Statistical modeling for on-farm experimentation with precision agricultural technology

BRUNO CECILIA; CÓRDOBA MARIANO; BALZARINI MÓNICA; PACCIORETTI PABLO; BULLOCK DAVID

Montpellier

Congreso; The 12th European Conference on Precision Agriculture; 2019

On-farm experiments are conducted in producers? fields using precision technologies that facilitate trial set up without additional tasks other than the necessary ones for crop development. Current availability of precise machinery equipped with proximal and remote sensors enable automation of changes in the assignment of agricultural inputs within the field. The possibility of changing input rates automatically and monitoring associated yields enable estimations of the effect of variable rates of agronomical inputs on productivity functions at the farmer scale. Advances in statistical analyses of this type of on-farms trials are essential for rapid development, fine-tuning and evaluation/adoption of precision agriculture. On-farm estimated productivity function (yield as a response of treatments) for a given number of crop seasons, allows us to make environmentally and economically optimum prescriptions for crop management in the agricultural plots. The underlying spatial variability among sites in a plot (soil properties, topography, water availability, historical yields) can be used for identification of management zones, in which treatments are compared under uniform conditions or, rather, as covariables of yield prediction models as a function of treatments.The benefits of adopting a site-specific predictive model approach, including site characteristics as covariables as well as interactions between them with input rates, are discussed through 12 real datasets of on-farm trials of variable seed and nitrogen rate on maize and wheat. Additionally, we evaluate the relative accuracy of modern statistical algorithms: Random Forest, Generalized additive model, Gradient boosting machine (GBM), Partial least squares (PLS). The algorithms were fitted to estimate predictive models and compared in terms of root mean square prediction error (RMSPE). As reference model, multiple linear regression was fitted. All algorithms were adjusted with and without previous zonification and with and without accounting for spatial correlations. All algorithms accuracy increased in models including site covariables. The spatial models showed smaller RMSPE than the non-spatially restricted models. The use of continuous covariates auto-performed the use of management zones as classification factors. GBM and PLS with spatial correlation were the best models for fitting productivity function.