution in this area also has a set pattern. Correlate and map the temporal change rate of the predicted values of ice surface stability with the ice surface displacement rate and thermal distribution data at these spatial positions. If the temporal change rate of the predicted values shows a reasonable correlation with the thermal distribution data of the ice surface displacement rate in space, for example, in an area where the ice surface displacement rate is fast and the thermal distribution is abnormal, the temporal change rate of the predicted values of ice surface stability also shows an unstable trend, then the spatial consistency of the predicted values 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.

[0097] In a possible implementation manner, step S250 includes:

[0098] Step S251: Perform pixel normalization processing on each frame of the visible light image sequence to generate a visible light normalized tensor, and perform temperature calibration processing on the thermal infrared image sequence to generate a thermal infrared temperature tensor.

[0099] For the visible light image sequence acquired from the target glacier area, pixel normalization is performed frame by frame. Taking one visible light image frame 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. To ensure comparability between pixel values across frames, pixel normalization is performed. Assuming the original numerical range of the pixels in the image is from 0 to 255, existing normalization algorithms are used to convert them to a new numerical range, such as 0 to 1. This operation is performed for each visible light image frame, and all normalized frames form a visible light normalized tensor. For the thermal infrared image sequence, temperature calibration is performed. Since thermal infrared images reflect the temperature of the ice surface, the numerical representation of temperature may vary depending on the sensor or acquisition conditions. Temperature calibration converts the temperature values in the thermal infrared images 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.

[0100] In step S252 , the normalized visible light tensor and the thermal infrared temperature tensor are aligned according to timestamps, and the sensor viewing angle offset is corrected by an optical flow method to generate a spatiotemporally aligned visible light tensor and thermal infrared tensor.

[0101] Both the visible light normalized tensor and the thermal infrared temperature tensor contain data collected in chronological order. Since the data was collected simultaneously across the glacier area, the corresponding timestamps are hidden. Taking the data collected at a specific moment as an example, the normalized tensor of the visible light image and the temperature tensor of the thermal infrared image corresponding to that moment should reflect different characteristics of the glacier at that moment. However, in actual acquisition, sensor perspective offset may occur due to factors such as sensor installation position or slight vibration during acquisition. This perspective offset is corrected using the optical flow method, which estimates perspective changes by analyzing pixel motion within an image. For example, if the positions of certain feature points (such as designated markings on the ice surface or stable landforms) change between two adjacent visible light image frames, the perspective offset can be calculated based on the displacement of these feature points. This offset is then applied to the thermal infrared image, spatially aligning the visible light normalized tensor and the thermal infrared temperature tensor, ultimately generating temporally and spatially aligned visible light and thermal infrared tensors. The two tensors not only correspond in time, but also accurately reflect the glacier characteristics of the same area in space.

[0102] 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.

[0103] In this embodiment, the polygonal mask data can be converted into a binary weight matrix. For example, for the pixels in the fragile area of the ice surface, they are assigned a first weight value. Suppose the first weight value is 1, indicating that these pixels have a relatively high importance in subsequent calculations. For the pixels in the non-fragile area, they are assigned a second weight value. Suppose 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 non-fragile area of the ice surface are distinguished by different weight values.

[0104] Step S254: Multiply the spatio-temporally aligned visible light tensor and the binary weight matrix pixel by pixel to generate a visible light weighted tensor, and multiply the thermal infrared tensor and the binary weight matrix to generate a thermal infrared weighted tensor.

[0105] For the spatio-temporally aligned visible light tensor, perform a pixel-by-pixel multiplication operation with the binary weight matrix. Taking a certain frame image in the visible light tensor as an example, each pixel in this frame image is multiplied by the corresponding element in the binary weight matrix. If a pixel is located in the fragile area of the ice surface and its corresponding weight value is 1, then the result after multiplication is the original value of the pixel; if it is located in the non-fragile area and the weight value is 0, the result after multiplication is 0. Perform such operations on each pixel of the entire frame image to obtain a new frame image, and all the frame images processed in this way constitute the visible light weighted tensor. Similarly, for the thermal infrared tensor, also perform a pixel-by-pixel multiplication operation with the binary weight matrix to obtain the thermal infrared weighted tensor.

[0106] Step S255: Input the visible light weighted tensor into the first branch of the three-dimensional convolutional neural network, and extract the displacement gradient of the ice surface crack between adjacent frames through the three-dimensional convolutional layer to generate an ice surface texture deformation feature map.

[0107] In this embodiment, in the first branch of the three-dimensional convolutional neural network, the three-dimensional convolutional layer starts to process the visible light weighted tensor. Taking two adjacent frames of visible light weighted images as an example, the convolutional kernel of the three-dimensional convolutional layer performs convolutional operations in the spatial dimensions (x and y directions) and the temporal dimension (t direction). The convolutional kernel starts from the upper left corner of the image and slides sequentially according to the set stride, and performs convolutional calculations on the pixels within each sliding window. During this process, by analyzing the changes in pixel values between adjacent frames, the displacement information of the ice surface cracks can be extracted. For example, in a certain area, if the edge pixels of the ice surface cracks in the previous frame image are displaced from the corresponding pixels in the subsequent frame image, the convolutional layer can detect this displacement and calculate the magnitude and direction of the displacement, thereby obtaining the displacement gradient of the ice surface cracks between adjacent frames. Such operations are performed on all adjacent frames in the entire visible light weighted tensor, and finally an ice surface texture deformation feature map is generated. This ice surface texture deformation feature map visually shows the deformation of the ice surface texture at different time and spatial positions, especially the displacement changes of the ice surface cracks.

[0108] Step S256: Perform time-axis sliding window slicing on the ice surface texture deformation feature map, extract the mean and variance of the crack displacement gradient within each time window, and generate a deformation feature sequence of the ice surface texture over time.

[0109] For the generated ice surface texture deformation feature map, perform slicing using a sliding window on the time axis. Assume that the size of the time window is set to 3 time steps. Starting from the beginning position of the time axis, the first time window contains the parts of the ice surface texture deformation feature map corresponding to the 1st, 2nd, and 3rd time steps. Within this time window, perform statistical analysis on the displacement gradients of all ice surface cracks. Calculate the mean of the displacement gradients of each ice surface crack, that is, add up the numerical values of the displacement gradients of all ice surface cracks within this time window, and then divide by the number of cracks. For example, if there are 5 ice surface cracks within this time window, and the corresponding displacement gradients are 0.1, 0.2, 0.15, 0.18, and 0.22 respectively, add these values to get 1.05, and then divide by 5 to obtain a mean of 0.21. At the same time, calculate the variance of the displacement gradients. The calculation of the variance is to first find the square of the difference between each displacement gradient and the mean. For example, for a crack with a displacement gradient of 0.1, the difference from the mean of 0.21 is -0.11, and the square of the difference is 0.0121. Perform such calculations for all cracks, then add up the squares of these differences, and divide by the number of cracks to obtain the variance. Through such calculations, the mean and variance of the displacement gradients of the cracks within this time window are obtained. Then, in chronological order, move the sliding window backward by one time step in sequence, and perform the same mean and variance calculations for each new time window. Finally, generate a deformation feature sequence of the ice surface texture over time. This deformation feature sequence contains the deformation features of the ice surface texture over time in different time windows, and reflects the change situation of the displacement gradients of the ice surface cracks in the form of the mean and variance.

[0110] Step S257: Input the thermal infrared weighted tensor into the second branch of the three-dimensional convolutional neural network, and extract the diffusion direction and fluctuation intensity of the temperature field in the vulnerable area through the dilated three-dimensional convolutional layer to generate a thermal diffusion feature map.

[0111] Input the thermal infrared weighted tensor into the dilated three-dimensional convolutional 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 convolutional kernel of the dilated three-dimensional convolutional layer performs a convolutional operation in three-dimensional space (including the time dimension). By analyzing the changes in the temperature values in the thermal infrared weighted tensor at different spatial positions and time steps, information about the temperature field in the vulnerable area can be extracted. For example, within the vulnerable area, the convolutional layer can detect the propagation direction of the 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 situation of the temperature at different time steps, such as the amplitude change of the temperature first rising and then falling in a certain area, the fluctuation intensity of the temperature field is calculated. Perform such processing on the entire thermal infrared weighted tensor to generate a thermal diffusion feature map, which shows the distribution of the diffusion direction of the temperature field and the fluctuation intensity at different spatial positions and time steps within the vulnerable area.

[0112] Step S258: Perform a spatial pooling operation on the thermal diffusion feature map in the time dimension to extract the extreme values of temperature changes at each spatial position over consecutive time steps, and generate the thermal diffusion features of the vulnerable area.

[0113] Perform a spatial pooling operation on the generated thermal diffusion feature map. Taking a certain spatial position as an example, within consecutive time steps, the temperature value at this spatial position will change. Through the spatial pooling operation, statistical analysis is performed on the temperature values at this spatial position in the time dimension. Find the highest temperature value and the lowest temperature value at this spatial position over all time steps, and these two temperature values are the extreme values of temperature changes. For example, at this spatial position, after observing for 5 consecutive time steps, the temperature values are -2°C, -1.5°C, -1°C, -2.5°C, -3°C respectively, then the highest temperature value is -1°C and the lowest temperature value is -3°C. Perform such operations on each spatial position in the thermal diffusion feature map, and finally generate the thermal diffusion features of the vulnerable area. The thermal diffusion features of the vulnerable area contain the information of the extreme values of temperature changes at each spatial position in the vulnerable area over consecutive time steps, and can reflect the characteristics of thermal diffusion in the vulnerable area, such as the range of temperature fluctuations, etc.

[0114] Step S259: Output the sequence of deformation features of the ice surface texture over time and the thermal diffusion features of the vulnerable area as the feature extraction results of the three-dimensional convolutional neural network.

[0115] Finally, output the sequence of deformation features of the ice surface texture over time and the thermal diffusion features of the vulnerable area calculated previously 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 the thermal infrared image sequence. The sequence of deformation features of the ice surface texture over time reflects the deformation of the ice surface texture, especially the ice surface cracks, in time and space, while the thermal diffusion features of the vulnerable area reflect the diffusion and fluctuation characteristics of the temperature field in the vulnerable area. These features will be used for subsequent operations such as ice surface stability analysis.

[0116] In a possible implementation manner, step S130 includes:

[0117] Step S131: Divide the ice surface stability evaluation parameters into subsets of consecutive time series according to a preset time window, perform normalization processing on the parameters within each subset, and splice them into multi-dimensional feature vectors.

[0118] Taking the ice surface stability evaluation parameters obtained from previous monitoring in large glacier areas as an example, assume that the preset time window is one day. The entire ice surface stability evaluation parameters are divided according to a one-day time window, and multiple subsets of continuous time series are thus obtained. For the ice surface stability evaluation parameters within each subset, since these parameters may have different numerical ranges and dimensions, in order to facilitate subsequent unified processing and analysis, normalization processing is required. For example, a certain subset contains the stability values of different regions of the ice surface, and these values may be between 0 and 100. Through existing normalization algorithms, they are converted to the range of 0 to 1. After completing the normalization processing for all parameters in each subset, these normalized parameters are concatenated in the set order to form a multi-dimensional feature vector, which comprehensively incorporates the ice surface stability-related information within each time window.

[0119] Step S132: Perform time series decomposition on the multi-dimensional feature vector, extract the acceleration extreme values of the ice surface displacement rate and the positions of the mutation points of the crack expansion rate, and generate time-domain abnormal fluctuation features.

[0120] For the concatenated multi-dimensional feature vector, perform time series decomposition operations. Taking the ice surface displacement rate as an example, by analyzing the part of the multi-dimensional feature vector related to the ice surface displacement rate, determine its changes at different time points. In this process, find the acceleration extreme values of the ice surface displacement rate. For example, within the time period corresponding to a certain subset, the acceleration of the ice surface displacement rate will change significantly at some moments. By comparing the acceleration values at different time points, find the maximum and minimum values, which are the acceleration extreme values of the ice surface displacement rate. At the same time, for the crack expansion rate, analyze its changes in the time series and determine the positions of the mutation points of the crack expansion rate. For example, at consecutive time points, the crack expansion rate remains relatively stable, but suddenly increases or decreases at a certain moment, and this moment is the mutation point of the crack expansion rate. Combine the acceleration extreme values of the ice surface displacement rate and the positions of the mutation points of the crack expansion rate and other information to generate time-domain abnormal fluctuation features. These time-domain abnormal fluctuation features reflect the unstable factors of the ice surface in the time dimension and are important bases for judging the ice surface risk areas.

[0121] Step S133: Based on the spatial distribution density of the time-domain abnormal fluctuation features, use the density clustering algorithm to identify the geometric center coordinates and radiation radius of the ice surface risk area.

[0122] According to the generated time-domain abnormal fluctuation characteristics, consider their spatial distribution density in the glacier area. 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. Use the density clustering algorithm to perform clustering 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 this area reaches the set density, the algorithm identifies it as a potential ice surface risk area. During this process, determine the geometric center coordinates of this ice surface risk area. The geometric center coordinates are determined by calculating the average value of the coordinates of all time-domain abnormal fluctuation characteristic points within this ice surface risk area. For example, there are several time-domain abnormal fluctuation characteristic points in the area, and their coordinates are (x1, y1, z1), (x2, y2, z2), etc. Add the x values of these coordinates and divide by the number of characteristic points to obtain the x coordinate of the geometric center. Similarly, obtain the y coordinate and z coordinate. At the same time, determine the radiation radius according to the distribution range of the time-domain abnormal fluctuation characteristics within this ice surface risk area. For example, calculate the distance from the time-domain abnormal fluctuation characteristic point farthest from the geometric center within the area to the geometric center, and this distance is the radiation radius.

[0123] Step S134, map the geometric center coordinates to the set of spatial positions of ice body rupture events in the historical disaster database, and calculate the spatial overlap degree and displacement trend similarity between the current risk area and the historical rupture area.

[0124] Map the geometric center coordinates of the identified ice surface risk area to the set of spatial positions of ice body rupture events in the historical disaster database. Taking a certain ice body rupture event in the historical disaster database as an example, it records the spatial position information of the area where the ice body rupture occurred. Calculate the spatial overlap degree between the current ice surface risk area and this historical rupture area. For example, regard the current ice surface risk area as a geometric shape in a three-dimensional space (such as a sphere, determined by the geometric center coordinates and the radiation radius), and regard the historical rupture area as a similar geometric shape. Calculate the ratio of the volume of the intersection part of these two geometric shapes to the volume of the current ice surface risk area, and this ratio is the spatial overlap degree. At the same time, analyze the displacement trend similarity between the current ice surface risk area and the historical rupture area. Determine the displacement trend similarity by comparing the change trends of the displacement rates of the ice bodies in different directions within the two areas. For example, in a certain direction, the displacement rate of the ice body within the current ice surface risk area is gradually increasing, and the ice body in the historical rupture area also had a similar increasing trend of the displacement rate in the same direction before rupture. Calculate the displacement trend similarity according to the similarity degree of this trend.

[0125] Step S135, adjust the dynamic threshold interval of the preset risk level according to the spatial overlap degree and displacement trend similarity, and perform weighted scoring on the time-domain abnormal fluctuation characteristics based on the dynamic threshold interval.

[0126] Adjust the dynamic threshold interval of the preset risk level according to the calculated spatial overlap degree and displacement trend similarity. For example, the original threshold interval of the ice surface stability evaluation parameter corresponding to the preset low risk level is from 0.6 to 1. If the spatial overlap degree between the current ice surface risk area and the historical rupture area is high and the displacement trend similarity is also high, adjust this threshold interval to 0.7 to 1. Then, perform weighted scoring on the time-domain abnormal fluctuation characteristics 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 this acceleration extreme value is within the adjusted threshold interval, give corresponding weight scores according to its specific position within the interval. Assume that the closer the acceleration extreme value is to the lower limit of the threshold interval, the lower the weight score given, and the closer it is to the upper limit, the higher the weight score given. Perform weighted scoring on other time-domain abnormal fluctuation characteristics such as the mutation point position of the crack expansion rate in a similar manner, and obtain a weighted scoring result for the time-domain abnormal fluctuation characteristics by synthesizing these weighted scores.

[0127] Step S136: 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 the independent risk level of each risk area and the global ice surface comprehensive risk level.

[0128] Input the weighted scoring result into a preset fuzzy logic classifier, in which the membership functions corresponding to the ice surface risk level labels are preset. For example, for the low risk level, there is a membership function that defines the degree of belonging to the low risk level corresponding to different weighted scoring results. When the weighted scoring result is input, calculate the degree of belonging to the low risk level according to this membership function. Similarly, there are corresponding membership functions for the medium risk level and the high risk level. In this way, calculate the degree of belonging of each risk area to different risk levels respectively, so as to determine the independent risk level of each risk area. At the same time, comprehensively consider the situations of all risk areas and generate the global ice surface comprehensive risk level. For example, if most risk areas are judged to be at a higher risk level, then the global ice surface comprehensive risk level may also be at a higher risk level.

[0129] Step S137: Perform confidence check on the independent risk level and the global ice surface comprehensive risk level, and after removing the misdetected areas with confidence lower than the preset threshold, update the spatial distribution matrix of the ice surface risk level label.

[0130] Perform confidence checks on the obtained independent risk levels and the overall ice surface comprehensive risk level. For example, preset the confidence threshold to 0.6. For the independent risk level of a certain risk area, if its confidence is lower than 0.6, it indicates that there may be errors in the determination of this risk area, and it is regarded as a misdetected area and excluded. A similar check is also performed on the overall ice surface comprehensive risk level. After removing the misdetected areas, update the spatial distribution matrix of the ice surface risk level labels, which records the risk level label information of each ice surface area.

[0131] Among them, 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.

[0132] In the matching process between the spatial distribution matrix and 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 body displacement rate. For example, when the ice body displacement rate is relatively fast, it indicates that the ice surface state changes relatively violently. At this time, appropriately increase the size of the sliding window to more comprehensively consider the risk situation of the ice surface. On the contrary, when the ice body displacement rate is relatively slow, appropriately reduce the size of the sliding window 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 more accurately matched, so as to better judge whether the ice surface reaches the warning condition.

[0133] In a possible implementation manner, step S132 includes:

[0134] Step S1321, slide and divide the multi-dimensional feature vector on the time axis at a fixed step length into multiple subsequence segments, and perform decomposition processing on the trend term and the residual term for each subsequence segment.

[0135] Taking the previously obtained multi-dimensional feature vector of the large glacier area as an example, assume that the fixed step length is set to one hour on the time axis. Starting from the starting moment of the multi-dimensional feature vector, slide and divide the entire multi-dimensional feature vector into multiple subsequence segments at a fixed step length of one hour. For each subsequence segment, perform decomposition processing on the trend term and the residual term. Taking a certain subsequence segment as an example, through the existing time series analysis method, decompose the data in this subsequence segment into a trend term and a residual term. The trend term represents the main change trend of the data within this subsequence segment. For example, the ice surface displacement rate is generally gradually increasing or gradually decreasing 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.

[0136] Step S1322, detect the zero-crossing point of the second derivative of the ice surface displacement rate in the residual term, mark the acceleration direction reversal moment, and calculate the acceleration integral area between adjacent reversal moments.

[0137] Among the obtained residual terms, focus on the part related to the ice surface displacement rate. The second derivative of the ice surface displacement rate reflects the change rate of acceleration. By calculating the second derivative of the ice surface displacement rate in the residual terms, find the moment when its second derivative crosses zero. For example, at a certain moment t1, the second derivative of the ice surface displacement rate changes from positive to negative, and this moment is the moment when the acceleration direction reverses. After marking these moments when the acceleration direction reverses, calculate the integral area of acceleration between adjacent reversal moments. Taking t1 and the next moment t2 when the acceleration direction reverses as an example, calculate the integral area of acceleration in the time period from t1 to t2. The calculation of this integral area is obtained by accumulating the values of acceleration in this time period. Assume that there are several time points between t1 and t2, and the acceleration values at each time point are a1, a2, a3, etc. Accumulate these acceleration values according to the time interval, and the result obtained is the integral area of acceleration, which reflects the cumulative effect of the acceleration of the ice surface displacement rate in this time period.

[0138] Step S1323: Perform threshold filtering on the integral area of acceleration, retain the subsequence segments exceeding the preset acceleration threshold, generate a set of timestamps of acceleration extrema, and simultaneously perform wavelet transform on the residual term of the crack propagation rate, extract the energy peak points of the high-frequency components, and mark them as crack propagation mutation points.

[0139] In this embodiment, the acceleration threshold is set to a specific value, which can be determined according to historical data and the actual characteristics of the glacier. Perform threshold filtering on the integral area of acceleration calculated for each subsequence segment. If the integral area of acceleration of a certain subsequence segment exceeds the preset acceleration threshold, retain this subsequence segment. For example, among multiple subsequence segments, the integral area of acceleration of subsequence segment 1 exceeds the threshold, while the integral area of acceleration of subsequence segment 2 does not exceed the threshold, then retain subsequence segment 1. Mark the times corresponding to all the retained subsequence segments, and these time marks constitute a set of timestamps of acceleration extrema. At the same time, for the residual term of the crack propagation rate, perform wavelet transform. Wavelet transform can decompose the residual term of the crack propagation rate into components of different frequencies. Among these components, the high-frequency components contain the mutation information of the crack propagation rate. By analyzing the energy distribution of the high-frequency components, find the energy peak points. For example, in the energy distribution curve of the high-frequency components, the energy value corresponding to a certain moment t3 is the maximum value in the whole curve, and this moment t3 is the energy peak point of the high-frequency components, and mark this energy peak point as the crack propagation mutation point.

[0140] Step S1324: Align and match the set of timestamps of acceleration extrema with the timestamps of crack propagation mutation points, and calculate the standard deviation and co-occurrence probability of the time intervals between the two.

[0141] Align and match the set of timestamps of the acceleration extreme values and the timestamps of the crack expansion mutation points. For example, in the set of timestamps of the acceleration extreme values, there are timestamps t1, t2, etc., and in the set of timestamps of the crack expansion mutation points, there are timestamps t3, t4, etc. Arrange these timestamps in chronological order for alignment. Then calculate the standard deviation of the time intervals between them. 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 timestamps of the acceleration extreme values and the timestamps of the crack expansion mutation points appear simultaneously or appear successively within a short period of time. For example, count the proportion of the number of timestamp pairs with a time interval less than a certain set value (such as 10 minutes) in all timestamp pairs, and this proportion is the co-occurrence probability.

[0142] Step S1325, according to the standard deviation and the co-occurrence probability, perform probability weighting on the time-domain abnormal fluctuation intensity within each subsequence segment, and generate a time-domain abnormal fluctuation feature vector with timestamp marks. Wherein, the time resolution of the time-domain abnormal fluctuation feature vector is consistent with the update frequency of the three-dimensional spatio-temporal features of the ice surface state feature set.

[0143] According to the calculated standard deviation and co-occurrence probability, perform probability weighting on the time-domain abnormal fluctuation intensity within each subsequence segment. For example, the time-domain abnormal fluctuation intensity of a certain subsequence segment is I. According to the values of the standard deviation and the co-occurrence probability, calculate a weighting coefficient w through an existing weighting algorithm. Multiply the time-domain abnormal fluctuation intensity I by the weighting coefficient w to obtain the weighted time-domain abnormal fluctuation intensity Iw. Perform such operations on each subsequence segment, and then combine these weighted time-domain abnormal fluctuation intensities with the corresponding timestamps to generate a time-domain abnormal fluctuation feature vector with timestamp marks. The time resolution of this time-domain abnormal fluctuation feature vector is the same as the update frequency of the three-dimensional spatio-temporal features of the ice surface state feature set. For example, it is updated once per hour, which ensures that in subsequent analyses, the two can be accurately matched in time to comprehensively consider various state information of the ice surface.

[0144] In a possible implementation manner, step S140 includes:

[0145] Step S141, collect environmental noise data of the target glacier area in real time, and the environmental noise data includes glacier vibration signals, wind speed fluctuation frequencies, and precipitation pulse intensities.

[0146] In large glacier regions, dedicated sensors are set up to collect environmental noise data in real time. For glacier vibration signals, vibration sensors installed at different positions on the glacier are used to obtain them. For example, vibration sensors are installed at key positions such as the edge area and the central area of the glacier. These sensors can accurately detect the minute vibrations of the glacier and record information such as 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 sensors can measure the wind speed in the glacier region in real time and record the change of the wind speed over time, thereby obtaining the wind speed fluctuation frequency. For the precipitation pulse intensity, devices such as rain gauges are used to collect it. When precipitation occurs, the rain gauge can measure the precipitation amount per unit time, and determine the precipitation pulse intensity according to the change of the precipitation amount. 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 region, and these changes may affect the accuracy of the ice landslide monitoring model. Therefore, it is necessary to dynamically adjust the weights of the model parameters.

[0147] Step S142: Convert the environmental noise data into spectral feature vectors, input the spectral feature vectors into the noise suppression layer in the ice landslide monitoring model for feature dimensionality reduction, and then use an online learning algorithm to perform weighted fusion on the dimensionality-reduced spectral feature vectors and the current model parameters to generate parameter adjustment gradients.

[0148] First, convert the collected environmental noise data (glacier vibration signals, wind speed fluctuation frequency, precipitation pulse intensity) into spectral feature vectors. Taking the glacier vibration signal as an example, by performing a Fourier transform on the vibration signal, it is converted from a time-domain signal to a frequency-domain signal to obtain the spectral features of the vibration signal. Similarly, corresponding conversions are also performed on the wind speed fluctuation frequency and the precipitation pulse intensity, and the spectral features of these three environmental noise data are combined to form a spectral feature vector. Then input this spectral feature vector into the noise suppression layer in the ice landslide monitoring model. In the noise suppression layer, the spectral feature vector is subjected to feature dimensionality reduction through existing dimensionality reduction algorithms. For example, the principal component analysis (PCA) algorithm is used to reduce the multiple feature dimensions in the spectral feature vector to a few main dimensions, and these main dimensions can retain most of the information of the original spectral feature vector. After feature dimensionality reduction, an online learning algorithm is used to perform weighted fusion on the dimensionality-reduced spectral feature vectors and the current model parameters. The online learning algorithm assigns different weights to the dimensionality-reduced spectral feature vectors and the current model parameters according to pre-set rules, and then performs operations such as weighted summation on them to generate parameter adjustment gradients. This parameter adjustment gradient reflects the direction and magnitude of the adjustment of the model parameters according to the environmental noise data.

[0149] Step S143: Adjust the convolution kernel size and channel attention weights of the multi-scale feature fusion layer according to the parameters. The adjustment amplitude of the convolution kernel size is positively correlated with the rising rate of the ice surface risk level label.

[0150] Adjust the gradient according to the generated parameters to update the convolution kernel size and channel attention weights of the multi-scale feature fusion layer. Take the convolution kernel size as an example. If the parameter adjustment gradient indicates that the convolution kernel size needs to be increased, then increase the size parameters such as the side length of the convolution kernel according to the set rules. At the same time, the channel attention weights are also updated accordingly. The adjustment amplitude 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 amplitude of the convolution kernel size is larger. If the ice surface risk level label rises rapidly from a low risk level to a high risk level, then the convolution kernel size of the multi-scale feature fusion layer may increase significantly, and the channel attention weights will also be adjusted accordingly, so that the model can better focus on the key features related to ice avalanches and improve the model's monitoring ability of ice surface state changes.

[0151] In a possible implementation manner, step S142 includes:

[0152] Step S1421: Construct a mapping relationship table between the environmental noise data and the changes in model parameters in historical disaster events. The mapping relationship table contains model learning rate adjustment coefficients corresponding to different noise intensity intervals.

[0153] Taking the glacier vibration signal as an example, according to the data in historical disaster events, when the intensity of the glacier vibration signal is within a certain interval, record the adjustment of the model parameters at that time, especially the adjustment coefficient of the model learning rate. Similarly, conduct similar analyses on the wind speed fluctuation frequency and precipitation pulse intensity, and construct a mapping relationship table that includes 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 coefficients. 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 the wind speed fluctuation frequency and precipitation pulse intensity.

[0154] Step S1422: Match the optimal learning rate adjustment coefficient from the mapping relationship table according to the amplitude distribution of the dimension-reduced spectral feature vector.

[0155] Taking the spectral components corresponding to the glacier vibration signals in the spectral feature vector as an example, observe the magnitude distribution of their amplitudes. If the amplitude distribution is within the low-intensity range, according to the previously constructed mapping relation table, the corresponding model learning rate adjustment coefficient is matched as a1. Similarly, similar analyses are also 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 and can make the model more reasonable and accurate when adjusting parameters according to environmental noise.

[0156] Step S1423, calculate the iteration step size and direction vector of the parameter adjustment gradient based on the optimal learning rate adjustment coefficient. Among them, the direction vector is used to constrain the parameter update path of the online learning algorithm and avoid the model weights deviating from the core feature dimensions of disaster warning.

[0157] For example, assume that the optimal learning rate adjustment coefficient is α. According to the formula of the online learning algorithm, calculate the iteration step size of the parameter adjustment gradient. The iteration step size determines the magnitude of each adjustment of the model parameters. At the same time, calculate the direction vector. 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 this direction vector is to constrain the parameter update path of the online learning algorithm. For example, during the model parameter update process, without the constraint of the direction vector, it may lead to the model weights deviating 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 the environmental changes in the glacier area and accurately monitor the ice surface state.

[0158] For example, in a possible implementation manner, the method for generating the ice surface texture visualization data includes:

[0159] Step S151, select the ice surface visible light image and the corresponding ice surface displacement rate data of the current time period from the dynamic image sequence.

[0160] In this embodiment, for the large glacier area that has been monitored before, its dynamic image sequence contains glacier-related data for multiple time periods. In this embodiment, visible light images of the ice surface and corresponding ice surface displacement rate data for the current time period (for example, within the current 24 hours) are selected therefrom. The visible light images of the ice surface can intuitively display the appearance texture, fissures, etc. of the ice surface. These images are obtained by the multispectral imager previously set in the glacier area. Within the current time period, one frame of visible light image is collected every set time (such as every hour), and these images form the visible light image sequence of the ice surface for the current time period. At the same time, the corresponding ice surface displacement rate data is also measured by devices such as synthetic aperture radar within the same time period. The ice surface displacement rate data details the displacement speed of each position on the ice surface relative to the previous moment at each measurement moment. These data are key information for accurately quantifying the motion state of the ice surface, corresponding to the visible light images of the ice surface, and providing a basis for generating visual data in the subsequent process.

[0161] Step S152: Perform fissure edge enhancement processing on the visible light image of the ice surface to obtain an enhanced visible light image.

[0162] Taking the visible light image of the ice surface in the selected current time period as an example, since the edges of the ice surface fissures in the original image may not be clear enough, which is not conducive to intuitively observing and analyzing the condition of the ice surface, fissure edge enhancement processing is required. First, an existing image edge detection algorithm is used to process the visible light image of the ice surface. This algorithm identifies edges by analyzing the change in gray values of pixels in the image. For the edge part of the ice surface fissure, the gray values of the surrounding pixels usually change greatly. The algorithm enhances the contrast of this gray value change, making the fissure edge more obvious. For example, in the original image, the difference in gray values between the pixels at the fissure edge and the surrounding pixels may be small. After being processed by the edge detection algorithm, the gray values of the pixels at the fissure edge are adjusted to be darker or brighter (depending on the specific settings of the algorithm), thus forming a more distinct contrast with the surrounding pixels. After performing such processing on all the fissure edges in the entire visible light image of the ice surface, an enhanced visible light image is obtained. In this enhanced visible light image, the contour of the ice surface fissure is clearer, and information such as the distribution and trend of the fissures can be observed more conveniently.

[0163] Step S153: Convert the ice surface displacement rate data into a color gradient layer, where the color depth of the color gradient layer represents the displacement speed, and superimpose this color gradient layer on the enhanced visible light image to obtain a superimposed image.

[0164] According to the obtained ice surface displacement rate data, convert it into a color gradient layer. Assume a color mapping rule is set. For example, the color corresponding to the ice surface area with the slowest displacement rate is light blue, and as the displacement rate increases, the color gradually deepens, and the color corresponding to the area with the fastest displacement rate is dark blue. Taking a certain area in the ice surface displacement rate data as an example, the ice surface displacement rate of 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. Another area has an ice surface displacement rate of 0.5 meters per hour, and 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 overlay this color gradient layer on the previously obtained enhanced visible light image. During the overlay process, ensure that the color gradient layer and the enhanced visible light image are accurately spatially corresponding, that is, each pixel position in the color gradient layer is consistent with the corresponding pixel position in the enhanced visible light image. Through the overlay, an overlaid image is obtained. In this overlaid image, the texture and cracks of the ice surface (from the enhanced visible light image) can be clearly seen, and the speed of displacement of different areas of the ice surface can be intuitively understood (through the color gradient layer).[[ID=...]] [[ID=...]]

[0165] Step S154, mark the risk area corresponding to the ice surface risk level label in the overlaid image, and use a dynamic arrow to represent the moving direction and speed magnitude of the ice body within this risk area.[[ID=...]] [[ID=...]]

[0166] [[ID=...]] [[ID=...]]

[0167] In the already obtained overlaid image, mark the corresponding risk areas according to the previously generated ice surface risk level labels. For example, through a series of previous analyses, the high-risk areas, medium-risk areas, and low-risk areas of the ice surface are determined. For the high-risk areas, circle this area in the overlaid image with a set mark (such as a red border). At the same time, within each risk area, use a dynamic arrow to represent the moving direction and speed magnitude of the ice body. Taking a certain point in the high-risk area as an example, through 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 this speed magnitude and direction, draw a dynamic arrow at this point in the image. The direction of the arrow points to the northeast, indicating the moving direction of the ice body, and the length of the arrow is set according to the speed magnitude. The faster the speed, the longer the arrow. For other points within the risk area, draw dynamic arrows in the same way. In this way, in the overlaid image, by marking the risk areas and drawing dynamic arrows, the risk status of different areas of the ice surface and the movement of the ice body within these areas can be intuitively displayed.[[ID=...]] [[ID=...]]

[0167] Step S155, according to the range of the warning trigger threshold, mark a color scale bar on the edge of the overlaid image to visually compare the correlation between the current displacement rate and historical disaster data, and obtain the marked image.

[0168] For example, the warning trigger threshold sets the range of ice surface displacement rates corresponding to different risk levels. The displacement rate range corresponding to the lowest risk level is 0 to 0.1 meters per hour, the medium risk level is 0.1 to 0.3 meters per hour, and the high risk level is above 0.3 meters per hour. According to this 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 that previously converted 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 this color scale bar, the correlation between the current ice surface displacement rate and the warning trigger threshold set in the historical disaster data can be visually 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 that area has a higher risk. After such marking, the marked image is obtained.

[0169] Step S156, bind the marked image with the geographical coordinates of the ice landslide risk warning signal to generate a scalable map-style warning image. Among them, the boundary shape and arrow length of the risk area in the map-style warning image change in real time with the update of the dynamic image sequence and are displayed through the monitoring terminal.

[0170] For example, the central geographical coordinates of a high-risk area are (x1, y1, z1). Bind this geographical coordinate with the corresponding risk area in the marked image. In this way, a scalable map-style warning image is generated. In this map-style warning image, the boundary shape and arrow length of the risk area change in real time with the update of the dynamic image sequence. As time goes by, the state of the glacier will change, and the displacement of the ice surface, crack expansion, etc. will affect the boundary shape of the risk area and the moving speed of the ice body. For example, as the ice surface moves, a certain risk area may expand or shrink, and its boundary shape will change accordingly. At the same time, the moving speed of the ice body within the risk area may also change, resulting in a change in the arrow length representing the moving direction and speed magnitude of the ice body. This map-style warning image is displayed to relevant personnel through the monitoring terminal. Relevant personnel can perform zoom operations on the map-style warning image on the monitoring terminal to view the conditions of different areas of the ice surface in more detail, timely understand the risk status of the ice surface and take corresponding measures.

[0171] Figure 2Schematic diagram of the structure of the ice landslide monitoring and warning system 10 based on AI image recognition provided by an embodiment of the present invention, including a processor 102, a memory 104, and a bus 106. Among them, the memory 104 is used to store execution instructions, including internal memory and external memory. The internal memory can also be understood as the main memory, which is used to temporarily store the operation data in the processor 102 and the data exchanged with external memories such as hard disks. The processor 102 exchanges data with the external memory through the internal memory. When the ice landslide monitoring and warning system 10 based on AI image recognition runs, the processor 102 communicates with the memory 104 through the bus 106, so that the processor 102 executes the ice landslide monitoring and warning method based on AI image recognition according to the embodiment of the present invention.

[0172] For ease of description, only one processor is described in the ice landslide monitoring and warning system 10 based on AI image recognition. However, it should be noted that the ice landslide monitoring and warning system 10 in the present application may also include multiple processors. Therefore, the steps executed by one processor described in the present application may also be jointly executed or separately executed by multiple processors. For example, if the processor of the ice landslide monitoring and 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 jointly executed by two different processors or separately executed 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 jointly execute steps A and B.

[0173] In addition, an embodiment of the present invention also provides a readable storage medium, in which computer-executable instructions are preset. When the processor executes the computer-executable instructions, the above-mentioned ice landslide monitoring and warning method based on AI image recognition is implemented.

[0174] It should be noted that, in order to simplify the description of the present invention disclosure and thus help the understanding of one or more embodiments of the invention, in the previous description of the embodiments of the present invention, sometimes multiple features are merged 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: 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; 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; 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; 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; Sending the ice landslide risk warning signal and corresponding ice surface texture visualization data to a monitoring terminal; The training method of the ice landslide monitoring model includes: Obtain an image training set of the target glacier area before a historical ice landslide event, wherein the image training set includes a visible light image sequence and a thermal infrared image sequence of the ice surface and ice displacement rate measurement data at corresponding timestamps; Grayscale processing is performed on the visible light image sequence to extract the pixel coordinate set of the edge of the ice 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 suddenly changes; Spatially superimposing the pixel coordinate set of the edge of the ice 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; 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 breakage 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 characteristics and the thermal diffusion characteristics 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 steps; A bidirectional long short-term memory network is used to perform time series modeling on the aligned ice surface state evolution tensor, outputting the predicted ice surface stability value at each time step. Performing loss calculation on the ice surface stability prediction value and the actual ice surface stability prediction value, and adjusting weight parameters of the three-dimensional convolutional neural network and the bidirectional long short-term memory network by back propagation; During the training process, randomly masking pixel blocks in the ice crack area in the visible light image sequence, reconstructing the texture continuity of the masked area using a generative adversarial network, and adding the reconstructed image to the image training set; The bidirectional long short-term memory network is used to perform time series modeling on the aligned ice surface state evolution tensor, and outputs the predicted value of ice surface stability at each time step, including: Splitting the aligned ice surface state evolution tensor into continuous tensor subsequences according to time steps, each tensor subsequence containing a temporal deformation feature sequence of the ice surface texture and thermal diffusion features of the fragile zone at the current time step; Inputting the tensor quantum sequence into the forward propagation layer of the bidirectional long short-term memory network, calculating the hidden state of each time step in chronological order, and generating a forward hidden state sequence; Synchronously inputting the tensor quantum sequence into the backward propagation layer of the bidirectional long short-term memory network in reverse order, calculating the hidden state of each time step in reverse chronological order, and generating a backward hidden state sequence; 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; Inputting the bidirectional fused hidden state sequence into a fully connected layer, mapping the fused hidden state at each time step into an initial probability distribution of ice surface stability prediction values; Performing a sliding average filter on the initial probability distribution in the time dimension to eliminate instantaneous noise interference and generate a smoothed ice surface stability prediction value sequence; Performing point-to-point matching on the smoothed ice surface stability prediction value sequence and the time step of the acceleration change curve, and calculating the error gradient between the prediction value and the actual ice surface stability measurement value at each time step; Inversely 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 the deformation characteristics and thermal diffusion characteristics that are strongly related to the ice body fracture event are preferentially retained when the hidden state of the subsequent time step is updated; During the time series modeling process, when it is detected that the difference between the predicted value of 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 backpropagation layer is forcibly reset to avoid interference of historical irrelevant features on the prediction of the current time step; The ice surface stability prediction value for each time step is output, and the temporal change rate of the ice surface stability prediction value is spatially correlated with the thermal distribution data of the ice surface displacement rate to verify the spatial consistency of the ice surface stability prediction value.

2. The ice landslide monitoring and early warning method based on AI image recognition according to claim 1 is characterized in that: The method includes collecting a dynamic image sequence of the target glacier area, extracting three-dimensional spatiotemporal features of ice surface texture, crack distribution morphology, and displacement rate in the dynamic image sequence, and generating an ice surface state feature set, including: Multispectral imagers, thermal infrared sensors, and synthetic aperture radars are used to synchronously collect visible light image sequences, thermal infrared image sequences, and radar interferometry image sequences of the target glacier area within a continuous time window, and then perform spatiotemporal registration according to a unified timestamp. Adaptive histogram equalization is performed on the visible light image sequence to enhance the contrast between 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 to identify the boundary coordinates of pixel regions 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 a three-dimensional space coordinate system, and generating an 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 spatial coordinate system to construct 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 similarity between the fused feature vector and the standard feature template in the historical glacier deformation database is measured, 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 of extracting crack edge pixel clusters of the ice surface texture using a multi-scale edge detection algorithm and performing optical flow tracking on the crack edge pixel clusters of adjacent frames to generate crack expansion direction vectors includes: 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 macro cracks on the ice surface, while the second branch convolutional network uses a small-scale convolution kernel to detect the bifurcation endpoints of micro cracks. Perform 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. After removing pseudo-edge noise using 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 record the centroid coordinates and principal axis direction angle of each pixel cluster. Between 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; Generate the velocity components of each crack pixel cluster in three dimensions: 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 median filtered 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 method of inputting the visible light image sequence, the thermal infrared image sequence and the 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 includes: 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 spatiotemporally aligned visible light tensor and thermal infrared tensor; 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 a three-dimensional convolutional neural network, extracting the displacement gradient of 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, and the mean and variance of the crack displacement gradient in each time window are extracted to generate a deformation feature sequence of the ice surface texture over time; 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 feature map; Performing a spatial pooling operation on the thermal diffusion feature map in the time dimension to extract the extreme temperature change of each spatial position in consecutive time steps to generate 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.

5. The ice landslide monitoring and early warning method based on AI image recognition according to claim 1 is characterized in that: 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; Performing time series decomposition on the multidimensional 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; 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 risk area; Mapping the geometric center coordinates to the spatial location 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; Adjusting the dynamic threshold interval of the preset risk level according to the spatial overlap and displacement trend similarity, and performing a weighted score on the time domain abnormal fluctuation characteristics based on the dynamic threshold interval; The weighted scoring results are input into the preset fuzzy logic classifier to match the membership function corresponding to the ice surface risk level label to generate the independent risk level of each risk area and the global ice surface comprehensive risk level; Performing a confidence check on the independent risk level and the global ice surface comprehensive risk level, eliminating 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.

6. The ice landslide monitoring and early warning method based on AI image recognition according to claim 5 is characterized in that: The multi-dimensional feature vector is subjected to time series decomposition to extract the acceleration extreme value of the ice surface displacement rate and the mutation point position of the crack expansion rate, and generate time domain abnormal fluctuation characteristics, including: Sliding the multidimensional feature vector into multiple subsequence segments on a time axis at a fixed step size, and performing a decomposition process on a trend term and a residual term for 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 a wavelet transform 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; Aligning and matching the timestamp set of the acceleration extreme value with the timestamp of the crack expansion mutation point, and calculating the standard deviation and co-occurrence probability of the time intervals between the two; According to the standard deviation and co-occurrence probability, the time domain abnormal fluctuation intensity within 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 update frequency of the three-dimensional spatiotemporal characteristics of the ice surface state feature set.

7. 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: Real-time collection of environmental noise data in the target glacier area, including glacier vibration signals, wind speed fluctuation frequency, and precipitation pulse intensity; The environmental noise data is converted into a spectrum feature vector, and the spectrum feature vector is input into the noise suppression layer in the ice landslide monitoring model for feature dimensionality reduction, and the spectrum feature vector after dimensionality 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 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.

8. The ice landslide monitoring and early warning method based on AI image recognition according to claim 7 is characterized in that: The online learning algorithm is used to perform weighted fusion of the reduced-dimensional spectrum feature vector and the current model parameters to generate parameter adjustment gradients, including: 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; Matching an optimal learning rate adjustment coefficient from the mapping relationship table according to the amplitude distribution of the spectrum feature vector after dimension reduction; 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.

9. 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 8 above.

Images