The study of plant community structure and its causes has advanced in the last years, but it is far from complete answers to questions such as the relationship between richness and productivity or the main determinants of the compositional structure and the scale in which they act. The hump-shaped curve observed in relationship between richness and productivity is the main pattern found in plant communities, but with less frequency in the Amazon. Linear relationships or the lack of any relationship may result from lack of coverage of the full productivity gradient. Climate, edaphic factors and spatial variables are commonly identified as important predictors of changes in floristic composition at the regional scale, but with low power of explanation. The low predictive power of most models has at least 3 causes: 1) sampling errors, which led to a community representation that is only a fraction of the real community; 2) errors in the choice of predictors, which may not include all relevant variables and 3) analytical errors, by the use of models imcompatible with the biological structures under study. In this study, my objectives were 1) to understand the relationships between species density, and the floristic variation of Zingiberales communities, and climate, edaphic factors and spatial features in Central and Northern Amazon terra firme forests. 2) to assess whether the application of the method ISOMAP improves the performance of distance matrix models. Composition of Zingiberales, climate (measure by the Walsh index), soil texture and fertility data were acquired in 170 plots (250 x 2 m) distributed over 7 forests sites, covering 800 km along a north-south axis. We used multiple linear regression, ISOMAP, multiple regression of distance matrices and multivariate regression trees to assess the effect of spatial and environmental variables. At the coarse grain, the species density was affected by climate, after accounting for the site variation in soil fertility. At the fine grain, only fertility affected negatively the species density. The total variance explained by the models was twice greater when the dissimilarity matrices were transformed by ISOMAP. No soil texture effect was detected in any analyses. Geographic distance was the main predictor of variation in floristic distances (54%), followed by soil fertility (8,4%) and climate (6,9%). Since climate and geographic distances were highly correlated in this dataset, it is not possible to determine the real cause of the floristic pattern. Soil texture was not a significant predictor in any of the models, contrary to the observed in studies in the mesoscale, and this may reflect different relationships between soil drainage and topography in each region. In a more detailed analysis it was shown that soil fertility determines the formation of different floristic groups only in the region with wetter climates, indicating an interaction between soil fertility and climate. Conclusions: (1) climate and soil fertility can be used in predictive models of floristic variation in the brazilian Amazon, but models reflecting the interaction between these variables must be sought; (2) direct measures of key variables, such as soil drainage must be included in future models, and (3) new study sites must be choose in order to reduce the high correlation between climatic and geographic distances.
