Reference maps of soil phosphorus for the pan-Amazon region

Abstract. Phosphorus (P) is recognized as an important driver of terrestrial primary productivity across biomes. Several recent developments in process-based vegetation models aim at the concomitant representation of the carbon (C), nitrogen (N) and P cycles in terrestrial ecosystems, building upon the ecological stoichiometry and the processes that govern nutrient availability in soils. Thus, understanding the spatial distribution of P forms in soil is fundamental to initialize and/or evaluate process-based models that include the biogeochemical cycle of P. One of the major constraints for the large-scale application of these models is the lack of data related to the spatial patterns of the various forms of P present in soils, given the sparse nature of in situ observations. We applied a model selection approach based on random forest regressions models trained and tested for the prediction of different P forms (total, available, organic, and inorganic P) – obtained by the Hedley sequential extraction method. As input for the models, reference soil group and textural properties, geolocation, N and C contents, terrain elevation and slope, soil pH and mean annual precipitation and temperature from 108 sites of the RAINFOR network were used. The selected models were then applied to predict the target P forms using several spatially explicit datasets containing contiguous estimated values across the area of interest. Here, we present a set of maps depicting the distribution of total, available, organic, and inorganic P forms in the topsoil profile (0–30 cm) of the pan-Amazon region in the spatial resolution of 0.5 x 0.5 degrees. The random forest regression models presented a good level of mean accuracy for the total, available, organic, and inorganic P forms (77.37 %, 76,86 %, 75.14 %, and 68.23 % respectively). Our results confirm that the mapped area have generally very low total P concentration status with a clear gradient of soil development and nutrient content. Total N was the most important variable for the prediction of all target P forms and the analysis of partial dependence indicates several features that are also related with soil concentration of all target P forms. We observed that gaps in the data used to train and test the random forest models, especially in the most elevated areas, constitute a problem to the methods applied here. However, most of the area could be mapped with a good level of accuracy. Also, the biases of gridded data used for model prediction are introduced in the P maps. Nonetheless, the final map of total P resembles the expected geographical patterns. Our maps may be useful for the parametrization and evaluation of process-based terrestrial ecosystem models as well as other types of models. Also, they can promote the testing of new hypothesis about the gradient and status of P availability and soil-vegetation feedbacks in the pan-Amazon region.