Coastline Extraction And Object Classification From Remote Sensing Images Based On Deep Learning

Automatic boundary extraction and object classification from remote sensing images are widely used in urban planning,environmental protection,map mapping,resource management,military reconnaissance,land use and so on.The remote sensing images classification has always been an important topic of remote sensing research.Boundary extraction of remote sensing images plays an important role in the information retrieval and classification of remote sensing images study.Boundary extraction in images is a fundamental step for the subsequent feature representation and scene classification.The coastline is a special boundary in remote sensing images.To this extend,the traditional coastline extraction methods are based on hand-crafted features and regulation-based segmentation algorithms,which cannot be directly used.In particular,with the large diversity,increasing scale,and high resolution,the content of remote sensing images becomes more complicated.The hand-crafted features therefore cannot satisfy the corresponding task of boundary extraction and classification.Therefore,it is of great significance and value to exploit new methods of boundary extraction and classification for remote sensing images,especially those based on deep learning.In this thesis,in regard to the problems in boundary extraction and object classification of remote sensing images,we mainly contribute on the following three aspects:(1).Firstly,a superpixel based coastline extraction algorithm is proposed to extract coastline from SAR and satellite images,respectively.This method treats the super pixel block as the unit to extract features.This paper use grayscale mean and different color space histogram fusion as features of SAR and satellite images,respectively.The coastline is extracted by an adaptive threshold,by which the false alarm of the boundary extraction is filtered out by a morphological method.The efficiency of the coastline extraction can also be improved.Due to the existence of speckle noise in SAR images,a speckle noise removal algorithm is proposed,which is able to effectively enhance the robustness against noise.(2).Secondly,on the basic of coastline extraction,we further target at the problem that the traditional feature descriptor is difficult to describe the complex scene of remote sensing images.To this end,a multi-level neural network is introduced.This network can describe the features from different scales of remote sensing images.Therefore,it can contribute to extracting the edge information of complex scenes in remote sensing images.In this thesis,the accuracy of the boundary detection of objects at different scales can be improved by combining the features of different hierarchical networks and the results of boundary extraction in different layers.This multi-level neural network can detect boundary simultaneously in different sizes of objects in the remote sensing images.Then,the trained multi-level network is applied to the low-resolution remote sensing images to extract coastlines.(3).Finally,due to large data scale and complex scene of remote sensing images,it is difficult to distinguish different objects,and meanwhile the hand-craft features of the objects is not robust enough.In this paper,the fully convolution neural network(FCN)is used to classify the remote sensing images.However,the classification results of FCN are still not satisfied.In this paper,we further propose to optimize the classification results,which takes the boundary detection results as the constraint condition not only improves the classification accuracy,but also keeps the object boundary information better.