Productivity Estimation of Pampaean Soils By An Artificial Neural Network Approach.

Deductive productivity indices (PI) can be generated using information from empirical models for crop yield estimation. These type of PI’s are directly validated against yield data. A soil PI represents the rating of potential plant biomass production of soils. Empirical models attempt to determine functional relationships between soil, climate and management factors. The objective of this work was to develop a deductive PI for assessing regional soil productivity for wheat yield in the Pampas. Soil data from soil surveys and interpolated climate information were used. Wheat yield data of a period of forty years was used. The study area was of ca. 45 Mha and was divided in 41 geographical units and all information was aggregated up to this level. Four techniques for yield modeling were tested: blind guess, polynomial regression, regression trees, and artificial neural networks (ANN). The empirical model with the best performance was the blind guess model but soils could only be rated when yield data was available. Yield model based on the ANN approach had a good performance (R² = 0.614, RMSE = 548 kg ha^-1) and was used for regional productivity index development. This index could be extrapolated to soils for which yield data were not available and its validation with yield averages was optimal (R² = 0.728; P = 0.05). Regional high productivity was achieved for soil of medium to high contents of soil organic carbon and soil available water storage capacity. Both soil variables interacted positively. This regional estimation could not be validated at plot scale. Using the ANN approach for assessing regional could be applied in other regions of the World and for different crops.