Using the local Moran index to remove errors from crop yield maps

VEGA, A.; CÓRDOBA, M.; BALZARINI M.

Victoria

Conferencia; XXVIIIth International Biometric Conference; 2016

Yield mapping is one of the most widely used precision farming technologies due to its usefulness in both development and evaluation of precision management strategies. The value of these maps can be compromised by the presence of systematic and random errors in raw yield data. To obtain accurate yield maps, the different types of outliers must be removed. We developed a protocol to automate the depuration of several yields, through the application of a twostep sequence of data filters. We use the mean and standard deviation to identify probably erroneous data and consider local outliers as well. Spatial outliers are data that differ significantly from their neighborhood, but that are located within a general range of variation of the data set. We propose first eliminating the data that are outside the mean ± 3 SD and then using the local Moran index of spatial autocorrelation to identify and remove those data that are inconsistent with their neighbor points. The implementation of the two-step protocol was illustrated by processing 593 yield maps of grain crops from the Argentine Pampas region. The effects of primary and secondary step on the spatial structure of grain yield and the precision of yield maps were quantified. Experimentaland theoretical (exponential, spherical and Gaussian) semivariograms were adjusted and the one with the best fit wasselected to make the spatial prediction. For each yield map, k-fold cross-validations (k=100) were performed, and their results were evaluated in terms of the root mean squared error (RMSE) of prediction. The relative improvement ofyield map precision obtained after second step compared with conducting the first step only was calculated as:100*(RMSEstep1 ?RMSEstep2)/RMSEstep1. The results showed that, on average, a total of 11 % of the original yieldmonitor data was removed, with 99% of the removal occurring at step 2 of the filtering process. With this additional filtering, yield semivariograms with smaller nugget values were obtained. Nugget effects, expressed as the proportionof the overall yield variance, decreased by 20% using the two-step protocol with respect to the use of raw yield data for mapping. Furthermore, a relative increase in map precision of 40% was obtained compared with conductingprimary screening only. The proposed algorithm is easy to apply and robust enough for implementation in databases of many grain yield maps.