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

Abstract

The invention discloses a statistical screening and deep learning-based landslide susceptibility assessment system and method, and relates to the technical field of landslide susceptibility assessment, and the method comprises the steps: constructing an initial factor set of landslide influence factors, and initializing an alternative factor set; selecting the factor with the maximum statistical magnitude value, checking the variance ratio and the significance, and obtaining an optimal factor combination; layer spaces in the optimal factor combination are aligned and stacked, and a multi-channel input feature tensor is constructed; a spatial regression neural network of an encoder-decoder structure is constructed, the encoder takes a residual network as a trunk and adapts to multi-channel input features, and the decoder fuses the multi-scale features of the encoder through up-sampling operation and jump connection and outputs a susceptibility prediction map; and performing reasoning on the multi-channel input features, generating a landslide susceptibility map, dividing susceptibility grades, and outputting a partitioning result. The method is based on a plurality of influence factors, and supports the susceptibility evaluation of different types of landslides in different areas.

Description

Technical Field

[0001] This invention relates to the field of landslide susceptibility assessment technology, specifically a landslide susceptibility assessment system and method based on statistical screening and deep learning. Background Technology

[0002] Landslides, as highly destructive geological hazards, require susceptibility assessment as a core foundation of disaster prevention and mitigation systems. Traditional statistical methods, such as the analytic hierarchy process (AHP), rely excessively on expert experience, have highly subjective factor weighting, and cannot objectively quantify factors and their interactions. Traditional statistical screening methods can only identify linear correlations, ignoring higher-order nonlinear interactions between factors such as faults and rivers, and ground motion parameters and topography, resulting in assessment results lacking geoscientific interpretability. While deep learning methods can fit nonlinear relationships, their "end-to-end" black-box nature makes them unstable and difficult to reproduce, and they are insufficient in handling invalid values ​​commonly found in geographic data, reducing the accuracy of assessments in real-world geographic environments. Summary of the Invention

[0003] The purpose of this invention is to provide a landslide susceptibility assessment system and method based on statistical screening and deep learning to solve the problems raised in the prior art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a landslide susceptibility assessment method based on statistical screening and deep learning, the method comprising: S1. Construct an initial set of factors affecting landslides and initialize the candidate factor set to an empty set; S2. Screen the factors in the initial factor set based on the spatial heterogeneity statistic; select the factor with the largest statistic value and test its variance ratio and significance; if the significance level meets the standard and the variance ratio meets the optimization condition, add it to the candidate factor set; otherwise, perform geographic cross-coding with the remaining factors in the initial factor set to generate cross-factors and add them to the candidate factor set; repeat the screening process until all factors are processed and the optimal factor combination is obtained. S3. Align and stack the layers in the optimal factor combination to construct a multi-channel input feature tensor; generate a unified binary mask based on the effective pixel region, fill invalid values ​​in the features and labels with zeros and normalize them; S4. 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 the multi-scale features of the encoder through upsampling operations and skip connections to output a single-channel susceptibility prediction map with the same resolution. The spatial regression neural network is trained using a mask-weighted loss function. S5. Select the model weights with the minimum loss on the validation set as the prediction model, perform inference on the multi-channel input features of the entire domain, generate a landslide susceptibility map, divide the susceptibility levels, and output the partitioning results.

[0005] According to the above scheme, the landslide influencing factors include elevation, slope, aspect, peak ground acceleration, distance to the river, distance to the fault, vegetation cover index, and distance to the road. The raster layers corresponding to the landslide influencing factors are resampled using spatial coordinate system one and resolution to form an initial factor set with spatial consistency. 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 that have statistical significance obtained through screening.

[0006] According to the above scheme, step S2 includes: S2-1. Calculate the spatial heterogeneity statistic for each factor in the initial factor set, using the following formula: ; Where q represents the spatial heterogeneity statistic; L represents the total number of strata; h represents the stratum index; N h Let h be the sample size of the h-th layer; σ represents the within-layer variance of the h-th stratum; N represents the total sample size; σ 2 SSW represents the total variance; SST represents the sum of squares within groups; SST represents the total sum of squares; the 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 unrelated to the distribution of landslide susceptibility; The formula for the sum of squares within a group is as follows: ; Where k represents the number of groups; i represents the group index; n i Let x represent the sample size of group i; j represent the sample index within the group; x ij Let be the observation value of the j-th sample in the i-th group; Let it be the mean of the i-th group; The total sum of squares is calculated using the following formula: ; in, Expressed as the population mean; S2-2. Select the factor with the largest spatial heterogeneity statistic value, and calculate the variance ratio of the factor with the largest spatial heterogeneity statistic value, as shown in the following formula: ; Where VR represents the variance ratio; SSB represents the sum of squares between groups; k represents the number of groups; k-1 represents the numerator degrees of freedom; and Nk represents the denominator degrees of freedom. The formula for the sum of squares between groups is as follows: ; Where, n i It is the number of the i-th group; S2-3. When the statistical significance of a factor reaches the preset threshold and the variance ratio meets the optimization conditions, the factor is added to the candidate factor set. S2-4. When the statistical significance of a factor does not reach the preset threshold or the variance ratio does not meet the optimization conditions, the factor and each of the other factors in the initial factor set shall be geographically interactively encoded. The geographic interactive encoding includes taking the raster layer corresponding to the two factors, judging the position of each pixel in the raster, and if any pixel value is 0 or the corresponding raster has no data value, the raster is output as having no data value. Otherwise, the pixel is mapped to a unique positive integer label through bijective mapping, an interactive factor is generated, and the interactive factor is added to the candidate factor set. S2-5. Remove the factor with the largest spatial heterogeneity statistic 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.

[0007] According to the above scheme, the optimization conditions include 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 candidate factors in the current iteration step; the variance ratio of the previous step includes the variance ratio when candidate factors for this step have not yet been added in the previous iteration step.

[0008] According to the above scheme, step S2-4 includes: The raster layer corresponding to the currently selected factor is the first input raster R1, and the raster layer corresponding to the other factor to be interacted with in the initial factor set is the second input raster R2. If the pixel value v1 at position (i,j) of the first input raster R1 is equal to 0, or the pixel value v2 at position (i,j) of the second input raster R2 is equal to 0, or v1 is equal to the no-data value nodata1 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 The pixel location is marked as the no-data value nodata1 for the first input raster; Otherwise, the pixel value pair (v1,v2) is mapped to a unique positive integer label using the bijective mapping function L(v1,v2), and this label is used as the output raster R. out The value at that pixel location; The bijective mapping function L(v1,v2) is implemented through a dictionary mapping mechanism, which assigns a unique integer label starting from 1 to each newly appearing unique value pair (v1,v2). Add the generated interactive factor raster to the candidate factor set to complete the geographic interactive coding of the current factor with one factor in the initial factor set; repeat the interactive process until the current factor and all other factors in the initial factor set have completed geographic interactive coding.

[0009] According to the above scheme, step S3 includes: S3-1. Align and stack the raster layers corresponding to the single factors and the interaction factors in the optimal factor combination in the spatial dimension to construct a multi-channel input feature tensor. The number of channels in the multi-channel input feature tensor is the total number of single factors and interaction factors in the optimal factor combination. S3-2. Based on the effective pixel regions of the input layers and landslide susceptibility label layers corresponding to all multi-channel input feature tensors, a unified binary mask is created, wherein 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. S3-3. Fill invalid values ​​in the multi-channel input feature tensor and landslide susceptibility label layer with 0, and normalize the valid pixel data in the multi-channel input feature tensor and landslide susceptibility label layer.

[0010] According to the above scheme, step S4 includes: S4-1. Construct a spatial regression neural network with an encoder-decoder structure. The encoder uses a pre-trained residual network as the backbone network. Modify the first convolutional layer of the pre-trained residual network to adapt the first convolutional layer to the number of channels of the multi-channel input feature tensor and reuse the pre-trained weights of the pre-trained residual network. The encoder extracts multi-scale feature maps by downsampling step by step. S4-2. The decoder constructs an upsampling path symmetrical to the encoder. The upsampling path consists of multiple upsampling blocks. Each upsampling block contains one upsampling operation and two modules executed sequentially. Each module sequentially performs convolution operation, batch normalization processing, and activation function transformation. S4-3. In each upsampling stage of the decoder, the feature map of the encoder at the corresponding resolution is spliced ​​with the upsampled feature map in the channel dimension through skip connections, so as to fuse high-resolution spatial detail information and deep abstract semantic information. S4-4. After the decoder restores the original input resolution, the multi-channel feature map is compressed into a single channel through a 1×1 convolutional layer to generate a landslide susceptibility prediction map with the same resolution as the input feature tensor. S4-5. The spatial regression neural network is trained using a mask-weighted loss function, wherein the mask-weighted loss function calculates the prediction error based on the effective pixel region of a unified binary mask identifier.

[0011] According to the above scheme, the mask weighted loss function is the mask weighted mean square error loss function, and the formula is as follows: ; Where L represents the loss value, and Mask(i,j) represents the binary mask value at position (i,j). Y(i,j) represents the predicted 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.

[0012] According to the above scheme, step S5 includes: S5-1. During the training of the spatial regression neural network, after each training round, the root mean square error is calculated on the validation set using a mask-weighted loss function. The changes in the root mean square error of the validation set are monitored in real time. The model weights corresponding to the minimum root mean square error of the validation set are saved, and the model weights are used as the weights of the prediction model and loaded. S5-2. Apply the spatial regression neural network loaded with the prediction model weights to the multi-channel input feature tensor of the entire domain, and generate a normalized landslide susceptibility prediction map covering the entire domain through inference. S5-3. Post-process the normalized landslide susceptibility prediction map, reverse normalize it to the original landslide susceptibility index range, or use geospatial analysis methods such as the natural discontinuity method or the quantile method 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 assessment.

[0013] A landslide susceptibility assessment system based on statistical screening and deep learning, comprising a data management module, an intelligent screening module, a feature engineering module, a deep learning module, and an assessment output module; The data management module includes a factor set construction module and a candidate set management module. The factor set construction module is used to perform spatial coordinate system unification and resolution resampling on the raster layer corresponding to the landslide influencing factors to form an initial factor set with spatial consistency. The candidate set management module is used to create and maintain an empty candidate factor set and store the optimal factor and factor interaction combination with 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.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 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; 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. 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

[0015] Figure 1 This is a flowchart illustrating the steps of the landslide susceptibility assessment method based on statistical screening and deep learning in this invention. Figure 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

[0016] 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.

[0017] Example: Figures 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: S1. Construct an initial set of factors affecting landslides and initialize the candidate factor set to an empty set; Specifically, landslide influencing factors include elevation, slope, aspect, peak ground acceleration, distance to the river, distance to the fault, vegetation cover index, and distance to the road. These factors are selected based on the geological hazard formation mechanism and cover five major categories of evaluation indicators: topography, geological structure, hydrological conditions, seismic motion, and human activities. The raster layers corresponding to the landslide influencing factors are resampled using a spatial coordinate system and resolution to form a spatially consistent initial factor set. The candidate factor set is an empty set containing no landslide influencing factors or factor interactions, used to store the statistically significant optimal factors and factor interactions obtained through screening.

[0018] For example: select eight landslide influencing factors within the study area, including elevation, slope, aspect, peak ground acceleration, distance to the river, distance to the fault, vegetation cover index, and distance to the road; unify the raster layers corresponding to each factor to the same spatial coordinate system and resolution to form an initial factor set A containing eight layers; at the same time, create an empty alternative factor set B; this is only an example and is not a limitation. S2. Factors in the initial factor set are screened based on spatial heterogeneity statistics. This screening process uses a stepwise geographic detector and achieves optimal factor selection through a collaborative judgment mechanism of q-statistic and variance ratio (VR). The factor with the largest statistic 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 conditions, it is added to the candidate factor set. Otherwise, it is geographically interactively encoded with the remaining factors in the initial factor set to generate interactive factors, which are then added to the candidate factor set. The screening process is repeated until all factors are processed and the optimal factor combination is obtained. Specifically, step S2 includes: S2-1. Calculate the spatial heterogeneity statistic for each factor in the initial factor set, using the following formula: ; Where q represents the spatial heterogeneity statistic; L represents the total number of strata; h represents the stratum index; N h Let h be the sample size of the h-th layer; σ represents the within-layer variance of the h-th stratum; N represents the total sample size; σ 2 SSW represents the total variance; SST represents the sum of squares within groups; SST represents the total sum of squares; the 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 unrelated to the distribution of landslide susceptibility; The formula for the sum of squares within a group is as follows: ; Where k represents the number of groups; i represents the group index; n i Let x represent the sample size of group i; j represent the sample index within the group; x ij Let be the observation value of the j-th sample in the i-th group; Let it be the mean of the i-th group; The total sum of squares is calculated using the following formula: ; in, Expressed as the population mean; S2-2. Select the factor with the largest spatial heterogeneity statistic value and calculate the variance ratio of the largest factor, as shown in the following formula: ; Where VR represents the variance ratio; the variance ratio VR is statistically equivalent to the F statistic, and it tests the statistical significance of the stratification results by comparing the ratio of the variance between groups to the variance within groups; SSB represents the sum of squares between groups; k represents the number of groups; k-1 represents the degrees of freedom in the numerator; Nk represents the degrees of freedom in the denominator. The formula for the sum of squares between groups is as follows: ; Where, n i It is the number of the i-th group; S2-3. When the statistical significance of a factor reaches the preset threshold and the variance ratio meets the optimization conditions, the factor is added to the candidate factor set. The preset threshold for 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. This p-value is calculated from the variance ratio VR. The optimization conditions include 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 candidate factors in the current iteration step; the variance ratio of the previous step includes the variance ratio in the previous iteration step before adding candidate factors in this step. S2-4. When the statistical significance of a factor does not reach the preset threshold or the variance ratio does not meet the optimization conditions, the factor is geographically interactively encoded with each of the other factors in the initial factor set. This encoding process systematically assigns an incremental unique integer label to each unique factor value pair by establishing a bijective mapping relationship. The geographic interactive encoding includes taking the raster layer corresponding to the two factors, judging the position of each pixel in the raster, and if any pixel value is 0 or the corresponding raster has no data value, the raster is output as having no data value. Otherwise, it is generated as a unique positive integer label through bijective mapping, and the interactive factor is added to the candidate factor set. Furthermore, steps S2-4 include: The raster layer corresponding to the currently selected factor is the first input raster R1, and the raster layer corresponding to the other factor to be interacted with in the initial factor set is the second input raster R2. If the pixel value v1 at position (i,j) of the first input raster R1 is equal to 0, or the pixel value v2 at position (i,j) of the second input raster R2 is equal to 0, or v1 is equal to the no-data value nodata1 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 The pixel location is marked as the no-data value nodata1 for the first input raster; Otherwise, the pixel value pair (v1,v2) is mapped to a unique positive integer label using the bijective mapping function L(v1,v2), and this label is used as the output raster R. out The value at that pixel location; Bijective mapping is implemented through a dictionary mechanism to ensure a one-to-one correspondence between each unique (v1,v2) value pair and a unique integer label; the bijective mapping function L(v1,v2) is implemented through a dictionary mapping mechanism to assign a unique integer label starting from 1 to each newly appearing unique value pair (v1,v2); Add the generated interactive factor raster to the candidate factor set to complete the geographic interactive coding of the current factor with one factor in the initial factor set; repeat the interactive process until the current factor and all other factors in the initial factor set have completed geographic interactive coding.

[0019] S2-5. Remove the factor with the largest spatial heterogeneity statistic 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.

[0020] For example: Calculate the q value of each factor in the initial factor set A; the q values ​​of each factor are 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 a limitation. Select the factor with the largest q value and its distance to the river (0.25), calculate its variance ratio VR=36.15, and obtain the corresponding p value. p<0.01; the p value is less than the preset threshold of 0.05, and this is the first step of screening, with no variance ratio from the previous step, so the optimization condition is met by default; add the factor of distance to the river to the candidate set B; remove the factor of distance to the river from the initial factor set A. Repeat the above steps: After the new round of calculation, the q value of elevation is the largest among the remaining factors (0.17); calculate the new variance ratio VR after adding it to the current model, which is 18.49; compare and find that 18.49 < 36.15, which does not meet the optimization conditions; Trigger geographic interactive coding, and encode the elevation factor with each of the remaining factors in set A, including slope, aspect, peak ground acceleration, distance to fault, vegetation cover index, and distance to road. For example, when encoding with peak ground acceleration, take the two factor raster layers and check each pixel: if any pixel is 0 or has no data, mark the corresponding position in the output raster as having no data; otherwise, through dictionary mapping, such as elevation=1 and peak ground acceleration=1, map to label 1, elevation=1 and peak ground acceleration=2, map to label 2, and so on, generating unique positive integer labels to form an interactive factor raster of "elevation∩peak ground acceleration"; add all generated interactive factors to candidate set B; remove elevation from set A.

[0021] Continue iterating until A is an empty set; finally, the candidate set B contains the single factors that are directly added and the interactive factors generated through interactive coding. This candidate set B is the optimal factor combination.

[0022] S3. Align and stack the layers in the optimal factor combination to construct a multi-channel input feature tensor; generate a unified binary mask based on the effective pixel region, fill invalid values ​​in the features and labels with zeros and normalize them; Specifically, step S3 includes: S3-1. Align and stack the raster layers corresponding to the single factors and the interaction factors in the optimal factor combination in the spatial dimension to construct a multi-channel input feature tensor. The number of channels in the multi-channel input feature tensor is the total number of single factors and interaction factors in the optimal factor combination. S3-2. Based on the effective pixel regions of the input layers and landslide susceptibility label layers corresponding to all multi-channel input feature tensors, create a unified binary mask, setting the mask value corresponding to the effective pixel region to 1 and the mask value corresponding to the invalid pixel region to 0. S3-3. Fill all invalid values ​​in the multi-channel input feature tensor and landslide susceptibility label layer with 0, and normalize the valid pixel data in the multi-channel input feature tensor and landslide susceptibility label layer. The normalization process is Min-Max normalization, and the normalization formula is: Y norm =(YY min ) / (Y max -Y min ), where Y norm Y represents the normalized value, and Y represents the original value before normalization. min Y represents the minimum value in the original numerical values. max This represents the maximum value in the original numerical values.

[0023] For example: the optimal factor combination B contains 6 features: distance to the river, elevation ∩ peak ground acceleration, slope ∩ distance to the road, etc.; the raster layers corresponding to these 6 features are spatially aligned and stacked to construct a 6-channel input feature tensor X; Based on the effective pixel regions of all 6 input layers and the landslide susceptibility label layer Y, create a uniform binary mask M; M=1 for effective pixels and M=0 for invalid pixels; Invalid values ​​in the feature tensor X and label Y are uniformly filled with 0; the valid pixel data in X and Y are normalized using Min-Max; for example, if the original value range of label Y is [0, 150], and the original value of a certain pixel is Y=75, then after normalization Y norm =(75-0) / (150-0)=0.5.

[0024] S4. 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 the multi-scale features of the encoder through upsampling operations and skip connections to output a single-channel susceptibility prediction map with the same resolution. The network is trained using a mask-weighted loss function. Specifically, step S4 includes: S4-1. Construct a spatial regression neural network with an encoder-decoder structure. The encoder uses a pre-trained residual network as the backbone network. Modify the first convolutional layer of the pre-trained residual network to adapt the first convolutional layer to the number of channels of the multi-channel input feature tensor and reuse the pre-trained weights of the pre-trained residual network. The encoder extracts multi-scale feature maps by downsampling step by step. S4-2. The decoder constructs an upsampling path symmetrical to the encoder. The upsampling path consists of multiple upsampling blocks. Each upsampling block contains one upsampling operation and two modules executed sequentially. Each module sequentially performs convolution operation, batch normalization processing, and activation function transformation. S4-3. In each upsampling stage of the decoder, the feature map of the encoder at the corresponding resolution is spliced ​​with the upsampled feature map in the channel dimension through skip connections, so as to fuse high-resolution spatial detail information and deep abstract semantic information. S4-4. After the decoder restores the original input resolution, the multi-channel feature map is compressed into a single channel through a 1×1 convolutional layer to generate a landslide susceptibility prediction map with the same resolution as the input feature tensor. S4-5. The spatial regression neural network is trained using a mask-weighted loss function. The loss function calculates the prediction error based on the effective pixel region of the unified binary mask identifier. Specifically, the mask-weighted loss function is the mask-weighted mean square error loss function, as shown in the following formula: ; Where L represents the loss value, and Mask(i,j) represents the binary mask value at position (i,j). Y(i,j) represents the predicted 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.

[0025] For example: Construct a spatial regression neural network with an encoder-decoder structure; the encoder is based on a pre-trained ResNet-50, with its first convolutional layer modified to accept 6-channel input and pre-trained weights reused; the encoder outputs feature maps at 5 scales with downsampling rates of 2, 4, 8, 16, and 32; the decoder constructs a symmetrical upsampling path, with each upsampling block containing one upsampling operation and two convolution-normalization-activation modules; through skip connections, the encoder's feature maps at corresponding resolutions, such as 2, 4, 8, and 16, are concatenated during decoding upsampling; at the end of the decoder, a 1×1 convolution is used to compress the number of channels to 1, outputting a susceptibility prediction map with the same resolution as the input. Training is performed using a mask-weighted MSE loss function L; for example, if the sum of squared errors between the predicted and true values ​​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.

[0026] S5. Select the model weights with the minimum loss on the validation set as the prediction model, perform inference on the multi-channel input features of the whole domain, generate a landslide susceptibility map, divide the susceptibility level, and output the partitioning results. Specifically, step S5 includes: S5-1. During the training of the spatial regression neural network, after each training round, the root mean square error is calculated on the validation set using a mask-weighted loss function. The changes in the root mean square error of the validation set are monitored in real time. The model weights corresponding to the minimum root mean square error of the validation set are saved, and the model weights are used as the weights of the prediction model and loaded. S5-2. Apply the spatial regression neural network loaded with the prediction model weights to the multi-channel input feature tensor of the entire domain, and generate a normalized landslide susceptibility prediction map covering the entire domain through inference. S5-3. Post-process the normalized landslide susceptibility prediction map, reverse normalize it to the original landslide susceptibility index range, or use geospatial analysis methods such as the natural discontinuity method or the quantile method to classify it into high, medium and low susceptibility levels; output the landslide susceptibility map after classifying the susceptibility levels to complete the landslide susceptibility assessment.

[0027] For example: During training, after each round, the mask-weighted RMSE is calculated on the validation set; the RMSE obtained in the 50th round is the smallest, so the model weights for that round are saved; this optimal weight is loaded as the prediction model, and inference is performed on the 6-channel feature tensor X of the entire domain to generate a normalized landslide susceptibility prediction map with a value range of [0,1]; the prediction map is denormalized to restore it to the original susceptibility index range; the susceptibility index is divided into three levels: high, medium, and low using the natural breakpoint method; the final landslide susceptibility map is output; this is just an example and is not a limitation.

[0028] This invention provides another technical solution: a landslide susceptibility assessment system based on statistical screening and deep learning, comprising a data management module, an intelligent screening module, a feature engineering module, a deep learning module, and an assessment output module; The data management module includes a factor set construction module and a candidate set management module. The factor set construction module is used to perform spatial coordinate system unification and resolution resampling on the raster layers corresponding to landslide influencing factors to form a spatially consistent initial factor set. The candidate set management module is used to create and maintain an empty candidate factor set, and store the 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. The loss function calculates the prediction error based on the effective pixel region of a unified binary mask. 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 results.

[0029] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A landslide susceptibility assessment method based on statistical screening and deep learning, characterized by: The method includes: S1. Construct an initial set of factors affecting landslides and initialize the candidate factor set to an empty set; S2. Screen the factors in the initial factor set based on the spatial heterogeneity statistic; select the factor with the largest statistic value and test its variance ratio and significance; if the significance level meets the standard and the variance ratio meets the optimization condition, add it to the candidate factor set; otherwise, perform geographic cross-coding with the remaining factors in the initial factor set to generate cross-factors and add them to the candidate factor set; repeat the screening process until all factors are processed and the optimal factor combination is obtained. S3. Align and stack the layers in the optimal factor combination to construct a multi-channel input feature tensor; generate a unified binary mask based on the effective pixel region, fill invalid values ​​in the features and labels with zeros and normalize them; S4. 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 the multi-scale features of the encoder through upsampling operations and skip connections to output a single-channel susceptibility prediction map with the same resolution. The spatial regression neural network is trained using a mask-weighted loss function. S5. Select the model weights with the minimum loss on the validation set as the prediction model, perform inference on the multi-channel input features of the entire domain, generate a landslide susceptibility map, divide the susceptibility levels, and output the partitioning results.

2. The landslide susceptibility assessment method based on statistical screening and deep learning according to claim 1, characterized in that: The landslide influencing factors include elevation, slope, aspect, peak ground acceleration, distance to the river, distance to the fault, vegetation cover index, and distance to the road. The raster layers corresponding to the landslide influencing factors are resampled using spatial coordinate system one and resolution to form an initial factor set with spatial consistency. 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 that have statistical significance obtained through screening.

3. The landslide susceptibility assessment method based on statistical screening and deep learning according to claim 1, characterized in that: Step S2 includes: S2-1. Calculate the spatial heterogeneity statistic for each factor in the initial factor set, using the following formula: ； Where q represents the spatial heterogeneity statistic; L represents the total number of strata; h represents the stratum index; N h Let h be the sample size of the h-th layer; σ represents the within-layer variance of the h-th stratum; N represents the total sample size; σ 2 SSW represents the total variance; SSW represents the sum of squares within groups; SST represents the total sum of squares. S2-2. Select the factor with the largest spatial heterogeneity statistic value, and calculate the variance ratio of the factor with the largest spatial heterogeneity statistic value, as shown in the following formula: ； Where VR represents the variance ratio; SSB represents the sum of squares between groups; k represents the number of groups; k-1 represents the numerator degrees of freedom; and Nk represents the denominator degrees of freedom. S2-3. When the statistical significance of a factor reaches the preset threshold and the variance ratio meets the optimization conditions, the factor is added to the candidate factor set. S2-4. When the statistical significance of a factor does not reach the preset threshold or the variance ratio does not meet the optimization conditions, the factor and each of the other factors in the initial factor set shall be geographically interactively encoded. The geographic interactive encoding includes taking the raster layer corresponding to the two factors, judging the position of each pixel in the raster, and if any pixel value is 0 or the corresponding raster has no data value, the raster is output as having no data value. Otherwise, the pixel is mapped to a unique positive integer label through bijective mapping, an interactive factor is generated, and the interactive factor is added to the candidate factor set. S2-5. Remove the factor with the largest spatial heterogeneity statistic 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.

4. The landslide susceptibility assessment method based on statistical screening and deep learning according to claim 3, characterized in that: The optimization conditions include 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 candidate factors in the current iteration step; the variance ratio of the previous step includes the variance ratio when candidate factors for this step have not yet been added in the previous iteration step.

5. The landslide susceptibility assessment method based on statistical screening and deep learning according to claim 3, characterized in that: Steps S2-4 include: The raster layer corresponding to the currently selected factor is the first input raster R1, and the raster layer corresponding to the other factor to be interacted with in the initial factor set is the second input raster R2. If the pixel value v1 at position (i,j) of the first input raster R1 is equal to 0, or the pixel value v2 at position (i,j) of the second input raster R2 is equal to 0, or v1 is equal to the no-data value nodata1 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 The pixel location is marked as the no-data value nodata1 for the first input raster; Otherwise, the pixel value pair (v1,v2) is mapped to a unique positive integer label using the bijective mapping function L(v1,v2), and this label is used as the output raster R. out The value at that pixel location; The bijective mapping function L(v1,v2) is implemented through a dictionary mapping mechanism, which assigns a unique integer label starting from 1 to each newly appearing unique value pair (v1,v2). Add the generated interactive factor raster to the candidate factor set to complete the geographic interactive coding of the current factor with one factor in the initial factor set; repeat the interactive process until the current factor and all other factors in the initial factor set have completed geographic interactive coding.

6. The landslide susceptibility assessment method based on statistical screening and deep learning according to claim 1, characterized in that: Step S3 includes: S3-1. Align and stack the raster layers corresponding to the single factors and the interaction factors in the optimal factor combination in the spatial dimension to construct a multi-channel input feature tensor. The number of channels in the multi-channel input feature tensor is the total number of single factors and interaction factors in the optimal factor combination. S3-2. Based on the effective pixel regions of the input layers and landslide susceptibility label layers corresponding to all multi-channel input feature tensors, a unified binary mask is created, wherein 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. S3-3. Fill invalid values ​​in the multi-channel input feature tensor and landslide susceptibility label layer with 0, and normalize the valid pixel data in the multi-channel input feature tensor and landslide susceptibility label layer.

7. The landslide susceptibility assessment method based on statistical screening and deep learning according to claim 1, characterized in that: Step S4 includes: S4-1. Construct a spatial regression neural network with an encoder-decoder structure. The encoder uses a pre-trained residual network as the backbone network. Modify the first convolutional layer of the pre-trained residual network to adapt the first convolutional layer to the number of channels of the multi-channel input feature tensor and reuse the pre-trained weights of the pre-trained residual network. The encoder extracts multi-scale feature maps by downsampling step by step. S4-2. The decoder constructs an upsampling path symmetrical to the encoder. The upsampling path consists of multiple upsampling blocks. Each upsampling block contains one upsampling operation and two modules executed sequentially. Each module sequentially performs convolution operation, batch normalization processing, and activation function transformation. S4-3. In each upsampling stage of the decoder, the feature map of the encoder at the corresponding resolution is spliced ​​with the upsampled feature map in the channel dimension through skip connections, so as to fuse high-resolution spatial detail information and deep abstract semantic information. S4-4. After the decoder restores the original input resolution, the multi-channel feature map is compressed into a single channel through a 1×1 convolutional layer to generate a landslide susceptibility prediction map with the same resolution as the input feature tensor. S4-5. The spatial regression neural network is trained using a mask-weighted loss function, wherein the mask-weighted loss function calculates the prediction error based on the effective pixel region of a unified binary mask identifier.

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 the mask-weighted mean square error loss function, and the formula is as follows: ； Where L represents the loss value, and Mask(i,j) represents the binary mask value at position (i,j). Y(i,j) represents the predicted 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.

9. The landslide susceptibility assessment method based on statistical screening and deep learning according to claim 1, characterized in that: Step S5 includes: S5-1. During the training of the spatial regression neural network, after each training round, the root mean square error is calculated on the validation set using a mask-weighted loss function. The changes in the root mean square error of the validation set are monitored in real time. The model weights corresponding to the minimum root mean square error of the validation set are saved, and the model weights are used as the weights of the prediction model and loaded. S5-2. Apply the spatial regression neural network loaded with the prediction model weights to the multi-channel input feature tensor of the entire domain, and generate a normalized landslide susceptibility prediction map covering the entire domain through inference. S5-3. Post-process the normalized landslide susceptibility prediction map, reverse normalize it to the original landslide susceptibility index range, or use geospatial analysis methods such as the natural discontinuity method or the quantile method 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 assessment.

10. A landslide susceptibility assessment system based on statistical screening and deep learning, characterized in that: The system includes: 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 includes a factor set construction module and a candidate set management module. The factor set construction module is used to perform spatial coordinate system unification and resolution resampling on the raster layer corresponding to the landslide influencing factors to form an initial factor set with spatial consistency. The candidate set management module is used to create and maintain an empty candidate factor set and store the optimal factor and factor interaction combination with 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