Modeling Soil Test Phosphorus Changes under Fertilized and Unfertilized Managements Using Artificial Neural Networks

Abstract

Core Ideas An artificial neural network was developed to describe soil P dynamics. The model accurately predicts soil test P increases and decreases. A meta‐model was derived to apply the build‐up and maintenance philosophy. The build‐up and maintenance criteria have been introduced for P fertilizer management in the Pampas of Argentina. However, methods for predicting soil test P changes under contrasting fertilizer rates are not available. We performed a meta‐analysis using results from 18 local field experiments performed under the most common crop rotations, in which soil test P changes with and without P fertilization and soil P balance were assessed. We assembled 329 soil test P variation data sets corresponding to a period 12 yr and 129 P balance records. The P balance was not a good predictor of annual soil test P changes (R2 = 0.33). In 38% of the cases, the P balance and soil test P changes showed opposite trends. Polynomial regression and artificial neural networks were tested for soil test P modeling. The neural networks performed better than the regressions (R2 = 0.91 vs. 0.83; P < 0.01). The network that yielded the best results used the initial soil test P level, the P fertilization rate and time as inputs. According to the model, unfertilized crops growing in soils with low initial P levels (soil test P = 10 mg kg−1 or lower) were subjected to only small decreases in soil test P levels, whereas greater decreases occurred in soils with initial high P levels. For fertilized crops, the model showed that P‐rich soils were less enriched in P than P‐poor soils. A simple meta‐model was developed for the prediction of soil test P changes under contrasting fertilizer managements.