Integrating field sampling, spatial statistics and remote sensing to map wetland vegetation in the Pantanal, Brazil

Abstract

To improve the protection of wetlands, it is imperative to have a thorough understanding of their structuring elements and of the identification of efficient methods to describe and monitor them. This article uses sophisticated statistical classification, interpolation and error propagation techniques, in order to describe vegetation spatial patterns, map plant community distribution and evaluate the capability of statistical approaches to produce high-quality vegetation maps. The approach results in seven vegetation communities with a known floral composition that can be mapped over large areas using remotely sensed data. The relations between remotely sensing data and vegetation patterns, captured in four factorial axes, were formalized mathematically in multiple linear regression models and used in a universal kriging procedure to reduce the uncertainty in mapped communities. Universal kriging has shown to be a valuable interpolation technique because parts of vegetation variability not explained by the images could be modeled as spatially correlated residuals, increasing prediction accuracy. Differences in spatial dependence of the vegetation gradients evidenced the multi-scale nature of vegetation communities. Cross validation procedures and Monte Carlo simulations were used to quantify the uncertainty in the resulting map. Cross-validation showed that accuracy in classification varies according with the community type, as a result of sampling density and configuration. A map of uncertainty resulted from Monte Carlo simulations displayed the spatial variation in classification accuracy, showing that the quality of classification varies spatially, even though the proportion and arrangement of communities observed in the original map is preserved to a great extent. These results suggested that mapping improvement could be achieved by increasing the number of field observations of those communities with a scattered and small patch size distribution; or by including new digital images as explanatory variables in the model. By comparing the resulting plant community map with a flood duration map, we verified that flooding duration is an important driver of vegetation zonation. We discuss our study in the context of developing a mapping approach that is able to integrate field point data and high-resolution remote sensing images, providing new basis to map wetland vegetation and allowing its future application in habitat management, conservation assessment and long-term ecological monitoring in wetland landscapes.