A Landslide Susceptibility Assessment Method With Dynamic And Static Features Combined With InSAR Deformation Information

Assessing landslide susceptibility is crucial in understanding the relationship between landslides and disaster-prone environments.A comprehensive assessment can provide valuable insights into the probability of unknown landslides occurring in similar environments,making it an effective tool for sustainable management of land resources and landslide area positioning.The quantitative modeling process involves acquiring landslide and non-landslide samples,extracting relevant features,and constructing a model.However,due to the complex mechanism of landslides and numerous influencing factors,utilizing multiple features for reliable modeling remains a challenge.Existing susceptibility assessment methods mainly rely on short-term static features of influencing factors and fail to consider the temporal dynamic features during landslide occurrence,resulting in poor reliability of landslide susceptibility assessment results.Therefore,this paper proposes an integrated approach to landslide susceptibility assessment that combines dynamic of InSAR deformation information and static features.We attempt to use parallel deep learning networks to achieve the comprehensive utilization of these features,thereby improving the reliability of results and providing innovative ideas for landslide susceptibility assessment research.This paper takes the main urban area of Lanzhou City in Gansu Province as the research object,and utilizes SBAS-InSAR technology to obtain surface deformation information from2015 to 2020.By combining time-series InSAR deformation information,Sentinel-2A and Google Earth images,landslide identification rules for the study area are established to identify potential landslide samples.To address the issue of missed landslides and limited samples,a semi-supervised generative adversarial landslide identification method is established to supplement landslide samples.Based on the dynamic features of time-series InSAR deformation information and the static features of environmental influencing factors,a joint dynamic-static feature integration landslide susceptibility assessment method based on the InSAR deformation information is constructed to achieve reliable landslide susceptibility mapping of the main urban area of Lanzhou City.The main conclusions are as follows:(1)Between 2015 and 2020,the main urban area of Lanzhou City remained mostly stable,while significant deformation was observed in certain areas such as mountainous regions,cultivated land,land reclamation areas,and mining sites,making them high-risk areas for landslides.Based on an InSAR average deformation rate threshold of ±10mm/year,combined with optical images and field investigations,landslide identification constraint conditions were established to preliminarily identify potential landslides in the main urban area of Lanzhou.A total of 152 landslide bodies were identified,with a total area of approximately 7.189 km~2.(2)A semi-supervised generative adversarial network(SSGAN)method for landslide recognition was constructed,which includes labeled and unlabeled samples to supplement missing landslide samples.The constructed SSGAN landslide recognition method showed good adaptability in the combined dataset of optical images and influencing factor features.After training,the model’s discriminator reduced the error rate of landslide judgment by approximately 5.5% compared to using only a single feature dataset.The generated images by the generator were closer to the real sample distribution.The proposed SSGAN method comprehensively learned the complex landslide environmental features and outperformed traditional unsupervised adversarial learning models in terms of precision,F1 score,Kappa coefficient,and MIo U evaluation indices,indicating better overall performance.Based on the proposed SSGAN,a total of 160 landslides were identified in the main urban area of Lanzhou,with a total area of approximately 10.328 km~2.Eight landslides were added,resulting in an increase in the number of landslide samples.(3)The time distributed convolutional neural network(TD-CNN),bidirectional gated recurrent unit(Bi-GRU),and multi-scale convolutional neural network(MSCNN)were used to construct a deep learning model that combines dynamic and static features to assess landslide susceptibility in the main urban area of Lanzhou.By using identified landslide samples,this study proposed a fusion network method to achieve a better performance than traditional models that only use static features.The proposed method achieved an accuracy increase of 2.7% and an MSE decrease of 1.1% compared to traditional methods during the training process.In the confusion matrix,all indicators of the proposed method were higher than those of the traditional method,and the AUC of the test set reached 0.9089.The susceptibility results showed that the proposed method has an obvious characterization of areas with high susceptibility levels.In only 10.18% of the study area,it accurately covered84.79% of the historical landslide area.The subjective and objective results showed that the integrated model combining dynamic and static features proposed in this study can effectively improve the reliability of landslide susceptibility assessment by fully utilizing landslide features and ensuring the stability of the overall learning process.Landslide distribution in the main urban area of Lanzhou is concentrated in areas with complex geological structures,high soil moisture content,and frequent human engineering activities.To reduce disaster risk,it is suggested to increase vegetation coverage,control human engineering activities,strengthen infrastructure construction,and enhance geological disaster monitoring.