You are on page 1of 11

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/260554113

Agronomy Journal

Article  in  Agronomy journal · December 2013

CITATIONS READS
0 1,077

5 authors, including:

Marcos Rodrigues Annamaria Castrignanò


Universidade Federal do Vale do São Francisco (UNIVASF) Council for Agricultural Research and Agricultural Economy Analysis
20 PUBLICATIONS   65 CITATIONS    236 PUBLICATIONS   1,949 CITATIONS   

SEE PROFILE SEE PROFILE

Tom Mueller Eduardo Abel Rienzi


John Deere University of Kentucky
63 PUBLICATIONS   676 CITATIONS    38 PUBLICATIONS   155 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Soil wetting under different tillage system. Effect of kinetic energy in mobilization of particulate organic carbon View project

Use of reflectance spectroscopy for soil conservation and management. View project

All content following this page was uploaded by Eduardo Abel Rienzi on 06 March 2014.

The user has requested enhancement of the downloaded file.


Agronomy, Soils & Environmental Quality

A Spatial and Temporal Prediction Model of Corn Grain Yield


as a Function of Soil Attributes
Marcos S. Rodrigues,* José E. Corá, Annamaria Castrignanò, Tom G. Mueller, and Eduardo Rienzi

Abstract
Effective site-specific management requires an understanding of the soil and environmental factors influencing crop yield
variability. Moreover, it is necessary to assess the techniques used to define these relationships. The objective of this study was
to assess whether statistical models that accounted for heteroscedastic and spatial-temporal autocorrelation were superior to
ordinary least squares (OLS) models when evaluating the relationship between corn (Zea mays L.) yield and soil attributes in
Brazil. The study site (10 by 250 m) was located in São Paulo State, Brazil. Corn yield (planted with 0.9-m spacing) was measured
in 100 4.5- by 10-m cells along four parallel transects (25 observations per transect) during six growing seasons between 2001
and 2010. Soil chemical and physical attributes were measured. Ordinary least squares, generalized least squares assuming
heteroscedasticity (GLShe), spatial-temporal least squares assuming homoscedasticity (GLSsp), and spatial-temporal assuming
heteroscedasticity (GLShe-sp) analyses were used to estimate corn yield. Soil acidity (pH) was the factor that most influenced corn
yield with time in this study. The OLS model suggested that there would be a 0.59 Mg ha–1 yield increase for each unit increase
in pH, whereas with GLShe-sp there would be a 0.43 Mg ha–1 yield increase, which means that model choice impacted prediction
and regression parameters. This is critical because accurate estimation of yield is necessary for correct management decisions. The
spatial and temporal autocorrelation assuming heteroscedasticity was superior to the OLS model for prediction. Historical data
from several growing seasons should help better identify the cause and effect relationship between crop yield and soil attributes.

Y ield mapping is a simple, inexpensive tool for monitoring


crop yield at fine spatial resolutions; however, yield maps
have little or no value unless they can be used for decisions that
that influence crop yield variability can change from year to year
because of seasonal weather variation. For example, Timlin et
al. (1998) found that corn yields were correlated with soil P, K+,
will improve crop and soil management (Pierce and Nowak, and organic matter (OM) content only in dry years in a Typic
1999). Consequently, effective and meaningful spatial analysis of Fragiochrept soil. Bakhsh et al. (2000) pointed out that it may be
yield variability and yield-limiting factors has become a critical necessary to also include climate and management factors in yield
issue throughout the world (McBratney et al., 2005). More- prediction models for detailed diagnosis of yield-limiting factors.
over, yield maps can vary substantially from year to year, so that Regression is one of the most common ways to describe the
analysis of data from only 1 yr might potentially lead to incorrect relationship between crop yield and soil attributes; however, it
decisions. Thus, it is critical to assess the temporal variability and has a number of important assumptions that should be taken
stability of crop yields (McBratney et al., 2005). into account (e.g., normal distribution and independence of
Several factors can affect crop yield variability, such as climate, errors, linearity, lack of collinearity, reliability of measurements,
soil fertility, terrain properties, weeds, and diseases. Numerous and homoscedasticity) (Osborne and Waters, 2002).
studies have found that crop yields are frequently spatially Ordinary least squares (OLS) regression, the most common
correlated to soil fertility attributes, but it is critical that these statistical procedure used for yield estimation, assumes normality,
kinds of studies use similar sampling support (Rodrigues et al., independence of errors, and homoscedasticity. Yield residuals,
2012). According to Diker et al. (2004), the dominance of factors however, generally are often spatially autocorrelated (Lark, 2000;
Lobell et al., 2005). If the test statistic does not account for the
M.S. Rodrigues, Univ. Federal do Vale do São Franscisco (UNIVASF), autocorrelation, it will be too large and the probability values too
Campus Ciências Agrárias–BR 407, 12 Lote 543, Projeto de Irrigacão small; therefore, the test will reject more often than it should. Type
Nilo Coelho–S/N C1, 56300-990 Petrolina, PE, Brazil; J.E. Corá, Dep. I errors then tend to increase with OLS when spatial dependence
of Soil Science, Univ. Estadual Paulista (UNESP), Jaboticabal, SP, Brazil,
14884-900; A. Castrignanò, Consiglio per la Ricerca e Sperimentazione in (Schabenberger and Gotway, 2005) and heteroscedasticity are
Agricolture (CRA), Via Celso Ulpiani, N. 5, 70125 Bari, Italy; T.G. Mueller, ignored (Osborne and Waters, 2002). This could lead to critical
Intelligent Solutions Group, John Deere and Company, Urbandale, IA misinterpretation errors of yield variation and consequently lead to
50322; and E. Rienzi, Dep. of Plant and Soil Sciences, Univ. of Kentucky,
Lexington, KY 40546. Received 26 Nov. 2012. *Corresponding author improper management decisions.
(marcos.rodrigues@univasf.edu.br). Spatial dependence structure may change from year to year;
Published in Agron. J. 105:1878–1887 (2013)
therefore, it is critical to investigate crop yield and soil property
doi:10.2134/agronj2012.0456
Copyright © 2013 by the American Society of Agronomy, 5585 Guilford Abbreviations: AIC, Akaike’s information criterion; AICC, Akaike’s
Road, Madison, WI 53711. All rights reserved. No part of this periodical may information criterion corrected; BIC, Baysian information criterion; GLShe,
be reproduced or transmitted in any form or by any means, electronic or general least squares assuming heteroscedasticity; GLSsp-he, spatial-temporal
mechanical, including photocopying, recording, or any information storage assuming heteroscedasticity; GLSsp, spatial-temporal least squares assuming
and retrieval system, without permission in writing from the publisher. homoscedasticity; OLS, ordinary least squares; OM, organic matter.

1878 A g ro n o my J o u r n a l  •  Vo l u m e 10 5 , I s s u e 6  •  2 013
interactions with time. In these situations, generalized least
squares (GLS) regression with correlated errors should be
used (Flinn and De Datta, 1984) because it allows spatial and
temporal error correlation components to be assessed and then
filtered from the total residual term of the model to improve the
power of the statistical tests.
As an alternative to traditional regression analyses, spatial-
temporal mixed effect models could be used, which assume
normally distributed and spatially and temporally correlated
errors and include both fixed and random effects (Bolker et al.,
2009). The study of the spatial-temporal structure of the errors
is very important for monitoring and evaluating crop yield.
Generally, space and time effects are considered separately;
however, interactions are common across spatial and temporal
scales and modeling efforts should account for both effects
simultaneously (Landagan and Barrios, 2007).
Some studies (Hurley et al., 2004; Lambert et al., 2004) have
pointed out that mixed effect models, which incorporate spatial
variability, will help improve the understanding of the factors that
affect crop yield. For example, Hurley et al. (2004) found that
the best model to describe the N response in corn yield was one
that took into account heteroscedasticity and spatial correlation.
There are not many studies, however, that treat spatial-temporal
variability and heteroscedasticity of crop yield in the same model.
Therefore, the objective of this study was to assess whether
statistical models that account for heteroscedasticity and spatial-
temporal autocorrelation are superior to OLS models when
evaluating the relationship between corn yield and soil attributes
in a field in Brazil.

MATERIALs AND METHODS


Site Description Fig. 1. Sampling scheme for soil attributes and corn yield in a
Rhodic Hapludox under no-till management.
This experiment was conducted in the city of Jaboticabal, in
São Paulo state, Brazil (21°14¢5² S, 48°17¢9² W, and altitude
of 613 m asl). Climatologically, the area belongs to tropical/ plants had four to six pairs of leaves totally developed. The field
megathermal zone or Köppen Aw (tropical climate with dry was uniformly treated. Corn was harvested about 150 d after
winter and average temperature of the coldest month >18°C). planting with a one-row plot combine that deposited the grain
The mean annual rainfall (1971–2006) is 1417 mm, with the into a burlap bag. Grain weights were obtained for each plot
distribution peaking in the period of October to March and a with a manual balance in the field. The grain for each plot was
relatively dry season in the period of April to September. The subsampled for moisture, and grain yields were determined at
soil of the experimental area was classified as a clayey Rhodic 13% gravimetric moisture.
Hapludox. The experimental area was managed as a corn–fallow Each year, five soil subsamples were collected within each
rotation under no-till management for 12 yr. plot using a Dutch auger (0.1-m depth) and were composited.
One of the soil subsamples was collected in the middle of the
Yield and Soil Sampling and Climatic Data plot and the other four samples were collected 2 m apart from
The size of the experimental area was 18 by 250 m, with the the middle in all four cardinal directions from the centroid. The
longer dimension oriented in the north–south direction. Each area associated with soil measurements (support) was assumed
of the 100 experimental plots had a dimension of 10 by 4.5 m similar to the one for crop yield (45 m2). Each soil composite
and were arranged in a 25 by 4 grid. The experimental scheme is sample was analyzed for particle size (pipette method) (Gee and
depicted in Fig. 1. Or, 2002), pH (1:1 soil/water mixture), OM content (Walkley–
Before planting, nonselective herbicides were applied. Black method), P (ion‐exchange resin), and K+, Ca2+, and Mg2+
Corn (triple-hybrid Syngenta Master) was planted at 65,000 (1 mol L–1 NH4OAc extractable buffered at pH 7) according
plants ha–1 with 0.9-m row spacing in early December during to Page et al. (1982). From the analytical determinations,
the 2001 to 2010 growing seasons, but the data were collected the cation exchange capacity (CEC = K+ + Ca2+ + Mg2+ +
according to the sampling design depicted in Fig. 1 only in the H+ + Al3+) and percentage of soil base saturation [BS = (K+ +
2001–2002, 2002–2003, 2003–2004, 2007–2008, 2008–2009, Ca2+ + Mg2+/CEC) ´ 100] were calculated. These variables
and 2009–2010 growing seasons. The starter fertilization were chosen because they are commonly measured by Brazilian
consisted of 30 kg of N, 70 kg of P2O5, and 50 kg of K 2O ha–1. farmers to characterize soil fertility.
Nitrogen fertilizer (urea) was applied at 100 kg N ha–1 when the

Agronomy Journal  •  Volume 105, Issue 6  •  2013 1879


forward selection, each of the available predictors is added if it
meets the statistical criterion of entry, which is the significance
level (P < 0.15) for the increase in the r2 produced by addition
of the predictor. At each step, variables that are already in the
model are first evaluated for removal, and if any are eligible for
removal (P > 0.15), the one whose removal would cause the
least decrease in the r2 is removed. This procedure is repeated
until there are no more predictors that are eligible for entry or
removal.

Linear Mixed Effect Model


Four different regression models were compared to assess
the impact of soil attributes and growing seasons (temporal
variability) on corn yield: ordinary least squares (OLS),
generalized least squares assuming heteroscedasticity (GLShe),
spatial-temporal model assuming homoscedasticity (GLSsp), and
spatial-temporal model assuming heteroscedasticity (GLShe-sp)
approaches (Schabenberger and Gotway, 2005) were used.
The standard linear model (OLS) can be written as

p
y i = å x ij bj + e i , i = 1,..., N [1]
j= 1

where yi are N data points of the response variable (i.e., corn


yield), xij are the observations of p explanatory variables (j),
Fig. 2. Summary of climate data: (a) average monthly rainfall which can be continuous variables (i.e., soil variables) or dummy
values for the period December to April of the studied years
and 30-yr average; and (b) days with rainfall for the period variables declaring class membership of a categorical variable
December to April of the studied years. (i.e., growing season), b i … b p are fixed effect coefficients to be
estimated, and ei are unknown independent and identically
distributed normal random variables with mean 0 and variance
Monthly cumulative rainfall, growing degree days (base s2 .
temperature of 10°C), average temperatures, relative humidity, For convenience and simplicity, Eq. [1] can be written using
and number of days with rainfall were recorded by the matrix notation:
climatological station of São Paulo State University (21°14¢5² S,
48°17¢9² W and altitude of 615 m asl) from December 2001 to Y = Xb + e [2]
April 2010 (Fig. 2). The weather station was located 30 m from
the experimental site. where Y is the vector of the responses, X is the matrix of the
observations, b is the vector of the unknown fixed effect
Preliminary Statistical Analyses coefficients, and e is the vector of the independent and identically
Yield and soil attributes were tested for normality (Shapiro distributed normal random errors, or in symbols: e ? N(0,
and Wilk, 1965), and yield was tested for heteroscedasticity with s2I) where I is an identity matrix. If the error variance s2 is not
Bartlett’s test (Snedecor and Cochran, 1989). Yield was also constant but varies as a function of a class variable (growing
initially tested for spatial autocorrelation with Moran’s I test season), the model is called a generalized least-squares model
(Moran, 1950). Moran’s I test is a weighted correlation coefficient (GLShe).
used to detect departures from spatial randomness, that is, it
determines whether neighboring areas were more similar than Spatial Mixed Effect Model
would be expected under the null hypothesis. Moran’s I test was Many times the independence distributional assumption
used as a generalized measure of spatial autocorrelation because about Y is too restrictive and the linear mixed model extends the
it indicates the presence or absence of a stable pattern of spatial general linear model by allowing elements of Y to be correlated.
dependence by applying to the whole data set (Anselin, 1993). This was performed in two ways: through a specification of
Exploratory stepwise regression analysis was performed to the covariance of e as a function of the distance between two
select the best subset of regressors out of all the study variables locations, say e ? N(0, R), for spatial variability, and the addition
(clay, sand, silt, pH, OM, K+, Ca2+, Mg2+, CEC, and BS). The of a random effect and random coefficient in the analysis, giving
final variables selected for the mixed effects model were chosen rise to a Zu term in the model, where u is normal with mean 0
because they performed well with initial testing, they did not and variance G, for temporal variability and Z is a matrix, similar
show collinearity, and they were biophysically meaningful. to X, that captures the complex covariance structure of the
Stepwise selection is similar to the forward method, except that temporal factor.
variables already in the model do not necessarily stay there. In a

1880 Agronomy Journal  •  Volume 105, Issue 6  •  2013


The spatial-temporal linear mixed effect model (GLSsp) can hypothesis, ρ = 0 and s2l = 0 with two degrees of freedom, was
then be written as performed. Under the null hypothesis that the spatial model is not
different from the nonspatial model, the likelihood ratio statistic is
Y = Xb + Zu + e
[3] distributed as c2, with the number of degrees of freedom equal to
the difference in the number of parameters between the two models.
where e ? N(0, R), u ? N(0, G), Cov[u, e] = 0, which implies Therefore, because the fixed part is the same in the spatial and
the assumption that u and e are uncorrelated. Differently than b, nonspatial models, only the parameters in the variance–covariance
the vector u does not contain parameters but random variables. structure need to be considered.
The temporal relationship (growing season) was explored by For comparison, the bias, accuracy, and precision of the four
postulating an autoregressive structure of Order 1 for the matrix regression models were assessed by means of leave-one-out cross-
G, which has homogeneous variances and correlations that decline validation using the three cross-validation statistics suggested by
exponentially with the time series. The AR(1) covariance structure Carroll and Cressie (1996) as follows. If Y[i] is the observed corn
has two unknown parameters: the variance (st2) and the lag-one yield value removed at the ith iteration, Ŷ[i] is its corresponding
correlation (ρt). However, a temporal factor was included also in prediction obtained by fitting the model to the remaining N – 1
the model as a fixed effect (growing season) and a dummy variable points, e[i] = Y[i] – Ŷ[i] is the difference between the observed
in the matrix X to assess the systematic or trend component in and estimated values, and s[i] is the mean squared prediction
corn yield variation and then evaluate its stability with time. error of Ŷ[i], then the three cross-validation statistics are
The spatial relationship was modeled by using three different
isotropic covariance functions of the distance, such as spherical,
1 n e [i ]
exponential, and Gaussian (Littell et al., 2006), and adding an CV1 = å
additional parameter (nugget effect) to adequately account for n i=1 s[i ]
abrupt changes across relatively small distances. The resulting 1/2
covariance model has the form æ 1 n e [2i ] ö÷ [4]
CV2 = ççç å 2 ÷÷
çè n i=1 s[ i ] ÷ø
Var [e i ] = s2 +s12
1/2
æ1 n ö
CV3 = çç å e [2i ] ÷÷÷
Cov éëêe i , e j ùûú = s2 éêf (d ij )ùú çè n i=1 ø
ë û
where f(dij) is one of the geostatistical spatial covariance The CV1 was used to assess the unbiasedness of the predictor,
functions of distance dij between two observations i and j, and the optimal value should be approximately zero; CV2 was
using a parameter ρ for the spatial scale (range) and sl2 and used to assess the accuracy of the mean squared prediction error
s2 + sl2 corresponding to the geostatistical parameters nugget and should be approximately 1; CV3 was used to check the
and sill, respectively. The fitting process relies on an iterative goodness of prediction, and models with smaller values of CV3
procedure aimed at maximizing the log likelihood of the should be preferred because this means that fitted values are close
data by the restricted maximum likelihood method (REML) to the observed values (Carroll and Cressie, 1996).
(Littell et al., 2006). The fixed effects estimates are obtained All statistical analyses in this study were computed with SAS
as generalized least squares estimates evaluated at the REML (release 9.3, SAS Institute) and the linear spatial mixed effect
estimate of the covariance parameters. A further complexity in model was estimated with the PROC MIXED procedure (see
the spatial-temporal model was added by allowing crop yield the appendix).
variance to vary across the growing seasons because it can cause
heterogeneity in the covariance structure. The whole analysis RESULTS AND DISCUSSION
was then repeated using a different set of covariance function Description of Soil Attributes across Years
parameters for each growing season (GLShe-sp). All soil fertility data were classified as low, medium, or
high according to the criteria determined by Raij et al. (1997)
Model Comparison for the state of Sao Paulo as reported below for each variable.
To compare the different competing spatial covariance The soil base saturation values were medium (51–70%) in the
models, their modeling criteria were compared: the best model 2001–2002 and 2002–2003 growing seasons and low (26–50%)
was selected as the one whose –2 log likelihood, Akaike’s in the 2003–2004, 2007–2008, 2008–2009, and 2009–2010
information criterion (AIC), Akaike’s information criterion growing season. The levels of soil K+ (1.6–3.0 mmolc L–1) as
corrected (AICC), and Schwarz’s Bayesian information criterion well as the soil P levels were medium (16–40 mg L–1) in all the
(BIC) were the smallest (Littell et al., 2006). study years. The levels of soil Ca2+ were high (>7 mmolc L–1) in
Each model with no spatial correlation (OLS and GLShe), i.e., all study years, whereas soil Mg2+ was high (>8 mmolc L) in the
with R = s2I, was compared with the corresponding homoscedastic 2001–2002, 2002–2003, and 2003–2004 seasons and medium
(GLSsp) and heteroscedastic (GLShe-sp) spatial model with nugget (5–8 mmolc L–1) in 2007–2008, 2008–2009, and 2009–2010.
effect, respectively, whose R = s2F+ s2I, where F is an N ´ N Medium values of pH (5.1–5.5) were observed in the 2001–2002
matrix whose ijth element is f(dij). Because F reduces to I if the and 2002–2003 seasons and low values (4.4–5.0) in 2003–2004,
spatial parameter ρ = 0 and sl2 = 0, to compare the –2 log likelihood 2007–2008, 2008–2009, and 2009–2010. Medium soil OM
for spatial and nonspatial models, a likelihood ratio test for the null contents were observed in all years. Most variables showed large

Agronomy Journal  •  Volume 105, Issue 6  •  2013 1881


variability on the basis of their range of variation. In general, coefficients (GS 1–6), which represent the specific contribution
the soil fertility variables were not normally distributed and of each season to the overall yield average.
coefficients of asymmetry and kurtosis varied across years. This The goodness-of-fit statistics (i.e., –2 log likelihood, AIC,
did not prevent regression analyses, however, because normality AICC, and BIC) differed among the models, which were
is a requirement only for the response variable. The clay contents ordered as follows: OLS > GLShe > GLSsp > GLShe-sp (smaller
(mean = 332 g kg–1, standard deviation = 19 g kg–1) and sand is better) (Table 3). This order was established on the basis of the
contents (mean = 623 g kg–1, standard deviation = 17 g kg–1) of values of at least three out of the four criteria being consistent.
the surface samples did not vary substantially throughout the If the BIC had been the only one used to compare the models,
experimental area and were considered normally distributed OLS would have been considered better than GLShe, whereas
according to the Shapiro–Wilk test. The soil was classified as a a different conclusion was determined on the basis of the
sandy clay loam according to the USDA texture method. other criteria. This result confirmed warnings that multiple
criteria should be considered when comparing models (Littell
Preliminary Statistical Analyses for Corn Yield et al., 2006). The likelihood ratio test of spatial covariance was
Corn yields ranged from 6.0 to 7.6 Mg ha–1 (Table 1) and significant (P < 0.001) when the nonspatial homoscedastic
showed normal data distribution in all years excepted the 2003– (OLS) and heteroscedastic (GLShe) models were compared
2004 and 2008–2009 growing seasons, as indicated by the with the corresponding spatial models (GLSsp and GLShe-sp),
Shapiro–Wilk test (Table 1). Their coefficients of skewness and respectively, showing that the structured component of spatial
kurtosis were close to zero, however, indicating no substantial dependence was significant at the study site (Table 4). The results
departure from normality, and the yield was then assumed to be of the cross-validation analyses showed that the better model was
normal. Bartlett’s test was significant at P < 0.001, indicating the GLShe-sp in terms of unbiasedness (CV1), accuracy (CV2),
heteroscedasticity of the corn yield with time, which was treated and precision (CV3) of the prediction (Table 5), confirming the
in this study by fitting different spatial models of covariance results of the goodness-of-fit statistics (Table 3). Moreover, the
function differing in sill and range parameters. Moran’s I test results obtained for the GLShe-sp model can be considered very
was also significant (P < 0.001), indicating spatial association. satisfactory because the CV1 was zero (up to four digits), CV2
The best subset of soil variables selected by stepwise regression was quite close to 1 (0.92), and CV3 was less than the standard
included pH, K+, P, and clay content (data not shown), which deviation of corn yield in any season (Table 1). These results
were used as soil regressors into the mixed effect. demonstrated that spatial-temporal models that account for
The spatial and temporal changes in corn grain yield are heteroscedasticity of variance were more reliable than other
reported in Fig. 3. In general, the data showed a slight decrease crop prediction models. Additionally, the results of residual
in corn yield from south to north. In the last 3 yr, it is possible analyses for the best model (GLShe-sp) showed that the residual
to verify clearly that there is a low-corn-yield zone in the north had a mean close to 0 and standard deviation close to 1 (Fig. 4).
region (between 190 and 230 m) because there probably was a A histogram and a Q–Q plot indicated that the residuals were
low-soil-fertility zone in this region of the study area. normal, whereas the scattergram of the residuals vs. estimates
showed no trend (Fig. 4). These results further suggest that this
Linear and Spatial Mixed Effect Models Analyses model was not biased and the assumptions (Osborne and Waters,
The variables K+ and P were not significant (P > 0.05) for 2002) required for regression analysis were confirmed.
all the mixed effect models tested, whereas clay content was The estimated coefficients for fixed effects of the GLShe-sp
not significant in the models that took heteroscedasticity into model of grain yield are given in Table 2, showing that among
account (i.e., GLShe and GLShe-sp) (Table 2). This last result the soil attributes, only pH has a significant impact on yield
indicates that ignoring heteroscedasticity can distort the results because the main factor that impacts crop yields in most
and increase the possibility of falsely declaring significant effects Brazilian agricultural fields is soil acidity (Amado et al., 2009;
(i.e., Type I errors) (Osborne and Waters, 2002), which indicated Dalchiavon et al., 2011; Nogara et al., 2011). This can be
that accounting for heteroscedasticity of variance was critical in explained because low pH values are often associated with high
this study, in agreement with the results observed by Lambert et Al3+ contents in Latosols (Oxisols according to the U.S. soil
al. (2004) and Hurley et al. (2004). Soil pH was significant for taxonomy), which are prevalent throughout many of the cropped
all the models, indicating that in all the study growing seasons, regions of Brazil (Muniz et al., 2011). Crop season had a large
the spatial variability of corn yield was highly correlated with soil impact on yield, as indicated by the significant coefficients for
acidity (Table 2). year that varied substantially with time. This variability can
The covariance function that best described the spatial be ascribed to the total amount of rainfall and distribution of
dependence was the spherical model. The two parameters (st2 rainfall (days with rain) in March, as confirmed by correlating
and ρt) of the AR(1) model were not significant at P < 0.05 (data the average values of corn yield to the amount of rainfall in
not shown) in the GLSsp and GLShe-sp models. This indicated March (R2 = 0.58) and the distribution of rainfall in March
that the stochastic component of temporal variability of corn (R2 = 0.71). March is a critical month for corn in São Paulo state,
yield was not significantly different than 0, thus temporal Brazil, because it corresponds with the period in which corn is in
observations could be considered to be uncorrelated; however, its milky/dough stage and water stress may be critical.
the temporal fixed effect (i.e., growing season) was highly Growing season (temporal variability) was a significant effect
significant for all the models (Table 2). The temporal effect was in the GLShe-sp model as well as in all the other models, causing
considered to be a deterministic process, and yield was not stable a different coefficient for each year. The full spatial and temporal
during the six-season study period as shown by the changing nature of the model prevented us from testing the significance of

1882 Agronomy Journal  •  Volume 105, Issue 6  •  2013


Table 1. Descriptive statistics for corn yield (Mg ha –1) during six growing season in a Rhodic Hapludox under no-till management.
Growing season Mean SD Min. Max. Skewness Kurtosis Pr < W†
2001–2002 7.6 0.50 6.3 8.6 –0.35 0.08 0.20
2002–2003 7.0 0.54 5.7 8.3 –0.20 –0.20 0.51
2003–2004 6.0 0.49 5.3 7.3 0.96 0.64 0.00
2007–2008 7.3 0.83 5.4 9.0 –0.14 –0.73 0.11
2008–2009 7.8 0.78 5.6 8.8 –0.91 0.25 0.00
2009–2010 6.9 0.75 5.4 8.6 –0.10 –0.49 0.53
† Pr < W, result of Shapiro–Wilk normality test (P < W > 0.10).

Fig. 3. Corn yield mean of four transects from each reference point of the studied years. Error bars represent the standard error of
the mean (n = 4 transects) for each reference point.

Agronomy Journal  •  Volume 105, Issue 6  •  2013 1883


Table 2. Tests of fixed effects of ordinary least squares (OLS), generalized least squares assuming heteroscedasticity (GLShet),
spatial model assuming homoscedasticity (GLSsp), and spatial model assuming heteroscedasticity (GLS he-sp) approaches used to
assess the impact of soil properties and growing seasons (GS) from 2001–2002 to 2009–2010 on corn yield. (From the visual inspec-
tion of this table, it is clear that the only coefficient estimates varying with time are the ones related to growing seasons, whereas
the coefficients related to the soil parameters are constant.)
Statistic pH Clay GS 1 GS 2 GS 3 GS 4 GS 5 GS 6
OLS
Estimate 0.5901 0.00355 3.2899 2.7033 1.7551 3.2758 3.833 3.0939
SE 0.07302 0.00165 0.5106 0.5128 0.5143 0.5058 0.5049 0.5022
df 473 473 473 473 473 473 473 473
Pr > |t| <0.0001 0.0314 <0.0001 <0.0001 0.0007 <0.0001 <0.0001 <0.0001
Lower 95% CL 0.4466 0.0003 2.2867 1.6957 0.7445 2.2819 2.841 2.107
Upper 95% CL 0.7335 0.0068 4.2932 3.7109 2.7656 4.2696 4.8251 4.0808
GLShet
Estimate 0.5224 0.0007 4.578 3.9884 3.0377 4.5409 5.0899 4.3349
SE 0.06485 0.00155 0.4683 0.4704 0.4721 0.4696 0.4693 0.4663
df 473 473 473 473 473 473 473 473
Pr > |t| <0.0001 0.6334 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
Lower 95% CL 0.395 –0.00231 3.6578 3.0642 2.11 3.6182 4.1677 3.4185
Upper 95% CL 0.6498 0.003785 5.4982 4.9127 3.9654 5.4636 6.0121 5.2512
GLSsp
Estimate 0.2971 0.004709 4.6071 3.942 2.9365 4.3961 4.8854 4.123
SE 0.0852 0.002391 0.8328 0.8303 0.8288 0.8207 0.8164 0.814
df 375 91.7 84.2 84.5 84.8 82.7 82.1 81.9
Pr > |t| 0.0005 0.0519 <0.0001 <0.0001 0.0006 <0.0001 <0.0001 <0.0001
Lower 95% CL 0.1295 –0.00004 2.9511 2.291 1.2885 2.7636 3.2614 2.5037
Upper 95% CL 0.4646 0.009458 6.2631 5.593 4.5845 6.0286 6.5094 5.7423
GLShe-sp
Estimate 0.4343 0.00177 4.7604 4.1788 3.3036 4.7121 5.1693 4.4624
SE 0.06993 0.001882 0.6423 0.6364 0.6522 0.6652 0.6354 0.6761
df 279 118 96.4 105 69.4 103 114 59.7
Pr > |t| <0.0001 0.3487 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
Lower 95% CL 0.2967 –0.00196 3.4856 2.917 2.0026 3.3928 3.9106 3.1098
Upper 95% CL 0.572 0.005497 6.0353 5.4406 4.6045 6.0315 6.4279 5.8149

Table 3. Model fit statistics of likelihood ratio test (–2 Res Table 5. Cross-validation errors including unbiasedness of
log.), Akaike’s information criterion (AIC), Akaike’s informa- the predictor (CV1), accuracy of mean squared prediction er-
tion criterion corrected (AICC), and Schwarz’s Bayesian in- ror (CV2), and goodness of prediction (CV3) of ordinary least
formation criterion (smaller is better) (BIC) for the ordinary squares (OLS), generalized least squares assuming homosce-
least squares (OLS), generalized least squares assuming het- dasticity (GLS he), spatial model assuming homoscedastic-
eroscedasticity (GLS he), spatial model assuming homoscedas- ity (GLSsp), and spatial model assuming heteroscedasticity
ticity (GLSsp), and spatial model assuming heteroscedasticity (GLS he-sp) approaches used to assess the impact of soil attri-
(GLS he-sp) approaches used to assess the impact of soil attri- butes and growing seasons on corn yield.
butes and growing seasons on corn yield. Model fitted CV1 CV2 CV3
Model –2 Res log. AIC AICC BIC Mg ha–1
OLS 899.3 901.3 901.3 905.5 OLS –0.1295 53.76 0.559
GLShe 869.6 881.6 881.7 906.6 GLShe –0.0348 55.62 0.559
GLSsp 869.2 875.2 875.2 887.7 GLSsp –0.0093 0.91 0.538
GLShe-sp 781.3 811.3 812.4 808.2 GLShe-sp 0.0000 0.92 0.489

Table 4. Likelihood ratio test of spatial covariance between


ordinary least squares (OLS) and the spatial model assum-
ing homoscedasticity (GLSsp), and between generalized least
squares assuming heteroscedasticity (GLS he) and the spatial
model assuming heteroscedasticity (GLS he-sp) used to assess
the impact of spatial and temporal error correlation on corn
yield. All data are significant at P < 0.001.
Models df c2 Pr > c2
OLS vs. GLSsp 2 30.1 0.001
GLShe vs. GLShe-sp 2 87.9 0.001

1884 Agronomy Journal  •  Volume 105, Issue 6  •  2013


Fig. 4. Results of residual statistics for corn yield using a spatial-temporal model assuming heteroscedasticity (GLS he-sp).

the impact of individual climate variables (e.g., rainfall, growing spatial autocorrelation in this study had a larger impact on the pH
degree days, or humidity) on corn yield because only one datum parameter estimate than did accounting for heterogeneity of errors.
was recorded for each season. Moreover, the model failed to The literature suggests that Type I errors in OLS models
reach convergence with these regressors. To explore these can result from violations of the assumptions of independent
relationships, we propose that a study be conducted across a large observations (Schabenberger and Gotway, 2005) and
area and with a longer historical yield data set. homoscedasticity (Osborne and Waters, 2002). The significance
Our results stress also that it is critical to account for temporal (i.e., OLS) and near significance (i.e., GLSsp) of clay in the
variation of the most dynamic soil attributes on a fine spatial models (Table 2) that assumed homoscedasticity were considered
scale because they can be influenced by climate factors between to be or nearly be Type I errors because clay was clearly not
the growing seasons. Hence, the relationships between crop significant in the models that accounted for heteroscedasticity
yield and these soil attributes may become extremely variable (i.e., GLShe and GLShe-sp), and the GLShe-sp model was assumed
across the seasons, as confirmed by Lambert et al. (2006), who to be most accurate. It is important to note, however, that the
observed spatial and temporal variability of corn yield response coefficient for clay in the OLS and GLSsp models was only
(relationship) to P and N in a 5-yr corn–soybean [Glycine max 0.004 and 0.005 Mg ha–1 corn yield, respectively. Experienced
(L.) Merr.] rotation in Minnesota. Therefore, it is recommended agronomists would have correctly concluded that this effect was
that farms use a set of data collected over time not only for yield biologically insignificant.
data but also for the most temporally variable soil attributes to The novelty of this study consists in splitting the whole variation
better understand the cause–effect relationships. of yield into the different components: the deterministic and
stochastic components of variation in both space and time. The
Practical Implications of Model Selection approach is of the regressive type, but taking into account error
Model choice impacted prediction and regression coefficient correlation and its heteroscedasticity, estimated by using different
estimates, and this is critical above all for the soil attributes affected mathematical models. Actually, there are not many studies in
by management. For example, the OLS model suggested that there the scientific literature that have performed such an exhaustive
would be a 0.59 Mg ha–1 yield increase for each unit increase in pH analysis of yield variation. The only limitation of the work is
(see the parameter estimate in Table 2). This prediction by the OLS that the estimation model cannot be considered as a full spatial-
model was significantly different from that of the GLSsp model temporal model due to the small size of the field. Meteorological
(0.30 Mg ha–1) and that of the GLShe-sp model (0.43 Mg ha–1), as variables, such as rainfall, expected to impact yield were not
confirmed by comparing the confidence limits of the pH coefficient repeatedly measured at any sample location but were assumed
estimates of the corresponding models (Table 2). Accounting for

Agronomy Journal  •  Volume 105, Issue 6  •  2013 1885


as constant across the field. This precluded estimation of the MODEL Yield=PH Clay P K Year
interaction term of soil ´ climate. /Cl S noint outpred=pred_
SOLS
SUMMARY AND CONCLUSIONS DDFM=SATTERTH
The heteroscedastic spatial-temporal autocorrelation model ;
(GLShe-sp) was superior to the other models (OLS, GLShe, random year/Type=AR(1) subject=ID;
and GLSsp) for analyzing yield–soil attributes relationships as Repeated/subject=intercept Type=sp(sph)
determined with cross-validation analysis. This study supports (X Y) local;
parms (0.064) (53) (0.291);
the modeling of spatial residual errors and accounting for
;
heterogeneity of the variance with time to obtain more accurate
estimates of treatment effects for interpreting yield variation. Soil
/*4–Spatial-temporal model assuming
acidity, as assessed by pH, was the soil effect that most influenced
heteroscedasticity (GLShe-sp)*/
corn yield with time in this study. The full spatial and temporal
PROC MIXED DATA=DataSet
nature of the model prevented testing the significance of the PLOTS(ONLY)=ALL
impact of climate variables on yield; however, it was possible to METHOD=REML covtest SCORING=50;
verify the impact of the total amount and distribution of rainfall ;
in March on yield. Farms should use a sufficiently long historical CLASS Year ID
data set of crop yield and soil attributes to understand cause– ;
effect relationships between crop yield and soil attributes and to MODEL Yield=PH Clay P K Year
predict crop yield accurately. /Cl S noint outpred=pred_
SGLS
Appendix DDFM=SATTERTH
For the variables soil pH, clay content, P content, K content, ;
and growing season or year: random year/Type=AR(1) subject=ID;
/*1–Ordinary least squares (OLS)*/ Repeated/Type=SP(sph)(X Y)
PROC MIXED covtest DATA=DataSet subject=intercept local group=Year;
PLOTS(ONLY)=ALL parms (0.3) (0.18) (0.040) (57) (0.054)
; (39) (0.061) (127) (0.468) (44) (0.351)
CLASS Year ID (24) (0.248) (58) (0.175);;
;
ACKNOWLEDGMENTS
MODEL Yield=PH Clay P K Year
/Cl S noint outpred=pred_ We are grateful for a scholarship (Grant 8492-11-5) to M.S. Rodrigues
OLS; from CAPES Foundation, Ministry of Education of Brazil, Brasília.
;
REFERENCES
Amado, T.J.C., L.Z. Pes, C.L. Lemainski, and R.B. Schenato. 2009. Chemical
/*2–Generalized least squares assuming and physical attributes of Oxisols and their relation with irrigated corn and
heteroscedasticity (GLShe)*/ common bean yields. (In Portuguese, with English abstract.) Rev. Bras.
PROC MIXED covtest DATA=DataSet Cienc. Solo 33:831–843. doi:10.1590/S0100-06832009000400008
PLOTS(ONLY)=ALL Anselin, L. 1993. The Moran scatterplot as an ESDA tool to assess local instability
in spatial association. Paper presented at: GISDATA Specialist Meeting on
; GIS and Spatial Analysis, Amsterdam. 1–5 Dec. 1993. Paper 9330.
CLASS Year ID Bakhsh, A., D.B. Jaynes, T.S. Colvin, and R.S. Kanwar. 2000. Spatio-temporal
; analysis of yield variability for a corn–soybean field in Iowa. Trans. ASAE
MODEL Yield=PH Clay P K Year 43:31–38.
Bolker, B.M., M.E. Brooks, C.J. Clark, S.W. Geange, J.R. Poulsen, M.H.H.
/Cl S noint outpred=pred_GLS Stevens, and J.S.S. White. 2009. Generalized linear mixed models: A
; practical guide for ecology and evolution. Trends Ecol. Evol. 24:127–135.
Repeated/group = Year; doi:10.1016/j.tree.2008.10.008
; Carroll, S.S., and N. Cressie. 1996. A comparison of geostatistical methodologies
used to estimate snow water equivalent. Water Resour. Bull. 32:267–278.
doi:10.1111/j.1752-1688.1996.tb03450.x
/*3–Spatial-temporal model assuming Dalchiavon, F.C., M.P. Carvalho, O.S. Freddi, M. Andreotti, and R.
homoscedasticity (GLSsp)*/ Montanari. 2011. Spatial variability of the bean yield correlated with
PROC MIXED DATA=DataSet chemical attributes of a Typic Acrustox under no-tillage system. (In
Portuguese, with English abstract.) Bragantia 70:908–916. doi:10.1590/
PLOTS(ONLY)=ALL S0006-87052011000400025
METHOD=REML covtest SCORING=50; Diker, K., D.F. Heermann, and M.K. Brodahl. 2004. Frequency analysis of
; yield for delineating yield response zones. Precis. Agric. 5:435–444.
doi:10.1007/s11119-004-5318-9
CLASS Year ID
Flinn, J.C., and S.K. De Datta. 1984. Trends in irrigated-rice yields under
; intensive cropping at Philippine research stations. Field Crops Res. 9:1–15.
doi:10.1016/0378-4290(84)90002-9

1886 Agronomy Journal  •  Volume 105, Issue 6  •  2013


Gee, G.W., and D. Or. 2002. Particle-size analysis. In: J.H. Dane and G.C. Topp, Nogara Neto, F., G. Roloff, J. Dieckow, and A.C.V. Motta. 2011. Spatially
editors, Methods of soil analysis, Part 4. Physical methods. SSSA Book Ser. varied soil and crop attributes related to corn yield. (In Portuguese, with
5. SSSA, Madison, WI. p. 255–289. doi:10.2136/sssabookser5.4.c12 English abstract.) Rev. Bras. Cienc. Solo 35:1025–1036. doi:10.1590/
Hurley, T.M., G.L. Malzer, and B. Kilian. 2004. Estimating site-specific S0100-06832011000300036
nitrogen crop response functions. Agron. J. 96:1331–1343. doi:10.2134/ Osborne, J., and E. Waters. 2002. Four assumptions of multiple regression that
agronj2004.1331 researchers should always test. Pract. Assess. Res. Eval. 8(2).
Lambert, D.M., J. Lowenberg-Deboer, and R. Bongiovanni. 2004. A comparison Page, A.L., R.H. Miller, and D.R. Keeney, editors. 1982. Methods of soil analysis.
of four spatial regression models for yield monitor data: A case study from Part 2. Chemical and microbiological properties. 2nd ed. Agron. Monogr.
Argentina. Precis. Agric. 5:579–600. doi:10.1007/s11119-004-6344-3 9. ASA and SSSA, Madison, WI.
Lambert, D.M., J. Lowenberg-DeBoer, and G.L. Malzer. 2006. Economic analysis Pierce, F.J., and P. Nowak. 1999. Aspects of precision agriculture. Adv. Agron.
of spatial-temporal patterns in corn and soybean response to nitrogen and 87:1–85.
phosphorus. Agron. J. 98:43–54. doi:10.2134/agronj2005.0005 Raij, B.V., H. Cantarella, J.A. Quaggio, and A.M.C. Furlani. 1997.
Landagan, O.Z., and E.B. Barrios. 2007. An estimation procedure for a Recomendações de adubação e calagem para o Estado de São Paulo. Inst.
spatial–temporal model. Stat. Probab. Lett. 77:401–406. doi:10.1016/j. Agron., Campinas, São Paulo.
spl.2006.08.006 Rodrigues, M.S., J.E. Corá, and C. Fernandes. 2012. Spatial relationships
Lark, R.M. 2000. Regression analysis with spatially autocorrelated error: between soil attributes and corn yield in no-tillage system. Rev. Bras.
Simulation studies and application to mapping of soil organic matter. Int. Cienc. Solo 36:599–609. doi:10.1590/S0100-06832012000200029
J. Geogr. Inf. Sci. 14:247–264. doi:10.1080/136588100240831 Schabenberger, O., and C. Gotway. 2005. Statistical methods for spatial data
Littell, R.C., G.A. Milliken, W. Stroup, R.D. Wolfinger, and O. Schabenberber. analysis. CRC Press, Boca Raton, FL.
2006. SAS system for mixed models. SAS Inst., Cary, NC. Shapiro, S.S., and M.B. Wilk. 1965. An analysis of variance test for normality
Lobell, D.B., J.I. Ortiz-Monasterio, G.P. Asner, R.L. Naylor, and W.P. Falcon. (complete samples). Biometrika 52:591–611.
2005. Combining field surveys, remote sensing, and regression trees to Snedecor, G.W., and W.G. Cochran. 1989. Statistical methods. 8th ed. Iowa
understand yield variations in an irrigated wheat landscape. Agron. J. State Univ. Press, Ames.
97:241–249.
Timlin, D.J., Y. Pachepsky, V.A. Snyder, and R.B. Bryant. 1998. Spatial and
McBratney, A., B. Whelan, T. Ancev, and J. Bouma. 2005. Future temporal variability of corn grain yield on a hillslope. Soil Sci. Soc. Am. J.
directions of precision agriculture. Precis. Agric. 6:7–23. doi:10.1007/ 62:764–773. doi:10.2136/sssaj1998.03615995006200030032x
s11119-005-0681-8
Moran, P.A.P. 1950. Notes on continuous stochastic phenomena. Biometrika
37:17–23.
Muniz, B.M., N. Curi, G. Sparovek, A. Carvalho Filho, and S.R.H.G. Silva.
2011. Updated Brazilian’s georeferenced soil database: An improvement
for international scientific information exchanging. In: E.B.Ö. Güngör,
editor, Principles, application and assessment in soil science. InTech,
Rijeka, Croatia. p. 309–332.

Agronomy Journal  •  Volume 105, Issue 6  •  2013 1887

View publication stats