Hot melt collapse prediction method and system based on multispectral imaging and medium

Abstract

The invention provides a hot melt collapse prediction method and system based on multispectral imaging and a medium, and relates to the technical field of hot melt collapse prediction, and the method comprises the steps: obtaining a to-be-measured frozen soil region, carrying out the partitioning, and obtaining the soil data; extracting feature vectors through a plurality of trained feature extraction networks; feature weights are distributed, weighted fusion is carried out on the spectral feature vectors based on the feature weights, and fusion features are formed through serial connection; a hot melt collapse probability prediction model based on a multi-head attention mechanism Transform model is constructed; the fusion features are converted into a sequence mode, the fusion features are input into a corresponding hot melt collapse probability prediction model, the hot melt collapse probability of each subarea of the frozen soil area to be detected is obtained, a hot melt collapse risk distribution map is drawn, risk levels of different areas are marked, and early warning is carried out. According to the method, multiple spectral band data are fused and analyzed, the prediction precision and reliability are improved, the slump risk is comprehensively evaluated, the prediction model is constructed, and a scientific basis is provided for early warning and disaster prevention.

Description

Technical Field

[0001] The present invention relates to the technical field of sensitive thermal melt landslide prediction, and in particular to a thermal melt landslide prediction method, system and medium based on multi-spectral imaging. Background Art

[0002] The most notable feature of my country's permafrost is that it is mainly mountain permafrost. Affected by global warming and human disturbance, it is prone to thermal thaw landslides, which directly affect the operational safety of roads, railways and other infrastructure in high-altitude cold areas, and are accompanied by a large amount of carbon release. At present, the data sources for thermal thaw landslide areas include historical aerial photographs, satellite images, drone images, RTK measured data, other meteorological data and basic geographic information. Field surveys, remote sensing image interpretation, statistical analysis and other methods are used to understand the spatial distribution characteristics, temporal variation characteristics, influencing factors and future development trend predictions of thermal thaw landslides.

[0003] However, in the current field of frozen soil prediction, the existing prediction methods still have significant limitations. These methods mainly rely on the direct acquisition and analysis of frozen soil surface data. Although this approach is simple and direct, it fails to fully consider the multi-dimensional factors that affect the frozen soil state and largely ignores the profound impact of environmental data and soil and vegetation data on the prediction results. Summary of the invention

[0004] To solve the above problems, the present invention provides a thermal melt landslide prediction method based on multi-spectral imaging. The method improves the prediction accuracy and reliability by fusing and analyzing data from multiple spectral bands, comprehensively assesses the landslide risk, and constructs a prediction model to provide a scientific basis for early warning and disaster prevention.

[0005] To achieve the above object, the present invention provides the following technical solutions.

[0006] The present invention provides a method for predicting thermal melt collapse based on multispectral imaging, comprising the following steps:

[0007] The frozen soil area to be tested is divided into zones, and the spectral data, GIS data and deep soil data of different bands on the surface of the zones are obtained respectively;

[0008] Through multiple trained feature extraction networks, spectral feature vectors of different data representation types of spectral data in different bands, as well as feature vectors of GIS data and deep soil data are extracted respectively; feature weights are assigned according to the degree of influence of the data representation type of the spectral feature vector on the thermal melt landslide, and the spectral feature vectors are weighted fused based on the feature weights, and are concatenated with the feature vectors of GIS data and deep soil data to form fusion features; wherein the data representation types include surface temperature, vegetation index, soil moisture, surface water content and rock and soil type;

[0009] A Transformer model based on a multi-head attention mechanism is constructed, and multiple thermal melt landslide probability prediction models for different periods are obtained based on the thermal melt landslide datasets of different periods. The thermal melt landslide datasets are divided into datasets of different periods by month.

[0010] The fused features are converted into sequence patterns and input into the corresponding thermal thaw landslide probability prediction model according to the current period of the permafrost area to be tested, so as to obtain the thermal thaw landslide probability of each zone in the permafrost area to be tested, and a thermal thaw landslide risk distribution map is drawn according to the thermal thaw landslide probability of each zone, marking the risk levels of different areas and issuing early warnings.

[0011] Preferably, the step of respectively acquiring spectral data of different bands of the partitioned surface comprises the following steps:

[0012] The image of the frozen soil area to be tested is obtained by multi-spectral imaging equipment, and grid division is performed to form multiple sub-areas, and the sub-areas are numbered;

[0013] The sub-regions are identified in sequence according to the numbers. Based on the spectral features and texture features in the image, the threshold segmentation method based on pixel value differences and the trained partition recognition network are used to obtain the frozen soil area to be tested. Secondary partitioning is performed through the grid division method to obtain multiple secondary partition areas containing the frozen soil area grids to be tested.

[0014] The spectral data of different bands of each secondary sub-region of the frozen soil area to be tested are collected by a multi-spectral imaging device, and the spectral data are preliminarily processed by edge computing technology, including data cleaning, denoising and compression; the multi-spectral imaging device is erected on a high tower or carried on a drone above the frozen soil area to be tested; the multi-spectral imaging device is erected on a high tower or carried on a drone through a motor with an encoder, and the motor drives the multi-spectral imaging device to swing; the multi-spectral imaging device is provided with a zoom lens;

[0015] The collected spectral data is transmitted to the processing center via Apache Kafka;

[0016] In the data processing center, Apache Flink is used to perform radiation correction, geometric correction, and atmospheric correction on the acquired spectral data, and the spectral data is converted into spectral vectors and filtered, aggregated, and connected.

[0017] Preferably, the GIS data includes slope, aspect and elevation data; the deep soil data includes soil temperature data at different depths; the soil temperature data at different depths are obtained through deep soil temperature sensors arranged in the active layer, frozen soil layer and sliding layer, and the depth and thickness of the frozen soil layer, active layer and sliding layer are determined through the soil temperature data.

[0018] Preferably, the feature extraction network is based on a CNN network model, and is trained to obtain multiple feature extraction networks using labeled surface temperature and thermal infrared band datasets, vegetation index and red band and near-infrared band datasets, vegetation growth status and green band datasets, soil moisture and short-wave infrared band, surface moisture and visible light band and near-infrared band datasets, and rock and soil type and short-wave infrared band datasets, respectively, as feature extraction models for spectral feature vectors of different data characterization types; wherein each label of each dataset is obtained through a sensor system.

[0019] Preferably, the weighted fusion of spectral feature vectors based on feature weights comprises the following steps:

[0020] Determine the influence of surface temperature, vegetation index, soil moisture, surface water content and rock and soil type on thermal melt landslide, the correlation and complementarity between different eigenvectors and their contribution to the fusion results, and determine the importance ranking results of each spectral eigenvector based on the predicted results and actual results of the thermal melt landslide probability prediction model using a specific sample set according to the SHAP method;

[0021] According to the importance ranking results of each spectral feature vector, weights are proportionally allocated and a weight allocation strategy is determined;

[0022] According to the weight distribution strategy, the weight of each feature vector in the fusion result is calculated, and the feature vectors are weightedly concatenated and fused using the weight to generate fusion features.

[0023] Preferably, the Transformer model based on the multi-head attention mechanism includes an encoder and a decoder;

[0024] The encoder includes a plurality of stacked encoder layers, each of which includes a multi-head attention sublayer and a feedforward neural network sublayer; the multi-head attention sublayer and the feedforward neural network sublayer use residual connection and layer normalization;

[0025] The decoder includes a plurality of stacked decoder layers, each of which includes two multi-head attention sublayers and a feedforward neural network sublayer; the first multi-head attention sublayer is a self-attention sublayer; the second multi-head attention sublayer is an encoder-decoder attention sublayer; the multi-head attention sublayer and the feedforward neural network sublayer use residual connection and layer normalization;

[0026] The feedforward neural network includes two linear transformations and an activation function.

[0027] Preferably, the conversion of the fusion features into a sequence pattern comprises the following steps:

[0028] Map each spectral feature vector to a fixed embedding space through a linear layer to form an embedding vector;

[0029] A position code is added to each embedding vector, and these embedding vectors and the position code are combined into a sequence; wherein the sequence of each spectral feature vector is a plurality of vectors of equal length.

[0030] Preferably, the historical prediction data of the probability of thermal melt collapse in each sub-area in each month of each year is obtained, and the data is cleaned;

[0031] Use Matplotlib to build a comparison chart based on historical forecast data, use a line chart to show the forecast data of the same month in different years, and use color or line type to distinguish data from different years;

[0032] Based on historical prediction data, a prediction model based on neural network is selected to predict the probability of thermal melt collapse in each subarea in the future;

[0033] Plot the forecast results into a chart and compare them with historical forecast data to obtain possible future change trends.

[0034] The present invention also provides a thermal melt collapse prediction device based on multispectral imaging, comprising:

[0035] The data acquisition module is used to obtain the frozen soil area to be tested and divide it into zones, and obtain the spectral data of different bands on the surface of the zones respectively;

[0036] The data processing and analysis module is used to extract spectral feature vectors of different data representation types of spectral data in different bands through a feature extraction network after multiple rounds of training; it is also used to assign feature weights according to the data representation type of the spectral feature vector for the degree of influence of thermal melt landslide, and perform weighted fusion of the spectral feature vectors based on the feature weights to form fusion features; wherein the data representation types include surface temperature, vegetation index, soil moisture, surface water content and rock and soil types;

[0037] The prediction model building module is used to build a Transformer model based on a multi-head attention mechanism and obtain a thermal melt landslide probability prediction model based on the thermal melt landslide dataset;

[0038] The prediction module is used to convert the fusion features into a sequence pattern and input it into the thermal thaw landslide probability prediction model to obtain the thermal thaw landslide probability of each subarea in the frozen soil area to be tested;

[0039] The early warning release module is used to draw a thermal melt landslide risk distribution map based on the probability of thermal melt landslide in each zone, mark the risk levels of different areas and issue early warnings, and release early warning information to relevant departments and personnel through SMS, WeChat and websites.

[0040] The present invention also provides a thermal melt collapse prediction system based on multispectral imaging, the system comprising:

[0041] processor;

[0042] a memory having stored thereon a computer program executable on the processor;

[0043] Wherein, when the computer program is executed by the processor, the steps of the thermal melt landslide prediction method based on multispectral imaging are implemented.

[0044] The present invention also provides a computer-readable storage medium, on which a data processing program is stored. When the data processing program is executed by a processor, the steps of the thermal melt landslide prediction method based on multispectral imaging are implemented.

[0045] Beneficial effects of the present invention:

[0046] The present invention proposes a method, system and medium for predicting thermal melt landslide based on multispectral imaging. The method divides the frozen soil area to be tested, and obtains spectral data of different bands in each partition, thereby improving the accuracy of data acquisition and ensuring the comprehensiveness and representativeness of the data. By acquiring multi-band spectral data, information of various data characterization types such as surface temperature, vegetation index, soil moisture, surface water and rock and soil type is captured, and multiple trained feature extraction networks are used to perform in-depth feature extraction on spectral data of different bands, providing rich feature information. Then, according to the different degrees of influence of the data characterization type of the spectral feature vector on the thermal melt landslide, corresponding feature weights are assigned to it, fully considering the importance of different features in predicting thermal melt landslide, ensuring the accuracy and effectiveness of weighted fusion and the representativeness of features, and using the Transformer model with a multi-head attention mechanism for prediction, effectively improving the accuracy of the overall prediction. The method draws a distribution map of thermal melt landslide risk and indicates the risk levels of different regions, which not only provides intuitive prediction results for decision makers, but also provides a scientific basis for them to formulate targeted early warning and prevention measures. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flow chart of a method for predicting thermal melt collapse based on multispectral imaging according to an embodiment of the present invention;

[0048] Figure 2 is a schematic diagram of a multi-channel imager and a sensor arrangement area according to an embodiment of the present invention;

[0049] Figure 3 is a flow chart of weighted fusion of spectral feature vectors according to an embodiment of the present invention;

[0050] Figure 4 This is a flow chart of fusion feature sequence conversion according to an embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0052] Example

[0053] In the current field of frozen soil prediction, the existing prediction methods still have significant limitations. These methods mainly rely on the direct acquisition and analysis of frozen soil surface data. Although this approach is simple and direct, it fails to fully consider the multi-dimensional factors that affect the state of frozen soil. In particular, they largely ignore the profound impact of environmental data and soil vegetation data on the prediction results. Environmental data, including temperature changes, precipitation patterns, humidity conditions, wind direction and speed, are indispensable factors in shaping the dynamic changes of frozen soil. For example, climate change can directly affect the melting rate and refreezing process of frozen soil, while the increase or decrease in precipitation is related to the moisture content of the soil, which in turn affects its thermal conductivity. In addition, as an important component of the ground cover, the type, density and growth status of soil vegetation not only affect the reflectivity and heat exchange of the ground, but also regulate the soil structure and moisture dynamics through root activity, all of which indirectly or directly affect the stability of the frozen soil layer. To this end, the present invention proposes a method for predicting thermal melt and landslides based on multispectral imaging. The process is as follows: Figure 1 As shown, the following steps are included:

[0054] S1: Obtain the frozen soil area to be tested and divide it into zones, and obtain the spectral data, GIS data and deep soil data of different bands on the surface of the zones respectively.

[0055] S2: Through multiple trained feature extraction networks, spectral feature vectors of different data representation types of spectral data in different bands, as well as feature vectors of GIS data and deep soil data are extracted respectively; feature weights are assigned according to the degree of influence of the data representation type of the spectral feature vector on the thermal melt landslide, and the spectral feature vectors are weightedly fused based on the feature weights, and concatenated with the feature vectors of GIS data and deep soil data to form fusion features; among which, the data representation types include surface temperature, vegetation index, soil moisture, surface water content and rock and soil type.

[0056] S3: Construct a Transformer model based on a multi-head attention mechanism, and train multiple heat-melt landslide probability prediction models for different periods based on heat-melt landslide datasets of different periods; wherein the heat-melt landslide dataset is divided into datasets of different periods by month. Since the probability of landslides in spring, summer, autumn, and winter is different, the time factor (by month) is added to distinguish the data, and different probability prediction models are generated respectively.

[0057] S4: Convert the fusion features into a sequence pattern, and input it into the corresponding thermal thaw landslide probability prediction model according to the current period of the permafrost area to be tested, so as to obtain the thermal thaw landslide probability of each zone in the permafrost area to be tested, and draw a thermal thaw landslide risk distribution map according to the thermal thaw landslide probability of each zone, mark the risk level of different areas and issue early warning.

[0058] Specifically, obtaining spectral data of different bands of the partition surface respectively includes the following steps:

[0059] S1.1: Obtain images of the frozen soil area to be tested through multispectral imaging equipment, divide the images into grids, form multiple sub-areas, and number the sub-areas;

[0060] S1.2: Identify the sub-regions in sequence according to the numbers, and based on the spectral features and texture features in the image, use the threshold segmentation method based on pixel value differences and the trained partition recognition network to obtain the frozen soil area to be tested; perform secondary partitioning by grid division method to obtain multiple secondary partition areas containing the frozen soil area grids to be tested;

[0061] S1.3: Use multispectral imaging equipment to collect spectral data of different bands in each secondary partition area of ​​the frozen soil area to be tested, and use edge computing technology to perform preliminary processing on the spectral data, including data cleaning, denoising and compression; the multispectral imaging equipment is installed on a high tower or carried by a drone above the frozen soil area to be tested; the multispectral imaging equipment is installed on a high tower or carried on a drone through a motor with an encoder, and the motor drives the multispectral imaging equipment to swing; the multispectral imaging equipment is provided with a zoom lens.

[0062] The layout of multi-channel imagers is as follows Figure 2 shown.

[0063] S1.4: The collected spectral data is transmitted to the processing center via Apache Kafka.

[0064] S1.5: In the data processing center, Apache Flink is used to perform radiation correction, geometric correction, and atmospheric correction on the acquired spectral data, and the spectral data is converted into spectral vectors, and filtered, aggregated, and connected.

[0065] In order to extract features that can characterize surface temperature, vegetation index, soil moisture, surface water content and rock and soil type information from spectral data of different bands, this embodiment uses multiple CNN network models to perform feature extraction. Specifically: the feature extraction network is based on the CNN network model, and is trained to obtain multiple feature extraction networks through labeled surface temperature and thermal infrared band data sets, vegetation index and red band and near-infrared band data sets, vegetation growth status and green band data sets, soil moisture and short-wave infrared band, surface water content and visible light band and near-infrared band data sets, and rock and soil type and short-wave infrared band data sets, as feature extraction models for spectral feature vectors of different data characterization types; wherein each label of each data set is obtained through a sensor system, and the layout of the sensor is as follows: Figure 2 shown.

[0066] In addition, in order to achieve large-scale data collection, this embodiment uses a motor to drive the equipment to rotate, a high tower or a drone to increase the monitoring range, and a secondary identification partition to refine the identification area. The slope within the partition and the deep soil temperature data in the GIS data have an impact on landslides. Therefore, when extracting data features, the terrain data dimension and the permafrost data dimension are increased to improve the representativeness of the features.

[0067] Specifically, this embodiment proposes to perform weighted fusion of spectral feature vectors based on feature weights, such as Figure 3 As shown, the following steps are included:

[0068] S2.1: Determine the influence of surface temperature, vegetation index, soil moisture, surface water content and rock and soil type on thermal melt landslide, the correlation and complementarity between different eigenvectors and their contribution to the fusion results, and determine the importance ranking results of each spectral eigenvector based on the predicted results and actual results of the thermal melt landslide probability prediction model using a specific sample set according to the SHAP method.

[0069] S2.2: According to the importance ranking results of each spectral feature vector, weights are proportionally allocated and a weight allocation strategy is determined.

[0070] S2.3: According to the weight distribution strategy, the weight of each feature vector in the fusion result is calculated, and the weight is used to perform weighted concatenation and fusion on the feature vectors to generate fusion features.

[0071] In this embodiment, the Transformer model based on the multi-head attention mechanism includes an encoder and a decoder. The encoder includes a plurality of stacked encoder layers, each of which includes a multi-head attention sublayer and a feedforward neural network sublayer; the multi-head attention sublayer and the feedforward neural network sublayer use residual connections and layer normalization. The decoder includes a plurality of stacked decoder layers, each of which includes two multi-head attention sublayers and a feedforward neural network sublayer; the first multi-head attention sublayer is a self-attention sublayer; the second multi-head attention sublayer is an encoder-decoder attention sublayer; the multi-head attention sublayer and the feedforward neural network sublayer use residual connections and layer normalization. The feedforward neural network includes two linear transformations and an activation function. Specifically, in the encoder and decoder, the self-attention mechanism allows the model to consider the information of all positions at each position of the sequence at the same time, thereby capturing the long-distance dependencies within the sequence. The self-attention layer receives three parameters as input: query, key, and value, which are obtained from the input representation through linear transformation. In order to enhance the expressive power of the model, the Transformer model of this embodiment adopts a multi-head attention mechanism. It splits the input representation into multiple "heads", each of which independently computes self-attention and merges the results. Each head learns different aspects of the input, allowing it to capture richer sequence information. Specifically, the multi-head attention mechanism divides the query, key, and value parameters into N parts, each of which is processed independently by a head. All similar attention calculations are then combined to generate the final attention score. This mechanism enables the Transformer to more powerfully encode multiple relationships and nuances of each word.

[0072] Specifically, this embodiment constructs a Transformer model based on a multi-head attention mechanism, and trains multiple heat-melt landslide probability prediction models for different periods based on heat-melt landslide data sets of different periods. Since the landslide probabilities in spring, summer, autumn, and winter are different, a time factor (by month) is added to differentiate the data, and different probability prediction models are generated respectively.

[0073] The thermal melt landslide dataset includes positive samples and negative samples. The spectral data, GIS data and deep soil data sets of different periods with different degrees of thermal melt landslide are used as positive samples of training samples; the spectral data, GIS data and deep soil data sets of different periods without thermal melt landslide are used as negative samples of training samples, and the cross entropy loss function is used as the loss function.

[0074] Specifically, the input of the Transformer model is a sequence, so the fusion features need to be converted into a sequence pattern, including the following steps:

[0075] S4.1: Map each spectral feature vector to a fixed embedding space through a linear layer to form an embedding vector.

[0076] S4.2: Add position encoding to each embedding vector, and combine these embedding vectors and position encoding into a sequence; wherein the sequence of each spectral feature vector is a plurality of vectors of equal length.

[0077] First, the frozen soil area to be tested was scientifically and rationally divided, and spectral data of different bands were obtained in each partition. This step not only improves the accuracy of data acquisition, but also ensures the comprehensiveness and representativeness of the data. By acquiring multi-band spectral data, it is possible to capture information on various data characterization types such as surface temperature, vegetation index, soil moisture, surface water content, and rock and soil types, which are crucial for accurately predicting thermal melt landslides.

[0078] Next, multiple trained feature extraction networks were used to perform in-depth feature extraction on spectral data of different bands. These networks can extract spectral feature vectors of different data representation types, providing rich feature information for subsequent weighted fusion.

[0079] On the basis of feature extraction, the corresponding feature weights were assigned to the spectral feature vectors according to the different degrees of influence of the data representation types on the thermal melt landslide. This step fully considered the importance of different features in predicting thermal melt landslides and ensured the accuracy and effectiveness of weighted fusion. Subsequently, the spectral feature vectors were weighted fused based on the feature weights to form a fusion feature, which integrates the information of multiple data representation types and provides strong support for subsequent predictions.

[0080] In order to build an accurate thermal melt landslide prediction model, a Transformer model based on a multi-head attention mechanism was used and fully trained using a thermal melt landslide dataset. This model can capture the key information in the fusion features and accurately predict the thermal melt landslide probability of each partition in the frozen soil area to be tested.

[0081] Finally, the fusion features were converted into sequence patterns and input into the trained thermal melt landslide probability prediction model to obtain the thermal melt landslide probability of each partition. Based on these probability values, a thermal melt landslide risk distribution map was drawn, and the risk levels of different areas were marked. This achievement not only provides decision makers with intuitive prediction results, but also provides a scientific basis for them to formulate targeted early warning and prevention measures.

[0082] By adopting the above-mentioned technical means, the technical problem of ignoring the influence of environmental data and soil vegetation data on the prediction results was effectively solved, and accurate prediction and risk assessment of thermal thaw landslides in frozen soil areas were achieved.

[0083] Furthermore, the present invention also proposes historical data analysis and change trend analysis, including the following steps:

[0084] Obtain the historical prediction data of thermal melt collapse probability in each sub-area for each month of each year, and clean the data;

[0085] Based on the historical forecast data, Matplotlib is used to build a comparison chart. A line chart is used to display the forecast data of the same month in different years. Color or line type is used to distinguish data from different years. Based on the above content, data analysis is performed: Analyze the changing trend of the data in the chart, such as year-on-year growth, seasonal fluctuations, etc. Find outliers or mutation points in the chart and analyze possible reasons. For example, the mean, median, standard deviation, etc., quantify the range and stability of data changes.

[0086] Based on the historical prediction data, a prediction model constructed based on a neural network is selected to predict the probability of thermal melt collapse in each zone in the future.

[0087] Plot the forecast results into a chart and compare them with historical forecast data to obtain possible future change trends.

[0088] Through the analysis of the identification results and the corresponding relevant data, it is clear that the development of thermal thaw landslides is affected by many factors. Among them, the temperature change, precipitation pattern, slope and deep soil temperature have a significant impact on the development of thermal thaw landslides from 10 square meters to 1000 square meters. First, temperature change is an important driving force for the development of thermal thaw landslides. As the temperature rises, the ice in the permafrost begins to melt, causing the soil to lose support and collapse. Especially in summer, extreme high temperatures will accelerate the melting process of permafrost, thereby increasing the risk of thermal thaw landslides. Secondly, precipitation patterns also have a significant impact on thermal thaw landslides. The increase in rainfall will provide dynamic conditions for thermal thaw landslides, accelerate the wetting and softening of the soil, reduce its shear strength, and promote the occurrence of landslides. Slope is another important factor affecting thermal thaw landslides. Within a suitable slope range (such as 3° to 8°), the soil is prone to collapse under the action of gravity. A slope that is too large or too small is not conducive to the development of thermal thaw landslide, because a slope that is too large may make the soil too loose, while a slope that is too small may make the soil lack sufficient sliding force. Finally, the temperature of the deep soil is also a key factor affecting thermal thaw landslide. The temperature of the deep soil directly determines the melting speed and range of the frozen soil. When the temperature of the deep soil rises, the active layer of the frozen soil will become thinner and more likely to melt, thereby increasing the possibility of thermal thaw landslide.

[0089] The above is a method for predicting thermal melt collapse based on multispectral imaging provided by one embodiment of the present embodiment. Based on the same idea, the present embodiment also provides a corresponding system for predicting thermal melt collapse based on multispectral imaging. The specific definition of the system for predicting thermal melt collapse based on multispectral imaging can be found in the above definition of the method for predicting thermal melt collapse based on multispectral imaging, which will not be repeated here. Each module in the above-mentioned system for predicting thermal melt collapse based on multispectral imaging can be fully or partially implemented by software, hardware and a combination thereof. Each of the above-mentioned modules can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above-mentioned modules.

[0090] This embodiment also provides a computer-readable storage medium, which stores a computer program, which can be used to execute the above Figure 1 A thermal melt landslide prediction method based on multispectral imaging is provided.

[0091] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0092] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting thermal melt collapse based on multispectral imaging, characterized in that: The following steps are involved: The frozen soil area to be tested is divided into zones, and the spectral data, GIS data and deep soil data of different bands on the surface of the zones are obtained respectively; Through multiple trained feature extraction networks, spectral feature vectors of different data representation types of spectral data in different bands, as well as feature vectors of GIS data and deep soil data are extracted respectively; feature weights are assigned according to the degree of influence of the data representation type of the spectral feature vector on the thermal melt landslide, and the spectral feature vectors are weighted fused based on the feature weights, and are concatenated with the feature vectors of GIS data and deep soil data to form fusion features; wherein the data representation types include surface temperature, vegetation index, soil moisture, surface water content and rock and soil type; A Transformer model based on a multi-head attention mechanism is constructed, and multiple thermal melt landslide probability prediction models for different periods are obtained based on the thermal melt landslide datasets of different periods. The thermal melt landslide datasets are divided into datasets of different periods by month. The fused features are converted into sequence patterns and input into the corresponding thermal thaw landslide probability prediction model according to the current period of the permafrost area to be tested, so as to obtain the thermal thaw landslide probability of each zone in the permafrost area to be tested, and a thermal thaw landslide risk distribution map is drawn according to the thermal thaw landslide probability of each zone, marking the risk levels of different areas and issuing early warnings.

2. The method for predicting thermal melt collapse based on multispectral imaging according to claim 1 is characterized in that: The method of respectively acquiring spectral data of different bands of the partition surface comprises the following steps: The image of the frozen soil area to be tested is obtained by multi-spectral imaging equipment, and grid division is performed to form multiple sub-areas, and the sub-areas are numbered; The sub-regions are identified in sequence according to the numbers. Based on the spectral features and texture features in the image, the threshold segmentation method based on pixel value differences and the trained partition recognition network are used to obtain the frozen soil area to be tested. Secondary partitioning is performed through the grid division method to obtain multiple secondary partition areas containing the frozen soil area grids to be tested. The spectral data of different bands of each secondary sub-region of the frozen soil area to be tested are collected by a multi-spectral imaging device, and the spectral data are preliminarily processed by edge computing technology, including data cleaning, denoising and compression; the multi-spectral imaging device is erected on a high tower or carried on a drone above the frozen soil area to be tested; the multi-spectral imaging device is erected on a high tower or carried on a drone through a motor with an encoder, and the motor drives the multi-spectral imaging device to swing; the multi-spectral imaging device is provided with a zoom lens; The collected spectral data is transmitted to the processing center via Apache Kafka; In the data processing center, Apache Flink is used to perform radiation correction, geometric correction, and atmospheric correction on the acquired spectral data, and the spectral data is converted into spectral vectors and filtered, aggregated, and connected.

3. The method for predicting thermal melt collapse based on multispectral imaging according to claim 1, characterized in that: The GIS data includes slope, aspect and elevation data; the deep soil data includes soil temperature data at different depths; the soil temperature data at different depths are obtained through deep soil temperature sensors arranged in the active layer, frozen soil layer and sliding layer, and the depth and thickness of the frozen soil layer, active layer and sliding layer are determined through the soil temperature data.

4. The method for predicting thermal melt collapse based on multispectral imaging according to claim 1, characterized in that: The feature extraction network is based on a CNN network model, and is trained to obtain a plurality of feature extraction networks respectively through labeled surface temperature and thermal infrared band datasets, vegetation index and red band and near infrared band datasets, vegetation growth status and green band datasets, soil moisture and shortwave infrared band, surface moisture and visible light band and near infrared band datasets, and rock and soil type and shortwave infrared band datasets, as feature extraction models for spectral feature vectors of different data characterization types; wherein each label of each dataset is obtained through a sensor system.

5. The method for predicting thermal melt collapse based on multispectral imaging according to claim 1, characterized in that: The method of weighted fusion of spectral feature vectors based on feature weights includes the following steps: Determine the influence of surface temperature, vegetation index, soil moisture, surface water content and rock and soil type on thermal melt landslide, the correlation and complementarity between different eigenvectors and their contribution to the fusion results, and determine the importance ranking results of each spectral eigenvector based on the predicted results and actual results of the thermal melt landslide probability prediction model using a specific sample set according to the SHAP method; According to the importance ranking results of each spectral feature vector, weights are proportionally allocated and a weight allocation strategy is determined; According to the weight distribution strategy, the weight of each feature vector in the fusion result is calculated, and the feature vectors are weightedly concatenated and fused using the weight to generate fusion features.

6. The method for predicting thermal melt collapse based on multispectral imaging according to claim 1, characterized in that: The Transformer model based on the multi-head attention mechanism includes an encoder and a decoder; The encoder includes a plurality of stacked encoder layers, each of which includes a multi-head attention sublayer and a feedforward neural network sublayer; the multi-head attention sublayer and the feedforward neural network sublayer use residual connection and layer normalization; The decoder comprises a plurality of stacked decoder layers, each of the decoder layers comprising two multi-head attention sub-layers and a feed-forward neural network sub-layer; the first multi-head attention sub-layer is a self-attention sub-layer; The second multi-head attention sublayer is an encoder-decoder attention sublayer; the multi-head attention sublayer and the feedforward neural network sublayer adopt residual connection and layer normalization; The feedforward neural network includes two linear transformations and an activation function.

7. The method for predicting thermal melt collapse based on multispectral imaging according to claim 1, characterized in that: The fusion feature is converted into a sequence pattern, comprising the following steps: Map each spectral feature vector to a fixed embedding space through a linear layer to form an embedding vector; A position code is added to each embedding vector, and these embedding vectors and the position code are combined into a sequence; wherein the sequence of each spectral feature vector is a plurality of vectors of equal length.

8. The method for predicting thermal melt collapse based on multispectral imaging according to claim 1, characterized in that: Also includes: Obtain the historical prediction data of thermal melt collapse probability in each sub-area for each month of each year, and clean the data; Use Matplotlib to build a comparison chart based on historical forecast data, use a line chart to show the forecast data of the same month in different years, and use color or line type to distinguish data from different years; Based on historical prediction data, a prediction model based on neural network is selected to predict the probability of thermal melt collapse in each subarea in the future; Plot the forecast results into a chart and compare them with historical forecast data to obtain possible future change trends.

9. A thermal melt collapse prediction system based on multispectral imaging, characterized in that: include: The data acquisition module is used to obtain the frozen soil area to be tested and divide it into zones, and obtain the spectral data of different bands on the surface of the zones respectively; The data processing and analysis module is used to extract spectral feature vectors of different data representation types of spectral data in different bands through a feature extraction network after multiple rounds of training; it is also used to assign feature weights according to the data representation type of the spectral feature vector for the degree of influence of thermal melt landslide, and perform weighted fusion of the spectral feature vectors based on the feature weights to form fusion features; wherein the data representation types include surface temperature, vegetation index, soil moisture, surface water content and rock and soil types; The prediction model building module is used to build a Transformer model based on a multi-head attention mechanism and obtain a thermal melt landslide probability prediction model based on the thermal melt landslide dataset; The prediction module is used to convert the fusion features into a sequence pattern and input it into the thermal thaw landslide probability prediction model to obtain the thermal thaw landslide probability of each subarea in the frozen soil area to be tested; The early warning release module is used to draw a thermal melt landslide risk distribution map based on the probability of thermal melt landslide in each zone, mark the risk levels of different areas and issue early warnings, and release early warning information to relevant departments and personnel through SMS, WeChat and websites.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a data processing program, and when the data processing program is executed by a processor, the steps of the thermal melt landslide prediction method based on multispectral imaging are implemented as described in any one of claims 1 to 7.

Images