CN112200364A - Landslide prediction method, device, equipment and storage medium - Google Patents

Landslide prediction method, device, equipment and storage medium Download PDF

Info

Publication number
CN112200364A
CN112200364A CN202011066238.6A CN202011066238A CN112200364A CN 112200364 A CN112200364 A CN 112200364A CN 202011066238 A CN202011066238 A CN 202011066238A CN 112200364 A CN112200364 A CN 112200364A
Authority
CN
China
Prior art keywords
landslide
grid
self
historical
data
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202011066238.6A
Other languages
Chinese (zh)
Inventor
滕佳颖
郑增荣
江子君
宋杰
胡辉
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hangzhou Ruhr Technology Co Ltd
Original Assignee
Hangzhou Ruhr Technology Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Hangzhou Ruhr Technology Co Ltd filed Critical Hangzhou Ruhr Technology Co Ltd
Priority to CN202011066238.6A priority Critical patent/CN112200364A/en
Publication of CN112200364A publication Critical patent/CN112200364A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Abstract

The embodiment of the invention discloses a landslide prediction method, a landslide prediction device, equipment and a storage medium, wherein a self-step classification learning model is selected, a self-step factor of the self-step classification learning model is determined according to the number of sub-boxes and the iteration frequency of historical non-landslide data, and a down-sampling proportion is determined according to the self-step factor of each sub-box and the self-step factors of all sub-boxes, so that when the self-step factor and the down-sampling proportion are iteratively adjusted based on the historical landslide data and the historical non-landslide data, the down-sampling quantity in each sub-box is gradually and uniformly changed along with the increase of the iteration frequency, and the down-sampling quantity of the sub-box with small hardness can be ensured to be always higher than that of the sub-sampling quantity of the sub-box with large hardness while the down-sampling proportion of each box is uniformly changed, so that a landslide prediction model. The landslide prediction is carried out based on the trained landslide prediction model, and the precision and the reliability of the landslide prediction can be improved.

Description

Landslide prediction method, device, equipment and storage medium
Technical Field
The embodiment of the invention relates to the technical field of landslide monitoring, in particular to a landslide prediction method, a landslide prediction device, equipment and a storage medium.
Background
Landslide is one of the most common disasters, and has the characteristics of wide distribution range, high occurrence frequency, high multiplicity, regionality, severity and the like, and the landslide can cause a great amount of casualties and great environmental and infrastructure loss every year. The method has important significance in evaluating the easiness of landslide.
The existing landslide tendency prediction generally adopts a non-deterministic method according to the difference of theoretical bases on which the landslide tendency prediction is based. With the rapid development of computer technology and 3S technology in recent years, non-deterministic methods are widely applied, mainly including fuzzy logic methods, analytic hierarchy process, decision trees, and the like. However, the non-deterministic method has a serious imbalance in the positive and negative sample ratios of the obtained landslide influence factors, which affects the accuracy of the landslide occurrence probability and needs to be improved.
Disclosure of Invention
The invention provides a landslide prediction method, a landslide prediction device, equipment and a storage medium, so that a diversified landslide prediction model with high robustness and strong containment is obtained, the landslide occurrence probability is determined based on the landslide prediction model, and the effect of improving the landslide prediction accuracy is achieved.
In a first aspect, an embodiment of the present invention provides a landslide prediction method, including:
acquiring landslide influence factors of each grid of a target area every day within a set time period;
determining the occurrence probability of each grid landslide on the current day according to the landslide influence factors based on a landslide prediction model, wherein the landslide prediction model is a self-step classification learning model, the landslide prediction model is determined based on historical landslide data and historical non-landslide data iteration adjustment self-step factors and a down-sampling proportion, the self-step factors are determined according to the number of boxes of the historical non-landslide data and the iteration times, and the down-sampling proportion is determined according to the self-step factors of each box and the self-step factors of all boxes.
In a second aspect, an embodiment of the present invention further provides a landslide prediction apparatus, including:
the landslide influence factor acquisition module is used for acquiring landslide influence factors of each grid of the target area every day within a set time period;
and the landslide prediction module is used for determining the occurrence probability of each grid landslide according to the landslide influence factors based on a landslide prediction model, wherein the landslide prediction model is a self-step classification learning model, the landslide prediction model is determined based on historical landslide data and historical non-landslide data through iterative adjustment of self-step factors and a down-sampling proportion, the self-step factors are determined according to the number of boxes of the historical non-landslide data and the number of iterations, and the down-sampling proportion is determined according to the self-step factors of each box and the self-step factors of all boxes.
In a third aspect, an embodiment of the present invention further provides a landslide prediction apparatus, including:
one or more processors;
a memory 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 are caused to implement the landslide prediction method described in the first aspect of the present invention.
In a fourth aspect, embodiments of the present invention also provide a storage medium containing computer-executable instructions, which when executed by a computer processor, are configured to perform the method for predicting landslide of the first aspect of the present invention.
The technical scheme of the embodiment of the invention selects the self-step classification learning model, determines the self-step factor of the self-step classification learning model according to the box number and the iteration times of historical non-landslide data, determines the down-sampling proportion according to the self-step factor of each box and the self-step factors of all the boxes, therefore, when the self-step factor and the down-sampling proportion are iteratively adjusted based on the historical landslide data and the historical non-landslide data, along with the increase of the iteration times, the sampling proportion of each box gradually changes uniformly from the lower-hardness down-sampling quantity to the last lower-sampling quantity in each box, and can ensure that the lower-sampling quantity of the box with low hardness is always higher than the lower-sampling quantity of the box with high hardness while the down-sampling proportion of each box is uniformly changed, so as to obtain the landslide prediction model with diversity, high robustness and strong inclusiveness, therefore, after acquiring the influence factors of the landslide every day in the set time period of each grid of the target area, the landslide prediction model can improve the precision and the reliability of landslide prediction and achieve the effect of improving the accuracy of landslide prediction.
Drawings
Fig. 1 is a flowchart of a landslide prediction method according to a first embodiment of the present invention;
fig. 2 is a flowchart of a landslide prediction method according to a second embodiment of the present invention;
fig. 3 is a flowchart of a landslide prediction method according to a third embodiment of the present invention;
fig. 4 is a schematic structural diagram of a landslide prediction apparatus according to a fourth embodiment of the present invention;
fig. 5 is a schematic structural diagram of a landslide prediction apparatus according to a fifth embodiment of the present invention.
Detailed Description
The present invention will be described in further detail with reference to the accompanying drawings and examples. It is to be understood that the specific embodiments described herein are merely illustrative of the invention and are not limiting of the invention. It should be further noted that, for the convenience of description, only some of the structures related to the present invention are shown in the drawings, not all of the structures.
Example one
Fig. 1 is a flowchart of a landslide prediction method according to an embodiment of the present invention, where the embodiment is applicable to a case where an evaluation is performed on a landslide susceptibility, and the method may be performed by a landslide prediction device, as shown in fig. 1, where the method includes the following steps:
and S110, acquiring landslide influence factors of each grid in the target area every day in a set time period.
The target area is usually an area where a landslide occurs, and may be any designated area. Raster data is a form of data in which a space is divided into regular grids, each grid is a raster or cell, and each cell is assigned a corresponding attribute value to represent an entity. The landslide impact factor is raster data. The set period of time may be one day, three days, one week, one month, or other period of time. The size of the grid may be 50 × 50m, 30 × 30m, or other sizes, and specifically, the size of the grid may be determined according to the size of the target area.
And S120, determining the occurrence probability of each grid landslide in the current day according to the landslide influence factors based on the landslide prediction model.
The landslide prediction model is a self-step classification learning model, the landslide prediction model is determined by iteratively adjusting self-step factors and downsampling proportions on the basis of historical landslide data and historical non-landslide data, the self-step factors are determined according to the number of boxes of the historical non-landslide data and the iteration times, and the downsampling proportions are determined according to the self-step factors of each box and the self-step factors of all the boxes.
The landslide impact factor comprises a dynamic factor and a static factor, wherein the dynamic factor comprises at least one of rainfall and soil moisture, and the static factor comprises at least one of elevation, gradient, slope, plane curvature, section curvature, terrain moisture index, current intensity index, sediment transport index, terrain roughness index, distance from fault, distance from river, distance from road, lithology, land use, and vegetation coverage. Specifically, a plurality of monitoring points can be set in the target area to obtain landslide influence factors of each monitoring point in real time. And further combining the data collected by a preset department to form a landslide influence factor of each grid of the target area every day in a set time period.
It will be appreciated that the number of days a landslide occurs in any region is generally less than the number of days a landslide does not occur, resulting in an imbalance in the proportions of positive and negative samples, i.e. an imbalance in the proportions of historical non-landslide data and historical landslide data. When the landslide prediction model is trained, the self-step classification learning model is selected as the landslide prediction model, the self-step classification learning model can balance unbalanced positive and negative samples in proportion to obtain a landslide prediction model with better robustness and stronger containment, and therefore when the trained landslide prediction model is used for landslide prediction, a landslide prediction result with higher accuracy can be obtained.
Optionally, the training method of the landslide prediction model includes the following steps:
step a, obtaining an initial prediction model, respectively extracting the characteristics of the historical landslide data and the historical non-landslide data, and respectively determining labels corresponding to the historical landslide data and the historical non-landslide data;
b, determining initial classification hardness of the initial prediction model based on historical landslide characteristics, historical non-landslide characteristics, historical landslide labels and historical non-landslide labels, determining the number of boxes of the historical non-landslide data according to the initial classification hardness, and determining self-stepping factors of all the boxes of the initial prediction model based on the number of the boxes;
c, determining the down-sampling proportion of each sub-box based on the self-stepping factor, and determining down-sampled historical non-landslide data in each sub-box based on the down-sampling proportion;
step d, inputting the down-sampled historical non-landslide data and the historical landslide data into the initial prediction model, determining a loss function of the initial prediction model based on the landslide probability and the historical landslide probability output by the initial prediction model, and iteratively adjusting a self-stepping factor and the down-sampling proportion of each sub-box based on the training classification hardness of the loss function on a single sample;
and e, adjusting the initial prediction model according to the self-step factors adjusted by iteration and the down-sampling proportion of each sub-box until the loss function reaches a set threshold value, and obtaining the landslide prediction model.
The historical landslide characteristics can be the average value, the maximum value or the minimum value of the landslide influence factors in the historical landslide grid, and the historical non-landslide characteristics are the average value, the maximum value or the minimum value of the landslide influence factors in the historical non-landslide grid. For example, the average value of the historical rainfall amount and the average value of the soil humidity of the historical landslide grid are used as the historical landslide characteristic, and the average value of the historical rainfall amount and the average value of the soil humidity of the historical non-landslide grid are used as the historical non-landslide characteristic. The label refers to the number of the historical landslide data and the historical non-landslide data, for example, the label of the historical landslide data is 1, and the label of the historical non-landslide data is 0.
The classification hardness refers to a loss function of the initial prediction model on a single sample, and can be positively correlated with the number of the sub-boxes. For example, if the classification hardness is larger, a larger number of bins is determined, and if the classification hardness is smaller, a smaller number of bins is determined, and the classification hardness of each bin is different. The self-step factor can be used for reducing the down-sampling proportion of the bins with excessive sample number, the self-step factor is obtained by adding 1 to the sum of the number of the bins, the opposite number of the bin labels and the iteration number, and the down-sampling proportion of each bin is the ratio of the self-step factor of each bin to the self-step factors of all the bins.
Specifically, the calculation formula of the classification hardness of the self-classification learning model is as follows:
Hx＝H(x，y，F)
wherein x is the characteristics of the sample, namely x comprises historical landslide characteristics and historical non-landslide characteristics, y is the label of the sample, namely y comprises historical landslide labels and historical non-landslide labels, and F is the self-classification learning model.
Specifically, the calculation formula of the historical non-landslide data in each sub-box is as follows:
Figure BDA0002713826220000061
wherein, H (·) refers to H (x, y, F), k is the number of the boxes, and l is the number of the boxes.
Specifically, the calculation formula of the self-step factor is as follows:
αlk +1-l + i, where i is the number of iterations.
Specifically, the down-sampling ratio is:
Figure BDA0002713826220000071
therein, sigmamαmK is the step factor for all bins, m 1.
Specifically, when an initial prediction model is trained, the initial prediction model is divided into a plurality of base classifiers, a training set D and model parameters of the initial prediction model are determined, the training set D comprises a majority of samples (historical non-landslide data D1) and a minority of samples (historical landslide data D2), the model parameters comprise hardness functions H (x, y, F), a base classifier F, the number n of the base classifiers and the number k of boxes, and the historical non-landslide data in each box are B respectively1，B2，...，BkWithin each bin, the self-step factor αlK +1-l + i, the self-step factor α in each bin is divided intolSelf-step factor alpha with all binsmDetermining down-sampling proportion of each bin
Figure BDA0002713826220000072
Based on down-sampled proportions
Figure BDA0002713826220000073
Determining downsampled historical non-landslide data within each bin
Figure BDA0002713826220000074
Historical non-landslide data based on downsampling of each box
Figure BDA0002713826220000075
Training a plurality of base classifiers of the initial prediction model by historical landslide data D2, further iteratively adjusting the self-stepping factor and the down-sampling proportion of each box based on the landslide probability and the historical landslide probability output by the initial prediction model until the loss function of the initial prediction model reaches a set threshold value to obtain a plurality of target classifiers, and classifying the target classifiers by a plurality of meshesAnd the model formed by the standard classifier is used as a landslide prediction model. It should be noted that, when the self-step classification learning model is used for learning, the down-sampling number of the bins with low hardness is always higher than the down-sampling number with high hardness, so that the contribution of each bin to the classification hardness can be balanced, and the sum of the classification hardnesses in each bin after down-sampling is consistent, so that the influence of noise on the learning process is reduced while the importance of the boundary samples is improved.
For example, the number k of bins of the initial prediction model is 10, and the self-step factor of each bin in the first iteration is: 11, 10, 9, 8, 7, 6, 5, 4, 3, 2; the self-step factors of each bin in the second iteration are respectively: 12, 11, 10, 9, 8, 7, 6, 5, 4, 3; the self-step factors of each sub-box in the third iteration are respectively as follows: 13, 12, 11, 10, 9, 8, 7, 6, 5, 4; ... the self-step factor of each bin in the last iteration is: 20, 19, 18, 17, 16, 15, 14, 13, 12, 11. Thus, the down-sampling ratio for each bin in the first iteration is:
Figure BDA0002713826220000081
the down-sampling proportion of each sub-box in the second iteration is as follows:
Figure BDA0002713826220000082
the down-sampling proportion of each sub-box in the third iteration is as follows:
Figure BDA0002713826220000083
.., the down-sampling proportion of each box in the last iteration is:
Figure BDA0002713826220000084
therefore, as the number of iterations increases, the sampling proportion of each box gradually changes uniformly without changing too quickly, that is, each box gradually balances the number of downsamples with low hardness to the number of downsamples of the last box, and the downsampling proportion of each box with low hardness can be ensured to be higher than that of the boxes with high hardness all the time while the downsampling proportion of each box changes uniformlyThe number of downsamplings of the self-classification learning model enables the base classifiers of the trained self-classification learning model to be diversified and high in robustness, the inclusion of samples with unbalanced positive and negative proportions is strong, the precision and the reliability of landslide prediction can be further improved, and after the landslide influence factors of each grid in the target area within the set time period are obtained, the occurrence probability of each grid landslide with high accuracy can be obtained based on the trained self-classification learning model (namely the landslide prediction model).
Optionally, after obtaining the landslide influence factors of each day, performing feature extraction on the landslide influence factors of each day to obtain landslide feature data; and inputting the landslide characteristic data into the landslide prediction model, and determining the occurrence probability of a specific time point after the current day of each grid landslide.
Wherein the specific time point after the current day may be the third day, the fifth day, the seventh day, etc. after the current day. The landslide signature data may be an average, minimum or maximum of daily landslide data. The landslide prediction model can be obtained by training according to the landslide occurrence probability after the historical time point and a plurality of historical landslide data. For example, the historical time point is 20 days in 12 months in 2019, the average value of landslide data in 11 months in 2019 and 1 day in 12 months in 2019 and 15 days in 12 months in 2019 and the landslide occurrence probability in 30 days in 12 months in 2019 and 21 days in 12 months in 2019 are selected to train the landslide prediction model, and the trained landslide prediction model is obtained. Therefore, when predicting the occurrence probability of landslide after 7 days on 7 months at 7 months (7 days on 7 months at 2020) in 2020 based on the trained landslide prediction model, the average value of the landslide influence factors for each of 4 days on 7 months at 2020 to 6 days on 7 months at 2020 can be input to the landslide prediction model, and the occurrence probability of landslide on 7 days at 7 months at 2020 can be obtained. The landslide prediction model obtained by the method can predict the occurrence probability of a specific time point after the current day so as to expand the application range of the landslide prediction model.
The technical scheme provided by the embodiment is that a self-step classification learning model is selected, the self-step factors of the self-step classification learning model are determined according to the number of the sub-boxes of historical non-landslide data and the iteration times, and the down-sampling proportion is determined according to the self-step factors of each sub-box and the self-step factors of all sub-boxes, so that when the self-step factors and the down-sampling proportion are adjusted in an iteration mode based on the historical landslide data and the historical non-landslide data, along with the increase of the iteration times, the sampling proportion of each sub-box is gradually and uniformly changed from the number of the down-samples with low hardness to the number of the down-samples in each last sub-box, and the number of the down-samples of the sub-boxes with low hardness can be ensured to be always higher than the number of the down-samples of the sub-boxes with high hardness while the down-sampling proportion of each box is uniformly changed, so as to obtain the landslide prediction model with high diversity, the landslide prediction model can improve the precision and the reliability of landslide prediction and achieve the effect of improving the accuracy of landslide prediction.
Example two
Fig. 2 is a flowchart of a landslide prediction method according to a second embodiment of the present invention, which is a further refinement of the previous embodiment. Optionally, the determining, based on the landslide prediction model, the occurrence probability of each grid landslide of the current day according to the landslide influence factor includes: determining the occurrence probability of landslide every day according to the landslide influence factor based on a first preset model; and if the occurrence probability of the landslide of the current day is greater than the preset probability threshold value, determining the occurrence probability of the landslide of each grid of the current day according to the landslide influence factors of each grid based on the second preset model. In the method, reference is made to the above-described embodiments for those parts which are not described in detail. Referring specifically to fig. 2, the method may include the steps of:
s210, acquiring landslide influence factors of each grid of the target area every day in a set time period.
S220, based on the first preset model, determining the occurrence probability of landslide every day according to the landslide influence factors.
Similar to the previous embodiment, the landslide impact factors include dynamic factors including rainfall (precipitation) and soil moisture, and static factors including elevation, grade, slope, plane curvature, section curvature, terrain moisture index, current intensity index, sediment transport index, terrain roughness index, distance from fault, distance from river, distance from road, lithology, land use, vegetation coverage, and the like.
Optionally, the first preset model is a self-step classification learning model, the first preset model may be determined by iteratively adjusting a self-step factor and a down-sampling proportion based on historical landslide data and historical non-landslide data, the self-step factor is determined according to the number of binning and the number of iteration times of the historical non-landslide data, and the down-sampling proportion is determined according to the self-step factor of each binning and the self-step factors of all the binning. The training process of the first preset model is the same as the training process of the landslide prediction model described in the foregoing embodiment.
Optionally, the method for determining the occurrence probability of daily landslide according to the landslide influence factor based on the first preset model includes: calculating the construction characteristics of the dynamic factors of each grid every day, wherein the construction characteristics comprise the sum, the average value, the maximum value, the minimum value, the range, the quartile and the rainfall time of each dynamic factor of each grid, and obtaining a first characteristic matrix of the first preset model according to the construction characteristics and the static factors every day; and inputting the first feature matrix into the first preset model so as to determine the occurrence probability of landslide every day according to the first preset model.
Illustratively, the acquired landslide impact factor X for each grid per day is: x ═ X(1)，X(2)，…，X(d)In which the upper corner indicates days, X(c)C is 1, 2, …, d, representing the landslide impact factor for each grid in the target area of day c, where the matrix X(c)The rows of (1) represent a grid, the columns represent a landslide factor, matrix X(c)Is m × n, i.e. comprises m grids and n landslide factors, wherein i is 1, 2, … n1Denotes a dynamic factor (rainfall, soil humidity, etc.), i ═ n1+1，n1+2, … n, representing a static factor. Constructing feature set X in units of days1(first feature matrix) constructed as follows: for each X(c)Push gridCalculating n1The sum, the average value, the maximum value, the minimum value, the range, the upper quartile, the lower quartile and other structural characteristics of the dynamic factors are obtained, and the characteristic set X is obtained from the structural characteristics and the static factors1And obtaining a first feature matrix, inputting the first feature matrix to the trained first preset model, and outputting the occurrence probability of the landslide every day. Because the first preset model is a self-step classification learning model determined by iteratively adjusting self-step factors and down-sampling proportions based on historical landslide data and historical non-landslide data, the precision of the occurrence probability of landslide every day can be improved, the dynamic factors are subjected to feature construction, the effect of the dynamic factors in model prediction is increased, and the prediction precision of the first preset model is further improved.
And S230, if the occurrence probability of the landslide of the current day is greater than a preset probability threshold value, determining the occurrence probability of the landslide of each grid of the current day according to the landslide influence factors of each grid based on a second preset model.
The preset probability threshold may be 0.5, 0.6 or other values, and may also be expressed by a fraction or a percentage. When the occurrence probability of the landslide is greater than the preset probability threshold value, the occurrence probability of the landslide on the current day is high. The landslide influence factors are screened by setting the preset probability threshold value, and only when the occurrence probability of landslide on the current day is greater than the preset probability threshold value, the data on the current day are transmitted to the second preset model for further prediction, so that the data volume input by the model is greatly reduced, the processing efficiency is improved, and meanwhile, the prediction precision is improved.
Similarly to the foregoing embodiment, the second preset model is a self-step classification learning model, and the second preset model may be determined based on historical landslide data and historical non-landslide data by iteratively adjusting a self-step factor and a down-sampling proportion, where the self-step factor is determined according to the number of binning and the number of iterations of the historical non-landslide data, and the down-sampling proportion is determined according to the self-step factor of each binning and the self-step factors of all the binning. The training process of the second preset model is the same as the training process of the landslide and model described in the previous embodiment.
Optionally, the determining, based on the second preset model and according to the landslide influence factor of each grid, the occurrence probability of each grid landslide of the current day includes: performing feature extraction on the landslide influence factors of each grid corresponding to the current day to generate a second feature matrix of the second preset model; normalizing the second feature matrix; and inputting the normalized second feature matrix into the second preset model so as to determine the occurrence probability of each grid landslide in the current day according to the second preset model. The current day may be one or more, and may be determined according to an output result of the first preset model.
Specifically, assuming that the landslide occurrence probability on the c-th day is greater than the prediction probability threshold, the landslide impact factor X on the c-th day is determined(c)And sending the feature set to a second preset model, and constructing the feature set of the second preset model in a specific construction mode: calculating the structural characteristics of the sum, the mean, the variance, the median, the mean and the variance of the difference, the skewness, the kurtosis and the like corresponding to 3 days, 7 days, 15 days and 30 days of each dynamic factor of each grid on the day, simultaneously considering the characteristics of each grid in a set range taking the current grid as the center, such as a 3X 3 area range, whether the current grid is the maximum value or the minimum value in the set range, whether the current grid exceeds the mean value corresponding to the set range, whether the grid in the set range generates the excessive landslide, the number of times of the landslide of the current grid and the like, and finally obtaining a second characteristic matrix X of a second preset model2 fea。
In particular, the second feature matrix X may be normalized based on max-min2 feaThe respective features of (a) are normalized. Of course, other normalization algorithms can be selected for normalization. The embodiment of the invention does not limit the normalization algorithm of the first feature matrix and the second feature matrix. And further, inputting the normalized second feature matrix into a second preset model, and outputting the occurrence probability of each grid landslide of the current day.
According to the technical scheme of the embodiment of the invention, the landslide influence factor is obtained, particularly the dynamic factor in the landslide is predicted by taking the day as a unit, the time sequence information of the dynamic factor is fully considered, and the prediction precision is improved; the first preset model and the second preset model are both diversified landslide prediction models with high robustness and strong inclusion, landslide prediction is carried out in two stages by setting the two preset models, whether the probability of landslide occurring on the current day is predicted in the first stage is larger than a set value, if yes, landslide prediction in the second stage is carried out, and the landslide occurrence probability of a specific grid is carried out by taking the data of the current day as a reference, so that the waste of computing resources is greatly reduced, and the precision, the efficiency and the reliability of landslide prediction are improved; by constructing the characteristics of the dynamic factors, the functions of the dynamic factors in model prediction are increased, and the precision of the model prediction is further improved; through feature screening and normalization processing, the efficiency of model prediction is further improved.
EXAMPLE III
Fig. 3 is a flowchart of a landslide prediction method according to a third embodiment of the present invention. The present embodiment is used to analyze the entire flow of landslide prediction. In the method, reference is made to the above-described embodiments for those parts which are not described in detail. Referring specifically to fig. 3, the method may include the steps of:
s310, acquiring landslide influence factors of each grid of the target area every day in a set time period.
Optionally, before obtaining the landslide influence factors of each grid of the target area every day within a set time period, remote sensing image information and topographic information of the target area may also be obtained, based on a functional area classification model, each functional area of the target area is determined according to the remote sensing image information and the topographic information, and the target functional area is screened according to a label corresponding to each functional area.
The remote sensing image information refers to a satellite image and can be obtained through land satellite series earth surface reflectivity products, and the remote sensing image information can reflect the land feature types of each grid of a target area, such as a grassland type, a forest land type, a rice type and a building type; the topographic information may be obtained from a Digital Elevation Model (DEM) of the target area, and the topographic information reflects a maximum degree of variation of the grid information of a certain grid and surrounding grids of the target area. The terrain information may include elevation information, gradient information, curvature information, and the like, the elevation information may reflect a terrain feature elevation point of a certain grid of the target area, an elevation point near the important geographic target, or a key elevation point on the distribution range, the gradient information may reflect a gradient degree of the certain grid of the target area, and the curvature information may reflect a concave-convex condition of the certain grid of the target area.
Optionally, the method for determining each functional area of the target area includes: preprocessing the remote sensing image information, wherein the preprocessing comprises at least one of atmospheric correction, radiation correction and edge detection processing; respectively carrying out feature extraction on the preprocessed remote sensing image information and the preprocessed topographic information, inputting the remote sensing image features and the topographic features into the functional area classification model, and determining each functional area of the target area, wherein the remote sensing image features comprise the mean value, the maximum value and the minimum value of the remote sensing image information, the topographic information comprises elevation information, gradient information and curvature information, and the topographic features respectively comprise the mean value, the maximum value and the minimum value of the elevation information, the gradient information and the curvature information.
The functional area classification model is a random forest model, and the training method of the functional area classification model comprises the following steps: acquiring an initial forest model; inputting sample remote sensing image information and sample terrain information of each functional area in a historical time period into the initial forest model, and determining a prediction label and the probability of the prediction label of each functional area; and adjusting parameters of the initial forest model based on the prediction label, the probability of the prediction label and the sample labels of the functional areas until the prediction label is consistent with the sample labels and the probability of the prediction label reaches a set threshold value, so as to obtain the functional area classification model. And if the prediction label is consistent with the sample label and the probability of the prediction label reaches a set threshold value, the probability that the prediction label is the sample label is high, and the initial forest model under the iteration times is used as a functional area classification model. The set threshold may be a large value, for example, the set threshold is 0.9. Optionally, before the sample remote sensing image information is input into the initial forest model, preprocessing may be performed on the sample remote sensing image information, for example, atmospheric correction, radiation correction, edge detection, false color synthesis processing, and the like are performed on the sample remote sensing image information, so as to improve the training precision of the functional region classification model of the sample remote sensing information.
It can be understood that each functional area output by the functional area classification model includes all functional areas of the target area, including functional areas such as town buildings, urban green lands, water bodies, farmlands, bare soil, mountain forests and the like, and the landslide occurrence probability of each functional area is different. For example, urban buildings and urban greenbelts have a very low probability of landslide and water and mountain forests have a high probability of landslide. It should be noted that, by comparing the historical prediction probability corresponding to each functional area with the historical probability threshold, the target functional area is screened out, and the targeted landslide prediction is performed on the target functional area, so that the calculation amount can be reduced, and the landslide prediction efficiency of the target area can be improved.
If the functional areas of the target area are determined by the foregoing method, and the target functional areas are screened according to the labels corresponding to the functional areas, S310 may be replaced by: and acquiring the landslide influence factors of the target function area in each set time period of each grid every day.
In order to improve the characterization effect of the landslide impact factor in the target function area, after the target function area is determined, the target size of the grid of the target function area can also be determined. Alternatively, the target size may be determined by two methods. Optionally, a first method for determining the grid of the target size is to divide the grid corresponding to the target functional area uniformly. The uniform dividing method of the target size comprises the following steps: acquiring landslide information of a target functional area; determining a target size for each grid of the target area based on the landslide information, wherein the landslide information comprises a landslide location and a landslide area; wherein the determining a target size for each grid of the target area based on the landslide information comprises: determining the initial size and the initial grid number of each grid of the target area, and determining the landslide grid number and the non-landslide grid number according to the landslide information, wherein the landslide grid number is the grid number of landslides, and the non-landslide grid number is the grid number of non-landslides; if the ratio of the landslide grid number to the non-landslide grid number is in the set balance interval, taking the initial size as the target size; if the ratio of the landslide grid number to the non-landslide grid number is not in the set balance interval, adjusting the size of each grid of the target area, and re-determining the landslide grid number, the non-landslide grid number and the current grid number; and if the ratio of the number of the newly determined landslide grids to the number of the non-landslide grids is within the set balance interval and the ratio of the current grid number to the initial grid number is smaller than a set threshold, taking the adjusted size of each grid as the target size.
Wherein determining a target size for each grid of the target area based on the landslide information comprises: acquiring rainfall data of each rainfall station of the target area, the initial size and the initial grid number of each grid of the target area, and determining the landslide grid number and the non-landslide grid number according to the landslide information, wherein the landslide grid number is the grid number of landslides, and the non-landslide grid number is the grid number of non-landslides; calculating a rainfall prediction value of a rainfall station corresponding to any current grid according to rainfall data of a specific number of rainfall stations, and calculating a rainfall error based on the rainfall prediction value and an actual rainfall value of the rainfall station corresponding to the current grid; if the rainfall error is smaller than a set error threshold value and the ratio of the landslide grid number to the non-landslide grid number is within a set balance interval, taking the initial size as the target size; if the rainfall error is larger than a set error threshold value, adjusting the size of each grid of the target area, and recalculating the rainfall error according to the adjusted rainfall predicted value and the actually measured rainfall value of the rainfall station corresponding to any grid; and if the recalculated rainfall error is smaller than the set error threshold value, the ratio of the number of the redetermined landslide grids to the number of the redetermined non-landslide grids is in a set balance interval, and the ratio of the number of the current grids to the number of the initial grids is smaller than a set threshold value, taking the size of the adjusted grid as the target size.
In the embodiment, the number of landslide grids and the number of non-landslide grids are determined in the above manner, so that the target size of the grid is determined according to actual landslide information and rainfall, and further, the landslide influence factor has the best characterization effect under the target size, and the landslide occurrence probability prediction accuracy is improved.
Optionally, a second method for determining the target size is to divide the grid corresponding to the target functional area non-uniformly. The uneven dividing method of the grid with the target size comprises the following steps: acquiring a first grid size and a second grid size of a target area, and acquiring a landslide influence factor of the target area under the first grid size and the second grid size respectively, wherein the second grid size is larger than the first grid size; determining whether the second grid size is a desired grid size based on distribution characteristics of the landslide influence factor at the first grid size and the landslide influence factor at the second grid size, respectively, wherein the desired grid size is smaller than the second grid size and greater than or equal to the first grid size; if not, iteratively adjusting the second grid size based on a particular scale factor; and if the current grid size corresponding to the current iteration times is the expected grid size, inputting the daily landslide influence factors of each grid corresponding to the expected grid size into a landslide prediction model, and determining the occurrence probability of each grid landslide based on the landslide prediction model.
Wherein the determining whether the second grid size is the desired grid size based on the distribution characteristics of the landslide impact factor under the first grid size and the landslide impact factor under the second grid size, respectively, comprises: performing significance test on the distribution characteristics of the landslide influence factors under the second grid size based on the distribution characteristics of the landslide influence factors under the first grid size; determining the second grid size to be the desired grid size if the landslide image factor at the second grid size passes a saliency test, otherwise, the second grid size is not the desired grid size. Determining whether the second grid size is the desired grid size based on the distribution characteristics of the landslide influence factor under the first grid size and the landslide influence factor under the second grid size, respectively, includes: calculating first characteristic data based on the distribution characteristics of the landslide influence factors under the first grid size, and calculating second characteristic data based on the distribution characteristics of the landslide influence factors under the second grid size; wherein the first characteristic data and the second characteristic data comprise at least one of variance and mean; comparing a difference between the first characteristic data and the second characteristic data to a set threshold; determining the second grid size to be the desired grid size if the difference is less than the set threshold, otherwise, the second grid size is not the desired grid size.
Iteratively adjusting the second grid size based on a specific scale factor to obtain a current grid size under a current iteration number, including: and iteratively reducing the second grid size according to the specific scale factor to obtain the current grid size under the current iteration times.
By carrying out uneven division on the grid with the target size, the landslide influence factor has the advantages of simultaneously keeping rich characteristic information and simplifying the redundancy of data under the divided grid size; when the landslide probability is predicted subsequently, the landslide influence factors are input into the landslide prediction model by taking days as units according to the landslide influence factors in grids corresponding to different grid sizes, the landslide occurrence probability is predicted, and the landslide prediction accuracy is improved.
In this embodiment, the target size may be obtained by adjusting the initial size for a plurality of times, and the target size may be used as the first grid size. In order to improve the accuracy of landslide prediction, the first grid size can be resampled, and landslide prediction can be performed by combining the characteristic information of the resampled grid size. The specific method comprises the following steps: the method comprises the steps of obtaining a first grid size of a target area, determining a second grid size after resampling of the first grid size, wherein the second grid size is larger than the first grid size, dividing an initial grid corresponding to the second grid size into a plurality of grids according to the first grid size, extracting feature information of a landslide influence factor under the grid of the second grid size based on the first grid size, and inputting the feature information and the landslide influence factor into a first preset model and a second preset model to predict landslide. The extraction method of the characteristic information comprises the following steps: acquiring a characteristic value of a landslide influence factor in an eight-neighborhood grid of a current grid; and determining the characteristic information of the current grid according to the characteristic values in the eight-neighborhood grid.
In the embodiment, the occurrence probability of each grid landslide is determined by combining the characteristic information and the landslide influence factor, so that the data volume of the landslide influence factor can be increased, the occurrence probability of each grid landslide can be more accurately determined according to the characteristic information, and the accuracy of landslide prediction is improved.
Further, as described in the previous embodiments, the landslide impact factor includes a dynamic factor and a static factor. The dynamic factor includes at least one of rainfall and soil moisture. The rainfall amount can be determined by inputting the acquired geographic environment data of the target area into a rainfall interpolation model. The geographic environmental data includes geographic location data, atmospheric data, terrain data, and underlying surface data. The rainfall interpolation model can be a back propagation neural network (BP), a multilayer feedforward network based on an error back propagation algorithm and composed of nonlinear transformation units is adopted, the BP generally consists of an input layer, a hidden layer and an output layer, each layer also comprises N neurons, the neurons in the same layer are independent, and the output of the neurons in each layer only influences the input of the neurons in the lower layer after passing through a specific excitation function.
Specifically, the training method of the rainfall interpolation model comprises the following steps: acquiring an initial model, determining an initial weight matrix and an initial threshold of the initial model, and calculating the fitness of the initial model; inputting sample geographic environment information in a historical time period into the initial model, determining a predicted rainfall, iteratively adjusting an initial weight matrix and an initial threshold of the initial model based on a genetic algorithm according to the predicted rainfall and an actual rainfall of the historical time period; and based on the weight matrix threshold value after iterative adjustment, adjusting the initial model and calculating the fitness of the adjusted model until the recalculated fitness reaches an expected value, and taking the model corresponding to the fitness reaching the expected value as the rainfall interpolation model.
Specifically, the initial weight matrix may include a connection weight between the input layer and the hidden layer and a connection weight between the hidden layer and the output layer; the initial threshold may include a threshold of the hidden layer and a threshold of the output layer. The calculation formula of the fitness of the initial model is as follows:
Figure BDA0002713826220000201
where n is the number of samples, y (i) is the actual output of the ith sample of the initial model, and t (i) is the expected output of the ith sample of the initial model. Specifically, after sample geographic environment information in a historical time period is input into an initial model, a predicted rainfall is obtained, a root mean square error between the predicted rainfall and an actually measured rainfall is calculated, if the root mean square error is larger than a preset error threshold, weight coding and threshold coding of the initial model are adjusted in an iterative mode, an adjusted weight matrix and an adjusted threshold are determined based on the adjusted weight coding and threshold coding until the root mean square error is not smaller than the preset error threshold, the initial model is adjusted to a stable state, the fitness of the adjusted model is calculated by adopting the fitness calculation formula, and if the recalculated fitness reaches a desired value, a model corresponding to the fitness reaching the desired value is used as the rainfall interpolation model.
Optionally, the weight and the threshold are encoded by binary encoding, real number encoding or gray code encoding, and the adjusted weight matrix and threshold are determined according to the weight encoding and the threshold encoding. Optionally, the initial model may be adjusted to a stable state by calculating an average error, an average absolute error, or a linear correlation coefficient between the predicted rainfall and the measured rainfall, and iteratively adjusting the weight code and the threshold code of the initial model according to the average error, the average absolute error, or the linear correlation coefficient until the average error, the average absolute error, or the linear correlation coefficient is not less than a preset error threshold, and further calculating the fitness of the adjusted model and determining the rainfall interpolation model.
It should be noted that the sample geographic environment data in the historical time period is multidimensional, and when the rainfall interpolation model is trained, the accuracy and robustness of the rainfall interpolation model can be improved by the multidimensional sample geographic environment data, so that after the daily geographic environment data is obtained, the prediction accuracy of the rainfall probability can be improved based on the rainfall interpolation model prediction, and the landslide prediction accuracy can be improved.
And S320, determining the occurrence probability of the landslide every day according to the landslide influence factor based on the first preset model.
The characteristic information of the landslide influence factor is determined through S310. S320 may be replaced by: and based on the first preset model, determining the occurrence probability of landslide every day according to the landslide influence factor and the characteristic information.
Similarly to the foregoing embodiment, the first preset model may be a self-step classification learning model, and the first preset model is determined by iteratively adjusting a self-step factor and a down-sampling proportion according to historical landslide data and historical non-landslide data. Optionally, the method for determining historical landslide data and historical non-landslide data includes: determining labels corresponding to grids of a history area, and determining historical rainfall and specific correlation factors corresponding to a landslide grid and a non-landslide grid at each history time point; and generating a landslide data set comprising historical rainfall and specific correlation factors respectively corresponding to landslide grids and non-landslide grids based on the labels, and determining the historical landslide data and the historical non-landslide data according to the labels corresponding to the grids in the landslide data set.
S330, if the occurrence probability of the landslide of the current day is larger than a preset probability threshold value, determining the occurrence probability of the landslide of each grid of the current day according to the landslide influence factors of each grid based on a second preset model.
Optionally, the second preset model may also be a self-walking ensemble learning model. The determination manner of the historical landslide data for training the second preset model can refer to S320, and the step is not described in detail.
Optionally, after determining the occurrence probability of each grid landslide of the current day, the occurrence probability of each grid of the current day may be compared with the early warning threshold of the target area, and the occurrence level of the landslide of the target area is determined based on the early warning level corresponding to the obtained occurrence probability of each grid, where the early warning threshold is determined according to the first coefficient and the second coefficient of the grid landslide of each area.
The method for determining the first coefficient and the second coefficient comprises the following steps: acquiring historical probability of landslide occurring in a set time period of grid landslide of each region; determining a first determined probability interval and a second determined probability interval based on the distribution characteristics of the historical probabilities; respectively calculating the sum of products of probability threshold values and corresponding coefficients of the same risk level in the first determined probability interval and the second determined probability interval to obtain intermediate determined probability; and if the intermediate determined probability does not reach the set evaluation index, iteratively adjusting the corresponding coefficient of the probability threshold value of each risk level until the intermediate determined probability reaches the set evaluation index, and determining a first coefficient and a second coefficient according to the corresponding coefficient reaching the set evaluation index.
The method for determining the early warning threshold value comprises the following steps: determining a first prediction probability interval and a second prediction probability interval based on the distribution characteristics of the occurrence probability of each grid landslide; and calculating a first product of each risk level and the first coefficient in the first prediction probability interval, calculating a second product of each risk level and the second coefficient in the second prediction probability interval, and taking the sum of the first product and the second product as an early warning threshold value of each risk level.
Wherein the determining the first determined probability interval and the second determined probability interval based on the distribution feature of the historical probability comprises: determining the historical probability and the corresponding times of the historical probability; according to the historical probabilities and the landslide occurrence times corresponding to the historical probabilities, determining the density distribution characteristics and the breakpoint distribution characteristics of the historical probabilities; determining the first determined probability interval based on the intensity distribution characteristic and determining the second determined probability interval based on the breakpoint distribution characteristic.
It can be understood that, in the above manner, based on the historical probability of landslide occurring in the set time period of grid landslide of each region, the first determined probability interval and the second determined probability interval are determined according to the distribution characteristics of the historical probability, the first coefficient and the second coefficient of each region are determined according to the probability threshold values of each risk level in the first determined probability interval and the second determined probability interval and the set evaluation index, after the occurrence probabilities of different regions are obtained, the early warning threshold values of different regions can be flexibly determined according to the first coefficient, the second coefficient and the occurrence probability corresponding to different regions, which is beneficial to accurately determining the risk level of the region according to the early warning threshold values corresponding to the regions subsequently.
Determining the occurrence level of the landslide of the target area based on the early warning levels corresponding to the obtained occurrence probabilities of the grids, wherein the determining of the occurrence level of the landslide of the target area comprises the following steps of: screening a target grade greater than a first grade in the early warning threshold value; calculating the average grade of the grid corresponding to the target grade; and determining the occurrence grade of the landslide of the target area based on the average grade, the setting coefficient and the ratio of the number of grids corresponding to the target grade to the number of all grids in the target area.
The formula for calculating the occurrence grade of the landslide of the target area is as follows:
Figure BDA0002713826220000231
wherein alpha is a set coefficient and is obtained by carrying out Bayesian calculation on the historical occurrence level of the landslide of the target area,
Figure BDA0002713826220000232
and p is the ratio of the number of grids corresponding to the target level to the number of all grids in the target area.
Example four
Fig. 4 is a schematic diagram illustrating a result of a landslide prediction apparatus according to a fourth embodiment of the present invention, where as shown in fig. 4, the landslide prediction apparatus includes: a landslide impact factor acquisition module 410 and a landslide prediction module 420.
The landslide influence factor acquiring module 410 is configured to acquire landslide influence factors of each grid of the target area every day within a set time period;
and the landslide prediction module 420 is configured to determine the occurrence probability of each grid landslide according to the landslide influence factor based on a landslide prediction model, where the landslide prediction model is a self-step classification learning model, the landslide prediction model is determined based on historical landslide data and historical non-landslide data by iteratively adjusting a self-step factor and a downsampling proportion, the self-step factor is determined according to the number of binning of the historical non-landslide data and the number of iterations, and the downsampling proportion is determined according to the self-step factor of each binning and the self-step factors of all binning.
The technical scheme of the embodiment of the invention selects the self-step classification learning model, determines the self-step factor of the self-step classification learning model according to the box number and the iteration times of historical non-landslide data, determines the down-sampling proportion according to the self-step factor of each box and the self-step factors of all the boxes, therefore, when the self-step factor and the down-sampling proportion are iteratively adjusted based on the historical landslide data and the historical non-landslide data, along with the increase of the iteration times, the sampling proportion of each box gradually changes uniformly from the lower-hardness down-sampling quantity to the last lower-sampling quantity in each box, and can ensure that the lower-sampling quantity of the box with low hardness is always higher than the lower-sampling quantity of the box with high hardness while the down-sampling proportion of each box is uniformly changed, so as to obtain the landslide prediction model with diversity, high robustness and strong inclusiveness, therefore, after acquiring the influence factors of the landslide every day in the set time period of each grid of the target area, the landslide prediction model can improve the precision and the reliability of landslide prediction and achieve the effect of improving the accuracy of landslide prediction.
Optionally, the apparatus further comprises: a landslide prediction model training module; the landslide prediction model training module is used for acquiring an initial prediction model, respectively extracting the characteristics of the historical landslide data and the historical non-landslide data, and respectively determining labels corresponding to the historical landslide data and the historical non-landslide data;
determining initial classification hardness of the initial prediction model based on historical landslide characteristics, historical non-landslide characteristics, historical landslide labels and historical non-landslide labels, determining the number of bins of the historical non-landslide data according to the initial classification hardness, and determining self-stepping factors of each bin of the initial prediction model based on the number of bins;
determining the downsampling proportion of each sub-box based on the self-stepping factor, and determining downsampled historical non-landslide data in each sub-box based on the downsampling proportion;
inputting the downsampled historical non-landslide data and the historical landslide data into the initial prediction model, determining a loss function of the initial prediction model based on the landslide probability and the historical landslide probability output by the initial prediction model, and iteratively adjusting a self-stepping factor and a downsampling proportion of each sub-box based on training classification hardness of the loss function on a single sample;
and adjusting the initial prediction model according to the self-stepping factor of the iterative adjustment and the down-sampling proportion of each sub-box until the loss function reaches a set threshold value, thereby obtaining the landslide prediction model.
Optionally, the down-sampling proportion of each bin is a ratio of a self-step factor of each bin to self-step factors of all bins, and the self-step factor is obtained by adding 1 to a sum of the number of bins, a reverse number of the tags of the bins, and the number of iterations.
Optionally, the apparatus further comprises: a landslide prediction module at a particular time point; the landslide prediction module at a specific time point is used for carrying out feature extraction on the landslide influence factors every day to obtain landslide feature data;
and inputting the landslide characteristic data into the landslide prediction model, and determining the occurrence probability of a specific time point after the current day of each grid landslide.
Optionally, the landslide prediction module 420 is further configured to determine an occurrence probability of daily landslide according to the landslide influence factor based on a first preset model;
and if the occurrence probability of the landslide of the current day is greater than the preset probability threshold value, determining the occurrence probability of the landslide of each grid of the current day according to the landslide influence factors of each grid based on the second preset model.
Optionally, the landslide prediction module 420 is further configured to calculate a configuration feature of a dynamic factor of each grid every day, where the configuration feature includes a sum, an average, a maximum, a minimum, a range, a quartile and a rainfall time of each dynamic factor of each grid, and obtain a first feature matrix of the first preset model according to the configuration feature and a static factor every day;
and inputting the first feature matrix into the first preset model so as to determine the occurrence probability of landslide every day according to the first preset model.
Optionally, the landslide prediction module 420 is further configured to perform feature extraction on the landslide influence factors of each grid corresponding to the current day to generate a second feature matrix of the second preset model;
normalizing the second feature matrix;
and inputting the normalized second feature matrix into the second preset model so as to determine the occurrence probability of each grid landslide in the current day according to the second preset model.
The landslide prediction device provided by the embodiment of the invention can execute the landslide prediction method provided by any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the execution method.
EXAMPLE five
Fig. 5 is a schematic structural diagram of a landslide prediction apparatus according to a fifth embodiment of the present invention, as shown in fig. 5, the apparatus includes a processor 510, a memory 520, an input device 530, and an output device 540; the number of the device processors 510 may be one or more, and one processor 510 is taken as an example in fig. 5; the processor 510, the memory 520, the input device 530 and the output device 540 of the apparatus may be connected by a bus or other means, as exemplified by the bus connection in fig. 5.
The memory 520 may be used as a computer-readable storage medium for storing software programs, computer-executable programs, and modules, such as program instructions/modules corresponding to the landslide prediction method in the embodiment of the present invention (for example, the landslide influence factor obtaining module 410 and the landslide prediction module 420 in the landslide prediction device). The processor 510 executes various functional applications of the device and data processing by executing software programs, instructions and modules stored in the memory 520, that is, implements the landslide prediction method described above.
The memory 520 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application program required for at least one function; the storage data area may store data created according to the use of the terminal, and the like. Further, the memory 520 may include high speed random access memory, and may also include non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other non-volatile solid state storage device. In some examples, the memory 520 may further include memory located remotely from the processor 510, which may be connected to the device/terminal/server via a network. Examples of such networks include, but are not limited to, the internet, intranets, local area networks, mobile communication networks, and combinations thereof.
The input device 530 may be used to receive input numeric or character information and generate key signal inputs related to user settings and function control of the apparatus. The output device 540 may include a display device such as a display screen.
EXAMPLE six
Sixth embodiment of the present invention also provides a storage medium containing computer-executable instructions which, when executed by a computer processor, perform a method of predicting landslide, the method comprising:
acquiring landslide influence factors of each grid of a target area every day within a set time period;
determining the occurrence probability of each grid landslide on the current day according to the landslide influence factors based on a landslide prediction model, and taking the occurrence probability as a landslide prediction result of the target area, wherein the landslide prediction model is a self-step classification learning model, the landslide prediction model is determined based on historical landslide data, historical non-landslide data, iteration adjustment self-step factors and a down-sampling proportion, the self-step factors are determined according to the number of boxes of the historical non-landslide data and the iteration times, and the down-sampling proportion is determined according to the self-step factors of each box and the self-step factors of all boxes.
Of course, the storage medium containing the computer-executable instructions provided by the embodiments of the present invention is not limited to the method operations described above, and may also perform related operations in the landslide prediction method provided by any embodiment of the present invention.
From the above description of the embodiments, it is obvious for those skilled in the art that the present invention can be implemented by software and necessary general hardware, and certainly, can also be implemented by hardware, but the former is a better embodiment in many cases. Based on such understanding, the technical solutions of the present invention may be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as a floppy disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a FLASH Memory (FLASH), a hard disk or an optical disk of a computer, and includes several instructions for enabling a computer device (which may be a personal computer, a server, or a network device) to execute the methods according to the embodiments of the present invention.
It should be noted that, in the embodiment of the prediction apparatus using landslide, the included units and modules are only divided according to the functional logic, but are not limited to the above division as long as the corresponding functions can be realized; in addition, specific names of the functional units are only for convenience of distinguishing from each other, and are not used for limiting the protection scope of the present invention.
It is to be noted that the foregoing is only illustrative of the preferred embodiments of the present invention and the technical principles employed. It will be understood by those skilled in the art that the present invention is not limited to the particular embodiments described herein, but is capable of various obvious changes, rearrangements and substitutions as will now become apparent to those skilled in the art without departing from the scope of the invention. Therefore, although the present invention has been described in greater detail by the above embodiments, the present invention is not limited to the above embodiments, and may include other equivalent embodiments without departing from the spirit of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims (10)

1. A method of predicting landslide, comprising:
acquiring landslide influence factors of each grid of a target area every day within a set time period;
determining the occurrence probability of each grid landslide on the current day according to the landslide influence factors based on a landslide prediction model, wherein the landslide prediction model is a self-step classification learning model, the landslide prediction model is determined based on historical landslide data and historical non-landslide data iteration adjustment self-step factors and a down-sampling proportion, the self-step factors are determined according to the number of boxes of the historical non-landslide data and the iteration times, and the down-sampling proportion is determined according to the self-step factors of each box and the self-step factors of all boxes.
2. The prediction method according to claim 1, wherein the training method of the landslide prediction model comprises:
acquiring an initial prediction model, respectively extracting the characteristics of the historical landslide data and the historical non-landslide data, and respectively determining labels corresponding to the historical landslide data and the historical non-landslide data;
determining initial classification hardness of the initial prediction model based on historical landslide characteristics, historical non-landslide characteristics, historical landslide labels and historical non-landslide labels, determining the number of bins of the historical non-landslide data according to the initial classification hardness, and determining self-stepping factors of each bin of the initial prediction model based on the number of bins;
determining the downsampling proportion of each sub-box based on the self-stepping factor, and determining downsampled historical non-landslide data in each sub-box based on the downsampling proportion;
inputting the downsampled historical non-landslide data and the historical landslide data into the initial prediction model, determining a loss function of the initial prediction model based on the landslide probability and the historical landslide probability output by the initial prediction model, and iteratively adjusting a self-stepping factor and a downsampling proportion of each sub-box based on training classification hardness of the loss function on a single sample;
and adjusting the initial prediction model according to the self-stepping factor of the iterative adjustment and the down-sampling proportion of each sub-box until the loss function reaches a set threshold value, thereby obtaining the landslide prediction model.
3. The prediction method according to claim 1, wherein the down-sampling ratio of each bin is a ratio of a step factor of each bin to a step factor of all bins, the step factor being obtained by adding 1 to a sum of the number of bins, a reverse number of the number of bin labels, and the number of iterations.
4. The prediction method according to claim 1, further comprising:
performing feature extraction on the landslide influence factors every day to obtain landslide feature data;
and inputting the landslide characteristic data into the landslide prediction model, and determining the occurrence probability of a specific time point after the current day of each grid landslide.
5. The prediction method according to claim 1, wherein the determining the occurrence probability of each grid landslide of the current day according to the landslide influence factor based on the landslide prediction model comprises:
determining the occurrence probability of landslide every day according to the landslide influence factor based on a first preset model;
and if the occurrence probability of the landslide of the current day is greater than the preset probability threshold value, determining the occurrence probability of the landslide of each grid of the current day according to the landslide influence factors of each grid based on the second preset model.
6. The prediction method according to claim 5, wherein the determining the occurrence probability of daily landslide according to the landslide influence factor based on the first preset model comprises:
calculating the construction characteristics of the dynamic factors of each grid every day, wherein the construction characteristics comprise the sum, the average value, the maximum value, the minimum value, the range, the quartile and the rainfall time of each dynamic factor of each grid, and obtaining a first characteristic matrix of the first preset model according to the construction characteristics and the static factors every day;
and inputting the first feature matrix into the first preset model so as to determine the occurrence probability of landslide every day according to the first preset model.
7. The prediction method according to claim 5, wherein the determining the occurrence probability of landslide of each grid on the current day according to the landslide influence factor of each grid based on the second preset model comprises:
performing feature extraction on the landslide influence factors of each grid corresponding to the current day to generate a second feature matrix of the second preset model;
normalizing the second feature matrix;
and inputting the normalized second feature matrix into the second preset model so as to determine the occurrence probability of each grid landslide in the current day according to the second preset model.
8. A landslide prediction apparatus comprising:
the landslide influence factor acquisition module is used for acquiring landslide influence factors of each grid of the target area every day within a set time period;
and the landslide prediction module is used for determining the occurrence probability of each grid landslide according to the landslide influence factors based on a landslide prediction model, wherein the landslide prediction model is a self-step classification learning model, the landslide prediction model is determined based on historical landslide data and historical non-landslide data through iterative adjustment of self-step factors and a down-sampling proportion, the self-step factors are determined according to the number of boxes of the historical non-landslide data and the number of iterations, and the down-sampling proportion is determined according to the self-step factors of each box and the self-step factors of all boxes.
9. An apparatus for predicting landslide, the apparatus comprising:
one or more processors;
a memory for storing one or more programs;
when executed by the one or more processors, cause the one or more processors to implement a method of predicting landslide of any of claims 1-7.
10. A storage medium containing computer-executable instructions for performing the method of predicting landslide of any one of claims 1-7 when executed by a computer processor.
CN202011066238.6A 2020-09-30 2020-09-30 Landslide prediction method, device, equipment and storage medium Pending CN112200364A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011066238.6A CN112200364A (en) 2020-09-30 2020-09-30 Landslide prediction method, device, equipment and storage medium

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011066238.6A CN112200364A (en) 2020-09-30 2020-09-30 Landslide prediction method, device, equipment and storage medium

Publications (1)

Publication Number Publication Date
CN112200364A true CN112200364A (en) 2021-01-08

Family

ID=74012947

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011066238.6A Pending CN112200364A (en) 2020-09-30 2020-09-30 Landslide prediction method, device, equipment and storage medium

Country Status (1)

Country Link
CN (1) CN112200364A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113780378A (en) * 2021-08-26 2021-12-10 北京科技大学 Disease high risk group prediction device

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106776842A (en) * 2016-11-28 2017-05-31 腾讯科技（上海）有限公司 Multi-medium data detection method and device
CN111144651A (en) * 2019-12-26 2020-05-12 杭州鲁尔物联科技有限公司 Geological disaster prediction method, device and equipment

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106776842A (en) * 2016-11-28 2017-05-31 腾讯科技（上海）有限公司 Multi-medium data detection method and device
CN111144651A (en) * 2019-12-26 2020-05-12 杭州鲁尔物联科技有限公司 Geological disaster prediction method, device and equipment

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
ZHINING LIU ET AL.: "Self-paced ensemble for highly imbalanced massive data classification", 2020 IEEE 36TH INTERNATIONAL CONFERENCE ON DATA ENGINEERING (ICDE), pages 841 - 852 *
刘乾坤: "基于SPARK的四川省滑坡灾害成因分析与临界阈值研究", 中国优秀硕士学位论文全文数据库 基础科学辑, no. 07, pages 011 - 24 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113780378A (en) * 2021-08-26 2021-12-10 北京科技大学 Disease high risk group prediction device
CN113780378B (en) * 2021-08-26 2023-11-28 北京科技大学 Disease high risk crowd prediction device

Similar Documents

Publication Publication Date Title
CN112200358B (en) Landslide level prediction method, landslide level prediction device, landslide level prediction equipment and storage medium
CN112200354A (en) Landslide prediction method, device, equipment and storage medium
Kadavi et al. Landslide-susceptibility mapping in Gangwon-do, South Korea, using logistic regression and decision tree models
CN112200362B (en) Landslide prediction method, landslide prediction device, landslide prediction equipment and storage medium
CN108876027B (en) GIS-based rural residential point centralized residential area site selection and optimization method
Li et al. Data mining of cellular automata's transition rules
Akın et al. The impact of historical exclusion on the calibration of the SLEUTH urban growth model
CN111144651A (en) Geological disaster prediction method, device and equipment
Chu et al. Comparison of landslide susceptibility maps using random forest and multivariate adaptive regression spline models in combination with catchment map units
CN111784044A (en) Landslide prediction method, device, equipment and storage medium
CN111310898A (en) Landslide hazard susceptibility prediction method based on RNN
Yang et al. Enhancing weather-related power outage prediction by event severity classification
CN115688404B (en) Rainfall landslide early warning method based on SVM-RF model
CN112200356A (en) Landslide prediction method, device, equipment and storage medium
Meena et al. Assessing the importance of conditioning factor selection in landslide susceptibility for the province of Belluno (region of Veneto, northeastern Italy)
Biard et al. Automated detection of weather fronts using a deep learning neural network
CN112200363A (en) Landslide prediction method, device, equipment and storage medium
CN111539904A (en) Rainfall-based disaster vulnerability prediction method
CN111968019A (en) Landslide hazard susceptibility prediction method based on CNN and RNN integration
KR20200052398A (en) Method and apparatus for landslide susceptibility mapping using machine-learning architecture
Li et al. Uncertainties of collapse susceptibility prediction based on remote sensing and GIS: Effects of different machine learning models
Rudra et al. Artificial neural network for flood susceptibility mapping in Bangladesh
Bui et al. Landslide susceptibility prediction mapping with advanced ensemble models: Son La province, Vietnam
Vosper et al. Deep Learning for Downscaling Tropical Cyclone Rainfall to Hazard‐Relevant Spatial Scales
CN112200364A (en) Landslide prediction method, device, equipment and storage medium

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination