Landslide susceptibility assessment system and method based on statistical screening and deep learning

Abstract

The application discloses a landslide susceptibility assessment system and method based on statistical screening and deep learning, and relates to the technical field of landslide susceptibility assessment. The method comprises the following steps: constructing an initial factor set of landslide influencing factors, and initializing a candidate factor set; selecting a factor with the largest statistical value, testing the variance ratio and significance, and obtaining an optimal factor combination; aligning and stacking the layers in the optimal factor combination in space, and constructing a multi-channel input feature tensor; constructing a spatial regression neural network with an encoder-decoder structure, wherein the encoder takes a residual network as the main trunk, and is adapted to the multi-channel input feature; the decoder fuses the multi-scale features of the encoder through upsampling operation and skip connection, and outputs a susceptibility prediction map; the multi-channel input feature is inferred, a landslide susceptibility map is generated, the susceptibility level is divided, and a partition result is output. The application supports susceptibility assessment of different regions and different types of landslides based on various influencing factors.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of landslide susceptibility assessment, in particular to a landslide susceptibility assessment system and method based on statistical screening and deep learning. BACKGROUND

[0002] As a geological disaster with extremely strong destructive power, landslide susceptibility assessment is the core foundation of the disaster prevention and mitigation system. Traditional statistical methods, such as the analytic hierarchy process, rely too much on expert experience, and the factor weights are highly subjective, making it difficult to objectively quantify factors and their interactions. Traditional statistical screening methods can only identify linear correlations and ignore high-order nonlinear interactions between factors such as faults, rivers, and seismic parameters and terrain, resulting in a lack of geoscience interpretability in the assessment results. Although deep learning methods can model nonlinear relationships, the "end-to-end" black box nature makes them less stable and difficult to reproduce results, and they are not effective in handling invalid values that are common in geographic data, reducing the accuracy of assessments in real geographic environments. SUMMARY

[0003] The present application relates to the technical field of landslide susceptibility assessment, in particular to a landslide susceptibility assessment system and method based on statistical screening and deep learning.

[0004] To achieve the above-mentioned purpose, the present application provides the following technical scheme: a landslide susceptibility assessment method based on statistical screening and deep learning, the method comprising:

[0005] S1, constructing an initial factor set of landslide influencing factors and initializing the candidate factor set as an empty set;

[0006] S2, screening factors in the initial factor set based on spatial differentiation statistics; selecting the factor with the largest statistical value, testing its variance ratio and significance; if the significance level meets the standard and the variance ratio meets the optimization condition, adding it to the candidate factor set; otherwise, geographically interacting with the remaining factors in the initial factor set to generate interaction factors and add them to the candidate factor set; repeat the screening process until all factors are processed to obtain the optimal factor combination;

[0007] S3, aligning and stacking the layers in the optimal factor combination to construct a multi-channel input feature tensor; generating a uniform binary mask based on the effective pixel area, filling the invalid values in the features and labels with zeros and normalizing them;

[0008] S4, constructing a spatial regression neural network with an encoder-decoder structure, the encoder uses a residual network as the backbone and adapts to multi-channel input features, the decoder fuses the multi-scale features of the encoder through upsampling operations and jump connections, and outputs a single-channel susceptibility prediction map with the same resolution; using a mask weighted loss function to train the spatial regression neural network;

[0009] S5, select the model weight with the minimum loss on the validation set as the prediction model, perform inference on the global multi-channel input features, generate a landslide susceptibility map, and divide the susceptibility levels to output the partition result.

[0010] According to the above scheme, the landslide influence factors include elevation, slope, slope direction, peak ground acceleration, distance to river, distance to fault, vegetation cover index, and distance to road;

[0011] The grid layers corresponding to the landslide influence factors are spatially coordinated and resampled in resolution to form a spatially consistent initial factor set;

[0012] The alternative factor set is an empty set that does not contain any landslide influence factors and factor interaction combinations, and is used to store the optimal factors and factor interaction combinations with statistical significance obtained through screening.

[0013] According to the above scheme, step S2 includes:

[0014] S2-1, calculate the spatial differentiation statistical value of each factor in the initial factor set, and the formula is as follows:

[0015] ;

[0016] Wherein, q represents the spatial differentiation statistical value; L represents the total number of layers; h represents the layer index; N h represents the sample size of the hth layer; represents the intra-layer variance of the hth layer; N represents the total sample size; σ 2 represents the total variance; SSW represents the within-group sum of squares; SST represents the total sum of squares; the value range of q value is [0, 1], q=1 indicates that the factor completely controls the spatial distribution of landslide susceptibility, and q=0 indicates that the factor is completely irrelevant to the distribution of landslide susceptibility;

[0017] Within-group sum of squares, the formula is as follows:

[0018] ;

[0019] Wherein, k represents the number of groups; i represents the group index; n i represents the sample size of the ith group; j represents the intra-group sample index; x ij represents the observation value of the jth sample in the ith group; represents the mean value of the ith group;

[0020] Total sum of squares, the formula is as follows:

[0021] ;

[0022] wherein, is expressed as the overall mean;

[0023] S2-2, select the factor with the largest spatial heterogeneity statistic value, calculate the variance ratio of the factor with the largest spatial heterogeneity statistic value, and the formula is as follows:

[0024] ;

[0025] wherein, VR is expressed as the variance ratio; SSB is expressed as the between-group sum of squares; k is expressed as the number of groups; k-1 is expressed as the numerator degrees of freedom; and N-k is expressed as the denominator degrees of freedom;

[0026] The between-group sum of squares is calculated according to the following formula:

[0027] ;

[0028] wherein, n i is the number of the ith group;

[0029] S2-3, when the statistical significance of the factor reaches the preset threshold and the variance ratio meets the optimization condition, the factor is added to the candidate factor set;

[0030] S2-4, when the statistical significance of the factor does not reach the preset threshold or the variance ratio does not meet the optimization condition, the factor is respectively subjected to geographical interaction coding with each of the remaining factors in the initial factor set, and the geographical interaction coding includes: taking the grid layers corresponding to the two factors, judging each pixel position in the grid, if any pixel value is 0 or the no-data value of the corresponding grid, the pixel of the output grid is marked as the no-data value, otherwise, a unique positive integer label is generated through the bijective mapping, an interaction factor is generated, and the interaction factor is added to the candidate factor set;

[0031] S2-5, the factor with the largest spatial heterogeneity statistic value is removed from the initial factor set, and steps S2-1 to S2-4 are repeatedly executed until the initial factor set is empty, and the candidate factor set is obtained as the optimal factor combination.

[0032] According to the above scheme, the optimization condition includes that the variance ratio of the current step is greater than or equal to the variance ratio of the previous step; the variance ratio of the current step includes the variance ratio after adding the candidate factor in the current iteration step; and the variance ratio of the previous step includes the variance ratio before adding the candidate factor in the current step.

[0033] According to the above scheme, step S2-4 includes:

[0034] The grid layer corresponding to the currently screened factor is a first input grid R1, and the grid layer corresponding to another interaction factor in the initial factor set is a second input grid R2;

[0035] If the pixel value v1 of the first input raster R1 at position (i,j) is equal to 0, or the pixel value v2 of the second input raster R2 at position (i,j) is equal to 0, or v1 is equal to the no-data value nodatal of the first input raster R1, or v2 is equal to the no-data value nodata2 of the second input raster R2, then the output raster R out In this pixel position, the value is marked as the no-data value nodatal of the first input raster;

[0036] Otherwise, the pixel value pair (v1, v2) is mapped to a unique positive integer label by the bijective mapping function L(v1, v2), and the label is taken as the output raster R out In this pixel position, the value is marked as the no-data value nodatal of the first input raster;

[0037] The bijective mapping function L(v1, v2) is realized by a dictionary mapping mechanism, which assigns a unique integer label starting from 1 to each newly appearing unique value pair (v1, v2);

[0038] The interaction factor raster generated by encoding is added to the candidate factor set, and the geographic interaction encoding of the current factor with one factor in the initial factor set is completed; the interaction process is repeated until the geographic interaction encoding of the current factor with all other factors in the initial factor set is completed.

[0039] According to the above scheme, step S3 includes:

[0040] S3-1, aligning the coordinates and stacking the raster layers corresponding to the single factor in the optimal factor combination and the raster layers corresponding to the interaction factor in the spatial dimension, constructing a multi-channel input feature tensor, the number of channels of the multi-channel input feature tensor is the total number of single factors and interaction factors in the optimal factor combination;

[0041] S3-2, based on the effective pixel area corresponding to the input layer and the landslide susceptibility label layer of all multi-channel input feature tensors, a unified binary mask is created, the mask value corresponding to the effective pixel area is set to 1, and the mask value corresponding to the invalid pixel area is set to 0;

[0042] S3-3, fill the invalid values in the multi-channel input feature tensor and the landslide susceptibility label layer with 0 uniformly, and normalize the valid pixel data in the multi-channel input feature tensor and the landslide susceptibility label layer.

[0043] According to the above scheme, step S4 includes:

[0044] S4-1, construct a spatial regression neural network of an encoder-decoder structure, the encoder adopts a pre-trained residual network as a backbone network, modifies a first convolutional layer of the pre-trained residual network to adapt to the channel number of a multi-channel input feature tensor, and reuses the pre-trained weights of the pre-trained residual network; the encoder extracts multi-scale feature maps through step-by-step down-sampling;

[0045] S4-2, the decoder constructs an up-sampling path symmetric to the encoder, the up-sampling path is composed of a plurality of up-sampling blocks, each up-sampling block contains one up-sampling operation and two modules executed in sequence, each module sequentially completes convolution operation, batch normalization processing and activation function transformation;

[0046] S4-3, at each up-sampling stage of the decoder, the feature map of the corresponding resolution of the encoder is spliced with the up-sampled feature map in the channel dimension through a jump connection, to fuse high-resolution spatial detail information and deep abstract semantic information;

[0047] S4-4, after the decoder is restored to the original input resolution, a 1×1 convolutional layer is used to compress the multi-channel feature map into a single channel, to generate a landslide susceptibility prediction map with the same resolution as the input feature tensor;

[0048] S4-5, a mask weighted loss function is used to train the spatial regression neural network, the mask weighted loss function calculates the prediction error based on the effective pixel region identified by a unified binary mask.

[0049] According to the above scheme, the mask weighted loss function is a mask weighted mean square error loss function, and the formula is as follows:

[0050] ;

[0051] Wherein, L represents the loss value, Mask(i,j) represents the binary mask value at position (i,j), represents the prediction value of the model at position (i,j), Y(i,j) represents the normalized label value at position (i,j), and ε represents a very small positive value used to prevent the denominator from being zero.

[0052] According to the above scheme, step S5 includes:

[0053] S5-1, during the training of the spatial regression neural network, after each round of training is completed, the mask weighted loss function is used to calculate the root mean square error on the validation set, the change of the root mean square error of the validation set is monitored in real time, the model weight corresponding to the minimum root mean square error of the validation set is saved, and the model weight is loaded as the weight of the prediction model;

[0054] S5-2, apply the spatial regression neural network loaded with the prediction model weight to the global multi-channel input feature tensor, and generate a normalized landslide susceptibility prediction map covering the whole area through inference;

[0055] S5-3, post-process the normalized landslide susceptibility prediction map, and reverse normalize it to the original landslide susceptibility index range, or divide it into high, medium and low susceptibility grades by using the natural breakpoint method or the quantile method; output the landslide susceptibility map after dividing the susceptibility grades, and complete the landslide susceptibility assessment.

[0056] The landslide susceptibility assessment system based on statistical screening and deep learning includes a data management module, an intelligent screening module, a feature engineering module, a deep learning module and an evaluation output module.

[0057] The data management module includes a factor set construction module and an alternative set management module; the factor set construction module is used to unify the spatial coordinate system and resample the resolution of the grid layer corresponding to the landslide influencing factor, forming a spatially consistent initial factor set; the alternative set management module is used to create and maintain an empty alternative factor set, and store the optimal factors and factor interaction combinations with statistical significance obtained through screening;

[0058] The intelligent screening module includes a statistical quantification module and an interaction encoding module; the statistical quantification module is used to calculate the spatial differentiation statistical value and variance ratio of each factor, and test the statistical significance; the interaction encoding module is used to perform geographic interaction encoding operation, and generate interaction factors through bijective mapping;

[0059] The feature engineering module includes a tensor construction module and a mask processing module; the tensor construction module is used to spatially align and stack the grid layers corresponding to the single factors and interaction factors in the optimal factor combination, and construct a multi-channel input feature tensor; the mask processing module is used to create a unified binary mask based on the valid pixel area of all input layers and label layers;

[0060] The deep learning module includes a network architecture module and a training optimization module; the network architecture module is used to construct a spatial regression neural network with an encoder-decoder structure, the encoder takes a residual network as the backbone and adapts to multi-channel input features, and the decoder fuses multi-scale features through upsampling operation and skip connection; the training optimization module is used to train the network using a mask weighted loss function, and the mask weighted loss function calculates the prediction error based on the valid pixel area identified by the unified binary mask;

[0061] The evaluation output module includes a prediction module and a post-processing module. The prediction module is used to load the prediction model weights, perform inference on the multi-channel input feature tensor of the entire domain, and generate a landslide susceptibility prediction map. The post-processing module is used to perform inverse normalization processing and susceptibility level classification on the prediction map and output the final partitioning result.

[0062] Compared with the prior art, the beneficial effects of the present invention are:

[0063] 1. This invention achieves objective factor screening through spatial heterogeneity statistics and variance ratio tests, and actively mines high-order nonlinear interaction factors by combining geographic interaction coding, providing statistical and geographical theoretical basis for the evaluation results;

[0064] 2. This invention employs a unified binary mask and mask-weighted loss function to accurately shield against invalid data interference, ensuring that the model is optimized only in the effective pixel area, thereby improving model stability and global prediction accuracy.

[0065] 3. This invention is based on multiple influencing factors and supports the assessment of the susceptibility of landslides in different regions and of different types. The classification results are intuitive and clear. Attached Figure Description

[0066] Fig. 1 This is a flowchart illustrating the steps of the landslide susceptibility assessment method based on statistical screening and deep learning in this invention.

[0067] Fig. 2 This is a schematic diagram of the landslide susceptibility assessment system based on statistical screening and deep learning of the present invention. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] Example: Figs. 1-2 As shown, this invention provides a technical solution: a landslide susceptibility assessment method based on statistical screening and deep learning, the method comprising:

[0070] S1. Construct an initial set of factors affecting landslides and initialize the candidate factor set to an empty set;

[0071] Specifically, the landslide influencing factors include elevation, slope, aspect, peak ground acceleration, distance to river, distance to fault, vegetation cover index, and distance to road; the landslide influencing factors are selected based on the mechanism of geological disasters, and cover five categories of evaluation indexes, including topography, geological structure, hydrological conditions, seismic motion, and human activities; the grid layers corresponding to the landslide influencing factors are unified in the spatial coordinate system and resampled in resolution to form an initial factor set that is spatially consistent; the alternative factor set is an empty set that does not contain any landslide influencing factor or factor interaction combination, and is used to store the optimal factor and factor interaction combination with statistical significance obtained through screening.

[0072] For example, eight landslide influencing factors, including elevation, slope, aspect, peak ground acceleration, distance to river, distance to fault, vegetation cover index, and distance to road, are selected in the study area; the grid layers corresponding to the factors are unified in the same spatial coordinate system and resolution to form an initial factor set A containing eight layers; at the same time, an empty alternative factor set B is created; this is only an example for illustration and does not limit the scope of the application.

[0073] S2, screening factors in the initial factor set based on spatial differentiation statistics; the screening process uses a stepwise geographical detector to realize factor optimization selection through a cooperative judgment mechanism of q-statistics and variance ratio VR; the factor with the largest statistical value is selected, and its variance ratio and significance are tested; if the significance level meets the standard and the variance ratio meets the optimization condition, the factor is added to the alternative factor set; otherwise, the factor is interactively coded with the remaining factors in the initial factor set to generate an interaction factor and add it to the alternative factor set; the screening process is repeated until all factors are processed, and the optimal factor combination is obtained.

[0074] Specifically, step S2 includes:

[0075] S2-1, calculating the spatial differentiation statistics value of each factor in the initial factor set, as follows:

[0076] ;

[0077] Wherein, q represents the spatial differentiation statistics value; L represents the total number of layers; h represents the layer index; N h represents the sample size of the hth layer; represents the intraclass variance of the hth layer; N represents the total sample size; σ 2 represents the total variance; SSW represents the within-group sum of squares; SST represents the total sum of squares; the value range of q is [0, 1], q=1 indicates that the factor completely controls the spatial distribution of landslide susceptibility, and q=0 indicates that the factor is completely irrelevant to the distribution of landslide susceptibility.

[0078] The within-group sum of squares is calculated as follows:

[0079] ;

[0080] wherein, k represents the number of groups; i represents the group index; n i represents the sample size of the i-th group; j represents the sample index within the group; x ij represents the observation value of the j-th sample in the i-th group; represents the mean value of the i-th group;

[0081] The total sum of squares is calculated according to the following formula:

[0082] ;

[0083] wherein, represents the population mean;

[0084] S2-2, select the factor with the largest value of spatial heterogeneity statistic, calculate the variance ratio of the largest factor, according to the following formula:

[0085] ;

[0086] wherein, VR represents the variance ratio; the variance ratio VR is equivalent to the F statistic in statistics, and the statistical significance of the stratification result is tested by comparing the ratio of the inter-group variance to the intra-group variance; SSB represents the inter-group sum of squares; k represents the number of groups; k-1 represents the numerator degrees of freedom; N-k represents the denominator degrees of freedom;

[0087] The inter-group sum of squares is calculated according to the following formula:

[0088] ;

[0089] wherein, n i is the number of the i-th group;

[0090] S2-3, when the statistical significance of the factor reaches the preset threshold and the variance ratio meets the optimization condition, the factor is added to the candidate factor set; the preset threshold of the statistical significance is the significance level test value corresponding to the variance ratio, which is used to verify the statistical significance of the selected factor, and the p value is calculated from the variance ratio VR;

[0091] The optimization condition includes that the variance ratio of the current step is greater than or equal to the variance ratio of the previous step; the variance ratio of the current step includes the variance ratio after adding the candidate factor in the current iteration step; the variance ratio of the previous step includes the variance ratio before adding the candidate factor in the current step in the previous iteration step;

[0092] S2-4, when the statistical significance of the factor does not reach the preset threshold or the variance ratio does not satisfy the optimization condition, the factor is respectively interactively encoded with each of the remaining factors in the initial factor set. The encoding process systematically assigns an increasing unique integer label to each unique factor value pair by establishing a bijective mapping relationship. The interactive encoding includes taking the corresponding grid layers of the two factors, judging each pixel position in the grid, and if any pixel value is 0 or the corresponding grid has no data value, the pixel marker of the output grid is marked as no data value, otherwise a unique positive integer label is obtained through the bijective mapping, an interactive factor is generated, and the interactive factor is added to the candidate factor set;

[0093] Further, step S2-4 includes:

[0094] The current selected factor corresponds to the first input grid R1, and the other factor to be interacted in the initial factor set corresponds to the second input grid R2;

[0095] If the pixel value v1 of the first input grid R1 at position (i,j) is equal to 0, or the pixel value v2 of the second input grid R2 at position (i,j) is equal to 0, or v1 is equal to the no data value nodata1 of the first input grid R1, or v2 is equal to the no data value nodata2 of the second input grid R2, then the output grid R out The pixel position is marked as the no data value nodata1 of the first input grid;

[0096] Otherwise, the pixel value pair (v1, v2) is mapped to a unique positive integer label through the bijective mapping function L(v1, v2), and the label is taken as the output grid R out The value of the pixel position;

[0097] The bijective mapping is realized through a dictionary mechanism, ensuring that each unique (v1, v2) value pair has a one-to-one correspondence with a unique integer label. The bijective mapping function L(v1, v2) is realized through a dictionary mapping mechanism, assigning a unique integer label starting from 1 to each newly appearing unique value pair (v1, v2);

[0098] The interactive factor grid generated by the encoding is added to the candidate factor set, and the geographic interaction encoding of the current factor with one factor in the initial factor set is completed. The interaction process is repeated until the geographic interaction encoding of the current factor with all other factors in the initial factor set is completed.

[0099] S2-5, remove the factor with the largest spatial differentiation statistic value from the initial factor set, and repeat steps S2-1 to S2-4 until the initial factor set is empty, and obtain the candidate factor set as the optimal factor combination.

[0100] For example, the q value of each factor in the initial factor set A is calculated; the q values of each factor are obtained as follows: elevation (0.17), slope (0.01), aspect (0.01), peak ground acceleration (0.12), distance to river (0.25), distance to fault (0.09), vegetation cover index (0.10), and distance to road (0.14); this is only an example and is not limiting;

[0101] The factor with the maximum q value, distance to river (0.25), is selected, the variance ratio VR=36.15 is calculated, and the corresponding p value is obtained, p<0.01; the p value is less than the preset threshold value 0.05, and the current is the first step of screening, without a previous step of variance ratio, and the optimization condition is met by default; the distance to river factor is added to the candidate set B; the distance to river factor is removed from the initial factor set A;

[0102] The above steps are repeatedly executed: after a new round of calculation, the q value of the remaining factor, elevation, is the largest (0.17); the new variance ratio VR=18.49 is calculated after adding the factor to the current model; it is found that 18.49<36.15, which does not meet the optimization condition;

[0103] The geographical interaction coding is triggered, and the elevation factor is coded with each of the remaining factors in A, including slope, aspect, peak ground acceleration, distance to fault, vegetation cover index, and distance to road. For example, when coding with peak ground acceleration, take two factor raster layers and judge pixel by pixel: if any pixel is 0 or has no data, the corresponding position of the output raster is marked as no data; otherwise, through dictionary mapping, such as elevation=1 and peak ground acceleration=1, mapping to label 1, elevation=1 and peak ground acceleration=2, mapping to label 2, and so on, to generate a unique positive integer label, forming an “elevation∩peak ground acceleration” interaction factor raster; all generated interaction factors are added to the candidate set B; the elevation is removed from set A.

[0104] Continue to iterate until A is empty; finally, the candidate set B contains single factors directly added and interaction factors generated by interaction coding, and the candidate set B is the optimal factor combination.

[0105] S3, align and stack the layer spaces in the optimal factor combination to construct a multi-channel input feature tensor; generate a unified binary mask based on the effective pixel area, fill the invalid values in the features and labels with zero and perform normalization;

[0106] Specifically, step S3 includes:

[0107] S3-1, align the grid layers corresponding to the single factors in the optimal factor combination and the grid layers corresponding to the interaction factors in the spatial dimension, stack the layers, construct a multi-channel input feature tensor, and the number of channels of the multi-channel input feature tensor is the total number of single factors and interaction factors in the optimal factor combination;

[0108] S3-2, based on the effective pixel regions of the input layers corresponding to all multi-channel input feature tensors and the landslide susceptibility label layer, create a unified binary mask, the mask value corresponding to the effective pixel region is set to 1, and the mask value corresponding to the invalid pixel region is set to 0;

[0109] S3-3, fill the invalid values in the multi-channel input feature tensor and the landslide susceptibility label layer with 0, and normalize the effective pixel data in the multi-channel input feature tensor and the landslide susceptibility label layer. The normalization processing is Min-Max normalization, and the normalization formula is: Y norm =(Y-Y min ) / (Y max -Y min ), wherein Y norm represents the normalized value, Y represents the original value before normalization, Y min represents the minimum value in the original value, and Y max represents the maximum value in the original value.

[0110] For example, the optimal factor combination B includes 6 elements: distance to river, elevation ∩ peak ground acceleration, slope ∩ distance to road, etc. Align and stack the grid layers corresponding to the 6 elements in space to construct a 6-channel input feature tensor X;

[0111] Based on the effective pixel regions of all 6 input layers and the landslide susceptibility label layer Y, create a unified binary mask M; M=1 at the effective pixel, and M=0 at the invalid pixel;

[0112] Fill the invalid values in the feature tensor X and the label Y with 0; perform Min-Max normalization on the effective pixel data in X and Y; for example, the original value range of the label Y is [0, 150], and the original value Y of a pixel is 75, then the normalized Y norm =(75-0) / (150-0)=0.5.

[0113] S4, construct a spatial regression neural network with an encoder-decoder structure, the encoder takes a residual network as the main part and adapts to multi-channel input features, the decoder fuses the multi-scale features of the encoder through upsampling operation and jump connection, and outputs a single-channel susceptibility prediction map with the same resolution; use a mask weighted loss function to train the network;

[0114] Specifically, step S4 comprises:

[0115] S4-1, a spatial regression neural network of an encoder-decoder structure is constructed, the encoder adopts a pre-trained residual network as a backbone network, a first convolutional layer of the pre-trained residual network is modified to adapt to the channel number of the multi-channel input feature tensor, and the pre-trained weights of the pre-trained residual network are reused; the encoder extracts multi-scale feature maps through step-by-step down-sampling;

[0116] S4-2, the decoder constructs a symmetric up-sampling path with the encoder, the up-sampling path is composed of a plurality of up-sampling blocks, each up-sampling block contains one up-sampling operation and two modules executed in sequence, and each module sequentially completes convolution operation, batch normalization processing and activation function transformation;

[0117] S4-3, at each up-sampling stage of the decoder, the feature map of the corresponding resolution of the encoder is spliced with the up-sampled feature map in the channel dimension through a jump connection, and high-resolution spatial detail information and deep abstract semantic information are fused;

[0118] S4-4, after the decoder is restored to the original input resolution, a 1x1 convolutional layer is used to compress the multi-channel feature map into a single channel to generate a landslide susceptibility prediction map with the same resolution as the input feature tensor;

[0119] S4-5, a mask weighted loss function is used to train the spatial regression neural network, and the loss function calculates the prediction error based on the effective pixel area identified by the unified binary mask;

[0120] Specifically, the mask weighted loss function is a mask weighted mean square error loss function, and the formula is as follows:

[0121] ;

[0122] Wherein, L represents the loss value, Mask(i,j) represents the binary mask value at position (i,j), represents the prediction value of the model at position (i,j), Y(i,j) represents the normalized label value at position (i,j), and ε represents a small positive value used to prevent the denominator from being zero.

[0123] For example, a spatial regression neural network is constructed to build an encoder-decoder structure; the encoder is based on a pre-trained ResNet-50, the first convolutional layer of which is modified to accept a 6-channel input and reuse the pre-trained weights; the encoder outputs feature maps of 5 scales, with downsampling rates of 2, 4, 8, 16 and 32 respectively; the decoder is constructed to have a symmetric upsampling path, each upsampling block containing one upsampling and two convolution-normalization-activation modules; through a jump connection, the corresponding resolution feature maps of the encoder are spliced when decoding and upsampling, such as 2, 4, 8 and 16; the decoder uses a 1x1 convolution at the end to compress the channel number to 1, outputting a susceptibility prediction map with the same resolution as the input. The mask weighted MSE loss function L is used for training; for example, the sum of the squares of the errors between the predicted value and the true value Y of a batch of valid pixels is 2.5, and the total number of valid pixels is 1000, then L = 2.5 / (1000+1e-8) = 0.0025.

[0124] S5, select the model weight with the minimum loss on the validation set as the prediction model, and infer the multi-channel input features of the whole domain to generate a landslide susceptibility map and divide the susceptibility levels, and output the partition results;

[0125] Specifically, step S5 includes:

[0126] S5-1, during the training process of the spatial regression neural network, after each round of training is completed, the root mean square error is calculated on the validation set using the mask weighted loss function, the change of the root mean square error of the validation set is monitored in real time, the model weight corresponding to the minimum root mean square error of the validation set is saved, and the model weight is loaded as the weight of the prediction model;

[0127] S5-2, apply the spatial regression neural network with the loaded prediction model weight to the multi-channel input feature tensor of the whole domain to generate a normalized landslide susceptibility prediction map covering the whole domain through inference;

[0128] S5-3, post-process the normalized landslide susceptibility prediction map to denormalize it to the original landslide susceptibility index range, or use a natural breakpoint method or a quantile method for geographic spatial analysis to divide it into high, medium and low susceptibility levels; output the landslide susceptibility map after dividing the susceptibility levels to complete the landslide susceptibility evaluation.

[0129] For example: during the training process, the mask weighted RMSE is calculated on the validation set after each round is completed; the minimum RMSE is obtained at the 50th round, and the model weight of this round is saved; load this optimal weight as the prediction model to infer the 6-channel feature tensor X of the whole domain to generate a normalized landslide susceptibility prediction map with a value range of [0, 1]; denormalize the prediction map to restore it to the original susceptibility index range; use the natural breakpoint method to divide the susceptibility index into high, medium and low levels; output the final landslide susceptibility map; this is only an example for illustration and does not limit the scope.

[0130] The application provides another technical solution, a landslide susceptibility assessment system based on statistical screening and deep learning, a data management module, an intelligent screening module, a feature engineering module, a deep learning module, and an assessment output module.

[0131] The data management module comprises a factor set construction module and an alternative set management module; the factor set construction module is used for spatial coordinate unification and resolution resampling of a grid layer corresponding to a landslide influencing factor, to form a spatially consistent initial factor set; and the alternative set management module is used for creating and maintaining an empty alternative factor set, and storing optimal factors and factor interaction combinations with statistical significance obtained through screening.

[0132] The intelligent screening module comprises a statistical quantification module and an interaction encoding module; the statistical quantification module is used for calculating spatial differentiation statistical quantity values and variance ratios of factors, and testing statistical significance; and the interaction encoding module is used for performing geographic interaction encoding operations, and generating interaction factors through bijective mapping.

[0133] The feature engineering module comprises a tensor construction module and a mask processing module; the tensor construction module is used for spatial alignment and stacking of grid layers corresponding to single factors and interaction factors in the optimal factor combination, to construct a multi-channel input feature tensor; and the mask processing module is used for creating a unified binary mask based on valid pixel regions of all input layers and label layers.

[0134] The deep learning module comprises a network architecture module and a training optimization module; the network architecture module is used for constructing a spatial regression neural network with an encoder-decoder structure, the encoder takes a residual network as a backbone and adapts to multi-channel input features, and the decoder fuses multi-scale features through upsampling operations and skip connections; and the training optimization module is used for training the network by using a mask weighted loss function, and the loss function calculates a prediction error based on valid pixel regions identified by the unified binary mask.

[0135] The assessment output module comprises a prediction module and a post-processing module; the prediction module is used for loading prediction model weights, performing inference on a global multi-channel input feature tensor, and generating a landslide susceptibility prediction map; and the post-processing module is used for performing inverse normalization processing and susceptibility grade division on the prediction map, and outputting a final partition result.

[0136] It will be apparent to those skilled in the art that the application is not limited to the details of the above-exemplified embodiments and that the present application can be implemented in other particular forms without departing from the spirit or essential characteristics of the present application. The embodiments should therefore be considered in all respects as illustrative and not restrictive, the scope of the application being indicated by the appended claims rather than by the above description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein. No reference signs in the claims should be considered as limiting the scope of the claims with respect to the figures of the patent document.

Claims

1. A landslide susceptibility assessment method based on statistical screening and deep learning, characterized in that: The method comprises: S1, constructing an initial factor set of landslide influencing factors, and initializing the candidate factor set as an empty set; S2, screening factors in the initial factor set based on spatial differentiation statistics; selecting the factor with the largest statistical value, testing the variance ratio and significance thereof; if the significance level meets the standard and the variance ratio meets the optimization condition, adding the factor to the candidate factor set; otherwise, performing geographic interaction coding with the remaining factors in the initial factor set, generating an interaction factor and adding the interaction factor to the candidate factor set; repeatedly performing the screening process until all factors are processed, and obtaining an optimal factor combination; S3, aligning and stacking layers in the optimal factor combination in space to construct a multi-channel input feature tensor; generating a unified binary mask based on an effective pixel area, filling invalid values in the features and labels with zero and performing normalization; S4, constructing a spatial regression neural network with an encoder-decoder structure, wherein the encoder takes a residual network as the main trunk and is adapted to multi-channel input features, the decoder fuses multi-scale features of the encoder through upsampling operation and jump connection, and outputs a single-channel susceptibility prediction map with the same resolution; and using a mask weighted loss function to train the spatial regression neural network; S5, selecting the model weight with the minimum loss on the validation set as the prediction model, performing inference on the multi-channel input features in the whole domain to generate a landslide susceptibility map, dividing the susceptibility levels, and outputting the partition result.

2. The landslide susceptibility evaluation method based on statistical screening and deep learning according to claim 1, wherein: the landslide influencing factors include elevation, slope, aspect, peak ground acceleration, distance to river, distance to fault, vegetation cover index, and distance to road; the raster layers corresponding to the landslide influencing factors are subjected to spatial coordinate system alignment and resolution resampling to form a spatially consistent initial factor set; the candidate factor set is an empty set that does not contain any landslide influencing factors or factor interaction combinations, and is used to store the optimal factors and factor interaction combinations with statistical significance obtained through screening. 3.The statistical screening and deep learning based landslide susceptibility assessment method according to claim 1, characterized in that: Step S2 comprises: S2-1, calculating the spatial differentiation statistical value of each factor in the initial factor set, and the formula is as follows: ； where q represents a spatial heterogeneity statistic value; L represents a total number of layers; h represents a layer index; N h represents a sample size of the hth layer; represents an intra-layer variance of the hth layer; N represents a total sample size; σ 2 represents a total variance; SSW represents a within-group sum of squares; SST represents a total sum of squares; S2-2, selecting the factor with the largest spatial differentiation statistical value, calculating the variance ratio of the factor with the largest spatial differentiation statistical value, and the formula is as follows: ； wherein, VR represents the variance ratio; SSB represents the between-group sum of squares; k represents the number of groups; k-1 represents the numerator degrees of freedom; and N-k represents the denominator degrees of freedom; S2-3, when the statistical significance of the factor reaches a preset threshold and the variance ratio meets the optimization condition, adding the factor to the candidate factor set; S2-4, when the statistical significance of the factor does not reach the preset threshold or the variance ratio does not satisfy the optimization condition, the factor is respectively interactively encoded with each of the remaining factors in the initial factor set, the interactive encoding includes taking the grid layers corresponding to the two factors, judging each pixel position in the grid, if any pixel value is 0 or the no data value of the corresponding grid, the pixel label of the output grid is marked as no data value, otherwise it is uniquely mapped to a positive integer label through a bijective mapping to generate an interactive factor, and the interactive factor is added to the candidate factor set; S2-5, the factor with the largest spatial differentiation statistical value is removed from the initial factor set, and steps S2-1 to S2-4 are repeatedly executed until the initial factor set is empty, and the candidate factor set is obtained as the optimal factor combination.

4. The landslide susceptibility assessment method based on statistical screening and deep learning according to claim 3, characterized in that: The optimization condition includes that the variance ratio of the current step is greater than or equal to the variance ratio of the previous step; the variance ratio of the current step includes the variance ratio after adding the candidate factor in the current iteration step; and the variance ratio of the previous step includes the variance ratio before adding the candidate factor in the current step in the previous iteration step. 5.The statistical screening and deep learning based landslide susceptibility assessment method according to claim 3, characterized in that: Step S2-4 includes: The grid layer corresponding to the current screened factor is a first input grid R1, and the grid layer corresponding to another factor to be interacted in the initial factor set is a second input grid R2; If the pixel value vl of the first input raster Rl at position (i,j) is equal to 0, or the pixel value v2 of the second input raster R2 at position (i,j) is equal to 0, or vl is equal to the no data value nodatal of the first input raster Rl, or v2 is equal to the no data value nodata2 of the second input raster R2, then the output raster R out In this pixel position is marked the no data value nodatal of the first input raster; Otherwise, map the pixel value pair (v1, v2) to a unique positive integer label by the bijective mapping function L(v1, v2) and output the label as the output raster R out the value at that pixel location; The bijective mapping function L(v1, v2) is realized through a dictionary mapping mechanism, and a unique integer label starting from 1 is assigned to each newly appearing unique value pair (v1, v2); The interactive factor grid generated by encoding is added to the candidate factor set, and the geographic interaction encoding of the current factor and one factor in the initial factor set is completed; the interaction process is repeated until the geographic interaction encoding of the current factor and all other factors in the initial factor set is completed. 6.The statistical filtering and deep learning based landslide susceptibility assessment method according to claim 1, characterized in that: Step S3 includes: S3-1, aligning the coordinates and stacking the layers of the grid layer corresponding to the single factor in the optimal factor combination and the grid layer corresponding to the interactive factor in the spatial dimension to construct a multi-channel input feature tensor, the number of channels of the multi-channel input feature tensor is the total number of single factors and interactive factors in the optimal factor combination; S3-2, based on the effective pixel area of the input layer corresponding to all multi-channel input feature tensors and the landslide susceptibility label layer, a unified binary mask is created, the mask value corresponding to the effective pixel area is set to 1, and the mask value corresponding to the invalid pixel area is set to 0; S3-3, fill the invalid values in the multi-channel input feature tensor and the landslide susceptibility label layer with 0, and normalize the valid pixel data in the multi-channel input feature tensor and the landslide susceptibility label layer. 7.The statistical filtering and deep learning based landslide susceptibility assessment method according to claim 1, characterized in that: Step S4 includes: S4-1, a spatial regression neural network of an encoder-decoder structure is constructed, the encoder adopts a pre-trained residual network as a backbone network, a first convolutional layer of the pre-trained residual network is modified to adapt to the channel number of a multi-channel input feature tensor, and the pre-trained weights of the pre-trained residual network are reused; the encoder extracts multi-scale feature maps through step-by-step down-sampling; S4-2, the decoder constructs a symmetric up-sampling path with the encoder, the up-sampling path is composed of a plurality of up-sampling blocks, each up-sampling block contains one up-sampling operation and two modules executed in sequence, and each module sequentially completes convolution operation, batch normalization processing and activation function transformation; S4-3, at each up-sampling stage of the decoder, the feature maps of the corresponding resolution of the encoder are spliced with the up-sampled feature maps in the channel dimension through a jump connection, and high-resolution spatial detail information and deep abstract semantic information are fused; S4-4, after the decoder is restored to the original input resolution, a 1×1 convolutional layer is used to compress the multi-channel feature map into a single channel to generate a landslide susceptibility prediction map with the same resolution as the input feature tensor; S4-5, a mask weighted loss function is used to train the spatial regression neural network, and the mask weighted loss function calculates the prediction error based on the effective pixel area identified by a unified binary mask.

8. The landslide susceptibility assessment method based on statistical screening and deep learning according to claim 7, characterized in that: the mask weighted loss function is a mask weighted mean square error loss function, and the formula is as follows: ； wherein L represents a loss value, Mask(i,j) represents a binary mask value at position (i,j), represents a predicted value of the model at position (i,j), Y(i,j) represents a normalized label value at position (i,j), and ε represents a small positive value for preventing the denominator from being zero. 9.The statistical filtering and deep learning based landslide susceptibility assessment method according to claim 1, characterized in that: Step S5 includes: S5-1, during the training process of the spatial regression neural network, after each round of training is completed, the root mean square error is calculated on the validation set using the mask weighted loss function, the change of the root mean square error of the validation set is monitored in real time, the model weight corresponding to the minimum root mean square error of the validation set is saved, and the model weight is loaded as the weight of the prediction model; S5-2, the spatial regression neural network loaded with the prediction model weight is applied to the multi-channel input feature tensor of the whole domain to generate a normalized landslide susceptibility prediction map covering the whole domain through inference; S5-3, the normalized landslide susceptibility prediction map is post-processed and normalized to the original landslide susceptibility index range, or divided into high, medium and low susceptibility grades by using a natural breakpoint method or a quantile method of geographic spatial analysis; a landslide susceptibility map after susceptibility grading is output, and landslide susceptibility assessment is completed.

10. A landslide susceptibility assessment system based on statistical screening and deep learning, characterized in that: The system comprises a data management module, an intelligent screening module, a feature engineering module, a deep learning module and an evaluation output module; The data management module comprises a factor set construction module and an alternative set management module; the factor set construction module is configured to perform spatial coordinate unification and resolution resampling on grid layers corresponding to landslide influence factors to form an initial factor set that is spatially consistent; and the alternative set management module is configured to create and maintain an empty alternative factor set, and store optimal factors and factor interaction combinations that have statistical significance obtained through screening. The intelligent screening module includes a statistical quantification module and an interactive coding module; the statistical quantification module is used to calculate the spatial heterogeneity statistics and variance ratios of each factor and test statistical significance; the interactive coding module is used to perform geographic interactive coding operations and generate interactive factors through bijective mapping. The feature engineering module includes a tensor construction module and a mask processing module. The tensor construction module is used to spatially align and stack the raster layers corresponding to the single factors and interaction factors in the optimal factor combination to construct a multi-channel input feature tensor. The mask processing module is used to create a unified binary mask based on the effective pixel areas of all input layers and label layers. The deep learning module includes a network architecture module and a training optimization module. The network architecture module is used to construct a spatial regression neural network with an encoder-decoder structure. The encoder uses a residual network as its backbone and adapts to multi-channel input features. The decoder fuses multi-scale features through upsampling operations and skip connections. The training optimization module is used to train the network using a mask-weighted loss function, which calculates the prediction error based on the effective pixel region of a unified binary mask identifier. The evaluation output module includes a prediction module and a post-processing module. The prediction module is used to load the prediction model weights, perform inference on the multi-channel input feature tensor of the entire domain, and generate a landslide susceptibility prediction map. The post-processing module is used to perform inverse normalization processing and susceptibility level classification on the prediction map and output the final partitioning result.

Images