Extraction Of Residential Areas Based On Time-series Spectral Feature From GF-4 Satellite Images

With the promotion and implementation of One Belt and One Road Initiative and New Pattern Urbanization Strategy,China's development of the spatial pattern will change dramatically over the next few decades.The economy and society have changed a lot since reform and opening-up,and people's life also reached to a higher level,in the meanwhile,the residential areas also have expanded faster and faster.Recognizing and extracting residential areas quickly and accurately plays an important role in promoting national strategic decision,achieving the digital city as well as assisting in urban planning.As a significant part of implementing high resolution earth observation system in China,GF-4 satellite can recognize the change of land surface timely and effectively.It also supports the applications of natural disasters—such as earthquake,flood,drought and typhoon,research on climate change and environmental survey of forestry and water resources.This thesis takes advantage of high time-series spectral feature of GF-4 satellite,extracts and analyzes the differences in spectral feature between residential areas and other land cover types.It also recognizes and extracts the residential areas with time-series spectral feature.The thesis mainly launches from the following parts:First of all,the background and importance of the research as well as the research status and progress of remote sensing extraction and residential areas extraction techniques were introduced,the content and technical route of this research were drawn up.Then the GF-4 satellite images and the way of images' preprocessing were introduced.Afterwards,a method of recognizing and extracting residential areas,which is in terms of images' spectral feature by the way of decision trees,was put forward followed by analyzing spectral feature of residential areas,vegetation and waterbody.After that,an extraction method of residential areas based on time-series spectral feature in the way of decision trees can be obtained and analyzed.On this basis,a method using Fully Convolutional Network(FCN)based on time-series spectral feature was come up with to extract residential areas information.Finally,the results were compared and analyzed,in the meanwhile,some conclusions were summarized.Major conclusions from the research can be summarized as:(1)The solar altitude affects not only spectral feature,even influences spectrum change rate and the rate's change.In the meanwhile,the spectral feature of residential areas will have different amplitude changes as the solar altitude changes.(2)The method of residential areas extraction with time-series spectral feature will get a better result than that only with spectral feature when using decision trees.The precision increased from 89.85% to 93.38%?(3)The method using Fully Convolutional Network(FCN)will improve the precision when extracting residential areas based on time-series spectral feature.The precision increased from 93.38% to 95.15%?Besides,innovations of this research can be concluded as:(1)The relationship between time-series spectral feature and solar altitude is brought in the model of decision trees,which improves the precision compared with the method just using spectral feature.(2)The method using Fully Convolutional Network(FCN)based on time-series spectral feature will get a higher precision than that using decision trees.Researches on the extraction for residential areas based on time-series spectral feature of GF-4 satellite images not only provides dynamic and updated data to disaster prevention and reduction,urbanization,urban refined management and land and resources administration,but also provides support technology and a strong demonstration effect to the application of domestic high-resolution satellite remote sensing data products in China.