Integrating intermediate inputs from partially classified images within a hybrid classification framework: An impervious surface estimation example

a r t i c l e i n f o With the constant proliferation of computational power, our ability to develop hybrid classifiers has improved. Hybrid classifiers integrate results from multiple algorithms and often improve classification accuracy. In this paper, a hybrid classification framework was used to evaluate two research hypotheses: i) can manipulated results from prior classifiers (" intermediate inputs " (IIs)) improve classification accuracy in subsequent classification steps. and ii) is there an optimal dataset proportion for creation and usage of intermediate inputs. These additional intermediate inputs were based on spatial and texture statistics calculated on a partially classified image. The implementation of intermediate inputs on an impervious surface classification task using a 2001 Landsat ETM+ image from central New York was demonstrated. The results suggested that there was an average accuracy improvement of 3.6% (maximum 6.6%) by using intermediate inputs. These improvements were proved statistically significant by a Z-test and tended to increase as classification difficulty increased. The experiments in this paper also showed that there was an optimal point that balanced the number of pixels and pixel classification accuracy from prior steps used to produce intermediate inputs. Additionally, some traditional problems such as separation of impervious surfaces and soil were successfully tackled through intermediate inputs. The concept of the intermediate inputs may easily apply to other sensors and/or ground features. Impervious surfaces, defined as water impenetrable surfaces such as rooftops, roads, parking lots, sidewalks, and other man-made surfaces, have become a key indicator in urban environmental studies (Arnold & Gibbons, 1996; Schueler, 1994). Accurate estimation of imperviousness is of high significance for hydrology, land use planning, resource management and ecosystem studies. Remote sensing imagery provides a cost-efficient alternative to ground-based mapping and thus has been increasingly employed for impervious surface estimation. Previous research explored classification approaches in order to estimate imperviousness based on their spectral and/or spatial characteristics. Typical classification methods include multivariate regression models, spectral mixture models, machine learning models and integration with geographical information systems. Early in 1980, Forster (1980) applied multiple regression analysis to derive linear equations relating each Landsat band with percentage of land use in the Sydney metropolitan area. Later, spectral mixture analysis was devel-In these studies, imperviousness can be estimated using a linear summation of endmembers (spectral signatures of 'pure' materials). Bauer et al. (2005) developed a second-order polynomial regression model to estimate the relationship between the imperviousness and …