Modeling distribution of Amazonian tree species and diversity using remote sensing measurements

The availability of a wide range of satellite measurements of environmental variables at different spatial and temporal resolutions, together with an increasing number of digitized and georeferenced species occurrences, has created the opportunity to model and monitor species geographic distribution and richness at regional to continental scales. In this paper, we examine the application of recently developed global data products from satellite observations in modeling the potential distribution of tree species and diversity in the Amazon basin. We use data from satellite sensors, including MODIS, QSCAT, SRTM, and TRMM, to develop different environmental variables related to vegetation, landscape, and climate. These variables are used in a maximum entropy method (Maxent) to model the geographical distribution of five commercial trees and to classify the patterns of tree alpha-diversity in the Amazon basin. Maxent simulations are analyzed using binomial tests of omission rates and the area under the receiver operating characteristics (ROC) curves to examine the model performance, the accuracy of geographic distributions, and the significance of environmental variables for discriminating suitable habitats. To evaluate the importance of satellite data, we used the Maxent jackknife test to quantify the training gains from data layers and to compare the results with model simulations using climate-only data. For all species and tree alpha-diversity, modeled distributions are in agreement with historical data and field observations. The results compare with climate-derived patterns, but provide better spatial resolution and detailed information on the habitat characteristics. Among satellite data products, QSCAT backscatter, representing canopy moisture and roughness, and MODIS leaf area index (LAI) are the most important variables in almost all cases. Model simulations suggest that climate and remote sensing results are complementary and that the best distribution patterns can be achieved when the two data sets are combined.