Using polarimetric C-Band data to discriminate wetland vegetation in the Lower Paraná River floodplain: assesment of a supervised object-based Random Forests classifier

NATALIA SOLEDAD MORANDEIRA; LUIZ FELIPE DE ALMEIDA FURTADO

João Pessoa, Paraiba

Simposio; XVII Simpósio Brasileiro de Sensoriamento Remoto; 2015

Instituto Nacional de Pesquisas Espaciais

The Lower Paraná River floodplain wetlands are dominated by herbaceous communities. Dominant macrophyte species have been classified in Plant Functional Types, summarizing their main structural and functional features and their expected responses to the environment. In a previous work, a polarimetric RADARSAT-2 C-Band scene was classified with an unsupervised per-pixel approach on the coherence matrix (a progressive Wishart H/ classifier), but a relatively low global accuracy (58.2%) and Kappa index (50.4%) were obtained. In this work, we assessed a supervised object-based Random Forests classifier on the same scene. Based in previous works in other areas, we expected a higher accuracy for the Random Forests classifier than for the Wishart one. However, we obtained a even lower global accuracy (55.2%) and Kappa index (40.6%). Also, most of the areas were assigned to Plant Functional Type A (corresponding to bulrush marshes). We compared the classifiers and discuss possible reasons for the lower accuracy of the object based classifier. Our results suggest that most of the errors can be caused by the high simmilarities between the Plant Functional Type classes andbetween short grasses and Bare Soil. Other possible explanation of the low accuracy of the Random Forests classifier is that it does not follow the statistical distribution of the polarimetric data.