Soft urban water cover extraction using mixed training samples and Support Vector Machines

Abstract

In this article, a new strategy is proposed for soft urban land-cover extraction based on mixed training data and Support Vector Machines (SVMs). The strategy is applied to soft classification of urban water cover using Landsat 8 multispectral data in Beijing. The results are validated with extensive manual mapping data and are compared with the results derived from a well-trained SVM and the linear spectral unmixing method using only pure samples. Our experimental results indicate that the proposed strategy works effectively in extracting the water cover in the urban image. The results are better than those obtained by SVM and linear spectral unmixing with only pure samples. Our findings demonstrate that the combination of image-based mixed training data and SVM enhances the strengths of kernel-based approaches for soft mapping urban land cover from imaging data. Therefore, the proposed workflow constitutes a new flexible and extendable approach to soft mapping urban land cover. Future work includes the construction of a comprehensive image-based spectral library for urban areas, and the testing of the proposed strategy on additional land-cover types in urban areas such as grass, tree, and bare soil.