The Remote Sensing Analysis Of Surface Tectonic Deformation Of Fault-propagation Folds Based On Convolutional Neural Network

Different tectonic deformation areas usually experience different tectonic deformation processes and different tectonic deformation mechanisms,resulting in different tectonic traces.The different intersections between these tectonic traces and the ground form a variety of surface deformation textures.Some of these textures are tectonic styles that we know,but most are not.However,these currently unknown textures also record the relevant information of the tectonic deformation process and tectonic deformation mechanism.Thus,it is a new and challenging work to extract and utilize these currently unknown surface deformation textures to reveal tectonic deformation features and deformation mechanisms.In order to break through the inherent mode of traditional tectonic research and the limitation of small observation sample set,we propose a remote sensing analysis method for surface tectonic deformation of fault-propagation folds based on convolutional neural network.In this method,the satellite images truly reflect the surface deformation texture,which can be used to make a big sample set.As a big datadriven method,the convolutional neural network is applied to the remote sensing extraction of the surface deformation texture.The principal feature maps(PFM)and the class activation maps(CAM)are used to open the ”black box” of the convolutional neural network,which can intuitively display the surface deformation texture of different tectonic deformation areas,and further reveal the tectonic deformation mechanism of different tectonic deformation areas.The main results and understandings obtained in this paper are as follows:1.The research scale of geological applications is large and the terrain of the study area is complex.Due to the constraints of objective factors such as poor image quality and small number of images in the area,it is difficult to obtain complete remote sensing images of the study area.In order to overcome this problem,we propose a nonlinear radiometric normalization model(named NMAG)for SITS based on Artificial Neural Networks(ANN)and Greedy Algorithm(GA),and the SITS corrected by NMAG can be used to obtain a complete remote sensing image of the study area.This method suppresses the continuously changing noise in SITS,reduces the difference of radiation background values between images,makes the DN values of remote sensing images acquired at different times more comparable,and obtains more accurate SITS.2.The extraction of features in existing tectonic deformation analysis methods based on satellite images are supervised,which are highly subjective and rely more on expert experience.To solve this problem,we propose a more objective remote sensing extraction model of surface deformation texture based on convolutional neural network(named DFRs M).In this model,we use GF-2 satellite images to obtain a big sample set of the front limb and the back limb of faultpropagation fold.And use the convolutional neural network to extract the surface deformation textures of different tectonic deformation areas by training the big sample set obtained above.In the construction of DFRs M,in order to explore the influence of the network structure and the research scale of the sample set on the accuracy of model,we conducted comparative experiments of different CNNs(Le Net,Goog Le Net,Res Net)and sample sets with different scale(64*64,96*96,144*144)respectively.The accuracy comparison results shows that Res Net is the optimal convolutional neural network,144*144 is the best research scale of the sample set,and the best test accuracy of the model is 96.9%.This result indicates that DFRs M can accurately extract the surface deformation texture of the front limb and the back limb of faultpropagation fold.The prediction results of DFRs M show that the model can objectively obtain the spatial distribution of the front limb and the back limb of fault-propagation fold.And the abnormal areas in the prediction map can reflect the differences in local tectonic causes,thereby revealing local tectonic activities in the region.3.Convolutional Neural Network is a ”black box” model that is trained,tested,and predicted end-to-end.Insufficient understanding of the intermediate processes of convolutional neural networks will limit our understanding of the surface deformation texture and tectonic deformation mechanism of different tectonic deformation areas.In order to solve this problem,we propose a surface tectonic deformation analysis method based on the principal feature maps and class activation maps.In this method,the principal feature maps are used to visually display the surface deformation textures of different tectonic deformation areas,and further objectively reveal the tectonic deformation mechanisms corresponding to these surface deformation textures.In addition,we use the density of key pixels in the class activation maps to characterize the intensity of surface tectonic deformation.Therefore,the density distribution map of key pixels can objectively reflects the distribution pattern of the intensity of surface tectonic deformation in this area,and provides a strong planar data support for the deformation analysis of different tectonic deformation areas.