Evaluation of Machine Learning-Based Urban Surface Mapping Using a New Moderate-Resolution Satellite Imagery Dataset

Abstract

The machine learning algorithm support high efficiency in urban surface mapping based on the moderate resolution and multispectral satellite imagery. In this study, we evaluate eight widely used machine learning-based classification methods. For the specific modality of the moderate resolution and multispectral satellite imagery, nevertheless, even though many labeled datasets target for kinds of remote sensing applications are built recently, the labeled data for the moderate-resolution and multispectral image is still lacked currently. In this study, we carried out a moderate-resolution and multispectral images labeling work in two urban regions, and four-class urban surface and six-class urban surface in terms of the bio-physical and land use are labeled, respectively. Based on the labeled dataset, eight widely used machine learning-based supervised classifiers are selected for the methods evaluation. As the results obtained, the SVM, MLC and ANN achieved the highest accuracy of 88.2%, 85.4% and 84.1% in a simple four-class urban surface mapping experiment. With the increasing requirement of the remote sensing application, more advanced machine learning algorithms would be explored, and the evaluation in our study as well as the labeled dataset can provide a baseline for the future research.