Urban Impervious Surface Mapping from High Resolution WorldView Imagery Based on PUL Algorithm

刘冉,李文楷,刘小平,陈逸敏,刘珍环

DOI: https://doi.org/10.3969/j.issn.1672-0504.2018.01.007

2018-01-01

Abstract:Correctly mapping the urban impervious surfaces is important in the studies of urban environment,hydrothermal cycle and the heat island effect.In this study,the positive and unlabeled learning (PUL) algorithm was investigated,which trained a classifier on positive and unlabeled data,to map the urban impervious surfaces based on the high-resolution WorldView images.Different from traditional supervised classification methods,which require both labeled positive and negative training data,the PUL algorithm requires only labeled positive and unlabeled training data.Experimental results show that the PUL algorithm outperforms the One-Class Support Vector Machine (OCSVM) and the Maximum Entropy (MAXENT) methods.For all training sample sizes,PUL consistently produces higher accuracies (i.e.overall accuracy and kappa coefficient) than the other two methods.The highest overall accuracy and kappa coefficient obtained by PUL are 91.27 ％ and 0.8255,respectively.Thus,PUL is an efficient and promising method for extracting impervious surfaces from high-resolution remote sensing images.