Research Of Building Extraction Technology For High-Resolution Remote Sensing Images

Automatic building extraction technology for high-resolution remote sensing im-ages has been studied for decades.These techniques play an important role in remote sensing image analysis,and they are widely used in a diverse range of applications,such as digital city modelling,military reconnaissance,disaster evaluation and so on.However,it is significantly challenging to extract arbitrary-size buildings with large-ly variant appearances or serious occlusions.In this thesis,we propose an integrated system of building rooftop recognition and contour approximation.Comparing with previous related methods,our system not only improves the building extraction accura-cy,but also reduces approximation error of building contours.For very high-resolution aerial images,the total process of building detection and contour extraction only takes a few seconds.To the best of our acknowledgement,our system is the fastest system for building extraction.More specifically,the contributions of this thesis are two-folds as following:(1)A novel end-to-end hierarchically fused fully convolutional net-work,namely,HF-FCN,is proposed to effectively integrate the information generated from a group of neurons with multi-scale receptive fields.Our architecture takes a super-resolution aeri-al image as the input without warping or cropping and directly generates the building map.It is not only convenient to use,but also significantly reduces the algorithm com-plexity.Through the improvement of our network,its parameters are compressed to 50 MB.Meanwhile,this thesis also introduce three variants of the HF-FCN.By comparing the performances of our method with several state-of-the-art algorithms on semantic segmentation,building detection algorithms based on deep learning,the experimen-t results demonstrate that our approach possesses obvious advantages in terms of the algorithm complexity and recognition accuracy.(2)This thesis designs an efficient and robust contour fitting scheme,which con-sists of different post-processing and contour estimation methods for buildings in dif-ference scales.Especially for large-size building regions,we propose a novel iterative detection algorithm for structure points.It well preserves the structure information and ensures high compression rate simultaneously.We conduct a group of experiments to analyse the time cost and accuracy for each algorithm step.In order to certify its char-acteristic of structure-preserving,we make detailed comparisons with the polygonal simplification algorithm using a set of complex samples.Experimental results demon-strate that our method surpasses the polygonal simplification method on the building contour approximation task.