Sub-pixel classification

Sub-pixel approaches aim to address the mixed pixel problem that arises when coarse spatial resolution images are used for mapping small agricultural fields in heterogeneous landscapes. They do so by determining the fractional proportion of the different land cover types in a pixel on the basis of an appropriate training data set. Spectral mixture analysis (SMA) is a popular sub-pixel classification method. In SMA, fractional proportions are determined by comparing the spectral signatures of land cover types in a pixel with a set of “pure” reference spectra, also known as endmembers (Elmore et al., 2000; Lillesand et al., 2004). Variants of SMA have been introduced by different researchers (Elmore et al., 2000; Lobell and Asner, 2004; Ozdogan, 2010) (Zurita-Milla et al., 2011; Amorós-López et al., 2013)

Approaches to sub-pixel classification based on regression have become popular in recent years. They estimate fractional land covers based on multi-resolution data and regression trees (RTs) (DeFries et al., 1997; Gessner et al., 2013; Hansen and DeFries, 2004; Tottrup et al., 2007). The classification results of a high spatial resolution image (e.g. Quickbird, ASTER, Landsat, etc.) are combined with a coarse spatial, but temporally high resolution image (e.g. MODIS, MERIS, AVHRR) in a regression tree analysis to determine contribution proportions of land covers in each pixel of the coarse spatial resolution image. Some studies have pointed out the advantages of regression-based approaches over other methods, such as linear mixture modelling (DeFries et al. 1997, Smith et al. 2003).