CN115859801A - Landslide space-time risk assessment method combined with effective rainfall model - Google Patents
Landslide space-time risk assessment method combined with effective rainfall model Download PDFInfo
- Publication number
- CN115859801A CN115859801A CN202211492956.9A CN202211492956A CN115859801A CN 115859801 A CN115859801 A CN 115859801A CN 202211492956 A CN202211492956 A CN 202211492956A CN 115859801 A CN115859801 A CN 115859801A
- Authority
- CN
- China
- Prior art keywords
- landslide
- time
- data
- effective rainfall
- rnn
- 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
Links
Images
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention provides a landslide space-time risk assessment method combined with an effective rainfall model, which analyzes the relevance and importance degree of landslide occurrence based on an obtained landslide induction factor; constructing a landslide susceptibility model by adopting three mixed deep learning networks, and calculating the spatial probability of landslide; by embedding an effective rainfall model in a transform network coding layer and changing an original attention mechanism module, an ST-transform deep learning network is provided for predicting effective rainfall of a landslide in different time periods, and the highest-precision prediction time is selected to evaluate the occurrence time probability of the landslide; and changing the original landslide risk calculation formula, and quantitatively calculating the landslide risk of the optimal prediction time by combining the two results. The method can provide reliable decision basis for landslide risk assessment and risk management.
Description
Technical Field
The invention relates to the field of risk assessment, in particular to a landslide space-time risk assessment method combined with an effective rainfall model.
Background
In order to deal with the threat caused by landslide disasters, a large number of methods are used for landslide hazard assessment, wherein a physical-based method is used for assessing landslide hazard by analyzing slope stability from the inside of soil and geological structures, and the method is accurate but can be implemented under the conditions of complete landslide historical data literature records and sufficient manpower and physical force. In addition, a statistical method, a probability method, a GIS hierarchical analysis method and the like have certain better interpretability on landslide risk assessment of a research area, but most of the landslide risk assessment have certain human subjective judgment factors and are not suitable for landslide risk assessment in a large range (such as national scale). Nowadays, machine learning methods are widely applied in the field of landslide geological disasters, including methods such as logistic regression, support vector machine, artificial neural network, bayesian network and the like. The method is mainly characterized in that a landslide risk assessment model is constructed on the basis of a series of inducing factors such as geology, landform and rainfall of a landslide, belongs to landslide risk assessment of the past time state, and cannot assess and analyze future landslide risk; meanwhile, the rainfall in the artificial landslide is taken as a static factor, such as the average annual rainfall is taken as a rainfall factor, and the influence of the time change of the rainfall factor on the landslide risk is not considered. In some existing landslide risk assessment methods, time dynamic changes of rainfall factors are considered, a landslide susceptibility index is calculated by using a machine learning method, then a rainfall threshold (the rainfall reaches a certain lower limit, the landslide occurrence probability is high) or a former-stage rainfall index is obtained by using a statistical model method according to real-time rainfall data, and the former-stage rainfall index and the landslide susceptibility index are combined into a dynamic disaster matrix, so that near-real-time dynamic assessment of landslide risk is achieved. However, the method has the following two disadvantages:
(1) Single dimension rainfall analysis: real-time precipitation data are utilized to estimate the time probability of future occurrence of landslide, and a dynamic disaster matrix is formed by the real-time precipitation data and the landslide probability, so that landslide risk is evaluated. Although the change of the rainfall factor along with time is considered in the method, the near-real-time assessment effect is achieved to a certain degree, but the influence of the comprehensive effect (the time-space integral change effect) that the rainfall factor changes along with time and geographic space on the landslide risk is ignored;
(2) Time uncertainty of rainfall threshold calculation: in the method for evaluating the landslide hazard based on the rainfall threshold, in the process of calculating the rainfall threshold, the influence on the stability of the landslide only when the rainfall reaches a certain lower limit is ignored, and in addition, different landslide time prediction results can be generated when the rainfall threshold is calculated at different time values, so that the landslide hazard analysis result has uncertainty and the evaluation effect is influenced.
Disclosure of Invention
Aiming at the problems, the invention improves the traditional Transformer method and provides a landslide space-time risk assessment method combined with an effective rainfall model, the method introduces the effective rainfall model into a Transformer network to extract effective rainfall factor data based on the originally input rainfall time sequence data, on the basis, a convolutional neural network CNN and a bidirectional long-time memory neural network Bi-LSTM are embedded, and a new space-time attention mechanism module is provided to replace the original attention mechanism module, so that the time and space characteristics of rainfall are fused, and the rainfall prediction precision is improved.
In order to solve the technical problems, the invention adopts the following technical scheme:
a landslide space-time risk assessment method combined with an effective rainfall model comprises the following steps:
step 2, constructing a landslide susceptibility model by using three mixed deep learning networks, predicting the spatial probability of landslide and obtaining a landslide spatial probability map;
step 3, embedding an effective rainfall model, a CNN convolutional neural network and a Bi-LSTM network to construct an ST-Transformer deep learning network based on the deep learning Transformer network, and predicting different reproduction periods T of the landslide, namely the possibility of the landslide occurring in different time in the future;
and 4, quantitatively calculating the landslide hazard index of the future optimal time T value by combining the space probability and the landslide time probability of the slope acquired in the step 3 through an improved landslide hazard calculation formula.
Further, the landslide induction factors comprise terrain factors, geology factors, landform factors, hydrology factors and road factors, and the obtained landslide induction factors are processed into a uniform spatial resolution and a uniform spatial coordinate system.
Further, the step 1 comprises the following sub-steps:
step a1, obtaining a landslide induction factor;
step a2, constructing landslide and non-landslide binary data;
a3, analyzing the relation between the landslide induction factor and the landslide;
further, in the step a2, constructing landslide and non-landslide binary data, specifically implementing the following substeps:
a2.1, acquiring regional landslide data;
and a2.2, acquiring non-landslide data which is equal to the landslide data, and constructing landslide binary data.
And further, extracting the attribute of the acquired landslide induction factor based on the acquired landslide and non-landslide two-classification data points, and finally acquiring landslide sample data which comprises landslide two-classification coded data and the induction factor attribute data of the corresponding landslide and non-landslide points.
Further, the step a3 of analyzing the relationship between the landslide induction factor and the landslide includes the following sub-steps:
a3.1, analyzing the correlation among the landslide induction factors by adopting a co-linear test method for the obtained landslide sample data, and eliminating the landslide induction factors with co-linear relation;
a3.2, calculating the importance indexes of the independent landslide induction factors by using an information gain analysis method, removing the factors of which the importance indexes are less than 0.1, and performing descending arrangement on the obtained new landslide induction factors according to the importance indexes, namely obtaining new ordered landslide induction factors without co-linearity, and dividing the new ordered landslide induction factors into a training set and a testing set by combining corresponding landslide binary data.
Further, the step 2 comprises the following sub-steps:
step b1, extracting main characteristics based on the CNN network:
and b2, constructing a landslide incidence model by the RNN recurrent neural network, predicting the spatial probability of landslide and obtaining a landslide spatial probability map.
Further, in step b2, the following is specifically implemented:
based on the CNN network training result in the step b1, obtaining independent and ordered landslide induction factors as input data of an RNN network, and outputting data which are landslide two-classification coded data to construct a landslide susceptibility model, wherein the RNN comprises three recurrent neural networks LSTM, GRU and SRU, and the characteristic result extracted by the three RNN networks is R cnn As input data of the RNN network, the calculation process after entering the hidden layer is as follows:
RNN t ＝f RNN (w*R cnn +u*RNN t-1 +b RNN )
O t ＝g RNN (v*RNN t )
where RNN represents the three variant networks, i.e., RNN = { LSTM, GRU, SRU }, f RNN ，g RNN Activation functions of a hidden layer and an output layer respectively, u, v, w are network parameters, b is a bias parameter, RNN t Output result of the hidden layer for time t, O t The landslide characteristics extracted by the CNN network enter the three RNN networks to be trained to obtain landslide space probability indexes and generate three landslide space probability graphs as the result of the RNN network output layer
Further, the step 3 comprises the following sub-steps:
step c1, embedding an effective rainfall model in an ST-Transformer network to calculate effective rainfall capacity of input rainfall data in different periods so as to obtain effective rainfall time sequence data in different periods;
step c2, embedding a CNN and a Bi-LSTM network in an Encoder of a transform network, firstly extracting main characteristic information of effective rainfall through a CNN convolutional neural network, embedding the Bi-LSTM network in a position encoding structure on the basis of the extraction of the main characteristic information of the effective rainfall, extracting spatial characteristic information of effective rainfall data, and then extracting time characteristics of the effective rainfall by combining one-hot encoding and the CNN convolutional neural network;
and c3, providing a space-time attention mechanism, fusing space-time characteristics of the extracted space-time characteristics of the effective rainfall data to be used as input data of a decoding layer, predicting the time probability of the landslide, evaluating a prediction result by adopting three error indexes, namely an average absolute error, a mean square error and a root mean square error, and selecting a time T value with the minimum error index as the optimal time for predicting the landslide time probability.
Further, the specific implementation comprises the following steps:
d1, obtaining the landslide space probability maps respectively in the step 2 and the step 3
And temporal probability map +>
Are divided into the same C categories, respectively, wherein C = {1,2,3, ·. };
d2, quantitative calculation by FR frequency method
And &>
Divided C classes of faciesAnd calculating the FR value of each category, namely the weighted value, according to the proportion of the pixels occupied by each category and the proportion of the pixels occupied by the corresponding landslide points, wherein the calculation formula is as follows:
in the formula, FR i Weight value, N, representing class i i Representing the number of landslide grid elements of the ith category,
representing the total number of all landslide grid pixels in the research area; p i Is the number of the grid pixels of the ith category, is->
Representing the total number of all grid pixels covering the study area;
d3, calculating the landslide risk of the optimal time period T in the future by using a risk formula so as to predict the future landslide risk and the landslide risk L HH The calculation formula is as follows;
in the formula, A i ＝{A 1 ，...A C Is as
The weight size, B, assigned to each level classification i ＝{B 1 ，...B C Is } is>
Each level classification is assigned a weight magnitude.
Compared with the prior art, the method has the following beneficial effects:
the invention provides a novel landslide risk assessment method, which mainly comprises the following three contents: the method comprises the following steps of (1) a landslide incidence evaluation method, (2) a landslide time probability prediction method and (3) a landslide risk index calculation method. The landslide susceptibility evaluation method utilizes three mixed neural networks CNN-LSTM, CNN-GRU and CNN-SRU to carry out landslide susceptibility calculation, namely landslide space probability calculation, and is superior in performance and high in model precision. The landslide time probability prediction method is characterized in that an ST-Transformer method is adopted to predict effective rainfall in the future based on rainfall data sets at different times, the method is based on a traditional Transformer network framework, an effective rainfall model is embedded, a CNN and a Bi-LSTM network are combined with position coding, time dimension calculation is integrated, rainfall spatio-temporal characteristics are obtained finally, spatio-temporal characteristic information is fused by utilizing a proposed spatio-temporal attention mechanism, effective rainfall at different times T in the future is predicted finally, and the time probability of landslide is analyzed. And selecting the prediction time T with the minimum result error index as the time probability of landslide evaluation. Based on the effective rainfall prediction result of the landslide incidence and the optimal prediction time T, the landslide risk of the future time T is calculated by using an improved landslide risk formula, and then a landslide risk prediction grading graph of the future time T is generated. The method provided by the invention has the advantages of less errors and better robustness.
Drawings
FIG. 1 is a block diagram of a landslide hazard analysis method according to the present invention.
Fig. 2 is a landslide induction factor graph selected according to an embodiment of the present invention, (a) a slope direction, (b) a slope gradient, (c) a slope height, (d) a slope length, (e) a plane curvature, (f) a section curvature, (g) a ground surface cutting depth, (h) an elevation variation coefficient, (i) a ground roughness, (j) a terrain humidity index, (k) a water flow intensity index, (l) a relative slope position, (m) a drainage basin area, (n) a drainage basin slope gradient, (o) a drainage basin length, (p) a valley depth, (q) a relative river network distance, (r) a terrain relative convergence index,(s) a terrain landslide point index, (t) a terrain surface curvature, (u) a terrain surface texture, (v) a stratigraphic structure, (w) a relative fault distance, (x) a soil type, (y) a ground surface peak acceleration, (z) a land utilization type, and (aa) a normalized difference vegetation index.
Fig. 3 is a landslide susceptibility graph generated by using CNN-LSTM, CNN-GRU and CNN-SRU methods in the embodiment of the present invention, (a) the landslide susceptibility graph generated by using CNN-LSTM, (b) the landslide susceptibility graph generated by using CNN-GRU, and (c) the landslide susceptibility graph generated by using CNN-SRU.
FIG. 4 is a diagram of rainfall prediction in the future year using the ST-Transformer method, the Transformer, the CNN-LSTM and the LSTM according to the embodiment of the present invention, (a) a diagram of rainfall predictions in the future year generated by the ST-Transformer method, (b) a diagram of rainfall predictions in the future year generated by the Transformer method, (c) a diagram of rainfall predictions in the future year generated by the CNN-LSTM method, and (d) a diagram of rainfall predictions in the future year generated by the LSTM method.
FIG. 5 is a graph of the prediction of rainfall effectivenesses of 3 months into the future using the ST-Transformer method, transformer, CNN-LSTM and LSTM in accordance with the embodiment of the present invention, (a) a graph of the prediction of rainfall effectivenesses of 3 months into the future generated by the ST-Transformer method, (b) a graph of the prediction of rainfall effectivenesses of 3 months into the future generated by the Transformer method, (c) a graph of the prediction of rainfall of 3 months into the future generated by the CNN-LSTM method, and (d) a graph of the prediction of rainfall of 3 months into the future generated by the LSTM method.
FIG. 6 is a graph showing the prediction of rainfall 1 month into the future using the ST-Transformer method, the Transformer, the CNN-LSTM and the LSTM according to the embodiment of the present invention, (a) a graph showing the prediction of rainfall 1 month into the future generated by the ST-Transformer method, (b) a graph showing the prediction of rainfall 1 month into the future generated by the Transformer method, (c) a graph showing the prediction of rainfall 1 month into the future generated by the CNN-LSTM method, and (d) a graph showing the prediction of rainfall 1 month into the future generated by the LSTM method.
Fig. 7 is a landslide hazard map created by combining a landslide susceptibility map and a rainfall forecast map in different periods generated by using a CNN-LSTM method according to an embodiment of the present invention, (a) a landslide hazard map for one year in the future is predicted, (b) a landslide hazard map for 3 months in the future is predicted, and (c) a landslide hazard map for 1 month in the future is predicted.
Fig. 8 is a landslide hazard map created by combining a landslide susceptibility map and a rainfall forecast map in different periods generated by using a CNN-GRU method according to an embodiment of the present invention, (a) a landslide hazard map for one year in the future is predicted, (b) a landslide hazard map for 3 months in the future is predicted, and (c) a landslide hazard map for 1 month in the future is predicted.
Fig. 9 is a landslide hazard map created by combining a landslide susceptibility map and a rainfall forecast map in different periods generated by using the CNN-SRU method according to the embodiment of the present invention, (a) a landslide hazard map for one year in the future is predicted, (b) a landslide hazard map for 3 months in the future is predicted, and (c) a landslide hazard map for 1 month in the future is predicted.
Detailed Description
The technical solution of the present invention is described in detail below with reference to the accompanying drawings and examples.
The embodiment of the invention provides a novel landslide risk assessment method (figure 1), which is mainly divided into three aspects: the method comprises the steps of landslide incidence calculation, landslide time probability calculation and landslide hazard calculation. (1) landslide susceptibility calculation: the landslide space induction factors without co-linear relation and with high importance index are selected through multiple co-linear inspection and information gain analysis based on landslide space induction factors such as terrain, landform, geology and hydrology, landslide sample data are combined to form landslide sample data, and the landslide sample data are divided into a training set of 70% and a testing set of 30%. Constructing a landslide susceptibility model by using three mixed neural networks of CNN-LSTM, CNN-GRU and CNN-SRU according to a training set, verifying by using a test set, evaluating the model precision by using four indexes of ACC, AUC, KAPPA and MCC, and finally calculating the space probability of landslide occurrence to generate a landslide susceptibility graph
(2) And (3) calculating the probability of landslide time: on the basis of an original transform framework, a convolutional neural network CNN and a bidirectional neural network Bi-LSTM model are introduced, and spatial feature extraction is performed in combination with position encoding (Positional Embedding) of the original transform network, and meanwhile, a new space-time attention mechanism is provided to fuse time and spatial feature information, so that the effect of improving the time sequence predictive performance is achieved. The method utilizes rainfall data of a research area to predict rainfall of different periods in the future so as to predict the time probability of landslide occurrence, and finally generates a landslide time probability graph/>
Evaluating rainfall prediction accuracy by adopting three MSE, MAE and RMSE error indexes; (3) landslide risk calculation: and (3) quantitatively calculating the future landslide risk through a proposed landslide risk formula based on the results obtained in the steps (1) and (2). Taking the local area of the United states as a research area, the method comprises the following steps:
and 1, analyzing the relation between the landslide induction factor and landslide occurrence based on the obtained landslide induction factor. This step can be subdivided into three substeps:
step a1, obtaining a landslide induction factor;
in step a1, relevant landslide induction factors from various sources are obtained, which are mainly classified into terrain, geology, landform, hydrological condition data and the like, and relevant descriptions thereof are as follows:
through digital Terrain Elevation data (USGS, 2005) with spatial resolution of 30 meters (https:// earth-height plorer. USGS. Gov /), slope (Slope), aspect (Aspect), plane Curvature and section Curvature (Plan and Profile Curvature), terrain Relief (Relief), surface Cutting Depth (Surface Cutting Depth, SCD), elevation variation coefficient (Elevation variation coefficient, EVC), ground Roughness (Terain Roughness), terrain humidity Index (Topographic Wetness Index, TWI), water flow intensity Index (Stream Power Index, SPI) are extracted using ArcGIS 10.5; meanwhile, SAGA gis2.3.2 (survey et al 2015) software is used for extracting a river basin Area (catch Area), a river basin Slope (catch Slope), a Flow passage Length (Flow Path Length, FPL), a Valley Depth (Valley Depth), a Relative river Network Distance (Distance to Channel Network, DTCN), a terrain Relative Convergence Index (terrain Relative Index, TRCI), a Relative Slope Position (Relative Slope Position, RSP), a terrain landslide point Index (terrain Position Index, TPI), a terrain surface curvature (Convergence), a terrain surface Texture (Texture), a Slope Height (Slope Height), a Slope Length (Slope Length), and morphological Features (Slope Features);
geological factors: the method mainly comes from the American geological survey bureau (https:// earth x plorer.usgs.gov /), contains formation lithology and fault data, and then uses ArcGIS10.5 software to perform rasterization processing on the formation data, and uses Jenks natural interruption method (Chen et al.2013) to perform grade classification to obtain a formation grid grading diagram, and simultaneously obtains a fault Distance grade grading Diagram (DTF) based on Euclidean Distance analysis;
landform factors: MODIS data are downloaded from an American earth data center (https:// search. Earth data. Nasa. Gov /), MODQ13 data in a global range are obtained, an HEGtool is adopted to convert an image into raster data, namely Normalized Differential Vegetation Index (NDVI), and finally an NDVI map of a city or a county is obtained by utilizing a mask extraction and analysis function of ArcGIS; obtaining Land cover utilization vector data of a local area of the United states from a national geographic information resource directory service system (https:// www.webmap.cn/main.domethod = index), extracting Land cover utilization data of a research area of the Water City county by using ArcGIS10.5 software, and converting the data into a grid form, namely obtaining a Land cover utilization grid map (Land cover) of the Water county; the method comprises the steps of downloading soil data and land classification data of an American native area from an agricultural organization website (http:// www. Fao. Org/soil-portal/data-hub/soil-maps-and-data bases/soil-profile-data bases/en /) of the United nations, then extracting soil classification data of a water city county area by using ArcGIIS10.5 software, and converting the soil classification data into a grid form, namely acquiring a water city county soil classification grid grading map; carrying out rasterization processing on the land classification data in the same way to obtain a land grid grading map (Landform);
hydrological and road class factors: acquiring Water system and traffic Road data of the American area from a global Road Water system database (https:// www.globio.info/download-grip-dataset), acquiring Water flow data and traffic Road data of the American local area by using ArcGIS10.5 software, and performing Euclidean Distance analysis to acquire a Distance To Water (DTW) and a Distance To Road (DTR) grading map;
in summary, the present embodiment obtains 33 landslide induction factors native to the united states, which are processed into a grid map with uniform spatial coordinate system (WGS 1984) and spatial resolution (30 m × 30 m) using arcgis 10.5.
Step a2, constructing landslide and non-landslide binary data;
in step a2, the specific implementation comprises the following steps:
(1) The invention obtains 64,356 landslide points of a local area of the United states from 1900 to 2019 from a landslide database (https:// www.science base. Gov/catalog/item/5c7065b4e4b0fe48cb43fbd 7) and a global landslide database (https:// maps. Nccs. Na. Go/arcgis/apps/MapAndAppApplery/index. Htmlappid =574f26408683485799d02e857e5d 9521);
(2) To balance the positive and negative sample ratios, an equal amount of non-landslide point data needs to be generated. In the invention, in an area outside the range of 8,000km radius (not influenced by the landslide point outside the distance) of the existing landslide point, arcGISI 10.5 is utilized to randomly generate non-landslide data equal to that of a landslide sample, then the landslide and non-landslide points are respectively coded into 1 and 0, finally 128,712 landslide classification points are obtained, then the attribute value of the landslide induction factor is extracted based on the landslide classification sample points, and finally complete landslide sample data is obtained, wherein the complete landslide sample data comprises the attribute of the relevant landslide induction factor and the landslide and non-landslide binary data.
A3, analyzing the relation between a landslide induction factor and a landslide;
in step a3, the specific implementation comprises the following steps:
(1) And (3) detecting the correlation of the landslide induction factors by adopting a multiple collinear method. VIF (variance inflation factor) and TOL (tolerance) are two evaluation indicators for the results of the collinearity analysis, indicating that no correlation exists between the variables if VIF <10, TOL >0.1 or VIF <5, TOL >. As can be seen from Table 2, the VIF of the Relief (Relief) is greater than 10 and the TOL is close to 0, so this factor is not considered as a landslide induction factor;
(2) On the basis, the importance index analysis is carried out on the new landslide induction factor by using an Information Gain method, IG (Information Gain) is used for representing the importance index, and the closer the index is to 1, the higher the relevance between the index and the landslide is. In table 2, the IG indexes of Landforms, morphometric Features, DTR and DTW are close to 0, indicating that these indexes have little influence on the occurrence of landslide, so the present invention removes the collinearity relationship and the landslide induction factors with very low importance Index, and finally obtains 27 induction factors with high importance without collinearity relationship (fig. 2), namely, the Slope direction (Aspect) (fig. 2 a), the Slope (Slope) (fig. 2 b), the Slope Height (Slope Height) (fig. 2 c), the Slope Length (Slope Length) (fig. 2 d), the plane Curvature (Plan Curvature) (fig. 2 e), the Profile Curvature (Profile Curvature) (fig. 2 f), the Surface Cutting Depth (Surface Cutting Depth, SCD) (fig. 2 g), elevation Variation Coefficient (EVC) (fig. 2 h), ground Roughness (Terrain Roughness, TR) (fig. 2 i), terrain humidity Index (Topographic weather Index, TWI) (fig. 2 j), water Flow intensity Index (Stream Power Index, SPI) (fig. 2 k), relative Slope Position (RSP) (fig. 2 l), basin Area (catch Area) (fig. 2 m), basin Slope (catch Slope) (fig. 2 n), flow Path Length (FPL) (fig. 2 o), valley Depth (Valley Depth) (fig. 2 p), relative river Network Distance (Distance to Channel, DTCN) (fig. 2 q), terrain Convergence Relative Index (Terrain contrast Index, TRCI) (fig. 2 r), terrain landslide point Index (TPI) (fig. 2 s), terrain surface Convexity (roughness) (fig. 2 t), terrain surface Texture (Texture) (fig. 2 u), stratigraphic Lithology (Lithology) (fig. 2 v), distance To Fault (DTF) (fig. 2 w), soil Type (Soil Type) (fig. 2 x), ground Peak Acceleration (Peak Ground Acceleration, PGA) (fig. 2 y), land use Type (Land use) (fig. 2 z), normalized differential Vegetation Index (Normalized Difference Vegetation Index, NDVI) (fig. 2 aa).
TABLE 1 results of co-linearity test and significance analysis of landslide induction factors
Step 2, constructing a landslide susceptibility model by using three mixed deep learning networks, calculating the spatial probability of landslide and obtaining a landslide spatial probability map;
the step 2 comprises the following substeps:
step b1, extracting main characteristics based on the CNN network:
and b2, constructing a landslide incidence model by the RNN recurrent neural network, predicting the spatial probability of landslide and obtaining a landslide spatial probability map.
Further, in step b2, the following is specifically implemented:
based on the CNN network training result in the step b1, obtaining independent and ordered landslide induction factors as input data of an RNN network, and outputting data which are landslide two-classification coded data to construct a landslide susceptibility model, wherein the RNN comprises three recurrent neural networks LSTM, GRU and SRU, and the characteristic result extracted by the three RNN networks is R cnn As input data of the RNN network, the calculation process after entering the hidden layer is as follows:
RNN t ＝f RNN (w*R cnn +u*RNN t-1 +b RNN ) (1)
O t ＝g RNN (v*RNN t ) (2)
where RNN represents the three variant networks, i.e., RNN = { LSTM, GRU, SRU }, f RNN ，g RNN Activation functions of a hidden layer and an output layer respectively, u, v, w are network parameters, b is a bias parameter, RNN t Output result of the hidden layer for time t, O t The landslide characteristics extracted by the CNN network enter the three RNN networks to be trained to obtain landslide space probability indexes and generate three landslide space probability graphs as the result of the RNN network output layer
In step 2, the obtained landslide induction factors are arranged in a descending order according to the importance index to accord with the ordered input data structure of three mixed networks of CNN-LSTM, CNN-GRU and CNN-SRU, and the output data is classified data of landslide and non-landslide, so that the spatial probability of landslide is calculated. The landslide susceptibility results of the three hybrid deep learning networks are shown in fig. 3, the model accuracy is evaluated by four indexes of AUC, ACC, KAPPA and MCC, the closer the evaluation index is to 1, the higher the model prediction performance is, and the model accuracy evaluation results are shown in table 2.
TABLE 2 evaluation results of CNN-SRU, CNN-LSTM and CNN-GRU hybrid network models
FIG. 3 is a landslide susceptibility spatial probability map generated by the three CNN-LSTM, CNN-GRU and CNN-SRU mixed deep learning networks. As can be seen from FIG. 3, the landslide tendency of the three mixing methods CNN-LSTM, CNN-GRU and CNN-SRU is the same, and the region of the local region of the United states where landslide is likely to occur is characterized by the aggregation of the east and west regions and the sporadic distribution of the middle region. Meanwhile, the landslide susceptibility area in the west area is larger than that in the east area, and tends to spread to the east from the middle area, wherein the area of the area range corresponding to the extremely high grade of landslide susceptibility is larger. FIGS. 3 (a) - (c) are spatial probability maps of landslide susceptibility generated using CNN-LSTM, CNN-GRU and CNN-SRU networks. It is understood that the range of the region with high grade of landslide susceptibility in fig. 3 (a) and 3 (c) is almost the same, but is higher than the grade region corresponding to landslide susceptibility in the landslide susceptibility map generated by CNN-GRU in fig. 3 (b), and the region with high grade of landslide susceptibility in fig. 3 (c) is slightly larger than the region corresponding to fig. 3 (a).
Table 2 shows the evaluation results of four indicators of the Area of the working characteristic Curve (Area Under the Curve, AUC), accuracy (Accuracy, ACC), KAPPA coefficient (KAPPA) and Mohs Correlation Coefficient (MCC) of the subject. On the whole, the ACC indexes of the three methods are the same, and the AUC indexes are higher than 96%, which shows that the landslide tendency prediction performances of the three models are good. The KAPPA and MCC indexes of the CNN-SRU and the CNN-LSTM are the same, and the AUC index of the CNN-SRU method is the highest and is 0.968, which shows that in the three landslide susceptibility models, the landslide susceptibility model constructed by the CNN-SRU method has better prediction performance, and is consistent with the landslide susceptibility grades described in the graphs (a) - (c) of fig. 3, namely, the region range with high landslide susceptibility grade in the landslide susceptibility graph generated by the CNN-SRU network model is higher than the corresponding region range in the CNN-LSTM and CNN-GRU graphs.
Step 3, embedding an effective rainfall model, a CNN convolutional neural network and a Bi-LSTM network to construct an ST-Transformer deep learning network based on the deep learning Transformer network, and predicting different reproduction periods T of the landslide, namely the possibility of the landslide occurring in different time in the future; the step 3 comprises the following substeps:
step c1, embedding an effective rainfall model in an ST-Transformer network to calculate effective rainfall capacity of input rainfall data in different periods so as to obtain effective rainfall time sequence data in different periods;
step c2, embedding a CNN and a Bi-LSTM network in the coding structure, firstly extracting main characteristic information of effective rainfall through a CNN convolutional neural network, on the basis, embedding the Bi-LSTM network in the position coding structure, extracting spatial characteristic information of effective rainfall data, and then extracting time characteristics of the effective rainfall by combining one-hot coding and the CNN convolutional neural network;
and c3, providing a space-time attention mechanism, performing space-time feature fusion on the extracted space-time features of the effective rainfall data, using the space-time features as input data of a decoding layer, predicting the time probability of landslide, evaluating a prediction result by adopting an average absolute error, a mean square error and three error indexes, and selecting a time T value with the minimum error index as the optimal time for predicting the landslide time probability.
In step c1, the present invention predicts the time probability of occurrence of landslide from rainfall data. The data are derived from CHIRPS global rainfall data (https:// www.chc.ucsb.edu/data/chirp /), the latitude and longitude range of which is 50N-50S,180W-180E, the spatial resolution is 0.05 degree multiplied by 0.05 degree unit grid, and the three data comprise data of three different periods of year, season and month, all of which start from 1981, wherein the annual rainfall data comprise time span of 39 years, and the seasonal rainfall and the monthly rainfall time respectively end at the third quarter of 2021 and 9 months of 2021. On the basis, rainfall data of the local area of the United states is selected to serve as a training sample for calculating the probability of occurrence time of the future landslide. In the proposed ST-transformer network, the first step is to introduce an effective rainfall model to obtain effective rainfall time series data of different time periods at different time periods based on the input rainfall data, as the input data in step b 2.
In step c2, CNN and Bi-LSTM networks are introduced, position Encoding (Positional Encoding) is combined, spatial features of rainfall are extracted based on rainfall data, time sequence prediction is carried out by using the ST-transducer network model provided by the invention, all experiments are carried out in NVIDIA GeForce RTX2060 computer environment, the model is based on an adaptive motion optimizer, 300 batches of data size (batch _ size) are set by using Mean Square Error (MSE), and 30 times of iterative (epochs) training is carried out. The invention predicts the effective rainfall of 1 year, 1 month and 3 months in the future according to the rainfall data of the three periods, and compares the rainfall data with three deep learning time sequence prediction network models, namely a traditional Transformer network (Transformer), a CNN-LSTM network and an LSTM network. In order to evaluate the prediction accuracy of the model, three evaluation indexes are adopted: the Mean Square Error (MSE), root Mean Square Error (RMSE), and Mean Absolute Error (MAE) were evaluated using a model, and the results are shown in table 3. FIG. 4, FIG. 5, and FIG. 6 are the ST-Transformer network model and other three transformers, CNN-LSTM, and LSTM models proposed by the present invention, respectively, for predicting the effective rainfall in the future of 1 year, 3 months, and 1 month, and based on the effective rainfall value, generating effective rainfall maps of different prediction periods by using ArcGIIS10.5. Similar to the landslide susceptibility grading graph, different effective rainfall ranges in the rainfall graph generated by each model based on different prediction time represent the effective rainfall of different areas, and the grading range from low to high represents that the effective rainfall is higher and higher.
The specific process of step c2.1 is as follows:
although the rainfall is landslideExternal triggers occur, but only when effective rainfall conditions are reached will cause landslide to occur. The effective rainfall is the part of the rainfall causing the landslide, namely the effective rainfall is a key factor for inducing the landslide, so that rainfall data input initially in an ST-Transformer network is converted into effective rainfall time sequence data through an effective rainfall model structure introduced into an encoding structure to construct an effective rainfall time sequence data set required by landslide time probability prediction. The effective rainfall model calculation process is shown as a formula (3), wherein R b Representing effective rainfall, a represents the time of days before landslide occurs, R a The rainfall of the previous a days and K is the effective rainfall coefficient.
Suppose that a certain landslide point in the United states with a daily rainfall attribute is denoted as r n ，n∈N 0 ，N 0 To study the number of regional landslide points, based on the landslide point r n The obtained effective rainfall time sequence data of the R sequence lengths is represented as R b (r n R ) The calculation process of the rainfall efficiency model using the above equation (3) is shown in (4) and (5). In the formula
Indicates a point of landslide r n Effective rainfall time sequence data of ith T sequence length, r n i Is a landslide point r n Rainfall on day i.
Effective rainfall data table calculated by the effective rainfall model through different rainfall time T at the landslide pointShown as
Represents N 0 Each landslide point has time series data of R-dimensional effective rainfall attributes. The data set may be classified as->
Wherein
RD T ∈RD R Spatial and temporal characteristics, N, representing effective rainfall, respectively 0 And R represents the number of landslide points with an effective rainfall attribute and the time-series dimension of the effective rainfall signature, respectively. For the space vector RD obtained after passing through the input layer S And a time vector RD T Respectively expanded as->
And &>
Three-dimensional features, and spatial feature extraction is performed based on the data, specifically described as follows:
the spatial feature extraction network structure based on the effective rainfall data is position coding embedded with a Bi-LSTM network. A CNN convolutional neural network is added before position coding of the embedded Bi-LSTM network, and the network layer is used for realizing the feature extraction of main effective rainfall of effective rainfall time sequence data. RD based on the above acquisition S And RD T The extraction process of the main characteristic information of the effective rainfall of the CNN convolution neural network is represented as D S ＝f CNN (x S ，RD S ，RD T ) Wherein f is CNN Is a1 × 1 convolutional layerThe structure is the space characteristic RD of the effective rainfall data S And a time vector RD T Connected feature vector
Conversion into a fixed->
Feature vector of individual dimension->
Therefore, the spatial characteristic information extraction precision of the effective rainfall data is improved. D S The feature vector is used as input data for spatial feature extraction.
In order to learn the spatial characteristics of the rainfall time series data more efficiently, a Bi-LSTM network is embedded in a position coding structure, and a Bi-LSTM model consists of two LSTM network structures which fully consider the front and back information of the time series data. In the application of forecasting the effective rainfall time sequence data, the bidirectional characteristic information of the effective rainfall data is extracted in consideration of the front and back change information of the effective rainfall time sequence data, so that the precision of landslide time probability forecasting is improved. The position code Positional Encoding embedded in the Bi-LSTM network obtains the characteristics that
Meaning having pick/place in each time step>
Individual dimension effective rainfall spatial feature vector, POS t And bi _ lstm t Respectively a position-coding layer and a Bi-LSTM network layer function.
The space characteristic vector obtained by the position coding layer embedded into the Bi-LSTM network is P S Three subvector spaces Q added to the initial input data transformation S ，k S ，v S The calculation process is as follows:
in the formula
Represents N 0 Spatial attention characterization results of effective rainfall obtained at each landslide point, A S A spatial characteristic weight value representing a valid rainfall, softmax being an activation function of the spatial attention mechanism, ->
The scaling parameter is used for preventing an overlarge dot product operation result in the calculation of the spatial characteristic weight value of the effective rainfall. The above is in the single attention mechanismThe spatial attention result obtained in (1). Computing N using a multi-head attention mechanism model using a position code with an embedded Bi-LSTM network and a plurality of different potential subspaces 0 Different spatial attention characteristics of the individual landslide points, i.e. dynamic characteristics based on the time variation of the effective rainfall at the landslide points.
The specific process of step c2.2 is as follows:
the time data set of the effective rainfall based on the landslide point is expressed as
Through one-hot coding pair RD T The temporal data is encoded as temporal feature vectors. In view of the powerful feature extraction potential of the convolutional neural network CNN, the CNN network is adopted to extract the time features of effective rainfall data. And the effective rainfall time data set enters a CNN convolutional neural network after converting a time vector through one-hot coding, and a1 x 1 convolutional neural network is selected to extract the characteristic information of the effective rainfall time sequence data. Converting the time characteristic vector into a time characteristic vector with the same dimension as the space vector by using a one-dimensional convolution layer, wherein the calculation process is as follows:
D T ＝Cov1D(one_hot(RD T ) (9)
in the above formula, one _ hot (RD) T ) Represents the time characteristic of the rainfall in effect with thermal coding, cov1D represents the one-dimensional convolutional layer function of the convolutional neural network CNN of 1 × 1.
Similar to the extraction of the spatial attention feature of the effective rainfall, the obtained effective rainfall time feature vector is added to three sub-vectors of the original data obtained by an attention mechanism
Three new temporal feature vectors in the time space are obtained, and the calculation process is as follows:
since the effective rainfall amount of the future T days is predicted based on the R pieces of historical time data, namely the effective rainfall time sequence data based on the initial landslide point is RD T ＝{RT 1 ，RT 2 ，......RT R Predicting the rainfall of the future T time step, and calculating the following process:
RT R+T ＝f E (T 1 ，RT 2 ，......RT R ) (13)
it is known that the prediction of the effective rainfall amount based on the landslide point depends on the rainfall data of the entire historical time series length and the effective rainfall dependency characteristics before and after the time. Therefore, in order to obtain the effective rainfall time characteristics before and after the long-time sequence and the time, in the same way as the space attention mechanism, the multi-head attention mechanism is adopted to obtain the time dependence characteristics before and after the long-time sequence, so that the prediction of the effective rainfall is improved.
The specific process of step c2.3 is as follows:
considering that the sigmoid activation function is one of the commonly used activation functions in the machine learning network, the role of the sigmoid activation function is to calculate the spatial attention characteristic O of the effective rainfall by using the activation function sigmoid between 0 and 1 through logarithmic value scaling S And temporal attention feature O T The weight value of (3). In the formulae (14), (15), wherein g X ，g T The function of the linear function is to convert the space attention characteristic and the time attention characteristic of the effective rainfall into a one-dimensional characteristic vector linearly to obtain a weighted value W of the space-time attention characteristic of the effective rainfall R And calculating based on the weight value to obtain space-time fusion characteristic of effective rainfall
W R ＝sigmoid(g X (O S )+g T (O T )) (14)
Spatiotemporal fusion features of effective rainfall
Outputting the final effective rainfall feature extraction result in the ST-Transformer coding structure through a full connection layer>
Represents N 0 D of each landslide point in R time steps st And the space-time dimension characteristic vector is used as input data of a Decoder network structure of a decoding layer, and effective rainfall of T time steps in the future is predicted, so that the time probability of landslide occurrence is predicted.
As can be seen from fig. 4 to 6, the effective rainfall prediction graph generated by the ST-Transformer network model provided by the present invention has a larger area with a high effective rainfall than the area ranges of the corresponding levels of the other three methods, which indicates that the effective rainfall time sequence prediction model provided by the present invention has a better performance. Meanwhile, the effective rainfall prediction grading diagrams generated by all experimental methods show the distribution characteristic of cluster aggregation, areas with high effective rainfall are mainly distributed in the west and east areas of the United states, meanwhile, the effective rainfall distribution in the east area is concentrated, the west area is relatively dispersed, and the effective rainfall in the middle area is far smaller than that in the east area. The effective rainfall prediction range of fig. 4 is much smaller than that of fig. 5 and 6, which shows that the effective rainfall prediction effect with a long period is lower than that with a short prediction period. From fig. 4 (a), fig. 5 (a), and fig. 6 (a) are respectively effective rainfall prediction grading graphs generated for the next 1 year, 1 month, and 3 months based on the ST-transformer method proposed by the present invention, and the region range of fig. 5 (a) where the effective rainfall is high is far larger than the corresponding region areas of fig. 4 (a) and fig. 5 (a), which illustrates that the prediction performance of the present invention method is more excellent for the prediction period of one month. Meanwhile, in fig. 4 to 6, the area range with high effective rainfall in the effective rainfall prediction grading diagram generated by the method of the present invention is higher than the area ranges displayed by the other three methods, which shows that the area with high effective rainfall in the effective rainfall timing prediction model proposed by the present invention is higher than the area ranges corresponding to the other three models in different prediction periods.
Table 3 shows the evaluation results of the method of the present invention and the other three methods based on different prediction periods and using three indexes of MSE, RMSE, and MAE. As can be seen from the table, in the four effective rainfall time sequence prediction models, the three evaluation indexes of the ST-Transformer model provided by the invention are far lower than those of the other three model methods, and the superiority of the method provided by the invention is reflected. Meanwhile, in the hierarchical graph generated by the method, the area range with high effective rainfall is far larger than the corresponding distribution areas of the other three methods, and the hierarchical graph is consistent with the description. The traditional Transformer method is higher than CNN-LSTM and LSTM methods, and the time sequence prediction performance of the Transformer is reflected.
TABLE 3 evaluation results of effective rainfall time series predictions using different methods
Step 4, an improved landslide hazard formula is provided, and the landslide hazard of different reappearance periods is quantitatively calculated by combining the two results obtained in the previous step; first, the method generates a landslide susceptibility map using the method described above
And a landslide time probability map>
Then the ArcGISs 10.5 Jenks natural interrupt method is used to->
And &>
Are divided into C categories, respectively.
In step 4, the specific implementation comprises the following steps:
d1 landslide space probability map respectively obtained in step 2 and step 3
And temporal probability map +>
The classification into the same C categories, wherein = {1,2, 3. - };
d2 quantitative calculation by FR frequency method
And &>
Divided intoAnd calculating the FR value of each category, namely the weight value, according to the proportion of the pixels occupied by each category and the proportion of the pixels occupied by the corresponding landslide points, wherein the calculation formula is as follows:
in the formula, FR i Weight value, N, representing class i i Representing the number of landslide grid elements of the ith category,
the total number of all landslide grid pixels in the research area is represented; p i Is the number of the grid pixels of the ith category, is->
Representing the total number of all grid pixels covering the study area;
d3, calculating the landslide risk of the optimal time period T in the future by using a risk formula so as to predict the future landslide risk and the landslide risk L HH The calculation formula is as follows;
in the formula, A i ＝{A 1 ，...A C Is as
The weight size, B, assigned to each level classification i ＝{B 1 ，...B C Is } is>
Each level classification is assigned a weight magnitude.
On the basis of obtaining a landslide incidence graph and an effective rainfall prediction grading graph, the landslide risk calculation method provided by the invention is adopted to evaluate the landslide risk of the American area. 3-6, the landslide susceptibility graph and the effective rainfall prediction grading graph are divided into 5 classes by adopting a Jenks natural interruption method of ArcGIS10.5, the weight of each class is calculated by FR in the invention, and finally the landslide risk index is obtained based on the superposition analysis of the ArcGIS10.5, namely the landslide susceptibility graph generated by three mixed neural networks CNN-RNN is utilized. FIGS. 7-9 show landslide susceptibility diagrams generated by superimposing effective rainfall prediction diagrams of different periods generated by the method of the invention on landslide susceptibility diagrams generated by three CNN-RNN methods to obtain American landslide hazard diagrams of different prediction times in the future. As can be seen from fig. 7 to 9, in each of the american landslide hazard maps at different prediction times, the regions corresponding to the eastern region higher levels or higher have large and many areas, the edges of the leftmost regions in the western region present fewer regions with higher landslide hazard, the eastern regions mainly include regions with medium and low landslide hazard, and the rest regions are regions with extremely low landslide hazard. Based on the landslide susceptibility graph generated by superposing the same CNN-RNN, the method disclosed by the invention is used for predicting that the area range with extremely high landslide risk in the landslide risk graph of one month in the future is higher than the corresponding grade area range in the landslide risk graphs of other landslide risk graphs of 1 year and 3 months in the future, and the landslide risk graph with a short prediction period is more accurate. According to a future effective rainfall map of the same prediction period (such as a month), three CNN-RNN mixed methods are superposed to generate a landslide susceptibility map, three landslide risk maps of the same prediction period in the future are generated, namely, a map 7 (b), a map 8 (b) and a map 9 (b), the landslide risk degrees shown in the three landslide risk maps are almost not different, and the landslide risk trends of other landslides in the next year and 3 months are the same, which indicates that rainfall is a main factor for landslide occurrence.
According to the specific implementation, the landslide space-time risk assessment method combined with the effective rainfall model provided by the invention mainly comprises the following three points:
(1) Although the traditional Transformer method can perform parallel computation efficiently, the data characteristics in the input attention mechanism neglect the dependency of the front and back data characteristics in time sequence prediction, and effective data characteristic learning cannot be performed;
(2) The invention provides a novel space-time attention mechanism, which is characterized in that time and space attention results of landslide are respectively calculated, then weight values of the time and space attention results are calculated according to the original self-attention mechanism, finally the time and space attention results of the landslide are subjected to superposition calculation by using the weight values, space-time attention results and instant space characteristic fusion results are obtained, and the influence of rainfall time variable and space attribute characteristic comprehensive change action on landslide danger in landslide space-time danger prediction is not considered.
(3) The method adopts an improved Transformer method to predict the effective rainfall capacity of different periods by utilizing rainfall data of long time span and combining an effective rainfall model, then quantitatively calculates each class weight with the landslide susceptibility (landslide space probability) in the time range and carries out superposition calculation, so as to evaluate the future landslide risk.
The above-described embodiments are merely examples provided to illustrate the invention in detail, and are not intended to limit the generality of the methods and examples.
The above embodiments are merely illustrative of the present invention and are not to be construed as limiting the invention. It will be appreciated that modifications and variations are possible to those skilled in the art in light of the above teachings, and it is intended to cover all such modifications and variations as fall within the scope of the appended claims.
Claims (10)
1. A landslide space-time risk assessment method combined with an effective rainfall model is characterized by comprising the following steps of:
step 1, selecting a landslide induction factor and analyzing the relation between the landslide induction factor and a landslide;
step 2, constructing a landslide susceptibility model by using three mixed deep learning networks, predicting the spatial probability of landslide and obtaining a landslide spatial probability map;
step 3, embedding an effective rainfall model, a CNN convolutional neural network and a Bi-LSTM network to construct an ST-Transformer deep learning network based on the deep learning Transformer network, and predicting different reproduction periods T of the landslide, namely the possibility of the landslide occurring in different time in the future;
and 4, quantitatively calculating the landslide risk index of the future optimal time T value by combining the space probability and the landslide time probability of the slope obtained in the step 3 through an improved landslide risk calculation formula.
2. The landslide space-time risk assessment method in combination with an effective rainfall model according to claim 1, wherein: the landslide induction factors comprise terrain factors, geology factors, landform factors, hydrology factors and road factors, and the obtained landslide induction factors are processed into a uniform spatial resolution and a spatial coordinate system.
3. The landslide space-time risk assessment method in combination with an effective rainfall model according to claim 1, wherein: the step 1 comprises the following substeps:
step a1, obtaining a landslide induction factor;
step a2, constructing landslide and non-landslide binary data;
and a3, analyzing the relation between the landslide induction factor and the landslide.
4. The landslide space-time risk assessment method in combination with an effective rainfall model according to claim 3, wherein: in the step a2, constructing landslide and non-landslide binary data, and concretely realizing the following substeps:
a2.1, acquiring regional landslide data;
and a2.2, acquiring non-landslide data which is equal to the landslide data, and constructing landslide binary data.
5. The landslide space-time risk assessment method in combination with an effective rainfall model according to claim 4, wherein: and extracting the attribute of the acquired landslide induction factor based on the acquired landslide and non-landslide two-classification data points, and finally acquiring landslide sample data which comprises landslide two-classification coded data and the induction factor attribute data of corresponding landslide and non-landslide points.
6. The landslide space-time risk assessment method in combination with an effective rainfall model according to claim 3, wherein: analyzing the relation between the landslide induction factor and the landslide in the step a3, wherein the concrete implementation comprises the following sub-steps:
a3.1, analyzing the correlation among the landslide induction factors by adopting a collinearity inspection method for the obtained landslide sample data, and eliminating the landslide induction factors with collinearity relation;
a3.2, calculating the importance indexes of the independent landslide induction factors by using an information gain analysis method, removing the factors of which the importance indexes are less than 0.1, and performing descending arrangement on the obtained new landslide induction factors according to the importance indexes, namely obtaining new ordered landslide induction factors without co-linearity, and dividing the new ordered landslide induction factors into a training set and a testing set by combining corresponding landslide binary data.
7. The landslide space-time risk assessment method in combination with an effective rainfall model according to claim 1, wherein: the step 2 comprises the following substeps:
step b1, extracting main characteristics based on the CNN network:
and b2, constructing a landslide incidence model by the RNN recurrent neural network, predicting the spatial probability of landslide and obtaining a landslide spatial probability map.
8. The landslide space-time risk assessment method in combination with an effective rainfall model according to claim 7, wherein: in step b2, the following is specifically implemented:
based on the CNN network training result in the step b1, obtaining independent and ordered landslide induction factorsThe RNN is used as input data of the RNN, the output data is landslide two-classification coded data, a landslide susceptibility model is constructed according to the landslide two-classification coded data, the RNN comprises three recurrent neural networks LSTM, GRU and SRU, and the characteristic result extracted through the three RNN networks is R cnn As input data of the RNN network, the calculation process after entering the hidden layer is as follows:
RNN t ＝f RNN (w*R cnn +u*RNN t-1 +b RNN )
O t ＝g RNN (v*RNN t )
where RNN represents the three variant networks, i.e., RNN = { LSTM, GRU, SRU }, f RNN ,g RNN Activation functions of a hidden layer and an output layer respectively, u, v, w are network parameters, b is a bias parameter, RNN t Output result of the hidden layer for time t, O t The landslide characteristics extracted by the CNN network enter the three RNN networks to be trained to obtain landslide space probability indexes and generate three landslide space probability graphs as the result of the RNN network output layer
9. The landslide space-time risk assessment method in combination with an effective rainfall model according to claim 1, wherein: the step 3 comprises the following substeps:
step c1, embedding an effective rainfall model in an ST-Transformer network to calculate effective rainfall capacity of input rainfall data in different periods so as to obtain effective rainfall time sequence data in different periods;
step c2, embedding a CNN and a Bi-LSTM network in an Encoder of a transform network, firstly extracting main characteristic information of effective rainfall through a CNN convolutional neural network, embedding the Bi-LSTM network in a position encoding structure on the basis of the extraction of the main characteristic information of the effective rainfall, extracting spatial characteristic information of effective rainfall data, and then extracting time characteristics of the effective rainfall by combining one-hot encoding and the CNN convolutional neural network;
and c3, providing a space-time attention mechanism, fusing space-time characteristics of the extracted space-time characteristics of the effective rainfall data to be used as input data of a decoding layer, predicting the time probability of the landslide, evaluating a prediction result by adopting three error indexes, namely an average absolute error, a mean square error and a root mean square error, and selecting a time T value with the minimum error index as the optimal time for predicting the landslide time probability.
10. The landslide space-time risk assessment method in combination with an effective rainfall model according to claim 1, wherein: in step 4, the specific implementation comprises the following steps:
d1, obtaining the landslide space probability maps respectively in the step 2 and the step 3
And temporal probability map +>
The method comprises the following steps of dividing the images into the same C categories, wherein C = {1,2,3, \8230; \8230 };
d2, quantitative calculation by FR frequency method
And &>
And calculating the FR value of each category, namely the weight value, according to the proportion of the pixels occupied by each category and the proportion of the pixels occupied by the corresponding landslide points, wherein the calculation formula is as follows:
in the formula, FR i Weight value, N, representing class i i Representing the number of landslide grid elements of the ith category,
the total number of all landslide grid pixels in the research area is represented; i number of grid pixels for the ith category>
Representing the total number of all grid pixels covering the study area;
d3, calculating the landslide risk of the optimal time period T in the future by using a risk formula so as to predict the future landslide risk and the landslide risk L HH The calculation formula is as follows;
in the formula, A i ＝{A 1 ,…A C Is as
The weight size, B, assigned to each level classification i ＝{B 1 ,…B C Is } is>
Each level classifies the weight magnitude given. />
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211492956.9A CN115859801A (en) | 2022-11-25 | 2022-11-25 | Landslide space-time risk assessment method combined with effective rainfall model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211492956.9A CN115859801A (en) | 2022-11-25 | 2022-11-25 | Landslide space-time risk assessment method combined with effective rainfall model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115859801A true CN115859801A (en) | 2023-03-28 |
Family
ID=85666660
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211492956.9A Pending CN115859801A (en) | 2022-11-25 | 2022-11-25 | Landslide space-time risk assessment method combined with effective rainfall model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115859801A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116108759A (en) * | 2023-04-11 | 2023-05-12 | 湖北省地质环境总站 | Landslide hazard evaluation method based on characteristic coupling |
| CN116151483A (en) * | 2023-04-14 | 2023-05-23 | 山东理工大学 | Regional rainfall landslide probabilistic prediction method and prediction terminal |
| CN116522764A (en) * | 2023-04-17 | 2023-08-01 | 华中科技大学 | Hot wave-flood composite disaster assessment method considering influence of climate change |
| CN116952943A (en) * | 2023-09-19 | 2023-10-27 | 吉林省林业科学研究院(吉林省林业生物防治中心站) | Forest land slope soil erosion measurement system based on oblique photography |
| CN117096875A (en) * | 2023-10-19 | 2023-11-21 | 国网江西省电力有限公司经济技术研究院 | Short-term load prediction method and system based on ST-transducer model |
-
2022
- 2022-11-25 CN CN202211492956.9A patent/CN115859801A/en active Pending
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116108759A (en) * | 2023-04-11 | 2023-05-12 | 湖北省地质环境总站 | Landslide hazard evaluation method based on characteristic coupling |
| CN116108759B (en) * | 2023-04-11 | 2023-06-30 | 湖北省地质环境总站 | Landslide hazard evaluation method based on characteristic coupling |
| CN116151483A (en) * | 2023-04-14 | 2023-05-23 | 山东理工大学 | Regional rainfall landslide probabilistic prediction method and prediction terminal |
| CN116151483B (en) * | 2023-04-14 | 2023-07-21 | 山东理工大学 | Regional rainfall landslide probabilistic prediction method and prediction terminal |
| CN116522764A (en) * | 2023-04-17 | 2023-08-01 | 华中科技大学 | Hot wave-flood composite disaster assessment method considering influence of climate change |
| CN116522764B (en) * | 2023-04-17 | 2023-12-19 | 华中科技大学 | Hot wave-flood composite disaster assessment method considering influence of climate change |
| CN116952943A (en) * | 2023-09-19 | 2023-10-27 | 吉林省林业科学研究院(吉林省林业生物防治中心站) | Forest land slope soil erosion measurement system based on oblique photography |
| CN116952943B (en) * | 2023-09-19 | 2023-12-08 | 吉林省林业科学研究院(吉林省林业生物防治中心站) | Forest land slope soil erosion measurement system based on oblique photography |
| CN117096875A (en) * | 2023-10-19 | 2023-11-21 | 国网江西省电力有限公司经济技术研究院 | Short-term load prediction method and system based on ST-transducer model |
| CN117096875B (en) * | 2023-10-19 | 2024-03-12 | 国网江西省电力有限公司经济技术研究院 | Short-term load prediction method and system based on spatial-Temporal Transformer model |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Roy et al. | Landslide susceptibility mapping using knowledge driven statistical models in Darjeeling District, West Bengal, India | |
| Pham et al. | Hybrid computational intelligence models for groundwater potential mapping | |
| Abijith et al. | Assessment of land use and land cover change detection and prediction using remote sensing and CA Markov in the northern coastal districts of Tamil Nadu, India | |
| Hong et al. | Improving the accuracy of landslide susceptibility model using a novel region-partitioning approach | |
| Tehrany et al. | The application of a Dempster–Shafer-based evidential belief function in flood susceptibility mapping and comparison with frequency ratio and logistic regression methods | |
| Cao et al. | Multi-geohazards susceptibility mapping based on machine learning—A case study in Jiuzhaigou, China | |
| Pourghasemi et al. | Landslide susceptibility mapping by binary logistic regression, analytical hierarchy process, and statistical index models and assessment of their performances | |
| Pourghasemi et al. | Random forests and evidential belief function-based landslide susceptibility assessment in Western Mazandaran Province, Iran | |
| Tang et al. | Flood susceptibility assessment based on a novel random Naïve Bayes method: A comparison between different factor discretization methods | |
| Nicu | Application of analytic hierarchy process, frequency ratio, and statistical index to landslide susceptibility: an approach to endangered cultural heritage | |
| Rahmati et al. | Identification of critical flood prone areas in data-scarce and ungauged regions: a comparison of three data mining models | |
| Pourghasemi et al. | GIS-based landslide susceptibility mapping with probabilistic likelihood ratio and spatial multi-criteria evaluation models (North of Tehran, Iran) | |
| Sangchini et al. | Assessment and comparison of combined bivariate and AHP models with logistic regression for landslide susceptibility mapping in the Chaharmahal-e-Bakhtiari Province, Iran | |
| CN115859801A (en) | Landslide space-time risk assessment method combined with effective rainfall model | |
| Bolibar et al. | Deep learning applied to glacier evolution modelling | |
| Althuwaynee et al. | A novel ensemble bivariate statistical evidential belief function with knowledge-based analytical hierarchy process and multivariate statistical logistic regression for landslide susceptibility mapping | |
| Ayalew et al. | Landslides in Sado Island of Japan: Part II. GIS-based susceptibility mapping with comparisons of results from two methods and verifications | |
| Rizeei et al. | Surface runoff prediction regarding LULC and climate dynamics using coupled LTM, optimized ARIMA, and GIS-based SCS-CN models in tropical region | |
| Duran-Llacer et al. | A new method to map groundwater-dependent ecosystem zones in semi-arid environments: A case study in Chile | |
| Yu et al. | Deep learning-based downscaling of tropospheric nitrogen dioxide using ground-level and satellite observations | |
| Yuan et al. | A novel method based on deep learning model for national-scale landslide hazard assessment | |
| Bopche et al. | Feature-based model for landslide susceptibility mapping using a multi-parametric decision-making technique and the analytic hierarchy process | |
| Khai-Wern et al. | Landslide hazard mapping of Penang island using probabilistic methods and logistic regression | |
| Correa-Muñoz et al. | Determination of landslide susceptibility in linear infrastructure. Case: aqueduct network in Palacé, Popayan (Colombia) | |
| Mawasha et al. | Detecting land use and land cover change for a 28-year period using multi-temporal Landsat satellite images in the Jukskei River catchment, Gauteng, South Africa |
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 |