Research On Building Extraction Method Based On Traditional Machine Learning And Full Convolutional Neural Network

The resolution of remote sensing image is constantly improving in the wake of remote sensing technology’s rapid development,at the same time the research of high resolution remote sensing image’s surface object classification is becoming increasingly significant.In remote sensing images,buildings are significant characteristics.Precise buildings’ collecting is very important for map updating,land use investigation,urban planning and construction,traffic management,change monitoring and so on.In the wake of remote sensing technology’s rapid development,image interpretation also faces severe challenges.In recent years,high-resolution images have become one of the geospatial information’s main origins.It has the advantages of more and more information,sufficient expression of element details,rich surface texture features,clear geometric structure,and clear expression of small surface features compared with medium and low spatial resolution images.However,the improvement of image spatial resolution cannot promote the accuracy of image interpretation and analysis.On the contrary,these characteristics of high-resolution images lead to the more prominent appearance of " different spectrum in the same thing" and " the same spectrum in the different thing",which makes it more difficult to extract buildings.At present,visual interpretation is a highly accurate method,but it takes a lot of time and effort to extract,at the same time it is unsuited for big image sets.Random forest,LS-SVM and rest machine learning ways have improved the extraction speed,but they can not effectively extract deep-level features of high-resolution images,and the classification effect is not well.Although the full convolutional neural network can deep level characteristics,there are phenomena of context information loss and pixel loss,and the integrity of the extracted buildings is poor.Aiming at the deficiencies in the above research,the article uses the Massachusetts building data set and shape,texture,and spectral characteristics of images to extract the building on the basis of machine learning and full convolutional neural network,and proposes an effective building recognition and extraction method.The main research contents of this paper are as follows:(1)In order to evaluate the effect of image segmentation more accurately,a variety of quantitative evaluation indexes were introduced to select the optimal parameters of multi-scale segmentation,first,the maximum area method was used to evaluate the optimal shape factor and compactness factor.The second method was to evaluate the optimal scale factor by combining the local variance analysis method with the neighborhood absolute mean difference standard ratio method(RMAS).A variety of quantitative evaluation indexes were used to improve the segmentation accuracy and lay the foundation for subsequent precise building extraction.(2)To solve the problem that image has many features and it is difficult to filter using machine learning,feature weight algorithm and fast correlation filtering algorithm were combined to effectively remove redundant features and weak correlation features in the process of building extraction,and the optimal feature subset was used to extract buildings,improving the accuracy and efficiency of building extraction.(3)In order to solve the problems of missing context information and pixel loss in extracting buildings by full convolutional neural networks,a Multiscale Atrous Convolution FCN8s(MAC-FCN8s)deep learning network was proposed by introducing multi-scale cavernous convolution based on full convolutional neural networks.The ordinary convolution was replaced by atrous convolution with different scales to enlarge the receptive field,reduced parameters quantity,obtained multi-scale feature information and context information of the image,and at the same time ensured the resolution and reduced the pixel loss.(4)The full convolutional neural network extraction method of pixel units still has the phenomenon of "pepper and salt",and the extracted boundary is fuzzy.Therefore,combining the object with the FCN network,inputting the multi-scale segmentation results into the FCN network,besides extracting the building at the object level can better eliminate the phenomenon of "pepper and salt",make the boundary clearer and improve the extraction accuracy.(5)Several experiments of building extraction were designed for comparison.First,the building was extracted by using the machine learning method of image spectral features in pixel unit.The second is to extract buildings by using machine learning method of multi-scale segmentation and feature optimization in object unit.Third,in pixel unit,the full convolutional neural network FCN8 s,FCN16s,FCN32 s and the full convolutional neural network integrating multi-scale cavity convolution were used to extract buildings.The fourth is to extract buildings in object unit by using full convolutional neural network.The optimal method for building extraction is the object-oriented full convolutional neural network method.