Improving the support vector machine-based method to map urban land of China using DMSP/OLS and SPOT VGT data

Abstract

Extracting urban land of China timely and accurately is essential for recognizing and understanding the urban pattern and urbanization process in China. Stable nighttime light data obtained by the Defense Meteorological Satellite Program/Operational Linescan System (DMSP/OLS) provides an economical and straightforward way to map the distribution of urban land. However, all the current methods of extracting urban land from DMSP/OLS data are difficult to effectively apply in the whole of China due to their inapplicability in large area with obvious regional variation. To address this problem, we proposed a stratified support vector machine-based method (SSVM). The urban land of China in 2008 extracted from DMSP/OLS and SPOT VGT NDVI data using SSVM showed that SSVM could extract urban land more effectively than the original support vector machine-based method (OSVM) in the nation where imbalance in economic development and regional variation were extremely obvious. The correlation coefficients between statistical data and urban land derived using SSVM (R>;0.90, p<;0.0001) were almost twice as much as those from OSVM (RO.48, p<;0.05). Meanwhile, the accuracy assessment using the Landsat ETM+ data with higher resolution also showed that SSVM effectively decreased the omission error and commission error of OSVM. The overall accuracy and Kappa of SSVM achieved 0.90 and 0.69, which were 0.09 and 0.17 higher than those of OSVM, respectively.