Predictive variables and sampling strategies for automatic classification of landcover at southwest subregion of the metropolitan region of São Paulo

Abstract

Alternatively of pixel-based, object-based classification have reached satisfactory results when applied to high resolution images, for landcover mapping. However, at semi-detailed level (1:50,000-scale) the best way to go remains unknown. Few studies compared those approaches using medium resolution images (~ 10 m). The effect of quantity and frequency distribution of training observations at each landcover classes are poorly investigated, as well as the efficiency of different kind of predictive variables. The objective of this study was to evaluate different sampling strategies for model training and the efficiency of predictive variables, comparing the results at pixel-based and object-based approaches. The mapping area was the southwest of metropolitan region of São Paulo, and the supervised classification was performed by Random Forest. Six databases were tested by changing quantity and frequency distribution of training observations at each landcover class. We used spectral information from Sentinel-2 mission, topographic information from SRTM, vegetation and water indexes, beyond of information about spatial-temporal variation of indexes. Maps performance was better when vegetation and water indexes, spatial variation and topographic information were added to training set data in pixel-based approach, reaching 82.44 % of accuracy and Kappa index of 0.80. At object-based approach, adding temporal variation of indexes and information about segments geometries were important for map improvement, reaching accuracy and Kappa index of 78.15 % e 0.75, respectively. The results of the maps depend of amount and frequency distribution of observation by classes used for training dataset. The both of approaches tested show similar quality. Toward to choose the better fit among classification results it is main to know the study region characteristics.