Multivariate spatial variability analysis in grain yields at fine scale

CÓRDOBA, M.; BRUNO, C.; BALZARINI, M.; COSTA, J.L.

Kobe

Conferencia; XXVIth International Biometric Conference; 2012

Biometric Society of Japan y Japanese Region of the International Biometric Society.

In modern agriculture georeferenced multivariate data are frequently obtained in crop fields, where several variables such yield and soil properties are measured at each site. Soil variables are frequently used to explain grain yield spatial variability. The spatial covariations of soil properties can be evaluated through classical principal component analysis (PCA). However as other multivariate descriptive methods, PCA was not specifically developed to handle spatial data. Recently developed multivariate analyses allow taking into account spatial autocorrelation among neighbouring sites. In this paper we use and compare PCA with spatially constrained multivariate analysis methods based on Moran´s Index (MULTISPATI-PCA) and their combinations with other multivariate methods like canonical correlation analysis (CCA) and fuzzy k-means clustering (FKM). These methods (PCA, MULTISPATI-PCA and CCA) allow generating new synthetic variables (axis) that maximize the explanation of variability and the autocorrelations among variables and variable groups. Both the synthetic variables obtained by each method and the original variables, were used as input for FKM to identify homogeneous zones in a multivariate sense, within the field. The number of cluster defined a priori was three. The CCA was used to evaluate the relationship between grain yield and the first three principal components (PC), as well as with the first three spatial principal components (PCe) of MULTISPATI-PCA. New synthetic variables were obtained from the first significant canonical correlation of the PC (CCA-PC) and PCe (CCA-PCe). The original soil variables, and the CCA-PC, CCA-PCe synthetic variables were categorized in three classes defined by the 33 and 66 percentiles, as another way to identify homogeneous zones within-field. Finally, we compared the average yield among the identified zones using a mixed model ANOVA with a spatial-type for the residual covariance matrix. In addition we used the standard deviation (SD) of the mean-adjusted yields as a measure to compare homogeneous zones obtained by different methods. The applications of these multivariate methods are illustrated in a data base at field scale. Apparent electrical conductivity at 30 cm (ECa30) and 90 cm (ECa90) deep, elevation, hardpan depth and soybean and wheat yields were measured at each georeferenced site by precision agriculture tools. Yield data were centered and averaged prior to analysis. The first axis score obtained by MULTISPATI-PCA showed a high and positive correlation with elevation (variable with the highest autocorrelation) whereas the same axis score obtained by PCA was correlated with electrical conductivity at 30 cm. Hardpan variable was correlated with the second MULTISPATI-PCA axis score and with the PCA third axis. The ANOVA results from soil variables showed significant differences between the three zones only to ECa30. The same results were obtained for multivariate methods combined with MULTISPATI-PCA and the combination of CCA and PCA. Furthermore, the multivariate methods that identify three areas were also those who had lower SD, mainly CCA-MULTISPATIPCA. The results show that MULTISPATI-PCA and its combination with other multivariate methods are a useful tool to map spatial variability and to identify homogeneous zones, in a multivariate sense, within fields.