Debris flow geological disaster monitoring and early warning method

Abstract

The invention discloses a debris flow geological disaster monitoring and early warning method, and belongs to the field of intelligent early warning, and the method comprises the following steps: carrying out the preprocessing of debris flow disaster impact factor data collected by a multi-source heterogeneous sensor, and obtaining a fusion data set under a unified spatial-temporal scale; performing abnormal value detection and data filling on the fused data set to obtain a high-quality data set; performing feature extraction on the high-quality data set by adopting a multi-scale convolutional neural network to obtain multi-scale features; establishing a debris flow disaster occurrence probability prediction model; and inputting the multi-scale features into the debris flow disaster occurrence probability prediction model to obtain debris flow geological disaster early warning. According to the invention, accurate early warning of debris flow geological disasters is realized, and powerful support is provided for disaster prevention and reduction.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent early warning, and in particular relates to a method for monitoring and early warning of debris flow-type geological disasters. Background Art

[0002] In the debris flow disaster monitoring and early warning system, how to effectively fuse the influencing factor data collected by multi-source heterogeneous sensors is a key technical issue. The data collected by different types of sensors have differences in temporal and spatial resolution, data format and semantic expression, resulting in many challenges for data fusion. In addition, the quality of data collected by sensors is uneven, and there may be problems such as noise, outliers and data missing, which directly affect the effect of data fusion. At the same time, the mechanism of debris flow disaster is complex, involving multiple factors such as rainfall, topography, and geology. How to extract effective features from massive multi-source heterogeneous data and build a high-precision debris flow disaster probability prediction model is also a technical problem that needs to be solved urgently. Traditional data fusion and feature extraction methods are difficult to fully explore the inherent correlation of multi-source heterogeneous data, resulting in insufficient generalization ability of the prediction model. Therefore, it is urgent to study the multi-source heterogeneous data fusion and feature extraction methods for debris flow disaster monitoring and early warning, improve the efficiency of data fusion and the accuracy of the prediction model, and provide strong support for the accurate early warning of debris flow disasters. Summary of the invention

[0003] In order to solve the above technical problems, the present invention proposes a debris flow type geological disaster monitoring and early warning method to solve the above problems existing in the prior art.

[0004] To achieve the above object, the present invention provides a method for monitoring and early warning of debris flow geological disasters, comprising:

[0005] The debris flow disaster impact factor data collected by multi-source heterogeneous sensors are preprocessed to obtain a fused data set at a unified spatiotemporal scale.

[0006] Performing outlier detection and data filling on the fused data set to obtain a high-quality data set;

[0007] Using a multi-scale convolutional neural network to extract features from the high-quality data set to obtain multi-scale features;

[0008] Establish a probability prediction model for debris flow disasters;

[0009] The multi-scale features are input into the debris flow disaster occurrence probability prediction model to obtain a debris flow type geological disaster warning.

[0010] Optionally, the debris flow disaster influencing factor data includes: rainfall data, topographic data and geological data;

[0011] The rainfall data includes: rainfall amount, rainfall intensity and rainfall duration;

[0012] The terrain data includes: slope, elevation and slope direction;

[0013] The geological data includes: lithology, soil type and structure.

[0014] Optionally, the process of obtaining a fused data set at a uniform spatiotemporal scale includes:

[0015] Collect data on debris flow disaster impact factors based on multi-source heterogeneous sensors;

[0016] Calculate the credibility weight of each sensor data based on the sensor data credibility evaluation model;

[0017] Multiply the debris flow disaster impact factor data of each sensor by the corresponding credibility weight to obtain weighted impact factor data;

[0018] The weighted impact factor data is subjected to a time-space scale conversion to obtain fused data at a unified time-space scale.

[0019] Optionally, the process of performing outlier detection and data filling on the fused data set to obtain a high-quality data set includes:

[0020] Using an isolation forest algorithm to perform outlier detection on the fused data set to obtain a data set after removing outliers;

[0021] A data filling algorithm based on multi-view learning is used to fill in the data after removing outliers to obtain a high-quality data set.

[0022] Optionally, the process of performing data filling on the data after removing outliers by using a data filling algorithm based on multi-view learning includes:

[0023] Processing the data after removing outliers based on a random sampling method to construct a number of incompletely overlapping data subsets;

[0024] estimating missing values ​​based on complementary information between the plurality of said incompletely overlapping data subsets;

[0025] The missing values ​​are filled based on temporal and spatial correlation to obtain a high-quality data set.

[0026] Optionally, a gated recurrent unit network is used to establish a debris flow disaster occurrence probability prediction model;

[0027] The debris flow disaster occurrence probability prediction model introduces a gating unit to capture the long-distance dependencies between features and combines the error back propagation algorithm to optimize the model parameters to obtain a trained prediction model.

[0028] The present invention also discloses a computer terminal device, comprising:

[0029] one or more processors;

[0030] A memory, coupled to the processor, for storing one or more programs;

[0031] When the one or more programs are executed by the one or more processors, the one or more processors implement a method for monitoring and early warning of debris flow-type geological disasters.

[0032] The present invention also discloses a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, a method for monitoring and early warning of debris flow-type geological disasters is implemented.

[0033] Compared with the prior art, the present invention has the following advantages and technical effects:

[0034] The present invention provides a method for monitoring and early warning of debris flow-type geological disasters. By comprehensively utilizing multi-source heterogeneous sensor data, the full-process automated processing from data preprocessing to disaster early warning is realized. The method first preprocesses the collected multi-source data, eliminates the data inconsistency between different sensors, and forms a unified fusion data set. Subsequently, the quality of the data set is further improved through outlier detection and data filling technology, ensuring the accuracy of subsequent analysis. Using a multi-scale convolutional neural network, the method can extract key multi-scale features from high-quality data sets, which comprehensively reflect the complex influencing factors of debris flow disasters. Finally, through the established prediction model, the probability of debris flow disasters can be accurately predicted, and effective early warning can be achieved, thereby providing a scientific basis for disaster prevention and mitigation, and significantly improving the accuracy and timeliness of early warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0036] Figure 1 This is a flow chart of a method for monitoring and early warning of debris flow-type geological disasters according to an embodiment of the present invention. DETAILED DESCRIPTION

[0037] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0038] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0039] Embodiment 1

[0040] like Figure 1 As shown, a method for monitoring and early warning of debris flow type geological disasters is provided in this embodiment, comprising the following steps: preprocessing the debris flow disaster influencing factor data collected by multi-source heterogeneous sensors to obtain a fused data set at a unified time and space scale; performing outlier detection and data filling on the fused data set to obtain a high-quality data set; using a multi-scale convolutional neural network to extract features from the high-quality data set to obtain multi-scale features; establishing a debris flow disaster probability prediction model; and inputting the multi-scale features into the debris flow disaster probability prediction model to obtain a debris flow type geological disaster early warning.

[0041] Step 1: Obtain the debris flow disaster influencing factor data collected by multi-source heterogeneous sensors, use an adaptive weighted fusion algorithm to dynamically adjust the weights according to the credibility of different sensor data, and obtain a fused data set at a unified spatiotemporal scale.

[0042] Furthermore, the debris flow disaster influencing factor data collected by multi-source heterogeneous sensors are obtained, and the corresponding data preprocessing methods are used for noise removal and data standardization for different types of sensor data. According to the pre-established sensor data credibility evaluation model, the credibility weight of each sensor data is calculated. The larger the weight value, the higher the credibility of the sensor data. The influencing factor data of each sensor is multiplied by its corresponding credibility weight to obtain the weighted influencing factor data. The weighted multi-source heterogeneous influencing factor data is converted into time and space scales to unify them to the same time and space scales. The adaptive weighted fusion algorithm is used to dynamically adjust the fusion weight according to the characteristics and credibility of the converted influencing factor data to achieve adaptive fusion of multi-source heterogeneous data. The fusion weight is continuously updated through the iterative optimization algorithm, so that the fusion result gradually converges, and finally the debris flow disaster influencing factor fusion data set under the unified time and space scale is obtained. The fused debris flow disaster influencing factor data is analyzed and modeled, and a debris flow disaster warning and risk assessment model is constructed to provide decision support for disaster prevention and control.

[0043] Furthermore, as a specific implementation of this embodiment, the data collection of debris flow disaster influencing factors involves multiple sensors, such as rainfall sensors collecting rainfall data, soil moisture sensors collecting moisture content data, and surface displacement sensors collecting landslide displacement data. According to the characteristics of different sensor data, corresponding preprocessing methods are adopted, such as using median filtering to remove mutation noise for rainfall data, and normalizing soil moisture data to unify the value range to between 0 and 1. The sensor data credibility evaluation model is based on factors such as the historical operating status of the sensor and data quality. For example, a rainfall sensor has stable operation and good data continuity, and its credibility weight is 0.8, while a soil moisture sensor has a credibility weight of 0.5 due to greater external interference. Multiply the sensor data with the corresponding weight, such as the rainfall value at a certain moment is 50 mm, multiplied by 0.8 to obtain a weighted value of 40 mm. In the process of spatiotemporal scale conversion, data at different time intervals are unified to the same sampling frequency, such as converting rainfall data at a ten-minute interval and soil moisture data at a thirty-minute interval to an hourly scale. In space, point observation data at different locations are converted into unified grid data through interpolation methods. The adaptive weighted fusion algorithm dynamically adjusts the fusion weight according to data characteristics. When a sensor data suddenly changes or is abnormal, its fusion weight is reduced; when the data is stable and reliable, its weight is increased. For example, if the rainfall data in a certain period suddenly changes, its fusion weight is reduced from 0.8 to 0.3. The weight is continuously adjusted through iterative optimization to gradually converge the fusion results. The fused data set can be used for debris flow warning. For example, by analyzing historical disaster cases, the critical rainfall and soil moisture thresholds are determined, and an early warning is issued when the fused data exceeds the threshold. In risk assessment, the hazard level is divided in combination with factors such as terrain and geology to provide a basis for disaster prevention and mitigation. For example, when the accumulated rainfall in a certain area exceeds 50 mm in one hour and the soil moisture content exceeds 85%, it is determined as a high-risk area and timely response measures are taken.

[0044] Step 2: For the fused data set, an improved isolation forest algorithm is used to detect outliers. The data is mapped to a high-dimensional space by constructing multiple random hyperplanes, and the average distance from the data point to the hyperplane is calculated. Data points whose distance exceeds a preset threshold are judged as outliers and removed to obtain a data set after removing outliers.

[0045] Further, based on the fused data set, an improved isolation forest algorithm is used for outlier detection. The specific steps are as follows: the fused data set is preprocessed, and data quality and consistency are ensured through data cleaning and standardization operations to prepare for subsequent outlier detection. An improved isolation forest model is constructed, and multiple random hyperplanes are constructed in a high-dimensional space by randomly selecting feature subsets and sample subsets, and the data are mapped to these hyperplanes. For each data point, its distance to all random hyperplanes is calculated, and the average distance is calculated as the abnormality metric value of the data point. According to the preset abnormal threshold, the data points whose abnormal metric values ​​exceed the threshold are judged as outliers and are removed from the data set. The quality of the data set after the outliers are removed is evaluated, and the effectiveness and reliability of the algorithm are evaluated by calculating the accuracy, recall rate and other indicators of the abnormal detection. If the evaluation result meets the expected requirements, the data set after the outliers are removed is used as the input for subsequent analysis and modeling; otherwise, the abnormal threshold is adjusted or the algorithm parameters are optimized, and the outlier detection is performed again. The final high-quality data set after removing outliers is applied to subsequent data analysis, machine learning modeling and other tasks to support business decision-making and optimization.

[0046] Furthermore, as a specific implementation of this embodiment, in the outlier detection of debris flow disaster influencing factor data, it is first necessary to preprocess the fused data set. For example, for rainfall data, by removing obvious error values ​​such as negative values ​​and standardizing the data to a zero mean unit variance distribution, data of different dimensions are made comparable. For terrain slope data, median filtering can be used to remove noise and then normalized. The improved isolation forest algorithm realizes anomaly detection by constructing random hyperplanes. Taking rainfall and slope as two examples, different combinations of these two features can be randomly selected to construct multiple two-dimensional planes. On each plane, the algorithm randomly selects segmentation points to divide the data into several regions. Normal data tends to cluster together and requires more segmentation to isolate it; while outliers are usually more scattered and easy to be quickly isolated. In terms of distance measurement, the Mahalanobis distance from the data point to the random hyperplane can be calculated. The rainfall of a data point is 80 mm per hour and the slope is 45 degrees, which is obviously deviated from the normal distribution. Its distance to most random hyperplanes is far away, and the average distance value is large, so it is judged as an outlier. Setting a reasonable anomaly threshold is crucial to the detection effect. By analyzing historical data, the 95th percentile of the anomaly metric value can be used as the initial threshold. If the proportion of detected anomalies is too high, it means that the threshold may be too strict and needs to be appropriately relaxed; otherwise, the threshold needs to be tightened. The cross-validation method can be used to evaluate the detection results. By manually annotating some known anomalies as a test set, the detection accuracy and recall rate are calculated. For example, in a certain evaluation, 85 of the 100 anomalies detected by the algorithm are consistent with the manual annotations, and 85 of the 90 anomalies are found, then the accuracy rate is 85% and the recall rate is 94%. The data set after outlier detection and removal is more reliable and can improve the accuracy of subsequent modeling. In the debris flow warning model, the rainfall threshold after removing the outliers more accurately reflects the critical conditions for the occurrence of debris flow, which helps to improve the accuracy of the warning and reduce false alarms and missed reports. At the same time, high-quality data sets also provide reliable data support for in-depth research on the mechanism of debris flow.

[0047] Step three, to address the data missing problem in the data set after removing outliers, a data filling algorithm based on multi-view learning is used. By constructing multiple incompletely overlapping data subsets, the missing values ​​are estimated using the complementary information between subsets, and the local linear embedding method is combined to maintain the data manifold structure to obtain a complete high-quality data set.

[0048] Furthermore, in the original data set, there are usually some outliers that affect the accuracy of subsequent analysis. For example, in meteorological data, there may be obvious erroneous values ​​of minus one hundred degrees or plus one hundred degrees, which need to be identified and eliminated through statistical methods such as box plot method. For multidimensional data, methods such as Mahalanobis distance can be used to detect outliers. When randomly sampling the processed data set, a stratified sampling method can be used. In order to maintain the essential characteristics of the data, the local linear embedding method is used to maintain the manifold structure of the data. Taking face image data as an example, when filling missing pixels, it is necessary to maintain the continuity and natural transition of facial features. By calculating the local relationship between data points, it is ensured that the filled data is still on the manifold structure of the original data. When filling the estimated missing values ​​into the corresponding subsets, it is necessary to consider the temporal and spatial correlation of the data. For example, in environmental monitoring data, the missing data of a monitoring station should maintain a certain correlation with the data of the adjacent stations, while considering the seasonal variation characteristics. In the subset fusion stage, it is necessary to resolve possible data conflicts. When different subsets produce different estimation results for the same missing value, they can be fused through methods such as weighted averaging, and the weights can be determined based on the reliability of each subset. When conducting the final quality assessment, it can be done from multiple dimensions such as completeness, accuracy, and consistency. The accuracy of the filling result can be verified by calculating the mean square error between the filled value and the true value; by comparing the data distribution, ensure that the distribution of the filled data is similar to the original data; and by verifying the business rules, ensure that the filling result meets the actual business constraints.

[0049] Step 4: In view of the complex mechanism of debris flow disasters and diverse influencing factors, a multi-scale convolutional neural network is used to extract multi-scale features such as rainfall, topography, and geology in the high-quality data set. By designing convolution kernels with different receptive fields, local and global features are adaptively extracted to obtain a comprehensive feature representation of the influencing factors of debris flow disasters.

[0050] Furthermore, for the preprocessed multi-source heterogeneous data, a multi-scale convolutional neural network model is designed, and local and global features in the data are extracted by setting convolution kernels of different sizes and shapes. In the multi-scale convolutional neural network, convolution kernels with different receptive fields are used to perform adaptive feature extraction on the data to obtain feature maps at different scales. The local and global features extracted by the multi-scale convolutional neural network are fused to obtain a comprehensive feature representation of the factors affecting debris flow disasters. According to the occurrence mechanism of debris flow disasters and the correlation between the influencing factors, a debris flow disaster prediction model is constructed, and the feature representation is used as the input of the model. The debris flow disaster prediction model is trained using a supervised learning method, and the model parameters are optimized to improve the model's prediction ability for debris flow disasters. In practical applications, the newly acquired rainfall, topography, geology and other data are input into the trained debris flow disaster prediction model to determine the possibility of debris flow disasters and provide decision support for disaster prevention and mitigation.

[0051] Furthermore, as a specific implementation of this embodiment, the multi-source heterogeneous data of debris flow disaster influencing factors include rainfall data, terrain data and geological data. Rainfall data mainly includes rainfall, rainfall intensity and rainfall duration, etc., which can be obtained through meteorological stations. Terrain data includes slope, elevation and aspect, etc., which can be obtained through digital elevation models. Geological data includes lithology, soil type and structure, etc., which can be obtained through geological surveys. The formats and scales of these data are inconsistent and need to be preprocessed, such as unifying the time resolution of rainfall data to hours and the spatial resolution to kilometers. The multi-scale convolutional neural network model extracts features by setting convolution kernels of different sizes. For example, for rainfall data, convolution kernels of sizes of three by three, five by five and seven by seven can be set to extract local rainfall features and regional rainfall features, respectively. For terrain data, convolution kernels of sizes of five by five, seven by seven and nine by nine can be set to extract terrain features of different scales. Smaller convolution kernels can capture detailed features, and larger convolution kernels can capture macro features. In the feature extraction process, convolution kernels of different receptive fields have different functions. Taking terrain data as an example, small-scale convolution kernels can extract micro-geomorphological features, such as valley morphology; large-scale convolution kernels can extract regional geomorphological features, such as mountain direction. Through multi-scale convolution, feature maps containing information of different scales can be obtained, such as terrain relief feature maps and surface fragmentation feature maps. The feature fusion stage combines features of different scales. For example, combining rainfall intensity features with terrain slope features can reflect the impact of rainfall on slope stability. Combining geological lithology features with terrain fragmentation features can reflect the impact of geological structure on debris flow occurrence. The construction of debris flow disaster prediction model is based on the occurrence mechanism. If the rainfall exceeds one hundred millimeters and lasts for more than twenty-four hours, and the slope is greater than twenty-five degrees, and the geological conditions are loose accumulation layers, the possibility of debris flow is greater. During the model training process, historical debris flow disaster data can be used as labels to improve the prediction accuracy by adjusting model parameters. In practical applications, the model can make predictions based on real-time monitoring data. For example, when heavy rainfall occurs in a certain area, the real-time rainfall reaches 80 mm, and the forecast shows that the rainfall will continue, and at the same time the slope of the area is steep and the geological conditions are prone to landslides, the model can provide timely warnings and provide decision-making basis for local disaster prevention and mitigation work.

[0052] This embodiment uses a supervised learning method to train a debris flow disaster prediction model, and improves the model's ability to predict debris flow disasters by optimizing model parameters. Obtain historical data on debris flow disasters, including meteorological, geological, hydrological and other related data, and construct training data sets and test data sets. Preprocess the training data, including data cleaning, feature selection and feature engineering, to extract key features related to debris flow disasters. Select a suitable supervised learning algorithm, such as a decision tree, support vector machine or neural network, to build a debris flow disaster prediction model. Input the training data into the model, and optimize the model performance by adjusting the model hyperparameters, such as the learning rate, regularization coefficient, etc. Use methods such as cross-validation to evaluate the prediction performance of the model on the test data, and select the model parameter combination with the best performance. Apply the optimized model to the actual debris flow disaster prediction, and judge the possibility of the occurrence of debris flow disasters based on the predicted probability output by the model. Continuously collect new debris flow disaster data, and regularly retrain and optimize the model to adapt to the changing environment and data distribution.

[0053] Step five: Based on the extracted multi-scale features, a gated recurrent unit network is used to establish a debris flow disaster probability prediction model. The long-distance dependencies between features are effectively captured by introducing the gated unit, and the model parameters are optimized by combining the error back propagation algorithm to obtain the trained prediction model.

[0054] Furthermore, an input data set is constructed based on multi-scale features and input into a gated recurrent unit network for feature learning and encoding. In the gated recurrent unit network, the transmission and update of information is controlled by introducing a forget gate and an input gate, and the long-distance dependency between multi-scale features is captured. The output of the gated recurrent unit network is used as a feature representation for predicting the probability of debris flow disasters, and a prediction model is constructed. The error back propagation algorithm is used to adaptively adjust and optimize the parameters of the gated recurrent unit network and the prediction model according to the error between the prediction result and the actual probability of occurrence. Through multiple rounds of iterative training, the model parameters are continuously optimized to improve the accuracy and generalization ability of the prediction model. The trained debris flow disaster probability prediction model is applied to actual scenarios, and the probability of debris flow disasters is predicted based on the input multi-scale features. The prediction results are analyzed and evaluated, and the risk level of debris flow disasters is judged according to the predicted probability, providing decision support for disaster warning and prevention.

[0055] Furthermore, as a specific implementation of this embodiment, the input data set is constructed according to the multi-scale features. First, it is necessary to standardize the multi-source heterogeneous data such as rainfall, topography, and geology to ensure the consistency and comparability of the data. For example, rainfall data can be converted into features of different time scales such as daily rainfall and monthly rainfall; terrain analysis data, terrain data can be quantified by indicators such as elevation and slope. After normalization, these features form a comprehensive multi-scale feature vector. The constructed input data set is input into the gated recurrent unit (GRU) network for feature learning and encoding. The GRU unit in the GR bed can effectively control the transmission and update of information by introducing a forget gate and an input gate. For example, the forget gate can determine which historical information needs to be retained or discarded, and the input gate is responsible for updating the current state. Assume that in a certain iteration, the forget gate outputs 0.8, which means that 80% of the historical information is retained; the input gate outputs 0.6, which means that 60% of the current input information is included in the new state. In this way, GRU can capture the long-distance dependencies between multi-scale features. The output of the GRU network is used as a feature representation for predicting the probability of debris flow disasters, and a prediction model is constructed. Assume that the output of the GRU network is a matrix containing multiple eigenvectors, which comprehensively reflect the influence of multi-scale factors such as rainfall, topography, and geology. These eigenvectors are input into the prediction fully connected layer, and nonlinear transformation is performed through the activation function (such as ReLU), and finally a probability value between 0 and 1 is output, indicating the possibility of debris flow disaster. The error back propagation algorithm is used to adaptively adjust and optimize the parameters of the GRU network and the prediction model according to the error between the prediction result and the actual probability of occurrence. For example, if the probability of debris flow in a certain area is predicted to be 0.3, and a debris flow actually occurs (the actual probability is 1), the prediction error is calculated, and the weights and biases in the network are adjusted through the back propagation algorithm to reduce the error of future predictions. Through multiple rounds of iterative training, the model parameters are continuously optimized to improve the accuracy and generalization ability of the prediction model. Assuming that after 100 rounds of training, the prediction accuracy of the model is increased from 70% to 90%, indicating that the model's prediction ability for debris flow disasters has been significantly enhanced. The trained debris flow disaster probability prediction model is applied to actual scenarios, and the probability of debris flow disasters is predicted based on the input multi-scale features. For example, a region has experienced heavy rainfall for several consecutive days, has a steep terrain slope, and has fragile geological conditions. The model predicts that the probability of a debris flow in the region is 0.8, indicating that the region has a high disaster risk. The prediction results are analyzed and evaluated, and the risk level of a debris flow disaster is determined based on the predicted probability. Assuming that 0.7 is set as the high-risk threshold, regions with a predicted probability higher than 0.7 will be marked as high-risk areas, and relevant departments can take disaster prevention and mitigation measures in advance, such as evacuating residents and reinforcing infrastructure.In this way, the application of multi-scale features combined with the GRU network can not only effectively capture complex multi-scale influencing factors, but also improve the accuracy and generalization ability of the prediction model through modeling of long-distance dependencies, providing strong decision-making support for debris flow disaster warning and prevention.

[0056] This embodiment uses the output of the gated recurrent unit network as a feature representation for predicting the probability of debris flow disasters to build a prediction model. Based on the historical data of debris flow disasters, the environmental factor features related to the occurrence of disasters are extracted; the extracted features are encoded using the gated recurrent unit network to obtain a deep representation of the features; the feature representation is input into the fully connected layer, and the probability of disaster occurrence is calculated through the Softmax function; a threshold is set to determine whether a debris flow disaster is likely to occur based on the calculated probability value; if the probability exceeds the threshold, a high-risk area for debris flow disasters is warned; environmental factor data is continuously monitored, and the prediction results of the probability of disaster occurrence are updated in real time; the prediction model is regularly evaluated and optimized to improve the accuracy and reliability of the prediction.

[0057] As a specific implementation of this embodiment, the historical data of debris flow disasters include environmental factors such as rainfall, terrain slope, and soil moisture content. Taking a mountainous area as an example, the rainfall data obtained by the meteorological station shows that the probability of debris flow increases significantly when the daily rainfall exceeds 100 mm. The terrain slope is obtained through remote sensing image analysis, and areas with a slope of more than 30 degrees are more likely to have debris flow. Soil moisture can be collected in real time through soil monitoring equipment. When the moisture content reaches saturation, the risk of debris flow rises sharply. When the gated recurrent unit network encodes these features, the information flow is controlled by the update gate and the reset gate. The update gate determines how much historical information is retained, and the reset gate determines how much historical information is ignored. For example, during continuous rainfall, the update gate will give a higher weight to recent rainfall, while the reset gate may reduce the impact of earlier rainfall data. After the feature is processed by the fully connected layer, the output is converted into a probability value through the Softmax function. Suppose the probability of debris flow in a certain area at a specific time is 0.75, and it exceeds the preset warning threshold of 0.6, the system will automatically issue a warning signal. The early warning information includes key information such as the specific location of the risk area and the scope of impact. In actual application, environmental monitoring equipment collects data every ten minutes and updates the prediction results. For example, if a monitoring shows that the cumulative rainfall reaches 80 mm within two hours and the soil moisture content has reached 85%, the system will immediately determine the area as a high-risk area. The evaluation of the prediction model uses confusion matrix analysis to analyze the prediction results. In a certain evaluation, the model correctly warned 85 times, missed 10 times, and falsely reported 5 times in the prediction of 100 debris flow events, indicating that the model has a high accuracy rate and a low false alarm rate. By regularly collecting new debris flow cases and updating the training data set, the model performance is continuously improved. For example, after one year of application of the model, by adding 20 new debris flow cases, the prediction accuracy of the model increased from 85% to 88%. In order to improve the reliability of the prediction, the system also integrates multi-source data. In addition to conventional environmental monitoring data, it also includes auxiliary information such as geological structure and vegetation coverage. These comprehensive information can more comprehensively reflect the potential risk of debris flow and improve the accuracy of early warning. For example, in areas where vegetation coverage is less than 30 percent, mudslides may occur even with low rainfall due to unstable soil structure.

[0058] Step six, to address the problem of insufficient training samples, a transfer learning method is used to initialize the prediction model using model parameters pre-trained on other relevant disaster data. The model is quickly adapted to the current debris flow disaster data through fine-tuning to obtain a debris flow disaster probability prediction model with stronger generalization ability. Debris flow-related geological disaster warnings are then issued based on the debris flow disaster probability prediction model.

[0059] Furthermore, historical data of debris flow disasters and other related disaster data are obtained, and the data are preprocessed to extract features. The transfer learning method is used to pre-train the deep learning model on other related disaster data to obtain the pre-trained model parameters. The debris flow disaster prediction model is initialized using the pre-trained model parameters, and the model is fine-tuned by setting appropriate learning rates and regularization parameters. The model performance is evaluated based on the validation set data, and the generalization ability and prediction accuracy of the model are improved by adjusting hyperparameters and optimizing algorithms. The trained debris flow disaster probability prediction model is applied to actual scenarios, and relevant environmental and geological parameters are input to determine the probability of debris flow disasters. If the predicted probability of debris flow disasters exceeds the preset threshold, the early warning mechanism is triggered, and early warning information is sent to relevant departments and residents through text messages, application push, etc. Continuously collect debris flow disaster data and environmental monitoring data, and regularly update and optimize the prediction model to adapt to environmental changes and improve the timeliness of early warnings.

[0060] Furthermore, as a specific implementation of this embodiment, the historical data of debris flow disasters include environmental parameters such as rainfall, terrain slope, soil moisture content, and information such as the time, location and scale of historical debris flow. The relevant disaster data may include geological disaster data such as landslides and collapses, which have similar triggering mechanisms and characteristics as debris flow. In the data preprocessing stage, missing values ​​need to be filled, and methods such as adjacent value averaging or time series interpolation can be used, while outliers are identified and processed. In terms of feature extraction, parameters such as slope, slope direction, and elevation can be extracted from terrain features, indicators such as rainfall intensity and accumulated rainfall can be extracted from meteorological data, and information such as lithology and structure can be extracted from geological conditions. For example, when the slope in a certain area is more than 30 degrees and the continuous rainfall exceeds 200 mm, the probability of debris flow increases significantly. The transfer learning method can make full use of the knowledge and experience of other geological disasters. For example, pre-training is performed on a landslide prediction model, and the model has learned the influence of factors such as terrain and rainfall on geological disasters. Through transfer learning, this knowledge can help the debris flow prediction model converge faster and better. During the fine-tuning of the model, the learning rate is set to a small value such as 0.001 to avoid destroying the features obtained by pre-training. At the same time, regularization terms are introduced, such as using a weight decay coefficient of 0.0001 to prevent the model from overfitting. In a certain study area, through transfer learning and fine-tuning, the prediction accuracy of the model has increased from 70% to 85%. The trigger threshold setting of the early warning mechanism needs to balance missed reports and false alarms. For example, the early warning threshold is set to 70%, that is, when the predicted probability of occurrence exceeds 70%, an early warning is issued. The early warning information should include specific information such as the time, location, and scope of impact that may occur, so that relevant departments and the public can take timely response measures. The model update optimization can adopt a sliding window method, such as using new data to update the model once every three months. This can adapt to the impact of factors such as climate change and changes in geological conditions. In practice in a certain area, through regular updates, the model's prediction accuracy for new debris flow events has increased by 10%. By establishing a complete closed-loop system of data collection, model training, early warning release and continuous optimization, the accuracy and timeliness of debris flow disaster early warnings can be continuously improved, providing strong support for disaster prevention and mitigation. In practical applications, the system has played an important role in many mountainous areas, effectively reducing casualties and property losses caused by mudslides.

[0061] In view of the complex influencing factors and heterogeneous data of debris flow disasters, this embodiment first uses an adaptive weighted fusion algorithm to fuse multi-source sensor data, then uses an improved isolation forest algorithm to detect and remove outliers, and solves the problem of missing data through a filling algorithm based on multi-view learning, and finally obtains a high-quality debris flow disaster data set. On this basis, the present invention uses a multi-scale convolutional neural network to extract multi-scale features, and combines a gated recurrent unit network to establish a debris flow disaster probability prediction model. In order to solve the problem of insufficient training samples, the present invention also introduces a transfer learning method to improve the generalization ability of the model. Through the above technical solution, the present invention realizes accurate early warning of debris flow geological disasters, providing strong support for disaster prevention and mitigation.

[0062] Embodiment 2

[0063] This embodiment also discloses a computer terminal device, including:

[0064] one or more processors;

[0065] A memory, coupled to the processor, for storing one or more programs;

[0066] When the one or more programs are executed by the one or more processors, the one or more processors implement a method for monitoring and early warning of debris flow-type geological disasters.

[0067] This embodiment also discloses a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, a method for monitoring and early warning of debris flow-type geological disasters is implemented.

[0068] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for monitoring and early warning of debris flow geological disasters, characterized in that: The following steps are involved: The debris flow disaster impact factor data collected by multi-source heterogeneous sensors are preprocessed to obtain a fused data set at a unified spatiotemporal scale. Performing outlier detection and data filling on the fused data set to obtain a high-quality data set; Using a multi-scale convolutional neural network to extract features from the high-quality data set to obtain multi-scale features; Establish a probability prediction model for debris flow disasters; The multi-scale features are input into the debris flow disaster occurrence probability prediction model to obtain a debris flow type geological disaster warning.

2. The method according to claim 1, characterized in that The data of factors affecting debris flow disasters include: rainfall data, topographic data and geological data; The rainfall data includes: rainfall amount, rainfall intensity and rainfall duration; The terrain data includes: slope, elevation and slope direction; The geological data includes: lithology, soil type and structure.

3. The method according to claim 1, characterized in that The process of obtaining a fused dataset at a unified spatiotemporal scale includes: Collect data on debris flow disaster impact factors based on multi-source heterogeneous sensors; Calculate the credibility weight of each sensor data based on the sensor data credibility evaluation model; Multiply the debris flow disaster impact factor data of each sensor by the corresponding credibility weight to obtain weighted impact factor data; The weighted impact factor data is subjected to a time-space scale conversion to obtain fused data at a unified time-space scale.

4. The method according to claim 1, characterized in that: The process of performing outlier detection and data filling on the fused data set to obtain a high-quality data set includes: Using an isolation forest algorithm to perform outlier detection on the fused data set to obtain a data set after removing outliers; A data filling algorithm based on multi-view learning is used to fill in the data after removing outliers to obtain a high-quality data set.

5. The method according to claim 4, characterized in that The process of using a data filling algorithm based on multi-view learning to fill the data after removing outliers includes: Processing the data after removing outliers based on a random sampling method to construct a number of incompletely overlapping data subsets; estimating missing values ​​based on complementary information between the plurality of said incompletely overlapping data subsets; The missing values ​​are filled based on temporal and spatial correlation to obtain a high-quality data set.

6. The method according to claim 1, characterized in that A probability prediction model for debris flow disasters is established using a gated recurrent unit network. The debris flow disaster occurrence probability prediction model introduces a gating unit to capture the long-distance dependencies between features and combines the error back propagation algorithm to optimize the model parameters to obtain a trained prediction model.

7. A computer terminal device, characterized in that: include: one or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the debris flow type geological disaster monitoring and early warning method as described in any one of claims 1-6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the debris flow type geological disaster monitoring and early warning method as described in any one of claims 1 to 6 is implemented.

Images