A spatiotemporal prediction method for landslide geological hazards in complex mountainous areas

Abstract

The present invention proposes a method for spatiotemporal prediction of landslide geological disasters suitable for complex mountainous areas. First, satellite remote sensing images of a large-scale monitoring area are obtained, and then differential interferometry D-InSAR technology is used to pre-process the remote sensing images of the monitoring area to remove the interference of mountain shadows caused by factors such as mountainous clouds and satellite shooting angles. On the basis of pre-processing the remote sensing images by D-InSAR, image recognition technology is used to identify landslide risk characteristics in the remote sensing images of the monitoring area to find out unstable areas prone to landslides. After delineating the landslide risk area in the complex mountainous area, the landslide risk area is monitored online, and the machine learning prediction model established by the present invention is used to predict the time of occurrence of landslide risks. The method of the present invention involves D-InSAR technology, image recognition technology, and machine learning technology, and is a composite cross-technical field for solving the spatiotemporal prediction problem of landslide geological disasters.

Description

Technical Field

[0001] The invention relates to geological disaster prediction technology, and in particular to a landslide geological disaster time-space prediction technology suitable for complex mountainous areas. Background Art

[0002] Complex mountainous areas have complex topography, fragile geological environment, and changeable meteorological conditions, which are prone to landslide geological disasters. The remote sensing satellite images of the monitoring area are affected by the complex mountainous area clouds and the satellite shooting angle, which will produce shadow noise problems. This will cause large errors in the identification of landslide risk feature targets in the monitoring area images.

[0003] At present, the spatiotemporal prediction of landslide geological disasters is mainly carried out through manual investigation and online monitoring of landslide risk areas. The advantage of this method is that it can accurately delineate landslide risk areas and monitor risk areas online, which requires manual surveys and the deployment of a large number of sensor equipment for monitoring, which is costly. For complex mountainous areas, manual investigation is very difficult, and the cost of deploying sensor equipment for online monitoring in a large monitoring area is very high. In addition, the existing landslide risk prediction methods in mountainous areas are not accurate enough for landslide risk prediction in mountainous areas with complex geological structures, and cannot meet the spatiotemporal prediction of landslide risks in complex mountainous areas. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a method for predicting the time of landslide geological disaster risks in areas prone to landslides at low cost in complex mountainous areas and in more accurately delineating risk areas.

[0005] The technical solution adopted by the present invention to solve the above technical problems is a spatiotemporal prediction method for landslide geological disasters suitable for complex mountainous areas, which uses D-InSAR technology to pre-process mountain remote sensing satellite images, and then uses image recognition technology to perform feature recognition processing, finds landslide risk areas from a large-scale monitoring area, and delineates the scope of the landslide risk area. In the delineated landslide risk area, the monitoring data is collected online, and the prediction model of the present invention is used to perform time series prediction of the surface deformation of the prone area. After obtaining the time series curve of the surface deformation prediction of the landslide-prone area, the time series curve data is analyzed and processed to achieve the time prediction of the landslide risk.

[0006] The following steps are involved:

[0007] 1. Collect remote sensing images of the monitoring area;

[0008] 2. Pre-process the remote sensing images of the monitoring area to eliminate shadows caused by cloudy weather and shooting angles;

[0009] 3. Construct a landslide-prone area identification model, and input the pre-processed remote sensing image of the monitoring area into the landslide-prone area identification model to obtain the landslide-prone area;

[0010] 4. Collect monitoring data in landslide-prone areas. The monitoring data is divided into static data and dynamic data. Use geological data as static data and regional meteorological data as dynamic data.

[0011] 5. Prediction of surface deformation in landslide-prone areas:

[0012] 5-1 First, use the differential interferometry short baseline set time series analysis SBAS-InSAR technology to calculate the surface shape variables of the InSAR image of the demarcated area;

[0013] 5-2 Use the moving average method to decompose the surface deformation variables at each moment into internal factors and external factors;

[0014] 5-3 The intrinsic factor item surface deformation variables and static data are combined into an intrinsic factor item feature vector, which is input into an intrinsic factor surface deformation variable prediction model based on a polynomial model. The factor surface deformation variable prediction model outputs an intrinsic factor item surface deformation time series prediction curve;

[0015] 5-4 Input the external factor item feature vector composed of the external factor item surface deformation variable and the dynamic data into the internal factor item surface deformation variable prediction model based on the long short-term memory LSTM model, and the external factor item surface deformation variable prediction model outputs the external factor item surface deformation time series prediction curve;

[0016] 5-5 The surface deformation time series prediction curve of the intrinsic factor item is added with the surface deformation time series prediction curve of the external factor item to obtain the surface deformation time series prediction curve;

[0017] 6. Calculate the tangent angle of the surface deformation time series prediction curve at each time point, and determine whether the tangent angle is greater than a threshold. If so, it is considered that there is a sign of landslide, and this time point is the predicted landslide risk time. Otherwise, it is considered that there is no sign of landslide within the predicted time.

[0018] Furthermore, the preprocessing in step 2 uses differential interferometric radar measurement D-InSAR technology.

[0019] Furthermore, in step 3, the improved Faster-RCNN model is used to complete the delineation of landslide-prone areas;

[0020] The improved Faster-RCNN model adds a refining network to the Fast-RCNN model; after the Fast-RCNN model generates the initial landslide area delineation result, it is judged whether the initial landslide area delineation result is the best fit. If so, the initial landslide area delineation result is used as the final landslide area delineation result to complete the construction of the landslide feature target recognition model. If not, the initial landslide area delineation result is output to the refining network to regress and refine the frame, and the landslide area is re-delineated to establish the landslide feature target recognition model. The best fit is to judge whether the area is a landslide prone area by delineating a small area around the delineated area. If it is, it belongs to the best fit, and if not, it needs to be retrained. This judgment method is set by the present invention based on actual conditions.

[0021] Specifically, the static data include geological structure, stratum lithology, slope, topography and vegetation coverage; the dynamic data include rainfall, surface temperature and soil moisture content.

[0022] Specifically, the surface shape variable A at each time t is t Decomposed into internal factor term α t and the external factor term β t The specific method is:

[0023]

[0024] Among them, A t-1 …A t-n represents the total surface deformation in the previous period, and n represents the number of previous moments included in the calculation.

[0025] Furthermore, the internal factor surface deformation variable prediction model and the external factor surface deformation variable prediction model are updated by updating the training set at set intervals. The specific steps are as follows:

[0026] 1) Temporarily store the surface deformation time series data and external data of the most recent set time in the sliding window;

[0027] 2) The decomposed external factor items are fused with external data and added to the external factor item training set, and the internal factor items are fused with static data and added to the internal factor item training set;

[0028] 3) Using the training set of intrinsic factor items to dynamically update the surface shape variable prediction model of intrinsic factor items based on the polynomial regression algorithm; using the training set of external factor items to dynamically update the surface shape variable prediction model of external factor items based on the LSTM neural network method;

[0029] 4) Wait for new surface deformation time series data and external data to enter the sliding window, and execute steps 1) to 3).

[0030] The beneficial effects of the present invention are:

[0031] The D-InSAR technology is used to preprocess the collected satellite images of the monitoring area to eliminate the shadow noise data of the satellite images of the monitoring area, thereby improving the accuracy of landslide risk feature recognition in the remote sensing images of the monitoring area.

[0032] The improved Fast-RCNN model is used to identify landslide feature targets in the pre-processed remote sensing images of the monitoring area, and the target area is circled on the image. This will help find landslide-prone points in a large monitoring area, thereby solving the problem of difficult manual surveys in complex mountainous areas and the cost of using sensors to monitor landslide risks in large areas.

[0033] The LSTM model is used to predict the impact of external factors on the surface deformation variable, and the PR model is used to predict the impact of internal factors on the surface deformation variable. In order to improve the prediction accuracy of the surface deformation variable, the present invention predicts the surface deformation variable based on internal and external factors respectively, and superimposes the prediction results of the two. In this way, the accuracy of landslide time prediction is improved.

[0034] According to the surface deformation variable prediction curve, the tangent angle of the surface deformation variable curve in the prone area is calculated and predicted, and the threshold is used to determine whether there is a possibility of landslide at each moment in the future. The surface deformation variable prediction model of landslide geological disasters is dynamically updated to ensure the prediction accuracy of surface deformation. This can improve the time prediction accuracy of landslide prone areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 Overall processing flow chart of spatiotemporal prediction of landslide geological hazards in complex mountainous areas;

[0036] Figure 2 Construction of landslide characteristic target identification model;

[0037] Figure 3 Landslide target identification results;

[0038] Figure 4 Surface shape variable prediction model learning and online prediction;

[0039] Figure 5 Predicted surface deformation tangent angle change curve;

[0040] Figure 6 The process of predicting the risk time of landslide-prone areas. Specific implementation plan

[0041] The process of accurate spatiotemporal risk prediction of landslide geological hazards in complex mountainous areas is as follows: Figure 1 As shown:

[0042] 1. Remote sensing image acquisition of monitoring area

[0043] The satellite SAR images of the area were downloaded by inputting the latitude and longitude of the monitoring area on the ASF system, a data sharing platform of the European Space Agency. According to the requirements of the Faster-RCNN method for SAR image processing, the image size was set to 600px×300px, and the InSAR images of the monitoring area were collected every 6 days.

[0044] 2. D-InSAR preprocessing of remote sensing images of the monitoring area

[0045] Due to the complex geological structure, poor geological conditions and large meteorological changes in complex mountainous areas, the InSAR images taken in the monitoring area are affected by shadow interference due to factors such as mountain fog and satellite shooting angle, which will affect the accuracy of landslide risk target identification in InSAR images. The embodiment uses existing D-InSAR technical tools to pre-process the SAR images of the monitoring area to eliminate image shadows.

[0046] Specifically, the SNAP software tool in the D-InSAR technical tool is used to perform orbit correction, pulse band removal and terrain correction on the remote sensing images, and output SAR images of the monitoring area with shadow noise removed, which prepares for subsequent SAR image recognition.

[0047] 3. Recognition of characteristic targets of landslide geological disasters in monitoring area images based on Faster-RCNN model method

[0048] There are many image recognition algorithms in various fields at present, but it is necessary to find an image recognition algorithm suitable for landslide features. In the field of image recognition, the algorithms that can support landslide feature recognition mainly include feature face algorithm, Fisher face algorithm, mtcnn algorithm and Faster-RCNN algorithm. Among them, Faster-RCNN is a very classic image feature recognition algorithm. As a two-stage algorithm, compared with the one-stage algorithm, its accuracy is higher, but the algorithm has a large time cost. For landslide image feature recognition, the most important thing is to be able to identify the landslide risk area with high accuracy, and there are not too many requirements for the speed of the algorithm. Therefore, the present invention realizes landslide image feature recognition based on Faster-RCNN optimization algorithm.

[0049] The landslide feature target recognition model is constructed by adding a refinement network to the existing standard Faster-RCNN model. The refinement network is composed of a deep convolutional neural network, a ROI pooling layer, a full connector and a bounding box regressor in series.

[0050] The construction process of the landslide feature target recognition model is as follows: Figure 2 shown.

[0051] Firstly, the InSAR image dataset of the landslide monitoring area is input into the Faster-RCNN model for training, and the deep convolutional neural network in the Faster-RCNN model is used to extract the feature attributes of the InSAR image. Then, the extracted feature attributes are input into the RPN proposal network and the ROI pooling layer respectively, where the RPN is used to generate candidate boxes and output the results to the ROI pooling layer. After comprehensive analysis, the ROI pooling layer generates and outputs the identification area to the bounding box regressor. The bounding box regressor is used to output the initial landslide area delineation result, and then judge whether the initial landslide area delineation result is the best fit. If so, the initial landslide area delineation result is used as the final landslide area delineation result to complete the construction of the landslide feature target recognition model. If not, the initial landslide area delineation result is output to the refinement network to regress and refine the bounding box, and the landslide area is re-delineated to establish the landslide feature target recognition model.

[0052] After the landslide feature recognition model is built, the SAR image of the monitoring area is input into the landslide feature recognition model to obtain the final landslide area delineation result, such as Figure 3 shown.

[0053] 4. Monitoring data collection in landslide-prone areas

[0054] After the landslide area is delineated, the surface deformation of the delineated area needs to be predicted, and the monitoring data of the area needs to be collected. The present invention divides the monitoring data into static data and dynamic data.

[0055] Geological data is static data, including geological structure, stratum lithology, slope, topography and vegetation cover, and regional meteorological data is dynamic data, including rainfall, surface temperature and soil moisture content. The collection websites of static data and dynamic data are shown in Table 1.

[0056] Table 1

[0057]

[0058]

[0059] Embodiment According to the website in the above table, write a python script to collect and store data.

[0060] 5. Prediction of surface deformation in landslide-prone areas

[0061] There are many factors that affect the land surface shape variables, which can be divided into two categories: internal factors and external factors. There is a large error in simply using the LSTM model method to predict the land surface shape variables. To this end, in order to improve the prediction accuracy of the land surface shape variables, the present invention combines the LSTM model with the PR polynomial regression model to learn the land surface variable prediction model, in which the LSTM model method will be used to predict the impact of external factors on the land surface shape variables, and the PR model method will be used for the impact of internal factors on the land surface shape variables.

[0062] First, use the differential interferometry short baseline set time series analysis SBAS-InSAR technology to calculate the surface deformation variables of the SAR image in the demarcated area. SBAS-InSAR technology is a prior art and will not be described here in detail. In order to improve the accuracy of the prediction of surface deformation variables, the present invention predicts the surface deformation variables from external and internal factors respectively, and then adds the two parts of the results. The embodiment learns to construct a surface deformation prediction model based on the long short-term memory-polynomial LSTM-PR combination method, and realizes the prediction of surface deformation variables in prone areas. The surface deformation prediction model for landslide-prone areas is as follows: Figure 4 As shown, it includes surface shape variable decomposition, internal factor prediction model and external factor prediction model.

[0063] 1) Decomposition of surface shape

[0064] The moving average method is used to convert the surface deformation variable A at each time t into t Decomposed into internal factor term α t and the external factor term β t .

[0065]

[0066] Among them, A t-1 …A t-n represents the total surface deformation in the previous period, and n represents the number of previous moments added to the calculation.

[0067] β t =A t -α t

[0068] Among them, A t represents the surface deformation at time t, α t Represents the intrinsic factor term at time t.

[0069] 2) Prediction model of surface deformation variables based on intrinsic factors

[0070] The intrinsic factor items and static data at time t are combined into the feature vector of the intrinsic factor surface deformation variable prediction model, and the polynomial-based intrinsic factor surface deformation variable prediction model outputs the internal factor item surface deformation time series prediction curve. The intrinsic factor surface deformation variable prediction model uses polynomial regression to complete learning during the training process, and each sample vector in the intrinsic factor item training set used in training is also a combination of intrinsic factor items and static data.

[0071] 3) Prediction of surface deformation variables based on external factors

[0072] The external factor item at time t and the dynamic data at time t are combined to form the feature vector of the external factor item surface deformation variable prediction model. The external factor item surface deformation variable prediction model based on long short-term memory (LSTM) outputs the external factor item surface deformation time series prediction curve. The external factor item surface deformation variable prediction model uses LSTM to complete learning during the training process. Each sample vector in the external factor item training set used during training is also a combination of external factor items and dynamic data. Compared with BP neural network and extreme learning machine (ELM), the main advantage of LSTM lies in time series data processing. The feature vector of the external factor item surface deformation variable prediction model is processed using the time series method.

[0073] 4) Surface deformation time series prediction curve

[0074] The surface deformation time series prediction curve of the internal factor item and the surface deformation time series prediction curve of the external factor item are added to obtain the surface deformation time series prediction curve. The surface deformation variable prediction model of the embodiment outputs the surface deformation time series prediction curve for the next seven days. For example, the change curve of the tangent angle of a landslide prone area predicted at 2021-07-01 00:00:00 in the next seven days is as follows: Figure 5 shown.

[0075] In order to ensure the adaptability and accuracy of the surface shape variable prediction model to the environment, the surface shape variable prediction model will update the model parameters by updating the training set at set intervals. The specific steps are as follows:

[0076] Step 1: Temporarily store the surface deformation time series data and external data of the last fourteen days in the sliding window. Static data does not belong to the data updated at the set time. The surface deformation variable is decomposed into internal factor items and external factor items using the moving average method;

[0077] Step 2: The decomposed external factor items are fused with the external data and added to the external factor item training set; the internal factor items are fused with the static data and added to the internal factor item training set.

[0078] Step 3: Use the training set of intrinsic factor items to dynamically update the intrinsic factor item surface shape variable prediction model based on the polynomial regression algorithm.

[0079] Step 4: Use the training set of external factor items to dynamically update the external factor item surface shape variable prediction model based on the LSTM neural network method.

[0080] Step 5: Wait for new surface deformation time series data and external data to enter the sliding window, and execute steps 1 to 4.

[0081] 6. Prediction of risk time in landslide prone areas

[0082] The flow chart of risk time prediction for landslide prone areas is as follows: Figure 6 As shown. The risk time prediction of landslide prone areas is based on the surface deformation variables, and the surface deformation time series prediction curve of the demarcated area is analyzed. The present invention predicts whether there will be a landslide risk in the future time period by judging the inflection point of the tangent angle of the predicted time series curve of the surface deformation variables in the landslide prone area. That is, the tangent angle of the curve at each moment is calculated to determine whether the tangent angle is greater than a threshold. If so, it is considered that there is a sign of landslide, and this time point is the predicted landslide risk time. Otherwise, it is considered that there is no sign of landslide in the next seven days.

[0083] from Figure 5 It can be seen that the tangent angle value of the tangent angle change curve in the landslide-prone area at 2021-07-04 09:00:00 is greater than 80 degrees, which means that the landslide may occur at 9:00 am on the 4th.

[0084] The embodiment can achieve a target recognition rate of 89% for landslide characteristics by delineating areas prone to landslides, meeting the needs of practical applications. For the identified areas prone to landslides, landslide risk monitoring data is collected, and machine learning is performed on these data to establish a surface deformation prediction model for the monitoring area. After obtaining the time series curve of the surface deformation prediction, the tangent angle of the surface deformation variable prediction time series curve is calculated to determine whether there is a curve inflection point to determine whether there is a landslide risk. In this way, a landslide geological disaster risk time prediction model is established. Compared with the previous model without delineating landslide risk areas, the accuracy of the predicted disaster risk time is improved by about 10%.

Claims

1. A spatiotemporal prediction method for landslide geological disasters suitable for complex mountainous areas, characterized in that: The following steps are involved: (1) Collect remote sensing images of the monitoring area; (2) Pre-process the remote sensing images of the monitoring area to eliminate shadows caused by foggy weather and shooting angles; (3) Construct a landslide-prone area identification model and input the preprocessed remote sensing image of the monitoring area into the landslide-prone area identification model to obtain the landslide-prone area; (4) Collect monitoring data in landslide-prone areas. The monitoring data are divided into static data and dynamic data. Use geological data as static data and regional meteorological data as dynamic data. (5) Prediction of surface deformation in landslide-prone areas: (5-1) First, use the differential interferometry short baseline set time series analysis SBAS-InSAR technology to calculate the surface shape variables of the InSAR image of the demarcated area; (5-2) Using the moving average method, the surface deformation variables at each moment are decomposed into internal factors and external factors; (5-3) The intrinsic factor item surface deformation variable and the static data constitute the intrinsic factor item feature vector and input it into the intrinsic factor surface deformation variable prediction model based on the polynomial model, and the factor surface deformation variable prediction model outputs the intrinsic factor item surface deformation time series prediction curve; (5-4) The surface deformation variables of the external factor items and the dynamic data constitute the feature vector of the external factor items and input it into the external factor surface deformation variable prediction model based on the long short-term memory (LSTM) model. The external factor surface deformation variable prediction model outputs the external factor item surface deformation time series prediction curve; (5-5) The surface deformation time series prediction curve of the intrinsic factor item is added to the surface deformation time series prediction curve of the external factor item to obtain the surface deformation time series prediction curve; (6) Calculate the tangent angle of the surface deformation time series prediction curve at each time point and determine whether the tangent angle is greater than a threshold. If so, it is considered that there is a sign of landslide and the time point is the predicted landslide risk time. Otherwise, it is considered that there is no sign of landslide within the predicted time.

2. The method according to claim 1, characterized in that: The preprocessing in step 2 uses differential interferometric radar measurement D-InSAR technology.

3. The method according to claim 1, characterized in that: In step 3, the improved Faster-RCNN model is used to delineate the landslide-prone areas; The improved Faster-RCNN model adds a refinement network to the Fast-RCNN model; after the Fast-RCNN model generates the initial landslide area delineation result, it is judged whether the initial landslide area delineation result is the best fit. If so, the initial landslide area delineation result is used as the final landslide area delineation result to complete the construction of the landslide feature target recognition model. If not, the initial landslide area delineation result is output to the refinement network to regress and refine the border, and the landslide area is re-delineated to establish the landslide feature target recognition model.

4. The method according to claim 1, characterized in that: The static data include geological structure, stratum lithology, slope, topography and vegetation cover; the dynamic data include rainfall, surface temperature and soil moisture content.

5. The method according to claim 1, characterized in that: The surface deformation variable A at each time t t Decomposed into internal factor term α t and the external factor term β t The specific method is: Among them, A t-1 …A t-n represents the total surface deformation in the previous period, and n represents the number of previous moments included in the calculation.

6. The method according to claim 1, characterized in that: The internal factor surface deformation variable prediction model and the external factor surface deformation variable prediction model are updated by updating the training set at set intervals. The specific steps are as follows: 1) Temporarily store the surface deformation time series data and external data of the most recent set time in the sliding window; 2) The decomposed external factor items are fused with external data and added to the external factor item training set, and the internal factor items are fused with static data and added to the internal factor item training set; 3) Using the training set of intrinsic factor items to dynamically update the surface shape variable prediction model of intrinsic factor items based on the polynomial regression algorithm; using the training set of external factor items to dynamically update the surface shape variable prediction model of external factor items based on the LSTM neural network method; 4) Wait for new surface deformation time series data and external data to enter the sliding window, and execute steps 1) to 3).

Images