An Automatic Land-cover Mapping Method Over Urban Areas Using Spectral Indices Based On Landsat Imagery

Land-cover mapping over urban areas using Landsat imagery has attracted considerable attention in recent years as it can promptly and accurately reflect the biophysical composition status of the urban landscape and allow further applications in fields such as urban planning and risk management.However,for global-scale urban area mapping,adequate training sample collection is still an unignorable hindering factor.In this paper,we present a novel Landsat image interpretation framework for automatically mapping urban areas into four categories(human settlement,vegetation,water,and bare soil).The main research work of this thesis is as follows:(1)A forward growth and backward elimination(FB)sample selection method,based on the counterpart land-cover spectral indices,is designed for the automatic collection of samples,aiming to meet the requirements for sample selection,including intra-class diversity and purity,inter-class separability,and reduced redundancy.(2)An automatic integrated classification scheme is then proposed by importing the automatically selected samples into a collaborative representation based classifier(CRC).To validate the effectiveness and robustness of the proposed scheme,a series of Landsat images over 39 representative cities from different biomes across the world were chosen.The automatic mapping results show a promising accuracy,with the overall accuracy(OA)ranging from 84.2%to 99.0%and kappa coefficients from 0.77 to 0.99.Moreover,the quality of the automatically selected samples is convincing,as the average correctness is 94.0%by a visual comparison,and the desirable inter-class separability of the automatically selected samples is proved with J-M distances ranging from 1.71 to 2.00.Compared to the results acquired by manually selected training samples,the proposed automatic approach can achieve a comparable mapping accuracy.Additionally,compared with the spectral indices based land-cover extraction approaches,the superiority of the proposed method is clearly demonstrated,particularly in discriminating between human settlement and bare soil.Due to the free access to Landsat data,the acceptable mapping accuracy,and the automation,the results of our study hint at a promising potential for large-scale mapping of urban areas in a cost-effective and automatic manner.