Potential of hyperspectral metrics and classifiers for mapping Brazilian savannas in the rainy and dry seasons

摘要: Land cover mapping of savannas in Brazil, a world's hotspot biodiversity, is still challenging due to the tree gradient and spectral similarity between some vegetation physiognomies. Here, we evaluated potential four classifiers (Decision Tree (DT), Random Forest (RF), Spectral Angle Mapper (SAM) Support Vector Machine (SVM)) for discriminating eight savanna physiognomies rainy dry seasons Brasilia National Park (BNP). Five sets Hyperion/Earth Observing One (EO-1) metrics (reflectance, first-order derivative reflectance; narrow-band indices (VIs); absorption band parameters; combination these attributes) were tested as input data each classifier. Before classification, Correlation-based Feature Selection (CFS) algorithm was applied reduce dimensionality. To evaluate agreement classifications different techniques, calculated Shannon entropy. Finally, Monte Carlo simulation determine presence statistical differences seasons. The results showed that greater confusion generally observed season compensated by selection number hyperspectral classification. SVM RF had highest overall classification accuracy (OA) Kappa values set metrics. reflectance VIs presented better discrimination capability than parameters data. When all considered analysis, gains 6% 8% OA obtained over first ranked (SVM with season; season). lowest entropy larger cover, while largest uncertainties noted grassland/shrub areas. From simulations, not statistically significant most at 99% confidence level. Variations brightness VIs, associated canopy structure, biochemistry physiology, therefore more important variations features, spectra shape bands.