Mapping Phosphorus Forms in the Pan-Amazon Region: A Machine Learning Approach

Phosphorus (P) is a key driver of terrestrial productivity. However, the lack of spatial data on various P forms in soils hinders the large-scale application of process-based vegetation models. To address this, we used a model selection approach based on Random Forest regression models to predict different P forms (total, available, organic, inorganic, and occluded P) in the pan-Amazon region. Our models were trained and tested using data from 108 sites of the RAINFOR network, including soil group and textural properties, geolocation, nitrogen (N) and carbon (C) contents, terrain elevation and slope, soil pH, and mean annual precipitation and temperature. The models were then applied to several spatially explicit datasets to predict the target P forms. The resulting maps depict the distribution of total, available, organic, inorganic, and occluded P forms in the topsoil profile (0 - 30 cm) at a spatial resolution of 5 arcminutes. Our models achieved a good level of mean accuracy (77.37 %, 76,86 %, 75.14 %, 68.23 %, and 64.62% for the total, available, organic, inorganic, and occluded P forms, respectively). Our results reveal a clear gradient of soil development and nutrient content, with the mapped area generally exhibiting very low total P concentration status. Total N was the most important variable for predicting all target P forms. Despite some gaps in the training and testing data, most of the area could be mapped with a good level of accuracy. Our maps can aid in the parametrization and evaluation of process-based terrestrial ecosystem models and promote the testing of new hypotheses about P availability and soil-vegetation feedbacks in the pan-Amazon region.