1.School of River and Ocean, Chongqing Jiaotong University, Chongqing 400074, China
2.Key Laboratory of Geological Disaster Reduction for Highway and Waterway Transportation in Mountainous Areas, Chongqing Municipal Education Commission, Chongqing Jiaotong University, Chongqing 400074, China
3.Institute of Mountain Hazards and Environment, CAS, Chengdu 610041, China
Objective Landslides represent a severe and frequent natural hazard, posing significant threats to human life and property. Current models for predicting landslide susceptibility exhibit two primary limitations: the inability to fully capture the spatial heterogeneity of environmental factors such as terrain, soil, and vegetation, and the failure to accurately distinguish between landslides induced by extreme and non-extreme rainfall events. These shortcomings hinder accurate forecasting and reduce the models' adaptability to diverse environmental conditions and rapidly changing climatic patterns. Therefore, this study introduces an innovative approach that combines Deep Embedded Clustering (DEC) with a Dynamic Rainfall Threshold (DRT) model based on a mixed distribution. In addition, a Multi-Task Learning Adaptive Neural Tree (MLANT) model has been developed to enhance model flexibility and prediction accuracy, particularly in varying environmental conditions and during extreme weather events. Methods This research applied three key methodologies to address the limitations of existing landslide susceptibility models. Deep Embedded Clustering (DEC): DEC was utilized to resolve spatial heterogeneity issues. Using deep learning techniques, environmental variables such as terrain, soil, and vegetation were embedded into a low-dimensional space to capture their complex, nonlinear relationships. The model clustered these representations to identify sub-regions with similar geological and environmental features. This clustering-based zoning enhanced the model's capability to analyze landslide susceptibility by managing spatial heterogeneity more effectively than traditional methods. Dynamic Rainfall Threshold (DRT) model: The DRT model was introduced to improve forecasting accuracy by distinguishing between the effects of extreme and non-extreme rainfall on landslide initiation. It integrated Gamma and Generalized Pareto Distributions (GPD) to represent rainfall events of varying intensities. Bayesian methods were employed to dynamically update model parameters, enabling real-time adaptation to changing rainfall conditions. This allowed for greater prediction precision and timeliness under both extreme and non-extreme rainfall events. Multi-Task Learning Adaptive Neural Tree (MLANT): The MLANT model was developed to address the inflexibility and limited adaptability of traditional prediction models. Unlike conventional models, which struggled to adjust to diverse prediction tasks and shifting environmental conditions, MLANT dynamically modifies its processing strategies to match specific tasks and environmental variations. This multi-task learning approach enhances the model's accuracy and flexibility in predicting landslides, particularly in nonlinear environments with complex terrain and rapidly changing weather conditions. The models were validated using real-world data from Tongjiang County, a region prone to landslides. Inputs such as terrain features, soil characteristics, vegetation indices, and rainfall patterns were utilized to train and evaluate the models. The DEC, DRT, and MLANT models were compared to traditional susceptibility models to assess improvements in prediction accuracy and adaptability. Results and Discussions The results demonstrated significant improvements in prediction accuracy when using the proposed models. DEC-Based Clustering: The DEC model predicted higher landslide densities in high- and very-high-susceptibility zones by capturing spatial heterogeneity. Specifically, it predicted landslide densities of 0.036 92 and 0.046 92 events per square kilometer in high and very high-risk zones, respectively, identifying a total of 59 landslide events. This marked an improvement over traditional models that did not consider spatial heterogeneity, such as the overall effective rainfall coefficient model, which predicted fewer events and lower densities. These findings highlighted the importance of integrating spatial heterogeneity in susceptibility modeling. Dynamic Rainfall Threshold Model (DRT): The DRT model further enhanced accuracy by effectively distinguishing the effects of different rainfall intensities. Its mixed-distribution approach, incorporating both Gamma and GPD distributions, improved the accuracy of landslide predictions under dynamic climatic conditions. In the practical application to Tongjiang County, the DRT model also predicted 59 landslide events in high and very high susceptibility zones, with densities of 0.036 92 and 0.046 92 events per square kilometer, respectively. This outperformed the DEC-based effective rainfall coefficient model, which exhibited lower accuracy and fewer correctly identified events. MLANT Performance: The MLANT model significantly improved flexibility and prediction performance across varied tasks. MLANT efficiently addressed various environmental conditions by dynamically adjusting its internal strategies, particularly in response to landslides triggered by extreme rainfall. Evaluation metrics, such as precision, F1 score, and ROC-AUC, demonstrated that MLANT outperformed traditional models that rely on static thresholds. In Tongjiang County, MLANT increased predicted landslide density from 0.038 events/km² and 44 events (using traditional methods) to 0.044 events/km² and 59 events, demonstrating superior performance in both frequent and rare landslide scenarios. Conclusions The models developed in this research effectively overcome the limitations of traditional landslide susceptibility prediction approaches by integrating Deep Embedded Clustering (DEC), a Dynamic Rainfall Threshold (DRT) model based on mixed distributions, and a Multi-Task Learning Adaptive Neural Tree (MLANT). Spatial heterogeneity is addressed through DEC clustering, which dynamically partitions regions based on environmental characteristics, improving prediction accuracy. Rainfall differentiation is achieved through the DRT model, which significantly improves the model's capability to predict landslides triggered by both extreme and non-extreme rainfall events by adapting to varying climatic conditions. The MLANT model provides flexibility and adaptability, delivering improved accuracy across multiple prediction tasks and in rapidly changing environmental conditions, outperforming traditional single-task models. These findings demonstrate the critical importance and effectiveness of integrating DEC, DRT, and MLANT models in enhancing landslide susceptibility prediction, particularly in areas characterized by complex environmental conditions and frequent extreme weather events. This research contributes both theoretical and practical advancements, providing valuable tools for enhancing landslide risk management and mitigation strategies.
WhiteleyJ S, ChambersJ E, UhlemannS,et al.Geophysical monitoring of moisture-induced landslides:A review[J].Reviews of Geophysics,2019,57(1):106‒145.
[2]
BasuT, PalS.A GIS-based factor clustering and landslide susceptibility analysis using AHP for Gish River Basin,India[J].Environment,Development and Sustainability,2020,22(5):4787‒4819.
[3]
AditianA, KubotaT, ShinoharaY.Comparison of GIS-based landslide susceptibility models using frequency ratio,logistic regression,and artificial neural network in a tertiary region of Ambon,Indonesia[J].Geomorphology,2018,318:101‒111.
[4]
ChenWenwu, ZhangShuai.GIS-based comparative study of Bayes network,Hoeffding tree and logistic model tree for landslide susceptibility modeling[J].CATENA,2021,203:105344.
[5]
HuangYu, ZhaoLu.Review on landslide susceptibility mapping using support vector machines[J].CATENA,2018,165:520‒529.
[6]
MerghadiA, AbderrahmaneB, Tien BuiD.Landslide susceptibility assessment at Mila Basin(Algeria):A comparative assessment of prediction capability of advanced machine learning methods[J].ISPRS International Journal of Geo-Information,2018,7(7):268.
[7]
YoussefA M, PourghasemiH R.Landslide susceptibility mapping using machine learning algorithms and comparison of their performance at Abha Basin,Asir Region,Saudi Arabia[J].Geoscience Frontiers,2021,12(2):639‒655.
[8]
HongHaoyuan, PradhanB, SameenM I,et al.Spatial prediction of rotational landslide using geographically weighted regression,logistic regression,and support vector machine models in Xing Guo area (China)[J].Geomatics,Natural Hazards and Risk,2017,8(2):1997‒2022.
[9]
FangZhice, WangYi, DuanGonghao,et al.Landslide susceptibility mapping using rotation forest ensemble technique with different decision trees in the Three Gorges Reservoir area,China[J].Remote Sensing,2021,13(2):238.
[10]
YangYang, YangJintao, XuChengdong,et al.Local-scale landslide susceptibility mapping using the B-GeoSVC model[J].Landslides,2019,16(7):1301‒1312.
[11]
DengMin, YangWentao, LiuQiliang,et al.Heterogeneous space-time artificial neural networks for space-time series prediction[J].Transactions in GIS,2018,22(1):183‒201.
[12]
WangQian, WangYi, NiuRuiqing,et al.Integration of information theory,K-means cluster analysis and the logistic regression model for landslide susceptibility mapping in the Three Gorges area,China[J].Remote Sensing,2017,9(9):938.
[13]
GodtJ W, BaumR L, SavageW Z,et al.Transient deterministic shallow landslide modeling:Requirements for susceptibility and hazard assessments in a GIS framework[J].Engineering Geology,2008,102(3/4):214‒226.
[14]
Nedumpallile VasuN, LeeS R, PradhanA M S,et al.A new approach to temporal modelling for landslide hazard assessment using an extreme rainfall induced-landslide index[J].Engineering Geology,2016,215:36‒49.
[15]
GuzzettiF, GalliM, ReichenbachP,et al.Landslide hazard assessment in the collazzone area,Umbria,central Italy[J].Natural Hazards and Earth System Sciences,2006,6(1):115‒131.
[16]
CasciniL, BonnardC, CorominasJ,et al.Landslide hazard and risk zon&ing for urban planning and development[C]//Hungr O,Fell R,Couture R,et al.Landslide risk management,2005:209‒246.
[17]
YiJingsong, ChengYingjian, XuYi,et al.Meteorological warning threshold of regional landslide disaster based on process rainfall[J].Bulletin of Soil and Water Conservation,2024,44(2):1‒8.
GladeT, CrozierM, SmithP.Applying probability determination to refine landslide-triggering rainfall thresholds using an empirical "antecedent daily rainfall model"[J].Pure and Applied Geophysics,2000,157(6):1059‒1079.
[20]
CaiHaojie, HanHaihui, ZhangYulian,et al.Convolutional neural network landslide recognition based on terrain feature fusion[J].Journal of Earth Sciences and Environmental Studies,2022,44(3):568‒579.
HuangFaming, CaoYu, FanXuanmei,et al.Influence of different landslide boundaries and their spatial shapes on the uncertainty of landslide susceptibility prediction[J].Chinese Journal of Rock Mechanics and Engineering,2021,40(Supp2):3227‒3240.
LiuHaizhi, XuHui, BaoHongjun,et al.Optimization experiment of early identification of landslides susceptibility areas in medium and small mountainous catchment based on ensemble learning[J].Engineering Science and Technology,2022,54(6):12‒20.
HuangFaming, ZhangYinlang, GuoZizheng,et al.Influence of different classification methods on regional landslide susceptibility zoning[J/OL].China Industrial Economics,2023:1‒13.[2023‒11‒23].
MuKe, XieWanli, LiuQiqi,et al.Research on landslide susceptibility evaluation based on logistic regression and LR coupling model[J].Disaster Science,2022,37(3):212‒218.
ChenXinyu, ShiYun, WenYongxiao,et al.Landslide susceptibility evaluation based on certainty factor and support vector machines[J].Science Technology and Engineering,2023,23(2):518‒527.
LiKang, NiTongguang, XueJing,et al.Deep soft clustering:Simultaneous deep embedding and soft-partition clustering[J].Journal of Ambient Intelligence and Humanized Computing,2023,14(5):5581‒5593.
[37]
XieJ, GirshickR, FarhadiA.Unsupervised deep embedding for clustering analysis[C]//Proceedings of the International Conference on Machine Learning.PMLR,2016.
[38]
LiJunxia, QianYuhua, MaGuoshuai,et al.Depth clustering algorithm for adaptive edge samples recognition[J].Journal of Southwest University(Natural Science Edition),2023,45(3):34‒46.
VessiaG, PisanoL, VennariC,et al.Mimic expert judgement through automated procedure for selecting rainfall events responsible for shallow landslide:A statistical approach to validation[J].Computers & Geosciences,2016,86:146‒153.
[43]
QiLyu.Statistical inference studies for Gompertz distribution under complex censored schemes[D].Beijing:Beijing Jiaotong University,2020.
[44]
吕琪.复杂删失方案下Gompertz分布的统计推断研究[D].北京:北京交通大学, 2020.
[45]
LiChangjiang, MaTuhua, SunLeling,et al.A new approach to calibrate antecedent precipitation for rainfall-triggering landslide forecast[J].Journal of Mountain Science,2011,29(1):81-86.
ChinkulkijniwatA, SaleeR, HorpibulsukS,et al.Landslide rainfall threshold for landslide warning in Northern Thailand[J].Geomatics,Natural Hazards and Risk,2022,13(1):2425‒2441.
[48]
TannoR.Adaptive neural trees[C]//Proceedings of the International Conference on Machine Learning.PMLR, 2019.
[49]
ChenYu.Application of artificial neural networks in the design of biomedical materials[J].E3S Web of Conferences,2021,233:02003.
[50]
GaoBinghai, HeYi, ZhangLifeng,et al.Dynamic evaluation of landslide susceptibility by CNN considering InSAR deformation:A case study of Liujiaxia Reservoir[J].Journal of Rock Mechanics and Engineering,2023,42(2):450‒465.
HuangFaming, ZengShiyi, YaoChi,et al.Uncertainties of landslide susceptibility prediction modeling:Influence of different selection methods of "non-landslide samples"[J].Advanced Engineering Sciences,2024,56(1):169‒182.