Analysis of fertilization experiments using spatial yield and soil monitoring in precision agriculture

CÓRDOBA, M.; PERALTA, N.; CASTRO, M.; COSTA, J.L.; APARICIO, V.; BALZARINI, M.

Florencia

Conferencia; XXVIIth International Biometric Conference; 2014

International Biometric Society

A goal of precision agriculture is to maximize the resource use efficiency of crops by varying the rate of input applications within field zones. The adoption of this technology demands on-farm experimentation for site specific impact assessments. To explore the fertilization response to site interactions within farmer´s fields, soil and terrain properties are measured at each site and different fertilization treatments are sparse all over the field in small plots under highly replicated block designs. To analyze yield responses from these on-farm experiments we evaluated two covariance modelling strategies: 1- A two-step procedure involving first the delineation of site classes (management zones), which is done by taking into account the multivariate spatial variability of several site properties, and then the treatment comparisons through statistical models making use of the delineated zones and the underlying plot structure; 2- A one-step procedure involving a site-specific yield regression model without previous zonification. To illustrate the procedure, we analyzed data from six on-farm experiments between 30 and 70-ha each, designed to assess the impact of different doses of nitrogen applied on winter cereal crops growing at the Southeast Pampas, in Argentina. Nitrogen was applied at low, medium, and high rates (treatments) in contiguous blocks of 60-by-30m plots over the field. The crop was harvested using a yield monitor, logging the spatial location and the associated yield data at 3 s intervals. In addition, apparent electric conductivity, elevation and soil deep were intensively measured from precise equipment for each plot. For the two-step strategy we first obtained to a two-cluster partition of sites using fuzzy k-means on the spatial principal components of site variables. Then, alternative mixed models including zone effect were fitted: homogeneous and heterogeneous randomized complete block and random block effects within zones (RB), and an RB model with spatially correlated errors (RBCE). For the one-step procedure we fitted a regression model including treatment effects as a class variable, site properties as covariables and spatially correlated errors (Exponential, Gaussian and Spherical), but no zone effects. In most experiments, the two-step strategy, fitting a RB within zone model, gave the best result with respect to the estimation of treatment significance. Models that include the random effect of block within zone had the lowest standard error for treatment comparisons in five of six datasets. Additionally, their results are easier to interpret and useful to the producers looking for the impact that decreasing nitrogen doses would have on yield at each potential management zone. The high amount of small blocks associated with spatial data for each plot, which allows multivariate zone delineation, and the opportunity of zone specific inferences lead us to recommend the two-step procedure rather than the alternative covariance model used for precision agriculture treatment evaluation.