| The landslide hazards in China cover a wide range,and their complex and diverse inducing environments greatly threatens people’s lives and properties and the safety of major national infrastructure construction.Precise evaluation of landslide susceptibility in largescale scene can provide theoretical guidance for accurate identification and prevention of large-scale landslide risks,regional planning and decision-making,reliable landslide monitoring and early warning.In recent years,data-driven methods have played an important role in landslide susceptibility analysis.On the one hand,the development of remote sensing information detection technology provides rich thematic information for landslide hidden danger detection.On the other hand,the development of deep learning theory provides solid theoretical support for landslide susceptibility analysis based on data-driven methods.However,when dealt with the task in a large-scale region,such as the construction of major projects in western administrative province of China and areas along the Sichuan-Tibet Railway,the existing data-driven landslide susceptibility prediction methods still face several problems: 1)There are few types of landslide thematic data,interrelated inducing factors,noise interference,which would result in poor discriminative model;2)In some areas,there is a lack of disaster-pregnancy knowledge to support landslide susceptibility analysis,the worse is that the model has poor transferability;and 3)The uneven distribution of landslide samples leads to the very lack of samples in some areas,and it is difficult to fully train the landslide susceptibility prediction model.In response to the above problems,this paper proposes a precise analysis of landslide susceptibility in large-scale scene based on representation learning,focusing on the study of unsupervised representation learning for landslide thematic data,Landslide-inducing knowledge transferring based on generative adversarial mechanism,and few-shot block-wise prediction of landslide susceptibility based on meta learning.This paper develops prototype systems,and selects several scenes in the middle and upper reaches of the Yangtze River as cases,to verify the validity of the proposed method by analyzing the landslide susceptibility prediction performance of Fengjie County,Qijiang District and Fuling District in Chongqing.The following gives the specific research contents:(1)Unsupervised representation learning for landslide thematic data.Aiming at the problem of poor discriminant ability of the model caused by the lack of types of landslide thematic data,interrelated hazard factors,noise interference,this paper stacks unsupervised modular restricted Boltzmann machines and denoising autoencoders,based on a greedy layerwise pre-training algorithm,so that the integrated model inherits the unique properties of each module,and improves the robustness,noise resistance and compactness of the model.The experimental results show that the unsupervised representation learned from the landslide thematic features is easier to be embedded into the discriminative manifold,and a more robust and reasonable landslide sensitive map is predicted.Compared with algorithms such as support vector machine,multilayer perceptron,and random forest,the accuracy is improved by 3%-5%.(2)Landslide-inducing knowledge transferring based on generative adversarial mechanism.Due to economic and geographical constraints,some areas in a wide area lack landslide-inducing knowledge to support landslide susceptibility analysis.Therefore,it is necessary to transfer the general concepts from rich knowledge areas.Through the training of generative adversarial mechanism,this paper enables the model to generate the representation that can confuse a strong discriminant model to recognize its domain label,which mines the shared landslide inducing factors and rises the domain adaption ability of the model.The experimental results show that the cross-domain transferring performance of the proposed model,trained based on the generative adversarial mechanism,is better than other models.Compared with the multi-layer perceptron with the same configuration,the training efficiency and prediction accuracy are greatly improved.(3)Few-shot block-wise prediction of landslide susceptibility based on meta learning.Various parts of the large-scale scene have different landslide-inducing environments,it is difficult for a single model to explain multiple landslide causative patterns at the same time.In addition,the uneven distribution of landslide records would cause landslide samples extremely scarce in some areas,which can lead to impossibility of well training a landslide susceptibility prediction model.For this reason,this paper decomposes the landslide susceptibility prediction task into multi-tasks,constructing a metadata set,and then metalearns an intermediate model that can be quickly adapted to each subtask with a very small number of samples and iterations,which solves the problem of predicting landslide susceptibility with insufficient training samples.The experimental results show that,compared with other algorithms,the proposed meta-learned intermediate model has the best few-shot adaption performance.At the same time,the proposed method alleviates the problem of inaccurate supervision to a certain extent. |
