CN-122020169-A - Cross-regional landslide vulnerability assessment method and system based on collaborative transfer learning

CN122020169ACN 122020169 ACN122020169 ACN 122020169ACN-122020169-A

Abstract

The invention provides a cross-regional landslide vulnerability assessment method and system based on collaborative transfer learning, and belongs to the technical field of geological disaster prediction and geographic information systems. The method comprises the steps of firstly obtaining landslide pregnancy factor data of a source domain and a target domain, and carrying out standardized pretreatment. Then, a collaborative migration learning framework is constructed, and the framework integrates a feature-based antagonism alignment module and an instance-based adaptive re-weighting module to realize global feature distribution alignment and local instance noise filtering between a source domain and a target domain. And then training a deep learning model by utilizing the data subjected to the collaborative migration processing, and generating a landslide vulnerability map of the target area. The invention effectively solves the problems of poor model generalization capability and negative migration caused by domain deviation and data scarcity in cross-regional application by the synergistic effect of a double migration mechanism, and remarkably improves the accuracy of landslide susceptibility assessment and the geographic space rationality.

Inventors

GUO FEI
DOU JIE
Dong Aonan
HUANG GAOYU
PANG HAO
Gui Xudong
CHEN CHANGQING
ZHANG DAGAO
LI ZONGHUI

Assignees

三峡大学
中国地质大学（武汉）

Dates

Publication Date: 20260512
Application Date: 20260128

Claims (10)

1. A cross-regional landslide susceptibility assessment method based on collaborative transfer learning is characterized by comprising the following steps: s1, acquiring landslide pregnancy factor data of a source domain and a target domain, and preprocessing the data in a standardized and spatial resolution unified way; S2, inputting the preprocessed source domain and target domain data into the collaborative transfer learning framework for processing, wherein the framework executes the following processing in the order of executing feature-based antagonism alignment firstly and executing instance-based adaptive weighting later: (1) Performing global feature space adaptation on the source domain and target domain data to construct a domain-invariant feature space; (2) Performing instance-based adaptive re-weighting processing, namely performing weight adjustment on source domain samples on the basis of the domain-invariant feature space so as to inhibit noise samples irrelevant to or conflicting with target domain tasks; s3, training a downstream landslide susceptibility prediction model by using the training data set after the S2 collaborative migration treatment; and S4, predicting the target domain area by using the prediction model trained in the step S3, and generating a landslide susceptibility map.
2. The method for estimating the susceptibility of a cross-regional landslide based on cooperative transfer learning of claim 1, wherein the landslide pregnancy factor comprises a topographic factor, a geological environment factor, a hydrological condition factor and an external trigger factor.
3. The cross-regional landslide susceptibility assessment method based on collaborative transfer learning according to claim 2, wherein the topographical factors comprise elevation, gradient, slope direction, plane curvature and section curvature, the geological environment factors comprise lithology and distance from fault, the hydrologic condition factors comprise topographical humidity index and distance from river, and the external trigger factors comprise distance from road, peak earth movement acceleration and peak earth movement speed.
4. The method for cross-regional landslide vulnerability assessment based on collaborative transfer learning according to claim 1, wherein the feature-based antagonism alignment process is implemented by using a VAE-GAN architecture based on a combination of a variational self-encoder and a generation antagonism network.
5. The method for cross-regional landslide vulnerability assessment based on collaborative transfer learning according to claim 4, wherein the VAE-GAN architecture maps data to a latent space through an encoder and optimizes the challenge training of a discriminator and a generator, wherein the objective of the challenge training is to minimize a complex loss function L: L=L_rec+D_KL+ L_GAN; wherein, L_rec is reconstruction error, D_KL is KL divergence, L_GAN is counterloss, Is a balanced superparameter.
6. The cross-regional landslide susceptibility assessment method based on collaborative transfer learning according to claim 1, wherein the instance-based adaptive weighting process is implemented by adopting TrAdaBoost algorithm.
7. The method for cross-regional landslide vulnerability assessment based on collaborative transfer learning of claim 6, wherein the key step of TrAdaBoost algorithm comprises iteratively training a weak classifier based on current sample weights, calculating a weighted error rate of the weak classifier over a subset of target domains T according to T calculating an adjustment factor And according to Updating sample weights of a source domain and a target domain, wherein when When t <0.5, the update rule is configured to decrease the source domain sample weight and increase the target domain sample weight.
8. The method for cross-regional landslide susceptibility assessment based on collaborative transfer learning of claim 7, wherein the adjustment factor is The calculation mode of (a) is as follows: = _t/(1- _t)。
9. The cross-regional landslide susceptibility assessment method based on collaborative transfer learning according to claim 1, wherein the downstream landslide susceptibility prediction model is a convolutional neural network, a bidirectional long-short-term memory network or a bidirectional gating circulation unit.
10. A co-migration learning landslide vulnerability assessment system for implementing the method of any one of claims 1 to 9, characterized in that the system comprises: The data acquisition and preprocessing module is used for acquiring landslide pregnancy factor data of a source domain and a target domain, and preprocessing the data in a standardized and spatial resolution unified way; The collaborative transfer learning module is used for inputting the preprocessed source domain and target domain data into the collaborative transfer learning framework for processing, and integrates a feature-based antagonism alignment sub-module and an instance-based self-adaptive weighting sub-module; The antagonism alignment submodule is used for carrying out global feature space adaptation on the source domain and the target domain data so as to construct a domain-unchanged feature space; the self-adaptive weighting sub-module is used for carrying out weight adjustment on the source domain samples on the basis of the domain-invariant feature space so as to inhibit noise samples irrelevant to or conflicting with the target domain tasks; The model training and predicting module is used for training a downstream landslide susceptibility prediction model by using the training data set after the collaborative migration treatment, and predicting a target domain area by using the trained prediction model to generate a landslide susceptibility graph; and the result output module is used for generating and displaying the landslide susceptibility map.

Description

Cross-regional landslide vulnerability assessment method and system based on collaborative transfer learning Technical Field The invention relates to the technical field of geological disaster prediction, geographic information system and artificial intelligence intersection, in particular to a cross-regional landslide vulnerability assessment method and system based on collaborative transfer learning. Background Landslide susceptibility assessment is a key link of regional geological disaster risk management and control. In a trans-regional assessment scene, particularly in an emerging disaster area with deficient data or an area with different geological conditions, the prior art faces two prominent limitations that firstly an assessment model is highly dependent on adequate and accurate landslide sample data of a target area to train, but the marked data are difficult to acquire in practice, so that the application of a data driving model is limited, secondly the conventional migration learning method adopts a single strategy, or only performs global feature alignment but cannot filter source domain noise, or only performs instance re-weighting but cannot bridge significant feature distribution differences, and cannot cooperatively cope with the complex challenges of global domain offset and local noise, so that the trans-regional generalization capability of the model is insufficient and the assessment precision is unstable. Therefore, the intelligent evaluation method capable of cooperatively utilizing the multi-source data and effectively overcoming the composite domain offset is developed, and has important significance for improving the accuracy, the robustness and the practical application value of landslide susceptibility evaluation. Disclosure of Invention The invention aims to overcome the defects and provide a cross-regional landslide vulnerability assessment method and a system based on collaborative transfer learning, which can realize collaborative mechanism of global feature alignment and local instance correction, the method effectively solves the problems of compound domain offset and data scarcity in cross-regional evaluation, and remarkably improves the prediction precision and space rationality of landslide susceptibility maps. In order to achieve the purpose, the technical scheme adopted by the invention is as follows, a cross-regional landslide susceptibility assessment method based on collaborative transfer learning comprises the following steps: s1, acquiring landslide pregnancy factor data of a source domain and a target domain, and preprocessing the data in a standardized and spatial resolution unified way; S2, inputting the preprocessed source domain and target domain data into the collaborative transfer learning framework for processing, wherein the framework executes the following processing in the order of executing feature-based antagonism alignment firstly and executing instance-based adaptive weighting later: (1) Performing global feature space adaptation on the source domain and target domain data to construct a domain-invariant feature space; (2) Performing instance-based adaptive re-weighting processing, namely performing weight adjustment on source domain samples on the basis of the domain-invariant feature space so as to inhibit noise samples irrelevant to or conflicting with target domain tasks; s3, training a downstream landslide susceptibility prediction model by using the training data set after the S2 collaborative migration treatment; and S4, predicting the target domain area by using the prediction model trained in the step S3, and generating a landslide susceptibility map. Preferably, the landslide pregnancy factor comprises a topography factor, a geological environment factor, a hydrologic condition factor and an external trigger factor. Preferably, the topography factors comprise elevation, gradient, slope direction, plane curvature and section curvature, the geological environment factors comprise lithology and distance from fault, the hydrologic condition factors comprise topography humidity index and distance from river, and the external trigger factors comprise distance from road, peak earth movement acceleration and peak earth movement velocity. Preferably, the feature-based resistance alignment process is implemented using a variational self-encoder based VAE-GAN architecture in combination with generating a resistance network. Preferably, the VAE-GAN architecture maps data to the latent space through an encoder and optimizes through the countermeasure training of the discriminators and generators, the objective of which is to minimize the composite loss function L: L=L_rec+D_KL+L_GAN; wherein, L_rec is reconstruction error, D_KL is KL divergence, L_GAN is counterloss, Is a balanced superparameter. Preferably, the instance-based adaptive re-weighting process is implemented using TrAdaBoost algorithm. Preferably, the key steps of the TrAdaBoost algorith