Cross-regional landslide susceptibility assessment method and system based on collaborative transfer learning

Abstract

The invention provides a cross-regional landslide susceptibility 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 following steps: firstly, acquiring landslide disaster-pregnancy factor data of a source domain and a target domain, and performing standardized preprocessing; then, a collaborative transfer learning framework is constructed, and the framework integrates a feature-based antagonism alignment module and an instance-based adaptive reweighting 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 using the data subjected to collaborative migration processing, and generating a landslide susceptibility map of the target area. Through the synergistic effect of double migration mechanisms, the problems of poor model generalization ability and negative migration caused by domain migration and data scarcity in cross-regional application are effectively solved, and the accuracy of landslide susceptibility assessment and the geographic space rationality are remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of geological disaster prediction, geographic information systems and artificial intelligence, and specifically to a cross-regional landslide susceptibility assessment method and system based on collaborative transfer learning. Background Technology

[0002] Landslide susceptibility assessment is a crucial step in regional geological disaster risk management. In cross-regional assessment scenarios, particularly in emerging disaster areas with scarce data or regions with vastly different geological conditions, existing technologies face two significant limitations: First, assessment models heavily rely on sufficient and accurate landslide sample data from the target region for training, but such labeled data is often difficult to obtain in practice, limiting the application of data-driven models. Second, existing transfer learning methods often employ single strategies, either performing only global feature alignment without filtering out source domain noise, or only reweighting instances without bridging significant differences in feature distribution. These methods fail to address the combined challenges of "global domain shift" and "local noise," resulting in insufficient cross-regional generalization ability and unstable assessment accuracy. Therefore, developing an intelligent assessment method that can collaboratively utilize multi-source data and effectively overcome composite domain shift is of great significance for improving the accuracy, robustness, and practical application value of landslide susceptibility assessment. Summary of the Invention

[0003] The purpose of this invention is to overcome the above-mentioned shortcomings and provide a cross-regional landslide susceptibility assessment method and system based on collaborative transfer learning. This method can effectively solve the problems of composite domain offset and data scarcity in cross-regional assessment through a collaborative mechanism of global feature alignment and local instance correction, and significantly improve the prediction accuracy and spatial rationality of landslide susceptibility maps.

[0004] To achieve the above objectives, the technical solution adopted by this invention is as follows: A method for assessing cross-regional landslide susceptibility based on collaborative transfer learning, comprising the following steps: S1: Obtain landslide-prone factor data from the source and target domains, and perform preprocessing on the data to standardize and unify spatial resolution; S2: The preprocessed source and target domain data are input into the collaborative transfer learning framework for processing. The framework performs the following processing in the order of first performing feature-based adversarial alignment, followed by instance-based adaptive reweighting: (1) Feature-based adversarial alignment: global feature space adaptation is performed on the source and target domain data to construct a domain-invariant feature space; (2) Instance-based adaptive reweighting: Based on the feature space that remains unchanged in the domain, the weights of the source domain samples are adjusted to suppress noisy samples that are irrelevant to or conflict with the target domain task. S3: Using the training dataset processed by S2 collaborative transfer, train a downstream landslide susceptibility prediction model. S4: Using the prediction model trained in S3, predict the target region and generate a landslide susceptibility map.

[0005] Preferably, the landslide-prone factors include: topographic factors, geological environment factors, hydrological conditions factors, and external triggering factors.

[0006] Preferably, the topographic factors include elevation, slope, aspect, plane curvature, and profile curvature; the geological environment factors include lithology and distance from faults; the hydrological condition factors include topographic humidity index and distance from rivers; and the external triggering factors include distance from roads, peak ground acceleration, and peak ground velocity.

[0007] Preferably, the feature-based adversarial alignment processing is implemented using a VAE-GAN architecture that combines a variational autoencoder and a generative adversarial network.

[0008] Preferably, the VAE-GAN architecture maps data to the latent space through an encoder and optimizes it through adversarial training between the discriminator and the generator; the goal of the adversarial training is to minimize the composite loss function L. L=L_rec+D_KL+ L_GAN; Where L_rec is the reconstruction error, D_KL is the KL divergence, and L_GAN is the adversarial loss. To balance the hyperparameters.

[0009] Preferably, the instance-based adaptive reweighting process is implemented using the TrAdaBoost algorithm.

[0010] Preferably, the core steps of the TrAdaBoost algorithm include iteratively performing the following operations: training a weak classifier based on the current sample weights; and calculating the weighted error rate of the weak classifier on a subset of the target domain. _t; according to _t calculates the adjustment factor ; and based on Update the sample weights of the source and target domains; where, when When _t < 0.5, the update rule is configured to reduce the weight of the source domain samples and increase the weight of the target domain samples.

[0011] Preferably, the adjustment factor The calculation method is as follows: = _t / (1- _t).

[0012] Preferably, the downstream landslide susceptibility prediction model is a convolutional neural network, a bidirectional long short-term memory network, or a bidirectional gated recurrent unit.

[0013] Another aspect of the present invention provides a collaborative transfer learning landslide susceptibility assessment system for implementing the aforementioned cross-regional landslide susceptibility assessment method based on collaborative transfer learning, the system comprising: The data acquisition and preprocessing module is used to acquire landslide-prone factor data from the source and target domains, and to perform preprocessing on the data to standardize and unify spatial resolution. The collaborative transfer learning module is used to input preprocessed source domain and target domain data into the collaborative transfer learning framework for processing. It integrates a feature-based adversarial alignment submodule and an instance-based adaptive reweighting submodule. The adversarial alignment submodule is used to perform global feature space adaptation on source and target domain data to construct a domain-invariant feature space. The adaptive reweighting submodule is used to adjust the weights of the source domain samples based on the feature space that remains unchanged in the target domain, so as to suppress noisy samples that are irrelevant to or conflict with the target domain task. The model training and prediction module is used to train a downstream landslide susceptibility prediction model using the training dataset after collaborative transfer processing, and to use the trained prediction model to predict the target domain area and generate a landslide susceptibility map. The results output module is used to generate and display the landslide susceptibility map.

[0014] The present invention has the following beneficial effects: 1. This invention achieves an organic combination of global feature distribution adaptation and local instance noise filtering by constructing a transfer learning framework that coordinates "feature-based adversarial alignment" and "instance-based adaptive reweighting". This collaborative mechanism fundamentally avoids the limitations of a single transfer strategy and can effectively suppress "negative transfer".

[0015] 2. This invention has low dependence on the number of labeled samples in the target domain and exhibits strong robustness in data-scarce scenarios. Even with a very small number of samples in the target domain, a high-performance evaluation model can still be obtained by fully mining and utilizing the knowledge of the source domain through the collaborative transfer mechanism. This solves the problem of deep learning models' dependence on massive labeled data in landslide susceptibility assessment, making it possible to quickly deploy high-precision evaluation models in data-scarce scenarios such as emerging disaster areas.

[0016] 3. This invention not only improves numerical accuracy but also ensures the geospatial rationality and physical reliability of the assessment results. The landslide susceptibility map generated by this invention exhibits a continuous, gradually changing unimodal distribution of susceptibility zones, avoiding the polarization or overestimation phenomena commonly found in existing methods. The spatial distribution pattern is more consistent with geomorphological and engineering geological laws, thus providing a more reliable and intuitive decision-making basis for disaster risk management. Attached Figure Description

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0018] Figure 1 This is a schematic diagram of the overall technical roadmap of the present invention.

[0019] Figure 2 This is a schematic diagram of the VAE-GAN architecture of the present invention.

[0020] Figure 3 This is a schematic diagram of the TrAdaBoost algorithm flow and its data distribution according to the present invention.

[0021] Figure 4 A comparison chart of ROC curves obtained from cross-regional evaluation using this invention.

[0022] Figure 5 The figure shows the landslide susceptibility evaluation results under different transfer learning strategies. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the embodiments of this invention will be described in detail below with reference to the examples. However, those skilled in the art will understand that the following examples are only for illustrating this invention and should not be regarded as limiting the scope of this invention.

[0024] Example 1: This invention provides a method and system for assessing cross-regional landslide susceptibility based on collaborative transfer learning. For example... Figure 1 As shown, the overall process of this method mainly includes the following four core steps: S1, data preparation and preprocessing; S2, collaborative transfer learning processing; S3, downstream model training; and S4, susceptibility graph generation.

[0025] Step S1: Data preparation and preprocessing; First, the study areas of the source and target domains are determined. In this embodiment, the source domain is a historical earthquake landslide area in Hokkaido, Japan, and the target domain is an earthquake landslide area in the Chuetsu region of Japan. Landslide inventories (location points of landslides that have occurred) and multi-source geospatial data are collected for both regions. A total of k landslide-prone factors are extracted for both regions. In this embodiment, k=12, specifically including: elevation, slope, aspect, plane curvature, profile curvature, lithology, distance from fault, topographic moisture index (TWI), distance from river, distance from road, peak ground acceleration (PGA), and peak ground velocity (PGV). All factors are generated into raster data using GIS software and resampled to a uniform spatial resolution (e.g., 30 meters). Subsequently, all raster data are standardized (e.g., Z-score standardization) to form a source domain dataset D_s (containing m samples, e.g., m=20,000) and a target domain dataset D_t (containing n samples, e.g., n=1,000 to 5,000).

[0026] Step S2: Collaborative transfer learning processing; This step is the core of the invention, and its purpose is to enable source domain knowledge to effectively adapt to the target domain through a dual-module collaborative framework. For example... Figure 2 As shown, the framework executes the following two modules sequentially: Feature-based adversarial alignment (global adaptation): This module adopts the VAE-GAN architecture, aiming to learn a "domain-invariant" common feature space. Specifically: (1) The encoder receives input data X (i.e., 12 LCFs), maps it to the latent space through a neural network, and outputs the mean of the distribution. and standard deviation .

[0027] The latent variable Z_p was obtained through sampling using the reparameterization technique. + ;in, The random noise is sampled from the standard normal distribution N(0,1) and is used to introduce randomness to achieve differentiable sampling. The decoder / generator receives Z_p and attempts to reconstruct X'.

[0028] The discriminator is trained to distinguish whether the input is "real data" (from D_s or D_t) or "generated data" (X').

[0029] This module performs adversarial training by minimizing the following composite loss function: L=L_rec+D_KL+ L_GAN; Where L_rec is the reconstruction error, ensuring information preservation; D_KL is the KL divergence, constraining the regularity of the latent space; L_GAN is the adversarial loss, driving the encoder to learn feature representations that make D_s and D_t indistinguishable in the latent space. To balance hyperparameters, this process transforms the data from the original space to an aligned feature space.

[0030] Instance-based adaptive reweighting (local correction): This module uses the TrAdaBoost algorithm for fine-grained correction based on the alignment features output by the adversarial alignment submodule. The process is as follows: Figure 3 As shown: Initialize the sample weights of the source and target domains; Perform G iterations. In the g-th iteration: a. Train a weak classifier C_g using the current weights (as in a training phase of the downstream CNN).

[0031] b. Calculate the weighted error rate of C_g on the target domain data. _t; c. Calculate the adjustment factor = _t / (1- _t).

[0032] d. Update weights: This is the core of this module. According to the formula: For source domain sample i: = For target domain sample i: = ;in, and These are the weights of the i-th sample in the source domain and the new weights of the i-th sample in the source domain at the (g+1)-th iteration, respectively. and These are the weights of the i-th sample in the target domain and the new weights of the i-th sample in the target domain at the (g+1)-th iteration, respectively. when When _t<0.5, <1, = The weights of all source domain samples will be reduced (with misclassified samples receiving a greater reduction). = This increases the weight of samples in the target domain. This mechanism dynamically reduces the impact of "noise" samples in the source domain that are mismatched with the target domain task.

[0033] Finally, a strong classifier is obtained by combining G weak classifiers through weighted voting.

[0034] The key to this invention lies in the fact that the two modules work together in strict coordination in the order of "alignment first, then reweighting". Module (1) first solves the macroscopic inter-domain distribution differences and builds a stable feature bridge; the adaptive reweighting submodule then performs accurate instance-level filtering on this pure feature space. The two complement each other and jointly solve the complex domain offset problem that a single strategy cannot handle.

[0035] Step S3: Downstream Model Training The final training dataset (containing reweighted source domain samples and all target domain samples) after the collaborative transfer processing in step S2 is input into a downstream deep learning model for end-to-end training. In this embodiment, a convolutional neural network (CNN) is selected for the downstream model to fully capture the spatial features and correlations of landslide-causing factors.

[0036] Step S4: Generate susceptibility result graph Using the CNN model trained in step S3, pixel-by-pixel prediction is performed across the entire target study area to obtain the landslide occurrence probability value (range 0-1) for each pixel. Based on the probability values, susceptibility levels are classified (e.g., extremely high, high, medium, low, extremely low), and a visual landslide susceptibility map (LSM) is generated using GIS software, such as... Figure 5 As shown.

[0037] To verify the effectiveness of the invention, comparative experiments were conducted in Hokkaido, Japan (source domain) and Chuetsu region, Japan (target domain). Figure 4 The ROC curves show that the AUC value of the method of this invention (IATL-CNN) reaches 0.891, which is significantly better than the non-transfer model and the single transfer strategy model. Figure 5 Spatial distribution comparison shows that the susceptibility map generated by this invention is continuous and reasonable in its partitioning, avoiding the polarization or overestimation problems existing in other methods, thus proving its superior accuracy and geospatial rationality.

[0038] Example 2: Corresponding to the above method, this invention also provides a collaborative transfer learning landslide susceptibility assessment system. The system specifically includes: The data acquisition and preprocessing module is used to execute step S1, which realizes the acquisition, factor extraction and standardization of multi-source data; The collaborative transfer learning module integrates the feature-based adversarial alignment submodule and the instance-based adaptive reweighting submodule, and is used to execute step S2; The model training and prediction module is used to execute steps S3 and S4 to complete downstream model training and global prediction. The results output and visualization module is used to generate and display the final landslide susceptibility map and analysis report.

[0039] Although the preferred embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many specific modifications under the guidance of the present invention without departing from the spirit of the invention and the scope of protection of the claims, and these modifications all fall within the scope of protection of the present invention.

Claims

1. A method for assessing cross-regional landslide susceptibility based on collaborative transfer learning, characterized in that, Includes the following steps: S1: Obtain landslide-prone factor data from the source and target domains, and perform preprocessing on the data to standardize and unify spatial resolution; S2: The preprocessed source and target domain data are input into the collaborative transfer learning framework for processing. The framework performs the following processing in the order of first performing feature-based adversarial alignment, followed by instance-based adaptive reweighting: (1) Feature-based adversarial alignment: global feature space adaptation is performed on the source and target domain data to construct a domain-invariant feature space; (2) Instance-based adaptive reweighting: Based on the feature space that remains unchanged in the domain, the weights of the source domain samples are adjusted to suppress noisy samples that are irrelevant to or conflict with the target domain task. S3: Using the training dataset processed by S2 collaborative transfer, train a downstream landslide susceptibility prediction model. S4: Using the prediction model trained in S3, predict the target region and generate a landslide susceptibility map.

2. The method for assessing cross-regional landslide susceptibility based on collaborative transfer learning according to claim 1, characterized in that, The landslide-inducing factors include: topographic factors, geological environment factors, hydrological conditions factors, and external triggering factors.

3. The method for assessing cross-regional landslide susceptibility based on collaborative transfer learning according to claim 2, characterized in that, The topographic factors include elevation, slope, aspect, plane curvature, and profile curvature; the geological environment factors include lithology and distance from faults; the hydrological conditions factors include topographic humidity index and distance from rivers; and the external triggering factors include distance from roads, peak ground acceleration, and peak ground velocity.

4. The method for assessing cross-regional landslide susceptibility based on collaborative transfer learning according to claim 1, characterized in that, The feature-based adversarial alignment process is implemented using a VAE-GAN architecture that combines a variational autoencoder and a generative adversarial network.

5. The method for assessing cross-regional landslide susceptibility based on collaborative transfer learning according to claim 4, characterized in that, The VAE-GAN architecture maps data to a latent space through an encoder and optimizes it through adversarial training between the discriminator and the generator; the goal of this adversarial training is to minimize the composite loss function L. L=L_rec+D_KL+ L_GAN; Where L_rec is the reconstruction error, D_KL is the KL divergence, and L_GAN is the adversarial loss. To balance the hyperparameters.

6. The method for assessing cross-regional landslide susceptibility based on collaborative transfer learning according to claim 1, characterized in that, The instance-based adaptive reweighting process is implemented using the TrAdaBoost algorithm.

7. The method for assessing cross-regional landslide susceptibility based on collaborative transfer learning according to claim 6, characterized in that, The core steps of the TrAdaBoost algorithm include iteratively performing the following operations: training a weak classifier based on the current sample weights; and calculating the weighted error rate of the weak classifier on a subset of the target domain. _t; according to _t calculates the adjustment factor ; and based on Update the sample weights of the source and target domains; where, when When _t < 0.5, the update rule is configured to reduce the weight of source domain samples and increase the weight of target domain samples.

8. The method for assessing cross-regional landslide susceptibility based on collaborative transfer learning according to claim 7, characterized in that, The adjustment factor The calculation method is as follows: = _t / (1- _t).

9. The method for assessing cross-regional landslide susceptibility based on collaborative transfer learning according to claim 1, characterized in that, The downstream landslide susceptibility prediction model is a convolutional neural network, a bidirectional long short-term memory network, or a bidirectional gated recurrent unit.

10. A collaborative transfer learning landslide susceptibility assessment system, used to implement the method according to any one of claims 1 to 9, characterized in that, The system includes: The data acquisition and preprocessing module is used to acquire landslide-prone factor data from the source and target domains, and to perform preprocessing on the data to standardize and unify spatial resolution. The collaborative transfer learning module is used to input preprocessed source domain and target domain data into the collaborative transfer learning framework for processing. It integrates a feature-based adversarial alignment submodule and an instance-based adaptive reweighting submodule. The adversarial alignment submodule is used to perform global feature space adaptation on source and target domain data to construct a domain-invariant feature space. The adaptive reweighting submodule is used to adjust the weights of the source domain samples based on the feature space that remains unchanged in the target domain, so as to suppress noisy samples that are irrelevant to or conflict with the target domain task. The model training and prediction module is used to train a downstream landslide susceptibility prediction model using the training dataset after collaborative transfer processing, and to use the trained prediction model to predict the target domain area and generate a landslide susceptibility map. The results output module is used to generate and display the landslide susceptibility map.

Images