Dual-season and full-polarimetric C band SAR assessment for vegetation mapping in the Amazon várzea wetlands

• Autor(a) principal: Furtado, Luiz Felipe de Almeida
• Data de Publicação: 2016
• Outros Autores: Silva, Thiago Sanna Freire [UNESP], Novo, Evlyn Márcia Leão de Moraes
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Repositório Institucional da UNESP
• Texto Completo: http://dx.doi.org/10.1016/j.rse.2015.12.013; http://hdl.handle.net/11449/172337
• topic: Mapping accuracy; Multitemporal; Polarimetric decomposition; PolSAR; Wetlands

This study answered the following questions: 1) Is polarimetric C-band SAR (PolSAR) more efficient than dual-polarization (dual-pol) C-band SAR for mapping várzea floodplain vegetation types, when using images of a single hydrological period? 2) Are single-season C-band PolSAR images more accurate for mapping várzea vegetation types than dual-season dual-pol C-band SAR images? 3) What are the most efficient polarimetric descriptors for mapping várzea vegetation types? We applied the Random Forests algorithm to classify dual-pol SAR images and polarimetric descriptors derived from two full-polarimetric Radarsat-2 C-band images acquired during the low and high water seasons of Lago Grande de Curuai floodplain, lower Amazon, Brazil. We used the Kappa index of agreement (κ), Allocation Disagreement (AD) and Quantity Disagreement (QD), and Producer's and User's accuracy measurements to assess the classification results. Our results showed that single-season full-polarimetric C-band data can yield more accurate classifications than single-season dual-pol C-band SAR imagery and similar accuracies to dual-season dual-pol C-band SAR classifications. Still, dual-season PolSAR achieved the highest accuracies, showing that seasonality is paramount for obtaining high accuracies in wetland land cover classification, regardless of SAR image type. On average, single-season classifications of low-water periods were less accurate than high-water classifications, likely due to plant phenology and flooding conditions. Classifications using model-based polarimetric decompositions (such as Freeman-Durden, Yamaguchi and van Zyl) produced the highest accuracies (κ greater than 0.8; AD ranging from 7.5% to 2.5%; QD ranging from 15% to 12%), while eigenvector-based decompositions such as Touzi and Cloude-Pottier had the worst accuracies (κ ranging from 0.5 to 0.7; AD greater than 10%; QD smaller than 10%). Vegetation types with dense canopies (Shrubs, Floodable Forests and Emergent Macrophytes), whose classification is challenging using C-band, were accurately classified using dual-season full-polarimetric SAR data, with Producer's and User's accuracies between 80% and 90%. We conclude that full polarimetric C-band imagery can yield very accurate classifications of várzea vegetation (κ ~0.8, AD ~3% and QD ~10%) and can be used as an operational tool for forested wetland mapping.