Testing the performance of spatial interpolation techniques for yield mapping with Big Data

CÓRDOBA, M.; BALZARINI, M.

Marrakech

Congreso; 61st World Statistics Congress; 2017

Geostatistical and non-geostatistical approaches have been used for mapping yield variation within agricultural fields. In this work, a non-geostatistical (inverse distance weighting, IDW) and a geostatistical approach (ordinary kriging, OK) were compared as interpolation methods for spatial prediction of yield at unsampled sites. We processed 1038 yield monitor datasets to assess the performance of both interpolation methods. All yield maps were automatically pre-processed to eliminate outliers and spatial outliers. For most yield maps, OK with a number of neighborhood points between 7 and 25, showed a slightly lower mean prediction error than IDW. Even though yield data artifacts were previously deleted, the performance of both interpolation techniques were linearly related with yield variation. The relative prediction error increases about 4% as the data coefficient of variation increases 10%. The good relative performance of IDW, the simplest interpolation method, suggest that it produces accurate yield predictions under high density data with regular samplings. IDW is easier to apply and lesser time-consuming than the geostatistical approaches for Big Data. We conclude that the prediction accuracy of yield at unsampled sites, in the context of high density and regular sampling designs as yield monitors produce, do not rely on the choice of a sophisticated spatial interpolation technique, but rather on data variation.