Subpixel analysis of landsat ETM ⁺ using Self-Organizing Map (SOM) neural networks for urban land cover characterization

This paper examines the subpixel analysis of Landsat ETM ⁺ data to estimate the percent cover of impervious surface, lawn, and woody tree cover in typical urban/suburban land-scapes. By combining Self-Organizing Map (SOM), Learning Vector Quantization (LVQ), and Gaussian Mixture Model (GMM) methods, the posterior probability of the various land cover components were estimated for each pixel as a means of sub-pixel analysis. The estimation of impervious surface and the differentiation of urban vegetation - grass versus woody tree cover - are the main objectives of this paper. Overall, the output estimates compared favorably with those obtained using higher spatial resolution aerial photograph and IKONOS satellite image and traditional hard classification techniques as independent reference. The SOM-LVQ-GMM model showed a moderate degree of similarity in the estimates of impervious surface [root mean-square errors (RMSEs) of <±12% for the aerial photo reference plots and <±18% for the IKONOS classified results at a 3 × 3 pixel scale]. The vegetation components of woody tree and grass cover had an RMSE of <±10% and 11% for the aerial photo reference plots and <±15% and 22% for the IKONOS classified results at a 3 × 3 pixel scale, respectively. While further work is needed to improve the partition of grass versus woody tree cover estimation, the SOM-LVQ-GMM approach was sufficient to effectively discriminate between those areas dominated by grass cover versus those with significant woody tree cover. The ability to separate the grass versus woody tree components in urban vegetation analysis provides a more nuanced view of urban ecosystems.