Comparison of segment and pixel-based non-parametric land cover classification in the Brazilian Amazon using multitemporal landsat TM/ETM+ imagery

This study evaluated segment-based classification paired with non-parametric methods (CART (R) and kNN) and inter-annual, multi-temporal data in the classification of an 11-year chronosequence of Landsat TM/ETM+ imagery in the Brazilian Amazon. The kNN and CART (R) classification methods, with the integration of multi-temporal data, performed equally well in the separation of cleared, re-vegetated, and primary forest classes with overall accuracies ranging from 77 percent to 91 percent, with pixel-based CART (R) classifications resulting in significantly lower variance than all other methods (3.2 percent versus an average of 13.2 percent). Segmentation did not improve classification success over pixel-based methods with the used datasets. Through appropriate band selection methods, multi-temporal bands were chosen in 38 of 44 total classifications, strongly suggesting the utility of inter-annual, multi-temporal data for the given classes and region. The land-cover maps from this study allow for an accurate annualized analysis of land-cover and landscape change in the region.