A method for landslide spatiotemporal hazard assessment combined with effective rainfall model

Abstract

This invention provides a method for assessing the spatiotemporal hazard of landslides using an effective rainfall model. This method analyzes the correlation and importance of acquired landslide-inducing factors with landslide occurrence. Three hybrid deep learning networks are used to construct a landslide susceptibility model and calculate the spatial probability of landslides. By embedding the effective rainfall model in the Transformer network coding layer and modifying the existing attention mechanism module, an ST-Transformer deep learning network is proposed to predict effective rainfall over different landslide time periods. The most accurate prediction time is then selected to assess the temporal probability of landslide occurrence. Furthermore, the existing landslide hazard calculation formula is modified, and the two previous results are combined to quantitatively calculate the landslide hazard at the optimal prediction time. This invention can provide a reliable decision-making basis for landslide hazard assessment and risk management.

Description

Technical Field

[0001] The invention relates to the field of risk assessment, in particular to a landslide spatiotemporal risk assessment method combined with an effective rainfall model. Background Art

[0002] To address the threat posed by landslides, numerous methods have been developed for landslide hazard assessment. Among them, physics-based methods assess landslide hazard by analyzing slope stability based on soil internals and geological structures. While these methods are relatively accurate, they require comprehensive documentation of historical landslide data and sufficient human and material resources. Statistical methods, probabilistic methods, and GIS-based hierarchical analysis methods offer some interpretability for landslide hazard assessments within a study area. However, most methods are subject to subjective judgment and are not suitable for large-scale landslide risk assessments. Machine learning methods, including logistic regression, support vector machines, artificial neural networks, and Bayesian networks, are currently widely used in the field of landslide hazard assessment. These methods generally construct landslide hazard assessment models based on a range of landslide-related geological, geomorphological, and rainfall-induced factors. These methods assess landslide hazard based on past conditions and are unable to assess future landslide hazard. Furthermore, they treat rainfall as a static factor, such as using annual average rainfall as the rainfall factor, without considering the impact of temporal variations in rainfall on landslide hazard. Some existing landslide hazard assessment methods consider the temporal dynamics of rainfall factors. These methods use machine learning to first calculate a landslide susceptibility index. Then, based on real-time rainfall data, they employ statistical models to derive a rainfall threshold (when rainfall reaches a certain lower limit, the likelihood of a landslide is high) or a previous rainfall index. These are combined with the landslide susceptibility index to form a dynamic hazard matrix, enabling near-real-time dynamic assessment of landslide hazard. However, this method suffers from two shortcomings:

[0003] (1) Single-dimensional rainfall analysis: Real-time precipitation data is used to estimate the probability of future landslide occurrence, and a dynamic hazard matrix is formed with landslide susceptibility to assess landslide hazard. Although this method takes into account the temporal variation of rainfall factors and achieves a certain degree of near-real-time assessment effect, it ignores the impact of the combined effect of rainfall factors changing simultaneously with time and geographic space (the overall temporal and spatial variation effect) on landslide hazard;

[0004] (2) Time uncertainty in the calculation of rainfall threshold: The above landslide hazard assessment method based on rainfall threshold ignores the fact that only rainfall reaching a certain lower limit will affect the stability of the landslide when calculating the rainfall threshold. In addition, different time values for calculating the rainfall threshold will produce different landslide time prediction results, resulting in uncertainty in the landslide hazard analysis results, which affects the assessment effect. Summary of the Invention

[0005] To address the above problems, the present invention improves the traditional Transformer method and proposes a landslide spatiotemporal hazard assessment method combined with an effective rainfall model. This method introduces an effective rainfall model into the Transformer network to extract effective rainfall factor data based on the original input rainfall time series data. On this basis, a convolutional neural network (CNN) and a bidirectional long short-term memory (Bi-LSTM) neural network are embedded, and a new spatiotemporal attention mechanism module is proposed to replace the original attention mechanism module, integrating the temporal and spatial characteristics of rainfall to improve the accuracy of rainfall prediction.

[0006] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0007] A landslide spatiotemporal hazard assessment method combined with an effective rainfall model comprises the following steps:

[0008] Step 1, selecting landslide inducing factors and analyzing the relationship between the landslide inducing factors and the landslide;

[0009] Step 2: Use three hybrid deep learning networks to build a landslide susceptibility model, predict the spatial probability of landslides, and obtain a landslide spatial probability map;

[0010] Step 3: Based on the deep learning Transformer network, the effective rainfall model, CNN convolutional neural network, and Bi-LSTM network are embedded to construct the ST-Transformer deep learning network to predict the different recurrence periods T of landslides, that is, the probability of landslides occurring at different time periods in the future.

[0011] Step 4: quantitatively calculate the landslide hazard index of the future optimal time T value by combining the improved landslide hazard calculation formula with the slope spatial probability and landslide time probability obtained in step 3.

[0012] Furthermore, the landslide inducing factors include terrain factors, geological factors, geomorphological factors, hydrological and road factors, and the acquired landslide inducing factors are processed into a unified spatial resolution and spatial coordinate system.

[0013] Furthermore, the step 1 includes the following sub-steps:

[0014] Step a1, obtaining landslide inducing factors;

[0015] Step a2, constructing landslide and non-landslide binary classification data;

[0016] Step a3, analyzing the relationship between landslide inducing factors and landslide;

[0017] Furthermore, in step a2, landslide and non-landslide binary classification data is constructed, and the specific implementation includes the following sub-steps:

[0018] a2.1 Obtain regional landslide data;

[0019] a2.2 Obtain the same amount of non-landslide data as landslide data and construct the landslide binary classification data.

[0020] Furthermore, based on the acquired landslide and non-landslide binary classification data points, the attributes of the acquired landslide inducing factors are extracted, and finally landslide sample data is obtained, which includes landslide binary classification coding data and inducing factor attribute data of corresponding landslide and non-landslide points.

[0021] Furthermore, in step a3, the relationship between landslide-inducing factors and landslides is analyzed, and the specific implementation includes the following sub-steps:

[0022] a3.1 Using the collinearity test method to analyze the correlation between the landslide-inducing factors in the acquired landslide sample data, exclude the landslide-inducing factors with collinear relationships;

[0023] a3.2 The information gain analysis method is used to calculate the importance index of independent landslide-inducing factors, and the factors with an importance index less than 0.1 are eliminated. The new landslide-inducing factors obtained are arranged in descending order according to the importance index. That is, new ordered landslide-inducing factors without collinearity are obtained. Combined with the corresponding landslide binary classification data, they are divided into training set and test set.

[0024] Furthermore, step 2 includes the following sub-steps:

[0025] Step b1, extract the main features based on the CNN network:

[0026] In step b2, the RNN recurrent neural network constructs a landslide susceptibility model, predicts the spatial probability of landslides and obtains a landslide spatial probability map.

[0027] Furthermore, in step b2, the specific implementation is as follows:

[0028] Based on the CNN network training results in step b1, independent and ordered landslide-inducing factors are obtained as the input data of the RNN network, and the output data is the landslide binary classification coding data, thereby constructing a landslide susceptibility model. The RNN includes three recurrent neural networks: LSTM, GRU and SRU. The feature results extracted by the three RNN networks are R cnn , as the input data of the RNN network, the calculation process after entering the hidden layer is as follows:

[0029] RNN t =f RNN (w*R cnn+u*RNN t-1 +b RNN )

[0030] O t =g RNN (v*RNN t )

[0031] In the formula, RNN represents its three variant networks, namely RNN = {LSTM, GRU, SRU}, f RNN , g RNN are the activation functions of the hidden layer and the output layer respectively, u, v, w are the network parameters, b is the bias parameter, RNN t is the output result of the hidden layer at time t, O t The landslide features extracted by the CNN network are trained in the three RNN networks mentioned above to obtain the landslide spatial probability index and generate three landslide spatial probability maps.

[0032] Furthermore, step 3 includes the following sub-steps:

[0033] Step c1: In the ST-Transformer network, an effective rainfall model is embedded to calculate the effective rainfall amount of the input rainfall data of different periods to obtain the effective rainfall time series data of different periods;

[0034] In step c2, a CNN and Bi-LSTM network are embedded in the encoder structure of the Transformer network. First, the main feature information of effective rainfall is extracted through the CNN convolutional neural network. Based on this, the Bi-LSTM network is embedded in the position encoding structure to extract the spatial feature information of effective rainfall data. Then, one-hot encoding and the CNN convolutional neural network are combined to extract the temporal features of effective rainfall.

[0035] In step c3, a spatiotemporal attention mechanism is proposed to fuse the spatiotemporal features of the effective rainfall data extracted above, and use them as the input data of the decoding layer to predict the time probability of landslide. The prediction results are evaluated by using three error indices: mean absolute error, mean square error, and root mean square error. The time T value with the smallest error index is selected as the optimal time for predicting the landslide time probability.

[0036] Furthermore, the specific implementation includes the following steps:

[0037] d1, landslide spatial probability map obtained in step 2 and step 3 respectively and time probability graph are divided into the same C categories, where C = {1, 2, 3, ...};

[0038] d2, quantitatively calculated using the FR frequency method and The corresponding weight values of the C categories are divided, and the FR value of each category, i.e., the weight value, is calculated by the ratio of pixels occupied by each category and the ratio of pixels occupied by the corresponding landslide points. The calculation formula is as follows:

[0039]

[0040] Where, FR i Represents the weight value of the i-th category, N i represents the number of landslide grid pixels in the i-th category, represents the total number of landslide grid pixels in the study area; P i is the number of grid pixels of the i-th category, Represents the total number of all raster pixels covering the study area;

[0041] d3, use the hazard formula to calculate the landslide hazard in the best time period T in the future, so as to predict the future landslide hazard, landslide hazard L HH The calculation formula is as follows;

[0042]

[0043] Where A i ={A1, ...A C}for The weight assigned to each level category, B i ={B1, ...B C}for The weight assigned to each level category.

[0044] Compared with the prior art, this application has the following beneficial effects:

[0045] The present invention proposes a new landslide hazard assessment method, which mainly includes three contents: (1) landslide susceptibility assessment method, (2) landslide time probability prediction method, and (3) landslide hazard index calculation method. The landslide susceptibility assessment method uses three hybrid neural networks CNN-LSTM, CNN-GRU and CNN-SRU to calculate landslide susceptibility, that is, landslide spatial probability calculation, which has superior performance and high model accuracy. The landslide time probability prediction method uses the ST-Transformer method to predict future effective rainfall based on rainfall data sets at different times. This method is based on the traditional Transformer network framework, embeds the effective rainfall model, combines CNN and Bi-LSTM networks with position encoding, and comprehensively calculates the time dimension. Finally, the rainfall spatiotemporal characteristics are obtained, and the proposed spatiotemporal attention mechanism is used to fuse the spatiotemporal feature information. Finally, the effective rainfall at different times T in the future is predicted, and the time probability of landslide is analyzed. The predicted time T with the smallest result error index is selected as the time probability of landslide occurrence. Based on landslide susceptibility and effective rainfall forecasts at the optimal prediction time T, an improved landslide hazard formula is used to calculate the landslide hazard at a future time T, thereby generating a landslide hazard prediction classification map for the future time T. This method can effectively capture the temporal dependence of long-term time series data and can predict the future landslide hazard of a region in the long or short term. The proposed method has low error and good robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a framework diagram of the landslide hazard analysis method implemented in the present invention.

[0047] Figure 2 Figure 2 shows the landslide inducing factor diagram selected for the embodiment of the present invention, including (a) slope direction, (b) slope gradient, (c) slope height, (d) slope length, (e) plan curvature, (f) profile curvature, (g) surface incision depth, (h) elevation variation coefficient, (i) ground roughness, (j) terrain moisture index, (k) flow intensity index, (l) relative slope position, (m) watershed area, (n) watershed slope gradient, (o) flow path length, (p) valley depth, (q) relative river network distance, (r) terrain relative convergence index, (s) terrain landslide point index, (t) terrain surface curvature, (u) terrain surface texture, (v) stratum lithologic structure, (w) relative fault distance, (x) soil type, (y) ground peak acceleration, (z) land use type, and (aa) normalized difference vegetation index.

[0048] Figure 3The landslide susceptibility maps generated by the CNN-LSTM, CNN-GRU, and CNN-SRU methods in the embodiments of the present invention are: (a) a landslide susceptibility map generated by the CNN-LSTM method, (b) a landslide susceptibility map generated by the CNN-GRU method, and (c) a landslide susceptibility map generated by the CNN-SRU method.

[0049] Figure 4 The embodiments of the present invention use the ST-Transformer method, Transformer, CNN-LSTM and LSTM to predict the rainfall forecast map for the next year. (a) The effective rainfall forecast map for the next year generated by the ST-Transformer method, (b) The effective rainfall forecast map for the next year generated by the Transformer method, (c) The effective rainfall forecast map for the next year generated by the CNN-LSTM method, and (d) The effective rainfall forecast map for the next year generated by the LSTM method.

[0050] Figure 5 The effective rainfall forecast map for the next three months is predicted using the ST-Transformer method, Transformer, CNN-LSTM and LSTM in the embodiments of the present invention. (a) The effective rainfall forecast map for the next three months generated by the ST-Transformer method, (b) The effective rainfall forecast map for the next three months generated by the Transformer method, (c) The effective rainfall forecast map for the next three months generated by the CNN-LSTM method, and (d) The effective rainfall forecast map for the next three months generated by the LSTM method.

[0051] Figure 6 The effective rainfall forecast map for the next month is predicted using the ST-Transformer method, Transformer, CNN-LSTM, and LSTM in the embodiments of the present invention. (a) The effective rainfall forecast map for the next month generated by the ST-Transformer method, (b) The effective rainfall forecast map for the next month generated by the Transformer method, (c) The effective rainfall forecast map for the next month generated by the CNN-LSTM method, and (d) The effective rainfall forecast map for the next month generated by the LSTM method.

[0052] Figure 7 Landslide hazard maps are created by combining a landslide susceptibility map generated using the CNN-LSTM method and effective rainfall forecast maps for different periods. (a) Landslide hazard map predicted for the next year, (b) Landslide hazard map predicted for the next three months, and (c) Landslide hazard map predicted for the next month.

[0053] Figure 8Landslide hazard maps are created by combining a landslide susceptibility map generated using the CNN-GRU method and effective rainfall forecast maps for different periods. (a) Landslide hazard map predicted for the next year, (b) Landslide hazard map predicted for the next three months, and (c) Landslide hazard map predicted for the next month.

[0054] Figure 9 Landslide hazard maps are created by combining the landslide susceptibility map generated by the CNN-SRU method in an embodiment of the present invention with effective rainfall forecast maps for different periods. (a) Landslide hazard map predicted for the next year, (b) Landslide hazard map predicted for the next three months, and (c) Landslide hazard map predicted for the next month. DETAILED DESCRIPTION

[0055] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0056] The embodiment of the present invention provides a new landslide hazard assessment method ( Figure 1 ), which is mainly divided into three aspects: landslide susceptibility calculation, landslide time probability calculation and landslide hazard calculation. (1) Landslide susceptibility calculation: First, based on the spatial landslide inducing factors such as topography, geomorphology, geology and hydrology, the landslide inducing factors with no collinearity relationship and high importance index are selected through multicollinearity test and information gain analysis, and then the landslide binary classification data are combined to form landslide sample data, which are divided into 70% training set and 30% test set. According to the training set, the landslide susceptibility model is constructed using three hybrid neural networks: CNN-LSTM, CNN-GRU and CNN-SRU. The test set is used for verification and the model accuracy is evaluated by four indices: ACC, AUC, KAPPA and MCC. Finally, the spatial probability of landslide occurrence is calculated to generate a landslide susceptibility map. (2) Calculation of landslide time probability: Based on the original Transformer framework, this paper introduces the convolutional neural network CNN and the bidirectional neural network Bi-LSTM model, and combines them with the positional embedding of the original Transformer network to extract spatial features. At the same time, a new spatiotemporal attention mechanism is proposed to fuse temporal and spatial characteristic information to achieve the effect of improving the time series prediction performance. This method uses the rainfall data of the study area to predict the rainfall in different periods in the future, so as to estimate the time probability of landslide occurrence and finally generate a landslide time probability map. The three error indices of MSE, MAE and RMSE are used to evaluate the accuracy of rainfall prediction; (3) Landslide hazard calculation: Based on the results obtained in steps (1) and (2), the proposed landslide hazard formula is used to quantitatively calculate the future landslide hazard. Taking the continental United States as the study area, the specific implementation of this method includes the following steps:

[0057] Step 1: Analyze the relationship between the acquired landslide-inducing factors and landslide occurrence. This step can be divided into three sub-steps:

[0058] Step a1, obtaining landslide inducing factors;

[0059] In step a1, relevant landslide-inducing factors were obtained from various sources, mainly including topographic, geological, geomorphological and hydrological data, which are described as follows:

[0060] Slope, aspect, plan and profile curvature, relief, surface cutting depth (SCD), elevation variation coefficient (EVC), terrain roughness, topographic wetness index (TWI), and stream power index (SPI) were extracted using digital terrain elevation data with a spatial resolution of 30 meters (USGS, 2005). Catchment area, catchment slope, flow path length (FPL), valley depth, distance to channel network (DTCN), and topographic relative convergence index (SPI) were also extracted using SAGA GIS 2.3.2 (Conrad et al., 2015). The landslide index (TRCI) is used as the basis for the landslide prediction, and the landslide prediction function is used as the basis for the landslide prediction. The landslide prediction function is used as the basis for the landslide prediction.

[0061] Geological factors: These are mainly derived from the United States Geological Survey and include stratigraphic lithology and fault data. The stratigraphic data were then rasterized using ArcGIS 10.5 software and classified using the Jenks natural break method (Chen et al. 2013) to obtain a stratigraphic grid classification map. Furthermore, a distance to fault (DTF) classification map was obtained based on Euclidean distance analysis.

[0062] Landform factors: MODIS data were downloaded from the Earth Data Center in the United States to obtain global MODQ13 data. The HEGtool tool was used to convert the images into raster data, namely the Normalized Difference Vegetation Index (NDVI). Finally, the mask extraction and analysis function of ArcGIS was used to obtain the NDVI map of Shuicheng County. Land cover utilization vector data of the contiguous United States were obtained from the National Geographic Information Resource Directory Service System. Land cover utilization data of the Shuicheng County study area were also extracted using ArcGIS 10.5 software and converted into raster form, thus obtaining the Shuicheng County land cover utilization raster map (Land cover). Soil data and land classification data of the contiguous United States were downloaded from the website of the United Nations Agricultural Organization. Soil type data of the Shuicheng County area were then extracted using ArcGIS 10.5 software and converted into raster form, thus obtaining the Shuicheng County soil type raster classification map. The land classification data were rasterized in the same way to obtain the land raster classification map (Landform).

[0063] Hydrological and road factors: We obtained water system and road data for the United States from the Global Road and Water System Database. We also used ArcGIS 10.5 to obtain water flow and road data for the continental United States. Euclidean distance analysis was performed to generate distance to water (DTW) and road (DTR) grid maps.

[0064] In summary, this example obtained 33 landslide-inducing factors in the continental United States. These factors were processed into raster maps with a unified spatial coordinate system (WGS 1984) and spatial resolution (30 m × 30 m) using ArcGIS 10.5.

[0065] Step a2, constructing landslide and non-landslide binary classification data;

[0066] In step a2, the specific implementation includes the following steps:

[0067] (1) The present invention obtained 64,356 landslide sites in the contiguous United States from 1900 to 2019 from the U.S. Geological Survey landslide database and the Global Landslide Database;

[0068] (2) In order to balance the ratio of positive and negative samples, an equal amount of non-landslide point data needs to be generated. In this paper, the present invention randomly generates an equal amount of non-landslide data as landslide samples in an area outside the radius of 8,000 km of the existing landslide point (the area outside this distance is not affected by the landslide point) using ArcGIS 10.5. Then, the landslide and non-landslide points are coded as 1 and 0, respectively, and finally 128,712 landslide classification points are obtained. Then, the attribute values of landslide-inducing factors are extracted based on these landslide classification sample points, and finally the complete landslide sample data is obtained, which includes the attributes of the relevant landslide-inducing factors and the landslide and non-landslide classification data.

[0069] Step a3, analyzing the relationship between landslide inducing factors and landslide;

[0070] In step a3, the specific implementation includes the following steps:

[0071] (1) The multicollinearity method was used to test the correlation of landslide-inducing factors. VIF (variance inflation factor) and TOL (tolerance) are two evaluation indicators used for collinearity analysis results. If VIF < 10, TOL > 0.1 or VIF < 5, TOL > 0.2, it indicates that there is no correlation between the variables. As shown in Table 2, the VIF of relief is greater than 10 and the TOL is close to 0, so this factor is not considered as a landslide-inducing factor.

[0072] (2) On this basis, the information gain method is used to analyze the importance index of the new landslide-inducing factors. The present invention uses IG (Information Gain) to represent the importance index. The closer the index is to 1, the higher the correlation with the occurrence of landslides. In Table 2, the IG index of Landforms, Morphometric Features, DTR and DTW is close to 0, indicating that these indices have little effect on the occurrence of landslides. Therefore, the present invention removes the landslide-inducing factors with collinear relationships and extremely low importance indexes, and finally obtains 27 ordered and non-collinear high-importance inducing factors ( Figure 2 ), respectively, Aspect ( Figure 2 a), Slope ( Figure 2 b), Slope Height ( Figure 2 c) Slope Length Figure 2 d) Plan Curvature Figure 2 e), Profile Curvature ( Figure 2f), Surface Cutting Depth (SCD) ( Figure 2 g), Elevation variation coefficient (EVC) ( Figure 2 h), Terrain Roughness (TR) Figure 2 i), Topographic Wetness Index (TWI) Figure 2 j), Stream Power Index (SPI) ( Figure 2 k), relative slope position (Relative Slope Position, RSP) ( Figure 2 l), Catchment Area ( Figure 2 m), Catchment Slope ( Figure 2 n), Flow Path Length (FPL) ( Figure 2 o), Valley Depth ( Figure 2 p), Distance to Channel Network (DTCN) Figure 2 q), Topographic Relative Convergence Index (TRCI) ( Figure 2 r), Topographic Position Index (TPI) ( Figure 2 s), terrain surface convexity (Convexity) ( Figure 2 t), terrain surface texture (Texture) ( Figure 2 u), lithology (Lithology) ( Figure 2 v), distance to fault (Distance to Faulty, DTF) ( Figure 2 w), Soil Type ( Figure 2 x), Peak Ground Acceleration (PGA) ( Figure 2 y), land use type (Land use)( Figure 2 z), Normalized Difference Vegetation Index (NDVI) ( Figure 2 aa).

[0073] Table 1 Results of collinearity test and importance analysis of landslide inducing factors

[0074]

[0075]

[0076] Step 2: Use three hybrid deep learning networks to build a landslide susceptibility model, calculate the spatial probability of landslides and obtain a landslide spatial probability map;

[0077] The step 2 includes the following sub-steps:

[0078] Step b1, extract the main features based on the CNN network:

[0079] In step b2, the RNN recurrent neural network constructs a landslide susceptibility model, predicts the spatial probability of landslides and obtains a landslide spatial probability map.

[0080] Furthermore, in step b2, the specific implementation is as follows:

[0081] Based on the CNN network training results in step b1, independent and ordered landslide-inducing factors are obtained as the input data of the RNN network, and the output data is the landslide binary classification coding data, thereby constructing a landslide susceptibility model. The RNN includes three recurrent neural networks: LSTM, GRU and SRU. The feature results extracted by the three RNN networks are R cnn , as the input data of the RNN network, the calculation process after entering the hidden layer is as follows:

[0082] RNN t =f RNN (w*R cnn +u*RNN t-1 +b RNN ) (1)

[0083] O t =g RNN (v*RNN t ) (2)

[0084] In the formula, RNN represents its three variant networks, namely RNN = {LSTM, GRU, SRU}, f RNN , g RNN are the activation functions of the hidden layer and the output layer respectively, u, v, w are the network parameters, b is the bias parameter, RNN t is the output result of the hidden layer at time t, O t The landslide features extracted by the CNN network are trained in the three RNN networks mentioned above to obtain the landslide spatial probability index and generate three landslide spatial probability maps.

[0085] In step 2, the landslide-inducing factors obtained above are arranged in descending order according to the importance index to meet the ordered input data structure of the three hybrid networks CNN-LSTM, CNN-GRU, and CNN-SRU. The output data is the landslide and non-landslide binary data, and the spatial probability of landslide is calculated. The landslide susceptibility results of the three hybrid deep learning networks are as follows: Figure 3 As shown in Figure 2, the model accuracy is evaluated using four indicators: AUC, ACC, KAPPA, and MCC. The closer the above evaluation indicators are to 1, the higher the model prediction performance. The model accuracy evaluation results are shown in Table 2.

[0086] Table 2 Evaluation results of CNN-SRU, CNN-LSTM and CNN-GRU hybrid network models

[0087]

[0088] Figure 3 It is the landslide susceptibility spatial probability map generated by the three CNN-LSTM, CNN-GRU and CNN-SRU hybrid deep learning networks mentioned above. Figure 3 It can be seen that the three hybrid methods, CNN-LSTM, CNN-GRU, and CNN-SRU, show similar landslide susceptibility trends. Furthermore, landslide-prone areas in the continental United States are concentrated in the east and west, with scattered distribution in the central region. Furthermore, landslide-prone areas in the west are larger than those in the east, with a trend of spreading eastward from the central region. Areas with extremely high landslide susceptibility are relatively large. Figure 3 (a)-(c) are the landslide susceptibility spatial probability maps generated by CNN-LSTM, CNN-GRU and CNN-SRU networks. Figure 3 (a) and Figure 3 The area with high landslide susceptibility level in (c) is almost the same, but higher than Figure 3 In the landslide susceptibility map generated by CNN-GRU in (b), the corresponding grade areas of landslide susceptibility are Figure 3 (c) In areas with high landslide susceptibility, the Figure 3 (a) The corresponding area.

[0089] Table 2 shows the evaluation results of four indicators: Area Under the Curve (AUC), Accuracy (ACC), KAPPA coefficient (KAPPA) and Matthews correlation coefficient (MCC). Overall, the ACC index of the three methods is the same, and the AUC index is higher than 96%, indicating that the three models have good landslide susceptibility prediction performance. The KAPPA and MCC indexes of CNN-SRU and CNN-LSTM are the same, and the AUC index of CNN-SRU method is the highest, which is 0.968. This shows that among the three landslide susceptibility models mentioned above, the landslide susceptibility model constructed using CNN-SRU method has better prediction performance and is comparable to the CNN-LSTM method. Figure 3 The landslide susceptibility levels described in (a)-(c) are relatively consistent, that is, the area with high landslide susceptibility level in the landslide susceptibility map generated by the CNN-SRU network model is higher than the corresponding area in the CNN-LSTM and CNN-GRU maps.

[0090] Step 3: Based on the deep learning Transformer network, an effective rainfall model, a CNN convolutional neural network, and a Bi-LSTM network are embedded to construct an ST-Transformer deep learning network to predict different landslide recurrence periods T, that is, the probability of landslide occurrence at different time periods in the future. Step 3 includes the following sub-steps:

[0091] Step c1: In the ST-Transformer network, an effective rainfall model is embedded to calculate the effective rainfall amount of the input rainfall data of different periods to obtain the effective rainfall time series data of different periods;

[0092] In step c2, a CNN and Bi-LSTM network are embedded in the encoding structure. First, the main feature information of effective rainfall is extracted through the CNN convolutional neural network. Based on this, the Bi-LSTM network is embedded in the position encoding structure to extract the spatial feature information of effective rainfall data. Then, one-hot encoding and the CNN convolutional neural network are combined to extract the temporal features of effective rainfall.

[0093] In step c3, a spatiotemporal attention mechanism is proposed to fuse the spatiotemporal features of the effective rainfall data extracted above and use them as the input data of the decoding layer to predict the time probability of landslide. The prediction results are evaluated by using the mean absolute error, mean square error, and three error indices. The time T value with the smallest error index is selected as the optimal time for predicting the landslide time probability.

[0094] In step c1, the present invention predicts the time probability of landslide occurrence through rainfall data. The data comes from CHIRPS global rainfall data, with a latitude and longitude range of 50N-50S, 180W-180E, a spatial resolution of 0.05°×0.05° unit grid, and contains data of three different periods: annual, seasonal, and monthly. These three types of data all start from 1981, of which the annual rainfall data contains a time span of 39 years, and the seasonal rainfall and monthly rainfall times end in the third quarter of 2021 and September 2021, respectively. On this basis, the present invention selects rainfall data from the continental United States as training samples for calculating the time probability of landslide occurrence in the future. 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 based on the input rainfall data at different time periods as input data in step b2.

[0095] In step c2, a CNN and Bi-LSTM network were introduced, combined with positional encoding, to extract spatial features of rainfall based on rainfall data. Time series forecasting was then performed using the ST-Transformer network model proposed in this paper. All experiments were conducted on an NVIDIA GeForce RTX2060 computer. The model was based on the Adam (adaptive moment estimation) optimizer, using the mean square error (MSE) loss function, with a batch size of 300 and 30 training iterations. The present invention predicted effective rainfall for the next one year, one month, and three months based on these three types of periodic rainfall data. The model was then compared with three deep learning time series forecasting network models: a traditional Transformer network, a CNN-LSTM network, and an LSTM network. To evaluate the model's prediction accuracy, three evaluation metrics were used: mean square error (MSE), root mean square error (RMSE), and mean absolute error (MAE). The model evaluation results are shown in Table 3. Figure 4 , Figure 5 , Figure 6The ST-Transformer network model proposed in this paper, along with three other Transformer, CNN-LSTM, and LSTM models, were used to predict effective rainfall for the next year, three months, and one month. Based on these effective rainfall values, effective rainfall maps for different forecast periods were generated using ArcGIS 10.5. Similar to the landslide susceptibility classification maps, the different effective rainfall ranges in the rainfall maps generated by each model at different forecast times represent the magnitude of effective rainfall in different regions, with the range from low to high indicating increasing effective rainfall.

[0096] The specific process of step c2.1 is as follows:

[0097] Although rainfall is an external trigger for landslides, it will only occur when effective rainfall is reached. This effective rainfall refers to the amount of rainfall that causes landslides, that is, effective rainfall is the key factor in inducing landslides. Therefore, the rainfall data initially input into the ST-Transformer network is converted into effective rainfall time series data through the effective rainfall model structure introduced in the encoding structure to construct the effective rainfall time series data set required for landslide time probability prediction. The calculation process of the effective rainfall model is shown in formula (3), where R b represents effective rainfall, a represents the number of days before the landslide occurs, R a is the rainfall in the previous day, and K is the effective rainfall coefficient.

[0098]

[0099] Assume that a landslide site in the United States has a daily rainfall attribute represented by r n , n∈N0, N0 is the number of landslide points in the study area, based on the landslide point r n The effective rainfall time series data of the above R sequence lengths are expressed as R b (r n R ), the calculation process of the effective rainfall model using the above formula (3) is shown in (4) and (5). Represents the landslide point r n The effective rainfall time series data of the i-th sequence length T, r n i is the landslide point r n The rainfall on day i.

[0100]

[0101]

[0102] The effective rainfall data calculated by the effective rainfall model at the above landslide point at different rainfall times T are expressed as It represents the time series data of N0 landslide points with R-dimensional effective rainfall attributes. This dataset can be divided into in RD T ∈RD R Represent the spatial and temporal characteristics of effective rainfall, N0 and R represent the number of landslide points with effective rainfall attributes and the temporal dimension of effective rainfall characteristics, respectively. S and time vector RD T Expanded to and Three-dimensional features, based on this data, spatial features are extracted, as described below:

[0103] The network structure for spatial feature extraction based on effective rainfall data is Positional Encoding embedded in Bi-LSTM network. CNN convolutional neural network is added before Positional Encoding embedded in Bi-LSTM network. This network layer is used to extract the main effective rainfall features from effective rainfall time series data. Based on the RD obtained above, S and RD T Spatiotemporal characteristics, the main feature information extraction process of effective rainfall by CNN convolutional neural network is expressed as D S =f CNN (x S , RD S , RD T ), where f CNN It is a 1×1 convolutional layer structure, which is the spatial feature RD of the above effective rainfall data. S and time vector RD T Concatenated feature vector Transformed into a fixed feature vector of dimension This improves the accuracy of extracting spatial feature information from effective rainfall data. S The feature vector serves as the input data for spatial feature extraction.

[0104] In order to more efficiently learn the spatial features of the above rainfall time series data, a Bi-LSTM network is embedded in the positional encoding structure. The Bi-LSTM model is composed of two LSTM network structures. The network structure of this model fully considers the before and after information of the time series data. In the application of effective rainfall time series data prediction, considering the before and after change information of the effective rainfall time series data, the bidirectional feature information extraction of the effective rainfall data is conducive to improving the accuracy of landslide time probability prediction. The features obtained by the positional encoding embedded in the Bi-LSTM network are Indicates that each time step has dimensional effective rainfall spatial feature vector, where POS t and bi_lstm t They are the position encoding layer and Bi-LSTM network layer functions respectively.

[0105] The spatial feature vector obtained by the position encoding layer embedded in the Bi-LSTM network is P S , added to the three sub-vector spaces Q transformed from the initial input data S , k S , v S , the calculation process is as follows:

[0106]

[0107] are the three sub-vectors of the newly obtained effective rainfall characteristics. The spatial characteristics of the effective rainfall at the landslide point changing with time are: The two sub-vectors are multiplied to obtain the spatial feature weight value of effective rainfall, and combined with The spatial characteristics of effective rainfall at each landslide point are calculated as follows:

[0108]

[0109]

[0110] In the formula The spatial attention feature results of the effective rainfall obtained from N0 landslide points, A S Represents the spatial feature weight value of effective rainfall, softmax is the activation function of the spatial attention mechanism, is a scaling parameter, which prevents excessively large dot product results when calculating the spatial feature weights of effective rainfall. The above spatial attention results are obtained using a single attention mechanism. Using a multi-head attention mechanism, we utilize position encoding with an embedded Bi-LSTM network and multiple latent subspaces to calculate different spatial attention features for N0 landslide sites. These features are based on the dynamic characteristics of effective rainfall at each landslide site over time.

[0111] The specific process of step c2.2 is as follows:

[0112] The above time dataset of effective rainfall based on landslide points is expressed as After one-hot encoding, RD T Time data is encoded as a time feature vector. Given the powerful feature extraction potential of convolutional neural networks (CNNs), CNNs are used to extract the time features of effective rainfall data. The effective rainfall time dataset is converted to a time vector using one-hot encoding and then fed into the CNN convolutional neural network. A 1×1 convolutional neural network is selected to extract the feature information of the effective rainfall time series data. A one-dimensional convolutional layer is used to convert this time feature vector into a time feature vector of the same dimension as the spatial vector. The calculation process is as follows:

[0113] D T =Cov1D(one_hot(RD T ) (9)

[0114] In the above formula, one_hot(RD T ) represents the time feature of effective rainfall for hot encoding, and Cov1D represents the one-dimensional convolution layer function of the 1×1 convolutional neural network CNN.

[0115] Similar to the spatial attention feature extraction of effective rainfall, the effective rainfall time feature vector obtained above is added to the three sub-vectors obtained by the attention mechanism of the above original data Three new time eigenvectors are obtained in the time space. The calculation process is as follows:

[0116]

[0117] As three new time feature vectors, using The time characteristic weight A is calculated by the time characteristic vectors of these two effective rainfall T , using the same scaling parameters as the spatial attention mechanism, combined with The time feature sub-vector obtains the time attention mechanism result O T , the calculation process is as follows;

[0118]

[0119]

[0120] Since the effective rainfall in the next T days is predicted based on R historical time data, that is, the initial effective rainfall time series data based on the landslide point is RD T = {RT1, RT2, ... RT R To predict the rainfall in the next T time steps, the calculation process is as follows:

[0121] RT R+T =f E (T1, RT2, ... RT R ) (13)

[0122] Therefore, the prediction of effective rainfall at the landslide point depends on the rainfall data of the entire historical time series and the effective rainfall dependency characteristics before and after the time. Therefore, in order to obtain the effective rainfall time characteristics of the long time series and the time before and after the time, the multi-head attention mechanism is used in the same way as the spatial attention mechanism to obtain the time dependency characteristics before and after the long time series, thereby improving the prediction of effective rainfall.

[0123] The specific process of step c2.3 is as follows:

[0124] Since the sigmoid activation function is one of the commonly used activation functions in machine learning networks, its role is to scale the value between 0 and 1 and use the activation function sigmoid to calculate the spatial attention feature of effective rainfall O S and temporal attention features O T The weight value of . In formula (14) and (15), g X , g T is a linear function, which is used to linearly convert the spatial attention features and temporal attention features of effective rainfall into a one-dimensional feature vector to obtain the weight value W of the effective rainfall spatiotemporal attention features. R Based on the weight value, the spatiotemporal fusion characteristics of effective rainfall are obtained.

[0125] W R =sigmoid(g X (O S )+g T (O T )) (14)

[0126]

[0127] Spatiotemporal fusion characteristics of effective rainfall After the fully connected layer outputs the final effective rainfall feature extraction result in the ST-Transformer encoding structure represents the d of N0 landslide points within R time steps st The spatiotemporal dimension feature vectors are used as the input data of the decoding layer Decoder network structure to predict the effective rainfall in the next T time steps, thereby predicting the time probability of landslide occurrence.

[0128] from Figure 4-6 It can be seen that the effective rainfall forecast map generated by the ST-Transformer network model proposed in this paper shows that the areas with high effective rainfall are larger than the corresponding areas of the other three methods, indicating that the effective rainfall time series forecast model proposed in this paper performs better. At the same time, the effective rainfall forecast classification maps generated by all experimental methods show a clustered distribution characteristic. The areas with high effective rainfall are mainly distributed in the western and eastern regions of the United States. At the same time, the effective rainfall distribution in the eastern region is more concentrated, while that in the western region is relatively dispersed. The effective rainfall in the central region is much smaller than that in the eastern and western regions. Figure 4 The effective rainfall forecast range is much smaller than Figure 5 and Figure 6 This shows that the effective rainfall prediction effect of a long period is lower than that of a short period. Figure 4 (a), Figure 5 (a), Figure 6 (a) Generate effective rainfall forecast classification maps for the next 1 year, 1 month, and 3 months based on the ST-transformer method proposed in this invention, Figure 5 (a) The area with high effective rainfall is much larger than Figure 4 (a) and Figure 5 The corresponding area of (a) shows that the method of the present invention has better prediction performance for a prediction period of one month. Figure 4-6 In the effective rainfall prediction classification diagram generated by the method of the present invention, the area with high effective rainfall is higher than the area shown by the other three methods, indicating that in different prediction periods, the area with high effective rainfall in the effective rainfall time series prediction model proposed by the present invention is higher than the corresponding area of the other three models.

[0129] Table 3 shows the evaluation results of the method of the present invention and three other methods based on different prediction periods, using the three indicators of MSE, RMSE, and MAE. As can be seen from the table, among the four effective rainfall time series prediction models, the three evaluation indicators of the ST-Transformer model proposed in the present invention are far lower than those of the other three model methods, demonstrating the superiority of the method of the present invention. At the same time, in the classification map generated by the method of the present invention, the range of areas with high effective rainfall is much larger than the corresponding distribution areas of the other three methods, which is consistent with this description. The traditional Transformer method is superior to the CNN-LSTM and LSTM methods, demonstrating the time series prediction performance of the transformer.

[0130] Table 3 Evaluation results of effective rainfall time series prediction using different methods

[0131]

[0132]

[0133] Step 4: propose an improved landslide hazard formula and combine the two results obtained above to quantitatively calculate the landslide hazard of different recurrence periods; first, the method uses the above method to generate landslide susceptibility maps respectively. and landslide time probability map Then, the ArcGIS 10.5 Jenks natural break method was used to and They are divided into C categories.

[0134] In step 4, the specific implementation includes the following steps:

[0135] d1 Landslide spatial probability map obtained in step 2 and step 3 respectively and time probability graph are divided into the same C categories, where = {1, 2, 3, ...};

[0136] d2 is quantitatively calculated using the FR frequency method and The corresponding weight values of the C categories are divided, and the FR value of each category, i.e., the weight value, is calculated by the ratio of pixels occupied by each category and the ratio of pixels occupied by the corresponding landslide points. The calculation formula is as follows:

[0137]

[0138] Where, FR i Represents the weight value of the i-th category, N i represents the number of landslide grid pixels in the i-th category, represents the total number of landslide grid pixels in the study area; Pi is the number of grid pixels of the i-th category, Represents the total number of all raster pixels covering the study area;

[0139] d3 uses the hazard formula to calculate the landslide hazard in the future optimal time period T, so as to predict the future landslide hazard, landslide hazard L HH The calculation formula is as follows;

[0140]

[0141] Where A i ={A1, ...A C}for The weight assigned to each level category, B i ={B1, ...B C}for The weight assigned to each level category.

[0142] Based on the landslide susceptibility map and the effective rainfall forecast classification map, the landslide hazard calculation method proposed in this invention is used to evaluate the landslide hazard in the United States. Figure 3-6 It can be seen that both the landslide susceptibility map and the effective rainfall prediction classification map are divided into five categories using the Jenks natural break method of ArcGIS10.5. The present invention uses FR to calculate the weight of each category, and finally obtains the landslide hazard index based on the overlay analysis of ArcGIS10.5. That is, the landslide susceptibility map generated by three hybrid neural networks CNN-RNN is used, and the effective rainfall time series prediction method proposed in the present invention generates different effective rainfall classification maps based on different prediction periods, and the weight ratio is distributed and overlaid for calculation. Figure 7-9 The effective rainfall forecast maps of different periods generated by the method of the present invention are superimposed on the landslide susceptibility maps generated by the three CNN-RNN methods to obtain the landslide hazard maps of the United States at different forecast times in the future. Figure 7-9 It can be seen that in each of the U.S. landslide hazard maps at different forecast times, the eastern region has a large and numerous area corresponding to high-level areas, the western leftmost edge of the region shows fewer areas with high landslide hazard, the eastward region is mainly composed of areas with medium and low landslide hazard, and the rest of the region is an extremely low landslide hazard area. Based on the landslide susceptibility map generated by superimposing the same CNN-RNN, the area with extremely high landslide hazard in the landslide hazard map predicted by the method of the present invention is higher than the corresponding grade area in the landslide hazard map of the next month than in other landslide hazard maps of the next 1 year or 3 months, indicating that the landslide hazard map with a short forecast period is more accurate. According to the future effective rainfall map of the same forecast period (such as month), the hybrid method of superimposing three CNN-RNNs is used to generate a landslide susceptibility map, and three landslide hazard maps of the same future forecast period are generated, namely, Figure 7 (b) Figure 8 (b) and 9(b), there is almost no difference in the degree of landslide hazard represented by the three landslide hazard maps, and the landslide hazard trends for the other future years and three months are the same, indicating that rainfall is the main factor in landslide occurrence.

[0143] Through the above specific implementation, it can be seen that the present invention proposes a landslide spatiotemporal hazard assessment method combined with an effective rainfall model. The contributions of this method are mainly divided into the following three aspects:

[0144] (1) Although the traditional Transformer method can perform efficient parallel computing, the data features in the input attention mechanism ignore the dependency of the previous and next data features in time series prediction, making it impossible to perform effective data feature learning. The present invention introduces an effective rainfall model, CNN, and Bi-LSTM network into the Transformer structure, which can effectively capture the features of long time series data, thereby improving the model prediction performance.

[0145] (2) The present invention proposes a new spatiotemporal attention mechanism, which calculates the time and space attention results of the landslide respectively, and then calculates the weight values of the two according to the original self-attention mechanism. Finally, the time and space attention results of the landslide are superimposed using the weight values to obtain the spatiotemporal attention result, that is, the spatiotemporal feature fusion result, which solves the problem that the spatiotemporal hazard prediction of landslides does not consider the impact of the comprehensive changes of rainfall time variables and spatial attribute characteristics on the landslide hazard.

[0146] (3) The present invention adopts an improved Transformer method to use rainfall data with a long time span, combined with an effective rainfall model to predict the effective rainfall in different periods, and then quantitatively calculates the weight of each category with the landslide susceptibility (landslide spatial probability) within the time range and performs superposition calculations to assess the future landslide hazard. At the same time, the short-term and long-term prediction scales can be arbitrarily set in this method, which can solve the problem of being unable to assess landslide hazard in the long term.

[0147] The above implementation examples are merely examples for describing the present invention in detail and are not intended to limit the generality of the methods and examples.

[0148] The above embodiments are only used to illustrate the present invention and are not intended to limit the present invention. It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the scope of protection of the appended claims of the present invention.

Claims

1. A landslide spatiotemporal hazard assessment method combined with an effective rainfall model, characterized in that: The following steps are involved: Step 1, selecting landslide inducing factors and analyzing the relationship between the landslide inducing factors and the landslide; Step 2: Use three hybrid deep learning networks to build a landslide susceptibility model, predict the spatial probability of landslides, and obtain a landslide spatial probability map; Step 3: Based on the deep learning Transformer network, the effective rainfall model, CNN convolutional neural network, and Bi-LSTM network are embedded to construct the ST-Transformer deep learning network to predict the different recurrence periods T of landslides, that is, the probability of landslides occurring at different time periods in the future. Step 4: quantitatively calculate the landslide hazard index of the future optimal time T value by combining the improved landslide hazard calculation formula with the slope spatial probability and landslide time probability obtained in step 3. The specific implementation includes the following steps: d1, landslide spatial probability map obtained in step 2 and step 3 respectively and time probability graph , are divided into the same C categories, where ; d2, quantitatively calculated using the FR frequency method and The corresponding weight values of the C categories are divided, and the FR value of each category, i.e., the weight value, is calculated by the ratio of pixels occupied by each category and the ratio of pixels occupied by the corresponding landslide points. The calculation formula is as follows: Where, Indicates the i The weight of the class, Indicates the i The number of landslide raster pixels in each category, represents the total number of landslide grid pixels in the study area; For the The number of raster pixels in each category, Represents the total number of all raster pixels covering the study area; d3, use the hazard formula to calculate the landslide hazard in the best time period T in the future, so as to predict the future landslide hazard. The calculation formula is as follows; Where, for The weight assigned to each level category, for The weight assigned to each level category.

2. The method for assessing the spatiotemporal risk of landslides in combination with an effective rainfall model according to claim 1, wherein: The landslide inducing factors include terrain factors, geological factors, geomorphic factors, hydrological factors and road factors, and the acquired landslide inducing factors are processed into a unified spatial resolution and spatial coordinate system.

3. The method for assessing the spatiotemporal risk of landslides in combination with an effective rainfall model according to claim 1, wherein: The step 1 includes the following sub-steps: Step a1, obtaining landslide inducing factors; Step a2, constructing landslide and non-landslide binary classification data; Step a3: Analyze the relationship between landslide-inducing factors and landslides.

4. The method for assessing the spatiotemporal risk of landslides in combination with an effective rainfall model according to claim 3, wherein: In step a2, landslide and non-landslide binary classification data is constructed. The specific implementation includes the following sub-steps: a2.1 Obtain regional landslide data; a2.2 Obtain the same amount of non-landslide data as landslide data and construct the landslide binary classification data.

5. The method for assessing the spatiotemporal risk of landslides in combination with an effective rainfall model according to claim 4, wherein: Based on the acquired landslide and non-landslide binary classification data points, the attributes of the acquired landslide inducing factors are extracted, and finally the landslide sample data is obtained, which includes the landslide binary classification coding data and the inducing factor attribute data of the corresponding landslide and non-landslide points.

6. The method for assessing the spatiotemporal risk of landslides in combination with an effective rainfall model according to claim 3, wherein: Step a3 analyzes the relationship between landslide-inducing factors and landslides, and the specific implementation includes the following sub-steps: a3.1 For the acquired landslide sample data, the collinearity test method is used to analyze the correlation between landslide-inducing factors and exclude landslide-inducing factors with collinear relationships; a3.2 The importance index of independent landslide-inducing factors was calculated using the information gain analysis method. Factors with an importance index less than 0.1 were eliminated, and the newly obtained landslide-inducing factors were sorted in descending order according to their importance index. This resulted in the acquisition of new, ordered landslide-inducing factors without collinearity. These factors were then combined with the corresponding binary landslide data and divided into training and test sets.

7. The method for assessing spatiotemporal landslide risk in conjunction with an effective rainfall model according to claim 1, wherein: The step 2 includes the following sub-steps: Step b1, extract the main features based on the CNN network: In step b2, the RNN recurrent neural network constructs a landslide susceptibility model, predicts the spatial probability of landslides and obtains a landslide spatial probability map.

8. The method for assessing the spatiotemporal risk of landslides in combination with an effective rainfall model according to claim 7, wherein: In step b2, the specific implementation is as follows: Based on the CNN network training results in step b1, independent and ordered landslide-inducing factors are obtained as the input data of the RNN network, and the output data is the landslide binary classification coding data, thereby constructing a landslide susceptibility model. The RNN includes three recurrent neural networks: LSTM, GRU and SRU. The feature results extracted by the three RNN networks are , as the input data of the RNN network, the calculation process after entering the hidden layer is as follows: In the formula, RNN represents its three variant networks, namely , , are the activation functions of the hidden layer and the output layer respectively, u , v , w is the network parameter, b is the bias parameter, for t The output of the hidden layer at this moment, The landslide features extracted by the CNN network are trained in the three RNN networks mentioned above to obtain the landslide spatial probability index and generate three landslide spatial probability maps. .

9. The method for assessing spatiotemporal landslide risk in conjunction with an effective rainfall model according to claim 1, wherein: The step 3 includes the following sub-steps: Step c1: In the ST-Transformer network, an effective rainfall model is embedded to calculate the effective rainfall amount of the input rainfall data of different periods to obtain the effective rainfall time series data of different periods; In step c2, a CNN and Bi-LSTM network are embedded in the encoder structure of the Transformer network. First, the main feature information of effective rainfall is extracted through the CNN convolutional neural network. Based on this, the Bi-LSTM network is embedded in the position encoding structure to extract the spatial feature information of effective rainfall data. Then, one-hot encoding and the CNN convolutional neural network are combined to extract the temporal features of effective rainfall. In step c3, a spatiotemporal attention mechanism is proposed to fuse the spatiotemporal features of the effective rainfall data extracted above, and use them as the input data of the decoding layer to predict the time probability of landslide. The prediction results are evaluated by using three error indices: mean absolute error, mean square error, and root mean square error. The time T value with the smallest error index is selected as the optimal time for predicting the landslide time probability.

Images