CN-122020563-A - Method for evaluating landslide susceptibility based on space-time multi-scale feature fusion deep learning of digital double driving

Abstract

A method for evaluating the susceptibility of a digital object-based double-drive space-time multi-scale feature fusion deep learning landslide is characterized in that a space-time fusion DL-LSA model of 'time sequence branch + Swin-transducer (upper branch) +CNN (lower branch)' is constructed, and a complete technical scheme of 'high-quality sample screening-deep space-time feature extraction-multi-scale self-adaptive fusion-dynamic probability prediction' is formed by combining an AFF multi-scale space fusion module and a TSF space-time fusion module, so that the geological disaster risk management is promoted to upgrade from single drive to digital object double drive and from static evaluation to dynamic early warning. Comprises 1, obtaining an influence factor (X). 2. An initial sample set (X-Y pair) is constructed. 3. Influence factor (X) pretreatment. 4. Influence factor encoding (generating normalized features X'). 5. Physical constraints (optimized sample set X' -Y) are provided based on the P-LSA model. 6. And constructing a space-time fusion DL-LSA model. 7. And designing a feature fusion module. 8. And outputting landslide susceptibility evaluation results.

Inventors

• HUANG YU
• XUE YUXUAN
• Liu kankan

Assignees

• 同济大学

Dates

Publication Date

20260512

Application Date

20260212

Claims (9)

1. 1. A method for evaluating the susceptibility of a landslide based on space-time multi-scale feature fusion deep learning of a plurality of objects is characterized by comprising the following steps: step S1, influencing factor X acquisition Determining 5 kinds of influence factors causing landslide disasters, and entering a step S2; Step S2, initial sample set X-Y pair construction Generating negative samples Y=0 by randomly selecting points in a buffer area, combining the positive samples and the negative samples to form an initial sample set X-Y pair, and entering step S3; Step S3 influence factor X pretreatment Judging the relevance by adopting the pearson correlation coefficient and the VIF variance expansion factor, and eliminating redundant influence factors until the absolute value of all influence factors in the initial sample set in the step S2 is <0.8 and the VIF is <10; s4, generating standardized characteristic X 'by influence factor coding' Coding two kinds of discrete influence factors and continuous influence factors respectively to generate a standardized characteristic X ', and verifying rationality by using a frequency ratio method to form a sample set X' -Y pair; step S5, providing a physical constraint optimization sample set X' -Y based on the P-LSA model The key mechanical parameters are associated with the characteristic X ', fs is solved through a correction formula, priori constraint is provided through a P-LSA model, the vulnerability grade is divided according to the magnitude of the Fs, positive/negative samples falling in the range of the extremely high/low vulnerability area in the step S4 are reserved, and a high-quality X' -Y sample set is formed; Step S6, constructing a space-time fusion DL-LSA model Based on the standardized feature X ' in the step S4, capturing time sequence features by adopting an LSTM network to generate a global time sequence feature vector S, extracting global Gl-local space features Ll by adopting ' Swin-transducer+CNN ' upper and lower branches, and carrying out multiple rounds of learning iteration; step S7, designing a feature fusion module On the output side of the space-time fusion DL-LSA model, feature fusion is further completed by serially connecting an AFF module and a TSF module, the AFF module realizes multi-scale space feature Gl and Ll fusion to form a space fusion feature Hl, and the TSF module realizes space-time feature fusion; Step S8, outputting landslide susceptibility evaluation results And outputting model prediction probability, drawing an evaluation result graph, grading a monitoring area, and verifying the result based on the sample set label in the step S5.
2. 2. The method for evaluating the susceptibility of a landslide based on the space-time multi-scale feature fusion depth learning of a digital double drive as claimed in claim 1, wherein in the step S1, the influence factor X acquires a focusing southeast coastal landslide disaster-causing mechanism, and 5 major categories of 16 LSA core influence factor (X) data are acquired.
3. 3. The method for evaluating the susceptibility of a landslide based on the space-time multi-scale feature fusion depth learning of a digital dual drive as set forth in claim 1, wherein in the step S2, an initial sample set X-Y pair is constructed by taking an X value+Y label as a core, a complete initial sample set is constructed, and the problem of sample scarcity is solved by S2.1, and positive sample Y=1 acquisition and expansion S2.1.1 taking landslide points of field investigation/history records of a research area as original positive samples, extracting corresponding 16 influence factors X to take values to form an original X-Y pair Y=1 S2.1.2 generating 29N positive samples using SMOTE oversampling S2.1.2.1 for each positive sample (I=1, 2,., N) find its k nearest positive sample neighbors in feature space using euclidean distances, which are given by: As a dimension of the features, , For the sample , At the position of The value of each feature; s2.1.2.2 randomly selecting a sample from k nearest neighbors New samples were synthesized by linear interpolation: Wherein, the Original positive samples; random neighbor samples, wherein delta is a random number between 0 and 1; S2.1.2.3 repeating the above process until 29N synthetic samples are generated; S2.2, negative sample y=0 selection S2.2.1, taking the geographical position of the extended positive sample as the center, and dividing an 800-meter buffer area; s2.2.2, randomly selecting points according to 5 times of the number of positive samples in the area outside the buffer area, and marking the points as negative sample labels Y=0; S2.3, generating an initial sample set, namely merging positive and negative samples to form an initial X-Y pair sample set of an influence factor value X+sample label Y, and providing basic data for subsequent S3 preprocessing.
4. 4. The method for evaluating the susceptibility of a landslide based on the space-time multi-scale feature fusion depth learning of a digital double drive as claimed in claim 1, wherein in the step S3, the influence factor X in the step S1 is preprocessed, and the influence factor actually suitable for entering the step S4 is screened out, and the method comprises the following specific steps: S3.1, pearson correlation coefficient screening S3.1.1, pearson correlation coefficient matrix calculation Calculating any number of influence factors in the influence factor dataset X And Pearson correlation coefficient between The formula is: Wherein: First of all Influencing factors in individual samples Taking a value; influence factor Is the average value of (2); the total number of samples; S3.1.2, correlation determination If | The I is more than or equal to 0.8, the two influence factors are proved to be in strong linear correlation, and factors with fuzzy physical significance are removed; S3.2, VIF collinearity screening S3.2.1 construction of a Linear regression model To influence the factor data X by each factor As dependent variables, all other factors are taken as independent variables, and a linear regression model is constructed: = + + +...+ + +...+ + Wherein: dependent variables; Independent variable; The total number of influencing factors; intercept items; Regression coefficients; Error terms; s3.2.2, calculating a determination coefficient R2 Wherein: Factor in kth sample Is a raw observation of (1); Calculated by linear regression model Is a function of the estimated value of (2); S3.2.3 to calculate the VIF value of each factor The calculation formula of the VIF is as follows: Wherein: determining coefficients of linear regression model in S3.2.2 S3.2.4, collinearity determination If VIF is more than or equal to 10, indicating that the factor and other factors have serious collinearity, and eliminating the influencing factor; S3.3 iterative checking Repeating steps 3.1-3.2 until all influencing factors | I <0.8 and VIF <10.
5. 5. The method for evaluating the susceptibility of a landslide based on space-time multi-scale feature fusion depth learning of a digital double drive as claimed in claim 1, wherein in the step S4, the influence factor set X after pretreatment in the step S3 is divided into two types of discrete influence factors and continuous influence factors according to the spatial distribution characteristics for coding to generate a standardized feature X', and the coding rationality is verified by adopting a frequency ratio method: s4.1, discrete influence factors, namely based on inherent classification system coding, the larger the coding value is, the stronger the landslide promotion effect is: s4.2, grading continuous influence factors by adopting a natural breakpoint method, wherein the larger the coding value is, the stronger the promotion effect is: s4.2.1 sorting the data, namely sorting all sample values of the continuous influence factors from small to large S4.2.2, breakpoint calculation, namely automatically dividing intervals by a statistical method of minimizing intra-class variance and maximizing inter-class variance; s4.2.3, standardized coding, namely assigning values to the divided intervals, wherein the larger the value is, the stronger the promotion effect is The rainfall intensity I reflects typical characteristics of storm in the southeast coastal typhoon period, and the calculating method comprises the following steps: Wherein the method comprises the steps of For a certain period of rainfall of mm, For a duration h of rainfall; Effective accumulated rainfall for 7 days Reflecting the accumulation effect (eliminating ineffective loss such as surface runoff, evaporation and the like), and calculating the method: Wherein, the (Daily decay coefficient of moisture content), Rainfall mm for day i; TWI calculation method: TWI = ln(As / tanβ) Wherein As is the water collecting area m2/m of unit contour length and reflects the upstream water collecting capacity, tan beta is the tangent value of the gradient and reflects the potential of the water flow speed; The SPI calculating method comprises the following steps: SPI = As×tanβ Wherein As is the water collecting area m2/m of unit contour line length, namely the flow accumulated value is multiplied by the grid resolution; NDVI formula: Wherein, NIR is near infrared band reflectivity, R is red band reflectivity If rationality needs to be enhanced, frequency ratio method FR is adopted for verification: Wherein is The number of landslide grids in the j-th class of the i-th influence factor, To investigate the total grid number of the landslide, The j-th class total grid number is the i-th influencing factor, S is the research area total grid number, The frequency ratio of the jth class of the ith influence factor; If it is 1, The interval is positively correlated with landslide; If 0.5 is less than or equal to Less than or equal to 1, weaker correlation, but no merging; If it is < 0.5. The section has little correlation with landslide and needs to be merged with the adjacent section.
6. 6. The method for evaluating the susceptibility of a landslide based on space-time multi-scale feature fusion depth learning based on a number double driving of a plurality of objects as claimed in claim 1, wherein in the step S5, a P-LSA model is constructed to carry out LSA, and reasonable positive/negative samples are screened from a very high/very low susceptibility region respectively to form a high-quality sample set X' -Y sample pair; the S5.1, P-LSA parameters are associated with the feature X '6 key mechanical parameters c' 、 、h、 、 All from the quantized assignments of feature X', S5.2, correction formula of P-LSA model (weight-based influence factor fusion comprehensive judgment value) Wherein: Land utilization correction coefficient (construction land taking-0.2, forest land taking +0.15); vegetation coverage correction coefficients; a rainfall intensity correction coefficient; S5.4, sample screening Reserving regions of extremely high susceptibility Positive samples y=1 in the range, and very low region Negative samples y=0 in the range, forming a high quality X' -Y sample set.
7. 7. The method for evaluating the susceptibility of a landslide based on space-time multi-scale feature fusion depth learning based on digital dual driving as claimed in claim 1, wherein in the step S6, a space-time fusion DL-LSA model is constructed into a time sequence branch and space dual-path branch architecture, a feature map is constructed based on standardized features X' generated in the step S4, deep space-time features are extracted autonomously, and dynamic dependency capture and full coverage of multi-scale space features are realized: S6.1, time sequence branching; s6.1.1, time series data structuring Converting the dynamic influence factors in S4, namely 7-day effective precipitation P e and standardized characteristics X' corresponding to rainfall intensity I into a t-moment structured characteristic vector: Wherein: Standardized characteristics corresponding to 7-day effective precipitation at t moment; the standardized characteristics corresponding to rainfall intensity at the moment t; is the characteristic dimension% ); S6.1.2, data normalization To the original influencing factor Performing min-max normalization, eliminating dimension difference, and adapting to LSTM input requirements: Wherein: s4, 7 days of effective precipitation and a standardized feature global minimum/maximum value corresponding to rainfall intensity; S6.1.3, sliding window method construction sequence: S6.1.3.1, setting the size w=7 days (adapting to the cumulative disaster effect of rainfall) of a sliding window, wherein the sliding step is 1 day, slicing the continuous time sequence features into input sequences with fixed length (7), each sequence corresponds to 1 evaluation day, and the window contains feature data of 7 continuous days: Wherein the method comprises the steps of Is an input sequence matrix of LSTM, (K=0, 1,.,. 6) is to evaluate the time series feature vector normalized by the current k days; to evaluate the time sequence feature vector normalized by the same day, m=2 is the feature dimension; s6.1.3.2 mapping the single-day feature vector in the window to the target dimension through the full connection layer: Wherein: Mapping a weight matrix; Is a bias vector; Activating a function for a ReLU; d is the feature dimension after mapping; S6.1.3.3, constructing an input matrix of LSTM: Wherein: The method comprises the steps that 7 days of mapped feature vectors in a sliding window are formed in time sequence, and LSTM is directly input to perform time sequence modeling; the single-day feature vector is mapped on the same day; S6.1.4, LSTM timing feature modeling The LSTM dynamically updates the cell state through a gating mechanism, captures the time sequence dependency of the single day feature in the window, and has the following formula: forgetting the door: an input door: Candidate cell status: Cell status update: Output door: Hidden state: Wherein: : a daily LSTM hidden state; Inputting a feature on the same day; : LSTM cell status; gating the weight matrix; Gating the bias vector; hadamard product; The Sigmoid activates a function that, A hyperbolic tangent function; After 7 days of data modeling in the sliding window, the LSTM layer outputs a full sequence hiding state set: Wherein: a full sequence hidden state matrix with dimension W x d, which comprises hidden states within 7 days of a window; the dimension is d; s6.1.5 attention mechanism Global timing feature aggregation By attention weighted aggregation, highlighting features of key dates within a window, global timing features are generated: s6.1.5.1 calculating daily hidden status Is a concentration score of (2): Wherein, the 、 、 K is the hidden layer dimension, which is the attention parameter; s6.1.5.2, converting the score into normalized weights by Softmax function: S6.1.5.3, weighted summation generates a global timing feature vector: S6.2 spatial branching Taking the multi-channel standardized feature map X ' generated in the step S4 as input, constructing a double-path structure of ' upper branch global space feature extraction and lower branch local space feature extraction ', wherein the double-path structure corresponds to mass-sending rules, cracks and deformation mechanisms of southeast coastal areas respectively; S6.2.1, upper branch, global spatial feature extraction S6.2.1.1, patch blocking and embedding Normalizing a multi-channel feature map For input, non-repeatedly blocked Is flattened into a vector sequence through sliding window partitioning, is arranged according to a space sequence, =Unfold(X’, kernel_size=P, stride=P) Wherein Unfold has an output shape of Kernel size is the kernel size, stride is the step size And then embedding a high-dimensional space into the linear mapping layer: The Linear layer adjusts the Patch vector dimension into a transform hidden dimension C and outputs ; S6.2.1.2, hierarchical Swin Transformer Block global feature extraction The global features G0, G1, G2, G3 are generated step by multi-stage combining PATCH MERGING downsampling (halving resolution per stage, doubling channel number): Wherein: is the first Stage input features; First of all A level global feature; s6.2.2 lower branch, local spatial feature extraction S6.2.2.1, hierarchical CNN local feature extraction To the same standardized characteristic diagram After P x P Patch blocking, local details are extracted layer by layer through a convolution block consisting of a convolution layer, a batch normalization layer BN+ReLU activation function, and the structure is as follows: wherein W is convolution kernel weight, b is bias term, Representing a convolution operation; output shape: Finally, multi-scale local features L0, L1, L2 and L3 aligned with the upper branch level are generated, the resolution of the multi-scale local features is consistent with the resolution of the corresponding global features G0-G3, and the complete adaption of the dimensions of subsequent spatial feature fusion is ensured.
8. 8. The method for evaluating the susceptibility to landslide based on the space-time multi-scale feature fusion depth learning of the digital dual drive as set forth in claim 1, wherein in the step S7, a feature fusion module is designed The serial AFF module realizes the space double-branch feature fusion and the TSF module realizes the space-time feature fusion; S7.1, spatial up-down branch feature fusion AFF module S7.1.1 core logic The global spatial features G0, G1, G2 and G3 output by the upper branch and the local spatial features L0, L1, L2 and L3 output by the lower branch are in one-to-one correspondence, and the serial AFF modules realize the spatial upper and lower branch feature fusion, wherein the front and rear of each AFF module are sequentially connected, and the fusion features output by the front AFF module are input together with the global spatial features and the local spatial features of the next stage to obtain the fusion features of the current stage; S7.1.2, initial feature fusion First, the same-level global features And local features Element-by-element addition to form a hierarchical initial fusion feature The formula is as follows: s7.1.3 Multi-scale channel attention extraction channel characteristics MS-CAM For initial fusion features Designing a global path and local path dual-path structure, respectively extracting channel characteristics, and providing a characteristic basis for subsequent AFF weight generation; s7.1.3.1 global path The light-weight design of dimension reduction-activation-dimension increase reduces the complexity of calculation and the interference of redundant information, highlights the key channel characteristics and enhances the effectiveness of the channel attention; S7.1.3.1.1 first pair Global average pooling GAP is performed, compressing the input features h×w×c to 1×1×c, each channel feature expressed as an overall spatial region average: i is a feature map row index, j is a feature map column index, and c is a feature channel index; Is that Is the space height and width of the channel, C is the number of channels; s7.1.3.1.2, reducing the dimension of the convolution of 1 multiplied by 1 channel by channel, and reducing the dimension of the channel number from C to C/r, wherein the formula is as follows: wherein W1 is a convolution weight matrix having the shape of (C/r) x C; s7.1.3.1.3, placing a ReLU activation function in the middle of each channel to realize nonlinear mapping, enhancing the expression capacity of the model on the characteristics, and adopting the following formula: BN represents a batch normalization operation; s7.1.3.1.4, channel 1×1 convolution upscale, re-upscale the channel number from C/r back to the original channel number C, recovering feature dimension, and obtaining global path output feature ; W2 is a convolution weight matrix having a shape of C× (C/r) S7.1.3.2 local paths Direct pairing without compressing spatial information Performing '1×1 convolution dimension reduction-BN normalization-ReLU function activation-1×1 convolution dimension increase' operation to obtain local path output characteristics ; The convolution weight C-C/r is 1 multiplied by 1; A convolution weight C/r- & gtC of 1X 1 liter dimension; s7.1.4, dual path output feature weighted fusion Based on a multi-scale channel attention MS-CAM mechanism, the channel weight A is generated by fusing dual-path output characteristics, so that dynamic weighting of initial fusion characteristics is realized, key characteristics are enhanced, redundant information is restrained, and dynamic fusion of global and local characteristics is realized; S7.1.4.1, generating channel attention weights Outputting the global path And local path output And respectively carrying out normalization and addition, and activating by a Sigmoid function to generate a channel attention weight A with the range of 0-1, wherein the formula is as follows: Wherein sigma (·) is a Sigmoid activation function; S7.1.4.2 channel weighting S7.1.4.3 residual connection By residual connection, the original global features And local features And S7.1.4.2 channel weighting features Adding, ensuring the integrity of original characteristic information, thereby enhancing the training stability of the network, and finally outputting the fusion characteristic of the current-level AFF module The formula is as follows: Wherein, the Fusion characteristics output by the current-stage AFF module and the global characteristics of the next layer And local features Together as input to the next AFF module; s7.2, space-time feature fusion TSF module The final spatial fusion feature H 3 output by the depth coupling S7.1 and the global time sequence feature S output by the S6.1.5 solve the problem of dimension isomerism of the spatial feature being a high-dimensional tensor and the time sequence feature being a one-dimensional vector, and output the unified representation Z of the fusion space-time disaster causing information; s7.2.1, channel alignment, semantic dimension matching Through the full connection layer, the global time sequence feature S output by S6.1.5 is projected to a channel dimension consistent with the final space fusion feature H 3 output by S7.1, so that semantic isomerism is eliminated: Wherein, the In order to project the weight matrix, Is a projection bias term; D is the dimension of the global time sequence feature S; S7.2.2 time sequence feature broadcast, space dimension expansion Time sequence characteristics after channel dimension alignment Broadcast replication in the spatial dimension, covering all spatial positions of H 3 , implementing pixel-level spatio-temporal feature matching, the formula is as follows: wherein, broadcast @ is ) For spatial broadcast operations, post-broadcast timing characteristics Example H 3 has a spatial resolution of 10×10×C, then S' is replicated 100 times; S7.2.3 initial fusion of spatio-temporal features To post-broadcast timing characteristics Adding the fusion characteristic H 3 to realize preliminary fusion to obtain a composite characteristic : S7.2.4, gating weighted fusion, dynamic feature screening S7.2.4.1 gating weight calculation Wherein, the A time sequence characteristic exclusive learnable channel weight matrix; spatial feature exclusive learnable channel weight matrix; Bias items; sigmoid activation function, output and output Gating weight G with dimension completely matched; s7.2.4.2, weighted fusion Weight is gated With initial fusion features Element-by-element dot multiplication, strengthening key space-time characteristics and inhibiting redundant information: Element-by-element dot product operation S7.2.5 residual connection Weighting gated features With original spatial features And timing characteristics Adding, retaining original information, enhancing fusion effect, and obtaining final space-time comprehensive characteristics ; 。
9. 9. The method for evaluating the susceptibility of landslide based on the space-time multi-scale feature fusion deep learning of the double driving of a plurality of objects as claimed in claim 1, wherein in the step S8, the susceptibility of landslide is evaluated by probability calculation, result drawing, grade division and verification analysis S8.1, evaluating each grid unit in the research area by using the space-time fusion DL-LSA model trained by the S6 and the S7, and updating to obtain a landslide occurrence probability value of each unit; s8.2, drawing LSA result according to the probability obtained in S8.1, grading, dividing and verifying analysis S8.2.1, corresponding landslide occurrence probability values to geographic space positions of a research area one by one, and generating a landslide susceptibility distribution map; S8.2.2, carrying out vulnerability classification on the result graph by adopting a natural breakpoint method, and dividing the result graph into five vulnerability classes of extremely low, medium, high and extremely high to obtain a final classification graph; s8.2.3 sorting the final compartmental map data into a table, high quality sample tags output as S5 And (3) carrying out validity verification according to the area ratio, the landslide point duty ratio and the landslide ratio of the final regional map, and obtaining a final LSA result after verification.

Description

Method for evaluating landslide susceptibility based on space-time multi-scale feature fusion deep learning of digital double driving Technical Field The invention belongs to the field of landslide geological disaster risk assessment. Background Landslide is one of the most damaging disasters worldwide, and seriously threatens the safety of human lives and properties. According to statistics (global natural disaster evaluation report), landslide disasters are easy to occur frequently in recent years, the number of dead people, the influence population and economic losses are obviously increased, but the found landslide disaster points are less than 20%, and the risk prevention and control situation is extremely severe. Landslide Susceptibility (LS) refers to the likelihood of an area to slip under certain geological conditions. Landslide susceptibility evaluation (LSA) provides key support for decision links such as regional homeland space planning, engineering site selection optimization and the like, plays an important role in disaster prevention and reduction, and according to an evaluation method, the LSA can be divided into a physical-based model (P-LSA) and a data-driven (D-LSA) model, wherein the D-LSA model takes the dominant role for a long time. Deep Learning (DL) is constructed by a multi-layer neural network, feature information (surface deformation features) can be automatically extracted from original data (such as SAR images) without or with little preprocessing, and nonlinear space-time correlation is captured, so that the method has become the most advanced D-LSA model core technology. The Convolutional Neural Network (CNN) is suitable for the fields of image recognition, target recognition and the like. A transducer model based on Attention mechanism (Attention) is adapted to multi-mode information fusion and is mainly applied to the field of computer vision. Long and Short Term Memory (LSTM) is outstanding in time sequence data processing, information flow is controlled through a gating mechanism, and model training efficiency and prediction effect are improved. The coastal areas of southeast China are landslide high-incidence gathering zones, hills are densely distributed, coastal zones are staggered, and soft soil and thick granite residual soil are widely distributed. Earthquake and extreme weather events (such as typhoons and storm) are easy to occur frequently, and engineering activities are active. Under the multi-factor coupling effect, landslide geological disasters have the characteristics of wide distribution, strong concealment and high time-varying property. Thus, the traditional LSA model is poorly adapted, and application in this area presents challenges: 1. landslide in southeast coastal areas is widely distributed and frequently occurs in remote mountain areas, and data are scarce; 2. Landslide is mainly shallow and small in volume, high in concealment, capable of overlapping a high-humidity vegetation environment, limited in optical/SAR image monitoring capability, easy to cause 'missing detection and misjudgment', and large in deviation from actual conditions; 3. The regional disaster-tolerant environment is complex, and the D-LSA model has 'black box effect', is easy to be fitted in a rare data scene and has low precision; 4. the geological condition of the area is fragile and is extremely easy to be influenced by environmental changes (mainly heavy rainfall), the area belongs to a time-varying system, the existing LSA model is mainly static evaluation, and the result is disjointed with the actual disaster-causing process; 5. the regional landslide simultaneously comprises a macroscopic mass-sending landslide law, a microscopic deformation and crack disaster-causing mechanism and obvious size difference, and the existing LSA model adopts a single grid size and cannot give consideration to the multi-scale characteristics; 6. The high-dimensional data (such as InSAR and LiDar data) causes the computational complexity of the traditional transducer model to increase in a quadratic way, and the operation efficiency is low, so that the engineering application is difficult. The prior art has the core defects of rare samples, large deviation, insufficient model interpretation, inconsistent static evaluation and dynamic disaster-causing process, difficult multi-scale feature fusion and large calculation load, and a high-precision LSA method suitable for southeast coastal areas is needed. Disclosure of Invention In view of the technical problems of the background technology, the invention provides a method for evaluating the susceptibility of landslide based on space-time multi-scale feature fusion deep learning of a digital object double-drive, which aims to break through the bottleneck of the prior art, solve a series of key problems faced by the prior LSA technology in practical application in the area, including the problems of data scarcity, large deviation, inconsistent sta