| The mountainous area of Southwest China is a high incidence area of landslides and other geological disasters.The local government has conducted several rounds of geological disaster investigation,but according to statistics,more than 70% of the landslide disasters causing serious losses are still not within the scope of potential landslides that have been investigated.One of the reasons is that the area has changeable terrain with high vegetation coverage,and the sliding source area is located at a high position with strong concealment.The traditional manual identification method is difficult to effectively identify the potential landslide.In recent years,the ‘three-step investigation’ system of the ‘Sapce-Air-Ground’ integration has realized the systematization and scientization of early identification and early warning of landslide hazards.As the basis of the ‘three-step investigation’ system,the purpose of ‘general investigation’ is to to conduct a comprehensive investigation of the target area and identify and locate potential landslides to the greatest extent.As the main method in the‘general investigation’ stage,optical remote sensing technology realizes the early identification of landslide through boundary features and texture information.However,due to the influence of image quality,vegetation coverage and unclear boundary of landslide itself,some potential landslides can not be effectively identified.As a technology that can continuously and widely observe surface deformation,InSAR can verify and supplement the potential landslides through the anomalous deformation area.However,there are various causes of surface deformation,which need to be further verified by subsequent "detailed investigation" stage and "verification" stage.Aiming at the problems and deficiencies in the identification and prediction of anomalous deformation area based on InSAR,this paper selects Maoxian County as the research object,combined with the InSAR deformation fusion,deep learning and grey prediction methods,and focuses on the key technologies in improving the deformation accuracy of InSAR monitoring and realizing the automatic identification and prediction of anomalous deformation areas.The main contents,conclusions and innovations are summarized as follows:(1)A fusion method of InSAR surface deformation results through time registration and space registration is proposed.In view of the deficiency that the direction of the traditional InSAR surface deformation calculation result is the line of sight direction of the satellite,which is not the real surface deformation.The time seriess urface deformation of ascending orbit and descending orbit are obtained by SBAS-InSAR technology,and the ascending orbit is selected as the main orbit and the descending orbit is selected as is the auxiliary orbit.Through comparative experiments,Akima spline interpolation method is selected as the registration method on the time scale,and geographic weighted regression method is selected as the registration method on the spatial scale.The deformation sequence of the auxiliary orbit is fused into the main orbit,and the fused deformation is transformed into slope deformation through spatial geometric relationship,which is more in line with the recognition requirements of anomalous deformation areas(2)An object detection model InSARNet for automatic recognition of InSAR anomalous deformation area is proposed.Aiming at the deficiency of long timeconsuming,low efficiency and inconsistent standards of traditional manual interpretation methods for the recognition of large-area InSAR deformation results.Based on deep learning,a two-stage object detection model InSARNet is constructed to realize the automatic and rapid extraction of InSAR anomalous deformation area.InSARNet takes the two-stage network Mask RCNN as the basic framework,and introduces a more efficient operator,Involution operator,and a deformable convolution pooling module with better effect on small-scale target recognition.In the training and testing on the data set of anomalous deformation area,InSARNet has achieved good recognition effect,and the overall accuracy is more than 90%.In parallel comparison with other object detection models,InSARNet achieves a good balance in operation speed,model complexity and model accuracy.At the same time,in order to meet the identification needs of potential landslides,the buffer zone is made centered on the rivers and roads in the study area,and the anomalous deformation area belonged to the buffer zones are regarded as possible potential landslides,which still need to be further determined through subsequent "detailed investigation" stage and "verification" stage.(3)The landslide displacement prediction method based on Grey Markov Verhulst model is improved.In view of the deficiency that the traditional grey prediction model only predicts through the displacement trend,the prediction curve will deviate greatly when facing the change of precipitation and temperature.Therefore,in this study,the displacement curve of potential landslide is decomposed into the trend term displacement component affected by its own gravity and the periodic termd isplacement component affected by external factors.The trend term displacement component is fitted by Grey Markov Verhulst dynamic model.For periodic term displacement component,the grey Markov Verhulst dynamic model is improved,and the factors of precipitation and temperature are added to the relative error fitting process of the model,The accuracy of deformation prediction is effectively improved.After accumulating the prediction results,through comparative analysis,the average relative error of the fitting results of the improved model to the original sequence is 2.53%,which is 2.36% higher than that before the improvement,and the average relative error of the prediction results of the improved model to the unknown sequence is 4.58%,which is 3.85% higher than that before the improvement.The results show that the accuracy of the improved prediction model is significantly improved and can effectively predict the anomalous deformation area. |
