CN-122022212-A - Cross-region landslide vulnerability evaluation method based on condition domain antagonistic neural network

CN122022212ACN 122022212 ACN122022212 ACN 122022212ACN-122022212-A

Abstract

The application provides a cross-regional landslide vulnerability evaluation method based on a condition domain antagonistic neural network. According to the method, historical landslide point data and non-landslide area sample data of a source area are obtained to construct a source area labeled sample data set, environment factor data of a target area are obtained to construct a target area unlabeled sample data set, then environment factors related to landslide occurrence are selected, unified spatial preprocessing is conducted on the environment factors to construct environment factor characteristic input data of the source area and the target area, a condition domain countermeasure neural network model is built, joint training is conducted on the basis of the source area labeled sample data set and the target area unlabeled sample data set, finally the trained condition domain countermeasure neural network model is applied to the environment factor characteristic input data of the target area, landslide occurrence probability values of evaluation units of the target area are obtained, and landslide occurrence probability values are based on the landslide occurrence probability values to evaluate the landslide occurrence probability of the target area.

Inventors

Zong Leli
LIU JIAN
CHEN ZI
NIU XIAONAN
ZHOU MO
SUN YANWEI
WANG SHANGXIAO
Xiao Shengjuan
ZHANG YIWU
TANG ZHIMIN

Assignees

中国地质调查局南京地质调查中心(华东地质科技创新中心)

Dates

Publication Date: 20260512
Application Date: 20260416

Claims (10)

1. A cross-regional landslide vulnerability evaluation method based on a condition domain antagonistic neural network is characterized by comprising the following steps: acquiring historical landslide point data and non-landslide region sample data of a source region to construct a sample data set with labels of the source region, and acquiring environmental factor data of a target region to construct a sample data set without labels of the target region; Selecting an environmental factor related to landslide occurrence, and carrying out unified spatial preprocessing on the environmental factor to construct environmental factor characteristic input data of the source region and the target region; constructing a conditional domain antagonistic neural network model, and performing joint training based on the labeled sample data set of the source region and the unlabeled sample data set of the target region; And applying the trained condition domain antagonistic neural network model to the environmental factor characteristic input data of the target area to obtain landslide occurrence probability values of all evaluation units of the target area, and evaluating the landslide susceptibility of the target area based on the landslide occurrence probability values.
2. The method of claim 1, wherein the condition domain antagonistic neural network model comprises a feature extraction module, a classification module, and a condition domain discrimination module; The feature extraction module is used for extracting feature representation from the environmental factor feature input data, the classification module is used for carrying out landslide/non-landslide classification on the source region labeled sample, and the condition domain judging module is used for carrying out domain judgment on the source region sample and the target region sample under the condition that the classification prediction result output by the classification module is introduced as the condition information.
3. The method of claim 1, wherein in the co-training of the conditional domain against neural network model, model parameters are optimized using a composite loss function comprising classification loss, domain against loss, against regularization loss, and conditional against entropy loss.
4. The method of claim 1, wherein the obtaining historical landslide point data and non-landslide region sample data for the source region to construct the source region labeled sample dataset comprises: Acquiring the space coordinates of the historical landslide points of the source region, and taking the historical landslide points as landslide positive samples; randomly selecting sample points with the same number as the positive samples as non-landslide negative samples in a stable area after removing rivers, water bodies and known landslide buffer areas; and extracting corresponding environmental factor characteristic values at positions of the landslide positive samples and the non-landslide negative samples, marking the landslide positive samples as landslide categories, and marking the non-landslide negative samples as non-landslide categories so as to construct a source region labeled sample data set.
5. The method of claim 1, wherein the obtaining the environmental factor data for the target region to construct a target region unlabeled exemplar dataset comprises: Acquiring environmental factor raster data in the target area range, and extracting corresponding environmental factor characteristic values by taking each evaluation unit or pixel in the target area as a unit under the unified spatial resolution; And taking the environmental factor characteristic value as a target area unlabeled sample data set for domain adaptation learning in the conditional domain antagonistic neural network model.
6. The method of claim 1, wherein the selecting the environmental factor associated with landslide occurrence and performing uniform spatial preprocessing on the environmental factor comprises: Selecting at least one of a topography factor, a geological factor, a hydrologic factor, a surface coverage factor and a human activity factor as an input factor for landslide susceptibility evaluation; performing unified projection, clipping and resampling processing on the environmental factor data to enable the environmental factor data of the source region and the target region to have a consistent spatial reference system and spatial resolution; and extracting corresponding environmental factor values at the source region sample position and the target region evaluation unit position according to a preset neighborhood window size, and constructing an input characteristic sample for the condition domain antagonistic neural network model.
7. The method of claim 1, wherein constructing the conditional domain antagonistic neural network model and performing joint training based on the source region labeled sample dataset and the target region unlabeled sample dataset comprises: Aiming at a sample with a label in a source region, a classification loss is defined by using cross entropy between a classification prediction result output by a classification module and a real classification label, and the classification loss is used for restricting landslide/non-landslide discrimination performance of a model on the source region; constructing domain countermeasures against losses by utilizing domain discrimination results output by a conditional domain discrimination module aiming at a source region sample in the source region labeled sample data set and a target region sample in the target region unlabeled sample data set; Constructing an anti-regularization loss, and taking the distance between feature centroids corresponding to the same category in the source region and the target region as a constraint; Constructing conditions against entropy loss, and taking entropy of the target area sample in the output space of the classification module as an optimization target; and linearly combining the loss terms according to preset weights to form a composite loss function, and simultaneously updating network parameters of the feature extraction module, the classification module and the condition domain judging module through a back propagation algorithm.
8. The method according to claim 1, wherein the applying the trained condition domain antagonistic neural network model to the environmental factor characteristic input data of the target area, obtaining a landslide occurrence probability value of each evaluation unit of the target area, and performing landslide susceptibility evaluation on the target area based on the landslide occurrence probability value, includes: inputting a trained condition domain antagonistic neural network model to all evaluation units or pixels in the target area range, and acquiring landslide occurrence probability of each evaluation unit as a landslide susceptibility index; Classifying the landslide susceptibility indexes by adopting a natural breakpoint method, and dividing a target area into a plurality of different landslide susceptibility grade areas; and generating a landslide vulnerability partition map of the target area based on the dividing result, wherein the landslide vulnerability partition map is used for assisting in landslide disaster risk identification.
9. An electronic device, comprising: Processor, and A memory for storing executable instructions of the processor; Wherein the processor is configured to perform the method of any one of claims 1 to 8 via execution of the executable instructions.
10. A computer readable storage medium having stored therein computer executable instructions which when executed by a processor are adapted to carry out the method of any one of claims 1 to 8.

Description

Cross-region landslide vulnerability evaluation method based on condition domain antagonistic neural network Technical Field The application relates to a data processing technology, in particular to a cross-regional landslide vulnerability evaluation method based on a condition domain antagonistic neural network. Background The landslide susceptibility evaluation aims to quantitatively characterize the occurrence probability of landslide in space, and is a key link of geological disaster risk identification and control decision. In the prior art, machine learning and deep learning methods such as random forests, support vector machines and deep neural networks can well learn complex nonlinear relations between landslide occurrence and environmental factors such as terrain, geology, hydrology, earth surface coverage and human activities in a single area, so that area landslide susceptibility evaluation with higher precision is realized. However, the method generally relies on landslide cataloging data with larger scale and higher quality in a research area to train and verify, when a model trained in a certain source area needs to be directly migrated to a target area with different geological environment and topography conditions, the traditional model is difficult to maintain the original discrimination capability due to the obvious difference of environmental factor distribution among different areas, and therefore accurate evaluation of landslide susceptibility of the target area is difficult to realize under the condition of lacking a landslide labeling sample of the target area. Disclosure of Invention The application provides a cross-region landslide susceptibility evaluation method based on a condition domain antagonistic neural network, which is used for outputting landslide occurrence probability distribution with higher spatial consistency and discriminant under the condition that a target region has few landslide labeling samples and even no labeling at all. In a first aspect, the present application provides a method for evaluating the susceptibility of a cross-regional landslide based on a condition domain countermeasure neural network, including: acquiring historical landslide point data and non-landslide region sample data of a source region to construct a sample data set with labels of the source region, and acquiring environmental factor data of a target region to construct a sample data set without labels of the target region; Selecting an environmental factor related to landslide occurrence, and carrying out unified spatial preprocessing on the environmental factor to construct environmental factor characteristic input data of the source region and the target region; constructing a conditional domain antagonistic neural network model, and performing joint training based on the labeled sample data set of the source region and the unlabeled sample data set of the target region; And applying the trained condition domain antagonistic neural network model to the environmental factor characteristic input data of the target area to obtain landslide occurrence probability values of all evaluation units of the target area, and evaluating the landslide susceptibility of the target area based on the landslide occurrence probability values. In a second aspect, the present application provides an electronic device comprising: Processor, and A memory for storing executable instructions of the processor; wherein the processor is configured to perform any one of the possible methods described in the first aspect via execution of the executable instructions. In a third aspect, the present application provides a computer-readable storage medium having stored therein computer-executable instructions which, when executed by a processor, are adapted to carry out any one of the possible methods described in the first aspect. According to the cross-region landslide susceptibility evaluation method based on the condition domain anti-neural network, through acquiring historical landslide point data and non-landslide region sample data of a source region, a source region labeled sample data set is constructed, environment factor data of a target region is acquired, a target region label-free sample data set is constructed, then environment factors related to landslide occurrence are selected, unified spatial preprocessing is conducted on the environment factors, environment factor characteristic input data of the source region and the target region are constructed, a condition domain anti-neural network model is constructed, combined training is conducted on the basis of the source region labeled sample data set and the target region label-free sample data set, finally, the trained condition domain anti-neural network model is applied to the environment factor characteristic input data of the target region, the landslide occurrence probability value of each evaluation unit of the target region is obtained, and lan