Uncertainties about the implications of land-cover heterogeneity on the Amazonian carbon (C) and water cycles are, in part, related to the lack of information about spatial patterns of key variables that control these fluxes at the regional scale. Leaf area index (LAI) is one of these key variables, regulating a number of ecosystem processes (e.g. evaporation, transpiration and photosynthesis). In order to generate a sampling strategy for LAI across a section of Amazonia, we generated a landscape unit (LU) map for the Tapajos region, Eastern Amazonia, as a basis for stratification. We identified seven primary forest classes, stratified according to vegetation and/or terrain characteristics, and one secondary forest class, covering 80% of the region. Primary forest units were the most representative, covering 62% of the total area. The LAI measurements were carried out in 13 selected LUs. In each LU, we marked out three 50 m x 50 m plots giving a total number of 39 plots (9.75 ha). A pair of LAI-2000 plant canopy analysers was used to estimate LAI. We recorded a total of 25 LAI measurements within each plot. We used the field data to verify the statistical distribution of LAI samples, analyse the LAI variability within and among sites, and show the influence of sample size on LAI variation and precision. The LAI showed a high coefficient of variation at the plot level (0.25 ha), from 5.2% to 23%, but this was reduced at the landscape unit level (three co-located plots, 1.8-12%). The level of precision was <10% and 15% at the plot and landscape unit level, respectively. The LAI decreased from a dense lowland forest site (5.10) to a secondary forest (3.46) and to a pasture site (1.56). We found evidence for differences in the scale of spatial heterogeneity of closed canopy forest versus open canopy forest and palm forests. Landscape variables could, in part, explain differences in LAI among forest sites, and land use is an important modifier of LAI patterns. The stratified LAI sampling proposed in the present study could cope with three important aspects of C and water fluxes modelling: (1) optimise the information obtained from field measurements, which is an advance for models parameterisation, compared to the usual random sampling; (2) generate information for a subsequent scaling up of point field