Research On The Building Extraction Technologies From High Resolution Remote Sensing Image Based On Deep Learning

In recent years,with the rapid development of sensor technology,the spatial resolution of available remote sensing images has been continuously improved,thus leads to the formulation and evolution of the High-Resolution Earth Observation System.At the same time,accurate and efficient remote sensing information extraction methods have become a key step in many remote sensing related applications.On the other hand,artificial intelligence technologies,especially deep learning algorithms,are leading the wave of intelligent revolution and has gradually become a key technology for national strategic competition and industry transformation breakthrough.At present,deep learning algorithms have achieved unprecedented results in the fields of image classification,natural language processing,dimensionality reduction,target detection,motion modeling,automatic driving,and robotics.Under this background,this paper adopts the frontier algorithm of deep learning field to carry out remote sensing information extraction research and exploits the effectiveness of deep learning technology in remote sensing field,with an intention to improve the accuracy and efficiency of remote sensing information extraction.With an aim to improve the intelligent information extraction performance of remote sensing images,the cutting-edge deep learning algorithms in the field of computer vision are used to study the building extraction technology on high resolution remote sensing images.This paper first introduces the basic principles of deep learning and convolutional neural networks(CNN),including the basic principles of deep learning and CNN,the characteristics and components of CNN,neural network optimization algorithms,and the basic principles of using transfer learning in CNN networks.For the over-fitting problems that may occur in network training,we introduce several commonly used methods for preventing over-fitting.Second,to address the challenges and issues in the current high-resolution remote sensing image classification methods,we designed and improved the existing deep learning models,and proposed two building extraction models based on edge-weighted CNN network and adversarial network.Specifically,in view of the problem of insufficient feature extraction ability of traditional CNN-based building methods,we designed a CNN model with more than one hundred convolutional layers for highly abstract feature extraction.Meanwhile,in order to maintain the high-frequency context features in this very deep network and ensure the predicted building boundaries can be aligned with the real ones,we designed an edge-weighted loss function to force our segmentation network to pay more attention to the boundary pixels during the training process.Third,traditional semantic segmentation models mostly use pixel-wise classification which neglects the high-order spatial correlation between pixels.To address this issue,we developed a novel deep adversarial network that jointly trains a deep convolutional neural network(generator)and an adversarial discriminator network for the robust segmentation of building rooftops in remote sensing images.The generator and discriminator compete with each other in an adversarial learning process until the equivalence point is reached to produce the optimal segmentation map of building objects.A soft weight coefficient is adopted to balance the operation of the pixel-wise classification and high-order structural feature learning.Meanwhile,in order to avoid the model collapse problem of traditional generative adversarial network,this paper adopted a relatively stable adversarial learning strategy to optimize the model.Finally,we verified the effectiveness of the above-mentioned building extraction models in complex urban scenarios.Results show that pixels in the predicted segmentation maps are broken and the building blocks are unclear.To address this problem,we proposed a new building extraction framework from the perspective of instance segmentation,which can achieve building instance extraction rather than pixel-wise segmentation.Our building instance extraction model starts from Mask R-CNN,which is one of the most popular instance segmentation algorithm in the field of computer vision,followed by polygon simplification and regularization.To sum up,the contributions of this paper mainly include:(1)In this paper,we newly designed a novel edge weighted fully convolutional Dense Net(EW-Dense Net)model for robust building footprint extraction on remote sensing images.Specifically,EW-Dense Net firstly leverages the start-of-the-art deep Dense Net architecture to learn highly abstract feature representations.Then,the proposed model gains advantages on computational efficiency and model compactness by considering the use of the bottleneck layers and compression strategy.More importantly,a novel edge weighted cost function is proposed to regularize the optimal semantic segmentation to be aligned well with the boundaries of objects.Experimental results on public remote sensing imagery data demonstrated that the superior performance of EW-Dense Net over other state-of-the-art techniques in building footprints extraction.(2)In this paper,we developed a novel deep adversarial network,that jointly trains a deep convolutional neural network(generator)and an adversarial discriminator network for the robust segmentation of building rooftops in remote sensing images.More specifically,the generator produces a pixel-wise image classification map using a fully convolutional Dense Net model,whereas the discriminator,tends to enforce forms of high-order structural features learned from ground truth label map.The generator and discriminator compete with each other in an adversarial learning process until the equivalence point is reached to produce the optimal segmentation map of building objects.Meanwhile,a soft weight coefficient is adopted to balance the operation of the pixel-wise classification and high-order structural feature learning.To avoid the optimization problem of the traditional adversarial network,a novel stable adversarial network training strategy is used to train the model.Experimental results show that our model can successfully detect and rectify spatial inconsistency on aerial images while archiving superior performance compared to other state-of-the-art building extraction methods.(3)This paper proposed a building extraction framework based on the instance segmentation algorithm.In the complex urban scene,the building maps extracted by the traditional segmentation algorithms have a large-scale fragmentation phenomenon,and it is difficult to obtain the building instances by using the post-processing algorithm.In this paper,our extraction method is based on instance segmentation and polygon regularization.More specifically,our building extraction framework first generates the building polygons from the Mask R-CNN instance segmentation model,followed by polygon simplification and regularization.