RS-GIS Based Development Characteristics And Susceptibility Assessment Of Landslides In The Anning River Active Fault Zone

The Anning River active fault zone is located in the alpine canyon area.Due to the strong neotectonic movement,complex geological environment,broken rock mass,strong weathering,and crisscross water system,landslide disasters occur frequently in the area.Based on the RS and GIS technology,this paper uses multi-source and multi-temporal satellite images such as Gaofen-1 and World View-2,combined with remote sensing interpretation and field investigation,to analyze the development characteristics of landslides,the influence of faults on landslides,and the relationship between landslide frequency and scale.Then,focusing on the landslide susceptibility mapping,the influence of data sampling,mapping unit,resolution,and integrated model on landslide susceptibility zoning is deeply discussed.The main contents and results of this paper are as follows:(1)The influence of the active fault zone on the landslide includes two aspects:(1)the steeply dipping fault fracture zone constitutes the landslide boundary;(2)the landslide is easy to form in the wide fracture zone.Combined with numerical simulation,it is found that the anti dip inner slope fault does not control the stability of the slope,while the anti dip outer slope fault has an important influence on the stability of the slope regardless of whether the fault dip angle is greater or less than the slope gradient.(2)The landslides are mainly concentrated in the ranges of elevation<2407 m,slope 15～35°,sunny slope,relief amplitude 103～268 m,cutting depth 50～142 m,drainage density 0.47～1.50 km/km~2,SPI<12.The distribution of landslides has a good correspondence with lithology,rainfall,fault zone,and other factors.Among them,Jurassic-Triassic and Sinian strata,24h rainfall of 80 mm,PGA of 0.3 g,fault zone distance of<1.2 km,and road density<0.65 km/km~2 are prone to landslides.(3)There is a power-law relationship between landslide area and frequency in the study area,and there is also a significant power-law dependence between landslide volume and area.The landslide area and probability density satisfy the three-parameter inverse gamma function.The landslide inventory in the area is incomplete.(4)The data sampling is optimized by combining RF,ANFIS,and CART models.The results show that among the three models,the prediction accuracy of the landslide scarp polygon is the highest,followed by the main scarp upper edge sampling and single-pixel sampling.Due to over-reliance on the centroid of the landslide surface,few sampling points,and a large degree of generalization factor layer,single-pixel sampling misestimates or omits important feature information,which eventually misleads the evaluation results and reduces the mapping accuracy.(5)The grid unit and the two kinds of slope units have obvious differences in the mapping details.The former is prone to fragmentation,the regional coherence is poor,and the warning object is not clear;the latter has good overall continuity,which can distinguish the stability of different slopes,clear warning objects,and have more practical significance.Combined with the classification results,the comprehensive prediction accuracy of the slope unit is better than the grid unit,and the prediction accuracy of the curvature watershed method is the highest among the two kinds of slope units.From the ROC curve,after selecting the landslide scarp polygon as the optimal sampling method,the AUC values of RF,ANFIS,and CART models are increased by 0.0221,0.0246,0.0238 and 0.0295,0.02 and 0.0199 respectively in the training and testing stages by combining with the optimized mapping units.(6)In general,among the three models of RF,ANFIS,and CART,the prediction accuracy of the model does not simply improve with the increase of resolution,and high-resolution data may not necessarily bring the best mapping effect.The optimum resolution is related to the overall area distribution of landslides in the study area,and it can better reflect the overall properties of landslides in the area.From the ROC curve,based on the most suitable mapping unit,the AUC values of RF,ANFIS,and CART models are increased by 0.0222,0.0259,0.0252,and 0.0223,0.0291,and0.0276 respectively in the training and testing stages by combining with the optimum resolution.(7)Given the weak generalization ability and strong variability of the single model,six integrated models are proposed in this paper.From the model classification results,the six new integrated models can objectively and accurately evaluate the landslide susceptibility in the study area,and the model with the highest classification accuracy is WPT-ACO-PSO-SVM,followed by ADTree-Ada Boost,WCLPI-FR-RF,ADTree-Bagging,WCLPI-EBF-RF,and WCLPI-IOE-RF.Similarly,based on the optimal data sampling,the most suitable mapping unit and the optimum resolution,compared with the aforementioned RF model with the best performance,the AUC values of WPT-ACO-PSO-SVM,ADTree-Ada Boost,WCLPI-FR-RF,ADTree-Bagging,WCLPI-EBF-RF,and WCLPI-IOE-RF models are increased by 0.0804,0.0623,0.049,0.0398,0.0333,0.022 and 0.0753,0.0575,0.0451,0.0368,0.0278,0.0157 respectively.Based on the above results,it is recommended to use the landslide susceptibility zoning divided by the WPT-ACO-PSO-SVM model as the resulting map in the study area.