Comparison of (Non)Linear Land-Cover Classification Methods in the Brazilian Amazon
Liesenberg, Veraldo¹; Souza Filho, Carlos Roberto^2
1TU Bergakademie Freiberg, GERMANY; ²Universidade Estadual de Campinas (Unicamp), BRAZIL

Forest inventory information has been a very important topic under the recent United Nations (UN) endorsed "Reducing Emissions from Deforestation and Degradation" (REDD+) protocols. In other words, frequently forest monitoring is needed, not only for planning the future forest management practices but also to monitor forested areas under the recent REDD+ policy. We evaluated the potential of airborne synthetic aperture radar (SAR) at P- and X-bands for the characterization of selected land use classes. Different study areas were selected in the Amazon biome. Multitemporal Landsat data were used for validation purposes. Intensity backscattering, polarimetric features and texture were used individually and combined in different scenarios. Nonlinear classification methods such as support vector machines (SVM), random forest (RF) and neural networks (NN) were tested. The selected nonlinear classification methods resulted in fine and accurate vegetation mapping. Classification accuracy varies according to the data input selection. The capabilities and limitations of exploring SAR single and dual frequency will be presented. The results may help to encourage the further development of joint techniques under the recent "Reducing Emissions from Deforestation and Degradation" (REDD+) protocols.